Lambert 2018

Optimization Applications in
Economics and Finance
R.Enkhbat, J.Enkhbayar, N.Tungalag,

D.Bayanjargal, S.Batbileg, Ch.Ankhbayar,
B.Ser-Od.
ii
Preface
Optimization plays an important role not only in science and technology,
operations research but also in economics and finance. This book discusses
applications of optimization theory and methods in Economics and Finance.
It is well known that the general equilibrium theory is based on convex op-
timization problems while finding Nash equilibriums is mainly focused on
nonconvex optimization problems. The purpose of this book is to reveal
applications of optimization theory and methods in economics and illus-
trate these on some examples of Mongolian economy. The authors assume
that the reader is familiar with mathematical programming, optimization
techniques, optimal control, economics and finance as well as game theory.
This book is based on the lectures on optimization courses given by authors
to graduate students in applied mathematics and business of the School
of Applied Science and Engineering, and the School of Business of Na-
tional University of Mongolia. The book is organized in 7 chapters. These
chapters are independed and devoted to various areas of economics such
as Solow growth model, consumer theory, game theory, cost minimization,
investment and population models, and mathematical model for finance. In
chapter 1, it is shown that the Solow growth model with nonconvex pro-
duction functions leads to parametric and global optimization problems.
We reduce the per capita consumption maximization problem to one vari-
able global optimization and provide with numerical results obtained by
a global search algorithm. In this chapter, an optimal control approach
to the Solow growth model has been proposed. Chapter 2 deals with the
nonconvex utility maximization problem. Depending on types of economic
market , we classify the problem as continuous, parametric, nonconvex and
discrete utility maximization problems. The corresponding optimization
methods for solving these problems are also discussed in the chapter. In
chapter 3, we present an algorithm for solving non-zero sum three-person
and N-person games. We reduce the problem of finding Nash equilibrium
points to a global optimization problem. Numerical results are provided for
3 and 4 person games. A cost minimization problem is discussed in chap-
ter 4. We show that the average cost function is pseudoconvex. Calculus
variations approach to economic order quantity model was also considered
in this chapter. In chapter 5, investment models, including the valuation
model and optimal control problem in financial modeling are discussed. The
conditional gradient method has been proposed for solving the optimal con-
trol problem and numerical results are provided. Population models are
described in chapter 6. We applied the deterministic and stochastic models
to the Mongolian population growth. Optimization applications in finance
presented and analyzed in chapter 7. Models of marketing advertisement for
insurance companies and credit risk evaluation models have been examined
on examples of Mongolian economy.
The authors
Contents
1 Solow Growth Theory 1

1.1 Parametric Optimization Approach to the Solow Growth The-
ory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Global Optimization Approach to the Solow Growth Theory 10
1.3 Optimal Control Approach to the Solow Growth Theory . . 15
2 Utility Maximization Problem 19

2.1 Continuous Utility Maximization
Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.2 Parametric Utility Maximization
Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3 Nonconvex Utility Maximization
Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.4 Discrete Utility Maximization Problem . . . . . . . . . . . . 33
3 Game Theory 38
3.1 The Curvilinear Search Algorithm for Solving Three-Person
Game . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.1.1 Non-zero sum three-person game . . . . . . . . . . . 39
3.1.2 The curvilinear global search algorithm . . . . . . . . 44
3.1.3 Computational results . . . . . . . . . . . . . . . . . 46
3.2 A Computational Method for Solving N -Person Game . . . 49
3.2.1 Nonzero sum N -person game . . . . . . . . . . . . . 49
3.2.2 Computational results . . . . . . . . . . . . . . . . . 53
4 Cost Minimization Problem 57

4.1 Pseudoconvexity Properties of Average Cost Function . . . 58
4.1.1 Average cost function . . . . . . . . . . . . . . . . . . 58
4.1.2 Average cost minimization problem. . . . . . . . . . . 59
4.1.3 Numerical results . . . . . . . . . . . . . . . . . . . . 60
4.2 Economic Order Quantity Model . . . . . . . . . . . . . . . 62
4.2.1 Calculus Variations Approach to EOQ . . . . . . . . 63
iv CONTENTS
5 Investment Model 69
5.1 The Basic Valuation Model . . . . . . . . . . . . . . . . . . 70
5.1.1 Optimization Approach . . . . . . . . . . . . . . . . . 71
5.1.2 Numerical Results . . . . . . . . . . . . . . . . . . . 73
5.2 Optimal Control Problem in Financial Modeling . . . . . . . 74
5.2.1 Conditional gradient method . . . . . . . . . . . . . . 75
5.2.2 Computational algorithm . . . . . . . . . . . . . . . . 76
5.2.3 Numerical Results . . . . . . . . . . . . . . . . . . . 77
6 Population Model 81
6.1 Application of Stochastic Differential Equations in Popula-
tion growth . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6.1.1 Population models . . . . . . . . . . . . . . . . . . . 82
6.1.2 Stochastic models . . . . . . . . . . . . . . . . . . . . 83
7 Optimization Applications in Finance 90

7.1 A Mathematical Model of Marketing Advertisement for In-
surance Companies . . . . . . . . . . . . . . . . . . . . . . . 91
7.1.1 Extension of advertising model . . . . . . . . . . . . 91
7.1.2 Problem solution . . . . . . . . . . . . . . . . . . . . 92
7.2 Mathematical Model for Credit Risk
Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
7.2.1 Model formulation . . . . . . . . . . . . . . . . . . . 96
7.2.2 Bank requirements and criteria . . . . . . . . . . . . 99
7.3 Optimization Problem of Foreign Reserves . . . . . . . . . . 101
7.3.1 Merton’s type consumption - investment
problem . . . . . . . . . . . . . . . . . . . . . . . . . 101
7.3.2 Foreign reserves problem . . . . . . . . . . . . . . . . 103
7.3.3 Estimation of econometric model . . . . . . . . . . . 105
1
Solow Growth Theory
2 Solow Growth Theory
1.1 Parametric Optimization Approach to the

Solow Growth Theory
Introduction
The production function model was applied to the study of growth problems
by R.Solow [5]. Solow developed a growth theory model within a neoclassical
economic framework. The Solow growth model assumes the maximization
of per capita consumption under economic equilibria or steady state ([1],[3]).
This model from view point of optimization problem was considered in [7].
Let’s consider briefly Solow growth model. The production function relating
output Y to capital K and labor L is
Y (t) = f (K(t), L(t)). (1.1)
We assume a fraction s of income is saved and invested. Then the standard
capital accumulation equation is
0
K (t) = sf (K(t), L(t)) − µK(t),
(1.2)
K(t0 ) = K0
where s the savings (saving rate), 0 ≤ s ≤ 1, and µ depreciation rate of
capital and the consumption C is
C(t) = (1 − s)f (K(t), L(t)). (1.3)
Let f (K, L) be a concave, differentiable homogeneous production func-
tion. Assume that the labor grows at exponential rate η which means that
L = L0 eηt . (1.4)
Define k as per capita capital function:
K(t)
k(t) = .
L(t)
Then
0
K 0 L − L0 K L0

0 K(t) 1 0 1 0
k (t) = = = K − K = K − ηK .
L(t) L2 L L L
If we substitute f (K, L) into this equation, we have

0 1
k = s(t)f (K, L) − µK − ηK
L
K K(t) K(t)
= s(t)f ,1 − µ −η
L L(t) L(t)
= s(t)ϕ(k) − (µ + η)k,
1.1 Parametric Optimization Approach to the Solow Growth Theory 3
where ϕ(k) is the Solow per capita production function. Then per capita
capital accumulation equation is
0
k = s(t)ϕ(k) − (µ + η)k
(1.5)
k(0) = k0 .
Economic equilibria condition is:
k 0 = s(t)ϕ(k) − (µ + η)k = 0. (1.6)
Now we consider per capita consumption function

C(t) (1 − s)f (K, L)
c(t) = = = (1 − s)ϕ(k) (1.7)
L(t) L
Assume that s is a constant function.
Let us consider the per capita consumption maximization problem subject
to economic equilibria. That is
c = (1 − s)ϕ(k) → max, (1.8)

sϕ(k) − (µ + η)k = 0. (1.9)
Then problem (1.8)-(1.9) is equivalent to the following one dimensional

problem:
max ϕ(k) − (µ + η)k .
k
∗
The solution k satisfies the equation
ϕ0 (k ∗ ) = (µ + η). (1.10)
So called the golden rule of level of accumulation is determined from (1.9)

as in [12]:
ϕ0 (k ∗ )k ∗
s∗ = . (1.11)
ϕ(k ∗ )
Economic growth with nonconvex production functions

In general, we consider a capital accumulation equation
0
K (t) = s(t)f (K(t), L(t), t) − µK(t)
(1.12)
K(tA ) = KA
when f (K, L, t) and L(t) are arbitrary continuously differentiable given

functions on their domains and s = s(t), µ = µ(t) are functions of t,
t ∈ [tA , tB ].
Then a per capita consumption function is
C(t) (1 − s(t))f (K(t), L(t), t)

c(t) = = . (1.13)
L(t) L(t)
Also, a per capita capital function is
K(t)
k(t) = . (1.14)
L(t)
Economic equilibria condition is written as

0
K 0 (t)L(t) − L0 (t)K(t)

0 K(t)
k (t) = = =0 (1.15)
L(t) L2
or
K0 L0
=
K L
which means that capital and labor growth rate must be equal.
Taking into account (1.12), we can write (1.15) as follows
K 0 (t)L(t) − L0 (t)K(t) = s(t)f (K, L, t) − µ(t)K(t) L(t) − L0 (t)K(t) = 0.

From this equation we find s(t):
µKL + L0 K
s(t) = . (1.16)
Lf (K, L, t)
Denote by KB the maximum of function K(t) defined by (1.12), i.e.,
KB = max K(t),
tA ≤t≤tB
and introduce the function φ(K, L, t) as follows
µKL + L0 K
φ(K, L, t) = f (K, L, t) − .
L
Definition 1.1. If the following condition
φ(K, L, t) ≥ 0, ∀t ∈ [tA , tB ]
holds then the interval [KA , KB ] is called an economic efficient interval.

From Definition 1.1 and (1.16), we can easily notice that 0 ≤ s(t) ≤ 1.
Now we consider per capita consumption maximization problem on a
given interval [tA , tB ].
(1 − s(t))f (K, L, t)
c(t) = → max, t ∈ [tA , tB ]
L
subject to
µKL + L0 K
s(t) = .
Lf (K, L, t)
This problem is equivalent to the following one dimensional parametric max-
imization problem.
f (K, L, t) µKL + L0 K
F (K, t) = − → max, t ∈ [tA , tB ]. (1.17)
L L2 K
We can write function f (K, L, t) as

µKL + L0 K
f (K, L, t) = + φ(K, L, t).
L
Further, we assume that [KA , KB ] is an efficient interval.
Then the problem (1.17) can be rewritten as
f (K, L, t) µKL + L0 K φ(K, L, t)
F (K, t) = − 2
= → max, t ∈ [tA , tB ].
L L L K
(1.18)
Problem (1.18) is a hard parametric optimization problem.
Lemma 1.1. The function F (K, t) satisfies the Lipschitz condition with
respect to t with constant M for each K ∈ [KA , KB ], i.e.,

F (K, t̂) − F (K, t) ≤ M t̂ − t, ∀t ∈ [tA , tB ].
Proof. Since f (K, L, t) is a continuously differentiable function with

respect to t, using Taylor expansion formula we can write down:
∂f (K, L, t + θ∆t)
f (K, L, t + ∆t) − f (K, L, t) = ∆t
∂t
where 0 < θ < 1, t + θ∆t ∈ [tA , tB ].
Now we have the following estimation

∂f (K, L, t + θ∆t)
f (K, L, t + ∆t) − f (K, L, t) = ∆t ≤
∂t

∂f (K, L, t)
≤ max ∆t.
KA ≤K≤KB ∂t
tA ≤t≤tB

∂f (K,L,t)
By setting M = maxKA ≤K≤KB ∂t , we obtain
t ≤t≤t A B

F (K, t̂) − F (K, t) ≤ M t̂ − t
which proves the lemma.

Lemma 1.2. Assume that the production function f (K, L, t) is a continu-
ously differentiable function with respect to t. Then for a given > 0, there
exists a discretization tA = t0 < t1 < ... < ti < ti+1 < ... < tN = tB
such that
|F (K ∗ (t), t) − F (K ∗ (ti ), ti )| < f or all t ∈ [tA , tB ] and certain ti ,
where
F (K ∗ (t), t) = max F (K, t), t ∈ [tA , tB ].
K∈[KA ,KB ]
Proof. We discretize [tA , tB ] in the following way

tB − tA
tA = t0 , ti = t0 + i , i = 1, 2, ..., N.
N
Clearly, for any t ∈ [tA , tB ], there exists j ∈ {1, 2, ..., N } such that t ∈
[tj , tj+1 ]. Consequently,
tB − tA
|t − tj | < . (1.19)
N
Due to Lemma 1, there exists M > 0 such that
|f (K(t̂), L(t̂), t̂) − f (K(t), L(t), t)| < M |t̂ − t|, ∀t, t̂ ∈ [tA , tB ].
Define > 0 as follows:

tB − tA
=M (1.20)
N
Now take any t ∈ [tA , tB ] and compute
tB − tA
|F (K ∗ (t), t) − F (K ∗ (tj ), tj )| ≤ M |t − tj | ≤ M =
N
which proves the lemma.
The above lemma allows us to find − approximate solution of problem
(1.18) by solving a finite number of nonlinear optimization problems.
Numerical Results
To reinforce the theoretical results, we solve the following problems nu-

merically on t ∈ [1, 5] for the given parameters of η = 0.012, L0 = 2, N =
40, = 0.001.
Example 1.1.
φ(K(t), L(t))
F (K(t), L(t)) = → max, t ∈ [1, 5]
L(t) K
where
φ(K, L) = −L2 K 4 + 8LK 3 − 9K 2 + 5, K ∈ [1; 4]
and the labor grows at exponential rate η.
Example 1.2.
φ(K(t), L(t))
F (K(t), L(t)) = → max, t ∈ [1, 5]
L(t) K
where
φ(K, L) = 0.000108K 5 − 0.00596LK 4 + 0.11365K 3 −

−0.889572K 2 + 2.986324K, K ∈ [0; 6]
Example 1.3.
φ(K(t), L(t))
F (K(t), L(t)) = → max, t ∈ [1, 4]
L(t) K
where
φ(K, L) = log(K 2 + LK)/K, K ∈ [1; 3]
The numerical results using MATLAB are given in the following tables,
respectively.
Table 1.1: Example 1

t L K∗ t L K∗ t L K∗
1.0 2.0404 2.5100 2.4 2.0983 2.4407 3.8 2.1579 2.3732
1.1 2.0445 2.5049 2.5 2.1025 2.4358 3.9 2.1622 2.3686
1.2 2.0486 2.4999 2.6 2.1068 2.4309 4.0 2.1666 2.3638
1.3 2.0527 2.4949 2.7 2.111 2.4260 4.1 2.1709 2.3591
1.4 2.0568 2.4899 2.8 2.1152 2.4242 4.2 2.1753 2.3543
1.5 2.0609 2.4850 2.9 2.1194 2.4164 4.3 2.1796 2.3497
1.6 2.065 2.4801 3.0 2.1237 2.4115 4.4 2.184 2.3449
1.7 2.0692 2.4750 3.1 2.1279 2.4067 4.5 2.1883 2.3403
1.8 2.0733 2.4701 3.2 2.1322 2.4019 4.6 2.1927 2.335
1.9 2.0775 2.4651 3.3 2.1365 2.3971 4.7 2.1971 2.3310
2.0 2.0816 2.4603 3.4 2.1407 2.3924 4.8 2.2015 2.3263
2.1 2.0858 2.4554 3.5 2.145 2.3876 4.9 2.2059 2.3217
2.2 2.09 2.4504 3.6 2.1493 2.3828 5.0 2.2103 2.3355
2.3 2.0941 2.4456 3.7 2.1536 2.3780

∗
t L K t L K∗ t L K∗
1.0 2.0404 2.4270 2.4 2.0983 2.4068 3.8 2.1579 2.3870
1.1 2.0445 2.4256 2.5 2.1025 2.4054 3.9 2.1622 2.3857
1.2 2.0486 2.4241 2.6 2.1068 2.4039 4.0 2.1666 2.3842
1.3 2.0527 2.4227 2.7 2.111 2.4025 4.1 2.1709 2.3828
1.4 2.0568 2.4212 2.8 2.1152 2.4011 4.2 2.1753 2.3814
1.5 2.0609 2.4198 2.9 2.1194 2.3997 4.3 2.1796 2.3800
1.6 2.065 2.4183 3.0 2.1237 2.3982 4.4 2.184 2.3786
1.7 2.0692 2.4168 3.1 2.1279 2.3969 4.5 2.1883 2.3773
1.8 2.0733 2.4154 3.2 2.1322 2.3954 4.6 2.1927 2.3759
1.9 2.0775 2.4139 3.3 2.1365 2.3940 4.7 2.1971 2.3745
2.0 2.0816 2.4125 3.4 2.1407 2.3927 4.8 2.2015 2.3732
2.1 2.0858 2.4110 3.5 2.145 2.3912 4.9 2.2059 2.3718
2.2 2.09 2.4097 3.6 2.1493 2.3898 5.0 2.2103 2.3704
2.3 2.0941 2.4082 3.7 2.1536 2.3884

t L K∗ t L K∗ t L K∗
1.0 2.0404 1.2131 2.4 2.0983 1.1874 3.8 2.1579 1.1619
1.1 2.0445 1.2112 2.5 2.1025 1.1856 3.9 2.1622 1.1602
1.2 2.0486 1.2097 2.6 2.1068 1.1838 4.0 2.1666 1.1583
1.3 2.0527 1.2075 2.7 2.111 1.1819 4.1 2.1709 1.1564
1.4 2.0568 1.2057 2.8 2.1152 1.1801 4.2 2.1753 1.1547
1.5 2.0609 1.2039 2.9 2.1194 1.1783 4.3 2.1796 1.1528
1.6 2.065 1.2021 3.0 2.1237 1.1765 4.4 2.184 1.1510
1.7 2.0692 1.2002 3.1 2.1279 1.1747 4.5 2.1883 1.1492
1.8 2.0733 1.1984 3.2 2.1322 1.1729 4.6 2.1927 1.1472
1.9 2.0775 1.1966 3.3 2.1365 1.1710 4.7 2.1971 1.1456
2.0 2.0816 1.1947 3.4 2.1407 1.1692 4.8 2.2015 1.1438
2.1 2.0858 1.1929 3.5 2.145 1.1674 4.9 2.2059 1.1420
2.2 2.09 1.1910 3.6 2.1493 1.1655 5.0 2.2103 1.1402
2.3 2.0941 1.1893 3.7 2.1536 1.1638
1.2 Global Optimization Approach to the Solow

Growth Theory
First, we consider problem (1.8)-(1.9) for the case when
 0
 ϕ (k) = γϕ(M − ϕ)
ϕ(0) = ϕ0 (1.21)
L = L0 eηt .

The solution of (1.21) is

ϕ0 M
ϕ(k) =
ϕ0 + (M − ϕ0 )−M γk
Then problem (1.8)-(1.9) reduces to the problem:

c = (1 − s)ϕ(k) → max,
(1.22)
sϕ(k) = (µ + η)k.
The problem (1.22) is equivalent to one variable maximization problem
c = ϕ(k) − (µ + η)k → max, k > 0. (1.23)
We find stationary points of the problem
c0 (k) = ϕ0 (k) − (µ + η) = 0.
It can be easily calculated that

M 2 γϕ0 (M − ϕ0 )e−M γk
ϕ0 (k) = 2 .
ϕ0 + (M − ϕ0 )e−M γk
If we substitute z = e−M γk and A = M 2 γϕ0 (M − ϕ0 ) then

Az
2 = µ + η.
ϕ0 + (M − ϕ0 )z
This reduces to the following quadratic equation
(M − ϕ0 )2 (µ + η)z 2 + 2ϕ0 (M − ϕ0 )(µ + η) − A z + (µ + η)ϕ20 = 0.

Denote by D:
2
D = 2ϕ0 (M − ϕ0 )(µ + η) − A − 4(µ + η)2 (M − ϕ0 )2 ϕ20 .

(1.24)
The solution z is found as

√
2ϕ0 (M − ϕ 0 )(µ + η) − A + D
z∗ = − 2
. (1.25)
2(µ + η)(M − ϕ0 )
Then
1
k∗ = − ln z ∗ .
Mγ
It can be checked that c00 (k ∗ ) = ϕ00 (k ∗ ) < 0. This means that z ∗ is a global
solution to the problem.
Lemma 1.3. 0 < z ∗ < 1.
The proof is obvious consequence of (1.24) an (1.25).
Problem (1.22) has been solved numerically for different parameters of
ϕ0 , M, µ, η, γ. We present the numerical results in the following table 1.
Table 1.
ϕ0 M µ η γ Solutions
∗
k c∗ s∗
1.5 10 0.01 0.023 0.2 4.06913 9.849 0.01345
0.3 5 0.2 0.014 0.5 2.7135 4.332 0.1182
4 13 0.043 0.032 0.7 0.89 12.924 0.005185
Finally, we consider the case when both functions ϕ(k) and L(t) are
logistic. It means that
 0

 ϕ (k) = γϕ(M − ϕ)
ϕ(0) = ϕ0

(1.26)

 L0 (t) = σL(T − L)
L(0) = L0 .

Capital accumulation equation is

K 0 (t) = sf (K, L) − µK(t).
Per capita capital function is
K(t)
k(t) = .
L(t)
Then
0
K 0 L − L0 K

0 K(t)
k = =
L(t) L2
0
0
K L K
= −
L L L
= sϕ(k) − µk − σ(T − L)k.
The stability condition is

sϕ(k) = µ + σ(T − L) k. (1.27)
Then problem (1.22)-(1.23) reduces to one variable parametric maximiza-

tion problem:

c(k) = ϕ(k) − µk − σ(T − L) k → max, t ∈ [0, +∞[. (1.28)
k>0
We solve the above problem numerically on [0, 1] for the given parameters
of ϕ0 , M, µ, γ L0 , T, σ and t = 0; t = 0.5; t = 1. The numerical results are
given in the following tables.
Table 2. t = 0
ϕ0 M µ γ L0 T σ Solutions
∗
k c∗ s∗
1 7 0.01 0.2 2 5 0.1 3.6992 5.6243 0.1694
2 12 0.2 0.13 3 8 0.21 2.6709 7.7980 0.2998
5 15 0.043 0.5 4 9 0.42 0.6153 13.390 0.0896
Table 3. t = 0.5
∗
k c∗ s∗
1 7 0.01 0.2 2 5 0.1 3.7783 5.7388 0.1554
2 12 0.2 0.13 3 8 0.21 2.9082 8.7641 0.2306
5 15 0.043 0.5 4 9 0.42 0.7794 14.4122 0.0336
Table 4. t = 1
∗
k c∗ s∗
1 7 0.01 0.2 2 5 0.1 3.8679 5.8580 0.1408
2 12 0.2 0.13 3 8 0.21 3.1960 9.6917 0.1648
5 15 0.043 0.5 4 9 0.42 0.9767 14.8359 0.0096
Now assume that the production function f (K, L) is differentiable and

homogenous, and satisfies the Lipschitz condition with respect to K with
the constant M > 0. It means that
|f (K 0 , L) − f (K 00 , L)| ≤ M kK 0 − K 00 k, ∀L > 0.
Then it is clear that the function ϕ(k) is nonconvex and satisfies the Lips-
chitz condition with the constant M .
The problem (1.8)-(1.9) becomes a nonconvex optimization problem and

can be solved by global optimization techniques. Consider the above prob-
lem on a sufficiently large interval [0, b].
F (k) = −ϕ(k) + (µ + η)k → min, k ∈ [0, b]. (1.29)
Problem (1.29) belongs to a class of global optimization and its solution

can be found by the method of piecewise linear function [4].
Now we present the algorithm of the method. Introduce the function
P (k1 , y) for any fixed y ∈ [0, b] in the following way:
P (k1 , y) = F (y) − M |k1 − y|, k1 ∈ [0, b].
Choose an arbitrary point k10 ∈ [0, b] and set y 0 = k10 . Define piecewise
linear function S1 (k1 ) = P (k1 , y 0 ).
Let y 1 be the solution of the problem:
S1 (k1 ) → min, k1 ∈ [0, b].
Find a point k11 such that
F (k11 ) = min F (k10 ); F (y 1 ) , k11 = k10 ∨ y 1 .

Construct piecewise linear function S2 (k1 ) by
S2 (k1 ) = max S1 (k1 ); P (k1 , y 1 ) .

Let y 2 be the solution of the problem:
S2 (k1 ) → min, k1 ∈ [0, b].
Define k12 so as to satisfy
F (k12 ) = min F (k11 ), F (y 2 ) , k12 = k11 ∨ y 2 .

Analogously, we can write the iteration process for the t-th step as follows:
F (k1t ) = min F (k1t−1 ); F (y t ) ), k1t = k1t−1 ∨ y t ,

where y t : St (k1 ) → min, k1 ∈ [0, b],
St (k1 ) = max St−1 (k1 ); P (k1 , y t−1 ) .

Consequently, the method of piecewise linear functions generates two se-

quences of points y t and k1t . It is well known [6] that sequences of points k1t
and y t produced by the above algorithm converges to the global minimizer

for the problem, i.e.,
lim F (k1t ) = lim St (k1t ) = min F (k).

t→∞ t→∞ k∈[0,b]
For example, η = 0.4, µ = 1.6, b = 5.4 and
ϕ(k) = − 0.0787k 7 + 1.5459k 6 − 12.2406k 5 + 49.8829k 4

− 109.764k 3 + 122.579k 2 − 51.4616k.
Then problem (1.22)-(1.23) reduces to the problem
c∗ = max ϕ(k) − 2k .

[0,5.4]
If we solve the problem by the above algorithm, then solutions will be:
k ∗ = 5.06; c∗ = 4.3277.
1.3 Optimal Control Approach to the Solow Growth Theory 15
1.3 Optimal Control Approach to the Solow

Growth Theory
Assume that per capita capital ratio is a increasing function on [0, T ], i.e,
K(t)
= x(t), x0 (t) ≥ 0, ∀t ∈ [0, T ]. (1.30)
L(t)
Then equation (1.30) reduces to
x0 (t)L(t) + x(t)L0 (t) = sF (x(t)L, L) − µx(t)L(t),
or equivalently,
L0 (t)
x0 (t) + x(t) = sF (x(t), 1) − µx(t). (1.31)
L(t)
Taking into account of homogeneousness of the function F (K, L), we denote

the following function by ϕ(k) as
ϕ(x(t)) = F (x(t), 1).
Then equation (1.31) can be written as
x0 (t) + x(t)η = sϕ(x(t)) − µx(t). (1.32)
A parameter s as a ratio of total consumption to output Y (t) is a no longer

constant and depends on t. On the other hand, s must satisfy condition
0 ≤ s(t) ≤ 1, t ∈ [0, T ]. (1.33)
Now from (1.32) and (1.33), we have
x0 (t) + x(t)η + µx(t) ≤ ϕ(x(t)). (1.34)
Now we formulate the per capita consumption maximization problem. Con-

sider per capita consumption function c(t)
C(t) (1 − s)F (K, L)

c(t) = = .
L(t) L
We can now easily check that
c(t) = ϕ(x(t)) − x0 (t) − (µ + η)x(t), ∀t ∈ [0, T ]

Now we consider the average per capita consumption over the period T .
1 T 1 T
Z Z
ϕ(x) − x0 − (µ + η)x dt

J= c(t)dt =
T 0 T 0
Then the average per capita consumption maximization problem has the
following form:
max J (1.35)
subject to:
x0 (t) + (µ + η)x(t) − ϕ(x(t) ≤ 0, ∀t ∈ [0, T ],

(1.36)
x0 (t) ≥ 0, ∀t ∈ [0, T ].
Introducing variables u1 ≥ 0 and u2 such that x0 (t) = u1 , u2 ≥ 0, we have
1 T
Z

max J = ϕ(x) − u1 − (µ + η)x dt
T 0


 x0 = u1
 u1 + x(µ + η) − ϕ(x) + u2 = 0, ∀t ∈ [0, T ]


u1 ≥ 0
u2 ≥ 0




x(0) = x0 ,

where the interval [0, T ] is fixed and x0 is not fixed.

This problem can be written also as
1 T
Z
max J = u2 dt (1.37)
T 0


 x0 = u1
 (µ + η)x − ϕ(x) + u1 + u2 = 0, ∀t ∈ [0, T ]


u1 ≥ M (1.38)
u2 ≥ 0




x(0) = x0 ,

This problem (1.37)-(1.38) is an optimal control problem with phase and

control constraints as well as unknown initial condition.
Remark 1.1. Now we consider a simple case when x(t) ≡ a is a constant.
Then the problem (1.37)-(1.38) reduces to the following problem:
1 T
Z

max J = ϕ(a) − (µ + η)a dt,
T 0
or
max J = ϕ(a) − (µ + η)a.
1.3 Optimal Control Approach to the Solow Growth Theory 17
A solution to this problem satisfies the condition
ϕ0 (a) = (µ + η). (1.39)
On the other hand, taking into account (1.32) we have
sϕ(a) = (µ + η)a. (1.40)
From (1.39)-(1.40), we obtain
ϕ0 (a)a
s=
ϕ(a)
which is called the golden rule of accumulation (1.2). If the production

function F (K, L) has the form F (K, L) = AK α L1−α (0 ≤ α ≤ 1), then it
can be easily shown that s∗ = α.
Remark 1.2. If the production function F has the form F (K, L, t), then the
average per capita consumption problem is formulated as follows.
1 T
Z
F (x, L, t)
max J = (1 − u) dt dt
T 0 L(t)

 x0 = uF (x, L, t) − µx
x(0) = x0
0 ≤ u(t) ≤ 1, u ∈ [0, T ],

with K(t) = x(t), s(t) = u(t), t ∈ [0, T ] and given function of L(t) and
parameter µ.
Bibliography
[1] David Romer.,(1996) Advanced Macroeconomics, University of Cali-

fornia, Berkeley .
[2] Enkhbat.R.,(2009)Quasiconvex Programming and its Applications,

Lambert Publisher,Germany.
[3] Gregory Mankiw.H.,(2000) Macro Economics, Harvard University,

Worth Publisher .
[4] Horst,R.,Panos M.Pardalos., N.Thoat.,(1995)Introduction to Global

Optimization, Kluwer Academic Publishers .
[5] Robert Solow.,(1956) A Contribution to the Theory of Economic

Growth, Quarterly Journal of Economics, 70 , pp.65-94.
[6] Vasiliev,O.V.,(1996) Optimization Methods, World Federation Pub-

lisher .
[7] R.Enkhbat, D.Bayanjargal and A.Griewank.(2010) Global Optimiza-

tion Approach to the Solow Growth Theory, Advanced Modeling
and Optimization, An Electronic International Journal, pp.133-140,
Vol.12, Number 2.
2
Utility Maximization Problem
20 Utility Maximization Problem

Problem
Introduction
U (x) → max, x ∈ D (2.1)
n
X
D = {x ∈ Rn+ pi xi = m, xi ≥ 0, i = 1, ..., n},
i=1
where x = (x1 , x2 , ..., xn ) is a vector of basket of produced goods (i, ..., s)

and resourced (j=s+1,. . .,n) to be consumed and supplied by a consumer,
U : Rn+ → R is the utility function representing the preference of an individ-
ual, p = (p1 , p2 , ..., pn ) is a vector of market prices of goods and resources,
and m is the individual income.
In problem (2.1) a consumer chooses a basket maximizing his utility

function over budget constrained set D. The utility function U (x) is al-
ways concave due to the marginal utility diminishing law [21], and typically
assumed to be differentiable over D, and variables xj are continuous, rep-
resenting divisible and continuous goods in the perfect competitive market.
In this context, problem (2.1) is referred to as the classical model of utility
maximization [7], [10], [11], [15], [19]. The utility maximization problem
plays a key role in the consumer theory which is an important part of the
general equilibrium in economics. Solution xj = xj (p, m), j = 1, ...s give de-
mands for goods of a consumer, and solutions xj = xj (p, m), j = s + 1..., m
constitute supplies of a resources by a consumer for firms. The sum of all
such demands and supplies over all consumer constitutes a market demand
and supply from the consumer side.
Depending on types of goods, market and times, the UMP further can
be split into:
1. Continuous utility maximization problem (CUMP)
2. Parametric utility maximization problem (PUMP)
3. Nonconvex utility maximization problem (NUMP)
4. Discrete utility maximization problem (DUMP)
Consider utility maximization problem:

U (x) −→ max, x ∈ D,
(CU M P ) : (2.2)
D = {x ∈ Rn ni=1 pi xi = m, xi ≥ 0, i = 1, ..., n}
P
Problem 21
U : Rn+ −→ R is concave and differentiable function.
Problem (CUMP) is equivalently presented as the following convex min-
imization problem so that one can apply well developed methods and algo-
rithms of convex programming for solving this problem.
f (x) = −U (x) −→ min, x ∈ D, (2.3)

Clearly f and D are convex. For solving problem (2.2) numerically, we
propose a method of feasible direction [6]. The algorithm of this method is
the following.
Algorithm CUMP 1.
Step 1. Choose a feasible point x0 ∈ D. Set k := 0

Step 2. Solve the following linear Programming problem:
0
hf (xk ), xi −→ min, x ∈ D.
Let xk be a solution of this problem.
0 k (xk )
(h,̇i̇) denotes the scalar product x and f (xk ) = ∂f∂x(x1 ) , · · · , f∂xn
.
Step 3. If hf 0 (xk ), x̄k − xk i = 0 then solution x∗ = xk and terminate.
Step 4. Solve one dimensional minimization problem:
ϕ(α) = f (xk + α)(x̄k − xk ) −→ min, α ∈ [0, 1].

Let αk be a solution to this problem.
Step 5.Construct a next approximation point
xk+1 as xk+1 := xk + α(x̄k − xk ).

Step 6. Let k := k + 1
and return Step 2.
The convergence of this algorithm is due to the following theorem.
Theorem 2.1. [6] Assume that f : Rn+ −→ R is convex and differentiable,

0
and the gradient f (x) satisfies the lipschitz condition on D,i.e.
kf 0 (x) − f 0 (y)k ≤ Lkx − yk, (x, y) ∈ D, L > 0

Then the sequence xk , k = 0, 1, ..., generated by Algorithm CUMP 1 is a
minimizing sequence for problem (2.2), that is lim f (xk ) = min f (x).
k−→∞ x∈D
Remark 2.1. Since the feasible set D is simple, the solution of linear pro-
gramming problem in step2 can be found as follows:
minx∈D hf 0 (xk ), xi = min{hf 0 (xk ), z 1 i, · · · , hf 0 (xk ), z n i},
where z j = (0, 0, · · · , pmj , 0, · · · , 0), j = 1, 2, ..., n
Remark 2.2. It is clear that the function ϕ(α) is convex and therefore, one
dimensional minimization problem in step 4 can be solved by the bisection
method [24].
Now we consider the continuous utility maximization problem with a
nondifferentiable continuous concave function U on Rn+ . Then problem (2.2)
reduces to nondifferential optimization problem. In order to propose an
algorithm for this case, we need to introduce some notions from convex
analysis. Let Q be a closed and convex set and x0 ∈ Rn . Consider the
convex minimization problem
1
ψ(x) = kx − x0 k2 −→ min, x ∈ Q, (2.4)
2
Definition 2.1. The solution to problem (2.4) is called projection of point
x0 onto the set Q.
Denote by kPD (x0 ) − x0 k = min kx − x0 k.
x∈D
Problem (2.4) can be also solved by the method of feasible direction
using Algorithm CUMP1 with functions ψ(x) and ψ 0 (x) = kx − x0 k.
Definition 2.2. Let f : Rn+ −→ R be a continuously concave function. A
vector g is called a subgradient of function f at point x0 ∈ Rn+ if ((x) −
f (x0 ) ≥ hg, x − x0 i holds for all x ∈ Rn+ . The set of all subgradients of f at
x0 , ∂f (x0 ), is called the subdifferential of function f at point x0 .
Theorem 2.2. [18]. ∂f (x0 ) is a convex compact set.
Example 2.1. Let functions fi : Rn+ −→ R, i = 1, ..., m be continuously
convex. Then f (x) = max fi (x) is continuous and convex with
1≤i≤m
n
∂f (x) = conv{∂fi (x)|i ∈ I(x)} = {z ∈ Rn |z = αi z i , z i ∈ ∂fi (x), i ∈
P
i=1
n
P
I(x), αi = 1, 0 ≤ αi ≤ 1}, where I(x) = {i : fi (x) = f (x)}
i=1
Now we are ready to present an algorithm for solving CUMP for the non
differentiable case.
Algorithm CUMP 2.
Problem 23
Step 1. Choose a x0 ∈ D and a sequence εk such that
∞
P
εk > 0, εk −→ 0, εk = ∞. Small parameter δ > 0 is given. Set k := 0
k=0
Step 2. Compute f (xk ) and find a g ∈ ∂f (xk ).
Step 3. Construct xk+1 as projection:
g
xk+1 := PD (xk − εk ).
kgk
Step 4. If |f (xk+1 ) − f (xk )| ≤ δ then terminate and x∗ := xk is a global
minimizer to problem (2.2)
Step 5.Set k := k + 1 and return Step2.
The rate of convergence of this algorithm is given by the following state-

ment.
Theorem 2.3. [18] Let f be Lipschitz continuous on D with constant L > 0
and
x0 ∈ B(x∗ , r) = (x ∈ Rn |kx − x∗ k < r)

.
Then
k
r2 + ε2i
P
i=1
f (xk ) − f (x∗ ) ≤ L k
P
2 εi
i=0
.
where x∗ ∈ arg min f (x).
x∈D
Example 2.2. We consider the following example solved by Algorithm CUPM1.
lnu(x) = 0.16121ln(x1 − 9.57493) + 0.16517ln(x2 − 2.7174)

+0.326ln(x3 − 8.494) + 0.2412ln(x4 − 11.79) → max,
5
X
p i xi = m
i=1
(p1 = 2492, p2 = 934, p3 = 4909, p4 = 703, m = 50000)
Solution: x∗ = {14.454, 7.457, 17.518, 12.041}

Example 2.3.
lnu(x) = 0.16121ln(x1 − 9.57493) + 0.16517ln(x2 − 2.7174)

+0.326ln(x3 − 8.494) + 0.1412ln(x4 − 11.79)+
0.205ln(x5 − 0.8779)) → max,
5
X
p i xi = m
i=1
(p1 = 2492, p2 = 934, p3 = 4909, p4 = 703, p5 = 554, m = 150000)
Solution: x∗ = {17.5558, 7.5558, 17.5558, 12.5558, 7.5558}

u∗ = 0.1084

Problem
Observations on economic variables such as utility, price and income made
over time lead to parametric optimization. In fact, the general equilibrium
theory was developed for economic processes operating per a unit time.
Consider one of the consumer theory problem is utility maximization prob-
lem in parameter (time) of t or its equivalent minimization problem:
f (x, t) = −U (x, t) → min, t ∈ [tA , tB ] (2.5)

n
X
pi xi = I, (2.6)
i=1
xi ≥ 0, i = 1, 2, ..., n, (2.7)
n
where consumer’s utility function U : R+ → R is strictly concave and
twice differentiable with respect to x at each t ∈ [tA , tB ], I(t) is con-
sumer’s income at moment t, pi (t) is price of goods i per unit. Function
pi : R+ → R+ , i = 1, 2, ..., n are assumed to be continuous. t is time as
parameter, t ∈ [tA , tB ]. xi , i = 1, 2, ..., n is quantity of good i a consumer
purchases at moment t. Parametric utility maximization problem is a con-
vex programming problem and has a unique solution at each t ∈ [tA , tB ],
I(t). This problem is a hard parametric optimization problem.
Problem 25
Lemma 2.1. Assume that the utility function U (x, t) is a differentiable with
respect to t and satisfies the Lipschitz condition with constant M for each
x ∈ D i.e.
|U (x, t̂) − U (x, t)| ≤ M |t̂ − t|, ∀t ∈ [tA , tB ]
Then for a given ε > 0, there exists discretization
tA = t0 < t1 < ... < ti−1 < ti < ti+1 < ... < tN = tB
such that
|U (x∗ (t), t) − U (x∗ (ti ), ti )| < ε, ∀ ∈ [tA , tB ] and certain ti
where
n
X
∗ n
U (x (t), t) = max U (x, t), t ∈ [tA , tB ], D = {x ∈ R | pi xi = m}.
x∈D
i=1
The proof is similar to the proof in Lemma 2.1 of Chapter 1.
Numerical Results
Consider the utility maximization problem:
n
Y
U M : u(x) = (xi − ci )αi → max, t ∈ [tA , tB ], (2.8)
i=1
n
X
pi (t)xi = I(t), (2.9)
i=1
xi ≥ 0, i = 1, 2, ..., n. (2.10)
where consumer’s utility function u(x) is strictly concave and twice differ-
entiable with respect to x at each t ∈ [tA , tB ], xi i = 1, 2, ..., n is a quantity
of good i and ci i = 1, 2, ..., n is minima required quantity of good i, I(t) is
consumer’s income at moment t, pi (t) is price of goods i per unit. Function
pi : R+ → R+ , i = 1, 2, ..., n are assumed to be continuous. t is time as
parameter, t ∈ [tA , tB ]. xi , i = 1, 2, ..., n is quantity of good i a consumer
purchases at moment t.
Problem (2.8) called Stone-Geary’s utility function (SG function) in mi-
croeconomic. We estimated parameters SG function for 5 goods using time
series data of statistics of Mongolia and solved for the following cases.
Example 2.4. Let pi (t) be a price of goods i and I(t) be a consumer’s bud-
get. Then we can formulate the following problem:
ln|u(x)| = −(0.16121ln(x1 − 9.57493) + 0.16517ln(x2 − 2.7174)

+0.326ln(x3 − 8.494) + 0.1412ln(x4 − 11.79)+ (2.11)
+0.205ln(x5 − 0.8779)) → min, t ∈ [2010; 2014],
5
X
pi (t)xi = I(t), (2.12)
i=1


 p1 (t) = 26.21t − 46170
p2 (t) = 48.11t − 95767




p3 (t) = 220.3t − 43934


 p4 (t) = 26.4t − 52361
p (t) = 419.9t − 83665

 5



I(t) = 1341.4t − 2.6727e + 006
xi ≥ 0, i = 1, 2, ..., n. (2.13)
Example 2.5.
ln|u(x)| = −(0.16121ln(x1 − 9.57493) + 0.16517ln(x2 − 2.7174)

+0.326ln(x3 − 8.494) + 0.1412ln(x4 − 11.79)+ (2.14)
+0.205ln(x5 − 0.8779))) → min, t ∈ [2011; 2015],
5
X
pi (t)xi = I(t), (2.15)
i=1


 p1 (t) = c1 t2 + c2 t + c3 (c1 = 5.2132, c2 = −20849, c3 = 2.0846e + 007)
p2 (t) = c4 t2 + c5 t + c6 (c4 = 10.849, c5 = −43383, c6 = 4.3371e + 007)




p3 (t) = c7 t2 + c8 t + c9 (c7 = 20.672, c8 = −82558, c9 = 8.243e + 007)


 p4 (t) = 0.91792t3 − 5513.6t2 + (1.104e + 007)t − 7.3678e + 009
 p5 (t) = 3.2037t3 + −19226t2 + (3.8459e + 007)t + −2.5644e + 010



I(t) = 10.019t3 − 60068t2 + (1.2005e + 008)t + −7.9975e + 010

xi ≥ 0, i = 1, 2, ..., n. (2.16)
The numerical results obtained on MATLAB are given in the following

tables, respectively.
Problem 27
Table 2.1: Example 2.4

time(t) x∗1 x∗2 x∗3 x∗4 x∗5 umax
2010 0.267 0.64 0.034 10.7636 0.0001 1.006
2011 0.0845 0.7381 0.0380 12.8164 0.0000 1.000
2012 0.1875 0.5654 0.0438 8.7568 0.0018 1.9985
2013 0.2817 0.7042 0.0420 12.8187 0.0000 1.0017
2014 0.2817 0.7042 0.0420 12.8187 0.0001 1.1495
Table 2.2: Example 2.5

time(t) x∗1 x∗2 x∗3 x∗4 x∗5 umax
2011 3.9798 1.7282 1.2753 4.6522 0.9159 1.6695
2012 5.1683 0.1901 2.5762 4.9632 1.8201 1.7825
2013 2.1996 0.6598 5.2251 4.5652 0.6817 1.0715
2014 2.0621 1.0025 5.2830 4.7399 1.1805 1.1495
2015 2.0621 1.0025 5.2830 4.7399 1.1805 1.1495

Problem
Sometimes nonconvexities arise in problem (CUMP) under certian non-
competitive market, the consumer influences the market price through the
amount that he purchases. Thus, if the market for good {i} is oligopsonistic
in this way, then pi is a function of xi , i.e, pi = pi (xi ). If markets for all
goods are oligopsonistic, then the utility maximization problem is:
U (x) −→ max, x ∈ D,
n
(N U M P ) : (2.17)
D = {x ∈ Rn | pi (xi )xi = m, xi ≥ 0, i = 1, ..., n}
P
i=1
where U : Rn+ −→ R is concave and functions pi : R+ −→ R, i = 1, ..., n,

are continuous, D is a subset in Rn+ .
We shall call this problem as the nonconvex, utility maximization prob-
lem (NUMP). (NUMP) can be equivalently presented as the minimization
problem of a convex function over a set D.
−f (x) = −U (x) −→ min, x ∈ D (2.18)

Problem (2.17) is nonconvex due to the nonconvex constraint D. In
general, there is no efficient algorithm for solving this problem globally.
However, under certain assumption, we have global optimality conditions
for problem (2.17), and then we can develop an algorithm based on this.
Consider the problem:
f (x) = U (x) −→ max, x ∈ D,

n
(2.19)
D = {x ∈ Rn | pi (xi )xi = m, xi ≥ 0, i = 1, ..., n}
P
i=1
where f : Rn+ −→ R is strongly concave and continuously differentiable,

pi : Rn+ −→ R, (i = 1, ..., n) are continuous and D is a compact set in Rn+ .
Denote by x∗ the unique global maximizer of f on Rn which exists due to
the strong convexity [4, 6] that is:
f (x∗ ) = maxn f (x).

x∈R
Define the set Ec (f ):
Ec (f ) = {y ∈ Rn |f (y) = c}.
Theorem 2.4. If x∗ ∈
/ D and max kf 0 (x)k > 0 then z ∈ D is a global
x∈D
maximizer of problem (2.19) if ad only if
hf 0 (x), x − yi ≤ 0 (2.20)
for all y ∈ Ef (z) (f ) and x ∈ D.
Proof. Necessity. Assume that z is a solution of problem (2.19).

Consider x ∈ D and y ∈ Ef (z) (f ). Then by the concavity of f , we have
0 < f (z) − f (x) = f (z) − f (x) ≤ hf 0 (x), y − xi.
Sufficiency. Let us prove the assertion by contradiction. Assume that

(2.20) holds and exists a point u ∈ D such that
f (u) > f (z).
Clearly, f 0 (u) 6= 0 by assumption max kf 0 (x)k > 0. Now define uα as follows

x∈D
for α > 0:
uα = u + αf 0 (u).
Then, by the concavity of f , we have
2
f (uα ) − f (u) ≤ hf 0 (u), uα − ui = α kf 0 (x)k ,
Problem 29
which implies
2
f (uα ) ≤ f (u) + α kf 0 (x)k < f (u).
Then find α = ᾱ such that
2
f (u) + α kf 0 (x)k = f (z),
that is,
f (u) − f (z)
ᾱ = > 0.
kf 0 (x)k2
Thus we get
2
f (uα ) ≤ f (u) + ᾱ kf 0 (x)k = f (z) < f (u).
Define a function h : Rn → R as
h(α) = f (u + αf 0 (u)) − f (z),
where R+ = {α ∈ R|α ≥ 0}. It is clear that h is continuous on [0, +∞).
Note that h(ᾱ) ≥ 0 and h(0) < 0. There are two cases with respect to the
value of h(ᾱ) which we should consider.
case 1. h(ᾱ) = 0 ( or f (u + αf 0 (u)) = f (z)), then
2
hf 0 (u), u − uᾱ i = hf 0 (u), ᾱf 0 ui = ᾱ kf 0 (u)k > 0
contradicting condition (2.20).
case 2. h(ᾱ) > 0 and h(ᾱ) < 0. Since h is continuous, there exists a point
α0 ∈ (0, ᾱ) such that h(α0 ) = 0 (or f (u + α0 f 0 (u)) = f (z).) Then we have
2
hf 0 (u), u − uα0 i = α0 kf 0 (u)k > 0,
again contradicting condition (2.20).
Thus, in both cases, we find contradictions, proving the theorem. 2
Remark 2.3. Theorem 2.4 states that, in order to conclude that a point
z ∈ D is not a solution of (2.19) if is sufficient to have a pair (v, y) ∈ Rn
such that
hf 0 (v), v − yi < 0, y ∈ Ef (z) f and v ∈ D.

Before we describe an algorithm for (NUMP), we define the function
g(y) an ψ(z) as follows:
g(y) = maxhf 0 (x), x − yi, y ∈ Rn ,

x∈D
ψ(z) = max g(y), z ∈ D.

y∈Ef (z) (f )
Now we can reformulate Theorem 2.4 in the following.

Theorem 2.5. Assume that f : Rn+ −→ R is strongly concave and continu-

ously differentiable. Let D be a compact set in Rn+ , and x∗ ∈
/ D. If ψ(z) = 0
then point z is a global maximizer of problem (2.19).
Proof. This is an obvious consequence of the following relations.
0 = ψ(z) ≥ p(y) ≥ hf 0 (x), x − yi,
which are fulfilled for all y ∈ Ef (z) (f ) and x ∈ D. The proof is complete. 2
Now we are ready to present an algorithm for solving problem (2.19).

We also suppose that one can efficiently solve the problem of computing
maxhf 0 (x), x − yi for any given y ∈ Rn .
x∈D
Algorithm NUMP
Step 1. Choose a feasible point x0 ∈ D. Set k:=0.

Step 2. Solve the following problem
g(y) −→ max, y ∈ Ef (xk ) (f ).
Let y k be a solution to this problem ,i.e.

0
g(y k ) = maxhf (x), x − y k i = max g(y),
x∈D y∈Ef (xk )
and let xk+1 realizes g(y k ):

0
ψ(xk ) = hf (xk+1 ), xk+1 − y k i.
Step 3. If ψ(xk ) = 0 the output x∗ := xk and terminate. Otherwise, let
k := k + 1 and return step2.
The convergence of this algorithm is based on the following theorem.
Theorem 2.6. [4] The sequence xk , k = 0, 1, ..., generetad by Algorithm

NUMP is a maximizing sequence for problem (2.19), that is,
lim f (xk ) = max f (x),

k−→∞ x∈D
and an accumulation of the sequence xk is a global maximizer of (2.19).
Proof. From the construction of the sequence {xk }, we have xk ∈ D

and f ∗ ≥ f (xk ) for all k, where f ∗ = f (x∗ ) = max f (x). Clearly, f 0 (x) 6= 0
x∈D
Problem 31
by assumption. Also, note that for all y ∈ Ef (z) (f ) and x ∈ D, we have
ψ(xk ) = max maxhf 0 (x), x − yi ≥ hf 0 (x), x − yi ≥ 0.

y∈Ef (z) (f ) x∈D
If there exists a k such that ψ(xk ) = 0 then, by Theorem 2.5, xk is a

solution to problem (2.20) and, in case the proof is complete. Therefore,
without loss generality, we can assume ψ(xk ) > 0 for all k, and prove the
theorem by contradiction. If the assertion is false, that is, xk is not a
maximizing sequence for problem (2.20), the following inequality holds.
lim sup f (xk ) < f ∗ (2.21)

k→∞
By the definition of ψ(xk ) and Algorithm NUMP, we have
p(y k ) = ψ(xk ) = max maxhf 0 (x), x − yi = hf 0 (xk+1 ), xk+1 − y k i

and f (xk ) = f (y k ). The concavity of f implies that
f (xk ) − f (xk+1 ) = f (y k ) − f (xk+1 ) ≤

(2.22)
≤ hf 0 (xk+1 ), xk+1 − y k i = ψ(xk ) < 0
Hence, we obtain f (xk ) < f (xk+1 ) for all k, and the sequence {f (xk )} is
strictly increasing. Since the sequence is bounded from below by f ∗, it has
and satisfies
lim (f (xk ) − f (xk+1 )) = 0 (2.23)

k→∞
Then, from (2.22) and (2.23), we obtain
lim ψ(xk ) = 0.
k→∞
From (2.21) we have f (xk ) < f ∗ for all k. Now define vα as follows:
vα = x∗ + αf 0 (x∗ ).
Then by the concavity of f , we have

2
f (vα ) − f (x∗ ) ≤ hf 0 (x∗ ), vα − x∗ i = α kf 0 (x)k ,
which implies
2
f (vα ) ≤ f (x∗ ) + α kf 0 (x)k < f (x∗ ). α > 0.
Choose α = αk such that

2
f (x∗ ) + αk kf 0 (x)k < f (xk )
that is,
f (xk ) − f (x∗ )
αk < < 0.
kf 0 (x)k2
Define a function hk : R+ → R as
hk (α) = f (x∗ + αf 0 (x∗ )) + f (xk ),
where R+ = {α ∈ R|α ≥ 0}. It is clear that hk is continuous on [0, +∞).

Note that hk (αk ) > 0 and hk (0) > 0. Since hk is continuous, there exists
a point α¯k ∈ (0, αk ) such that hk (α¯k ) = 0, that is f (vα¯k ) = f (xk ) and
vα¯k = x∗ + α¯k f 0 (x∗ ). Also, note that
ψ(xk ) = max maxhf 0 (x), x − yi ≥ hf 0 (x∗ ), x∗ − vα¯k i.

Taking into account vα¯k = x∗ + α¯k f 0 (x∗ ), we have
ψ(xk ) = hf 0 (x∗ ), x∗ −vα¯k i = kf 0 (x∗ )k kvα¯k − x∗ k ≤ max kf 0 (x∗ )k kvα¯k − x∗ k < 0.

x∈D
Since lim ψ(xk ) = 0, this implies

k→∞
lim vᾱk = x∗ .
k→∞
The continuity of f on Rn yields
lim f (xk ) = lim f (vᾱk ) = f (x∗ ). (2.24)

k→∞ k→∞
which is contradiction to (2.21).

Consequently, {xk } is a maximizing sequence for problem (2.18). Since D
is compact, we can always select the convergent subsequences {xkl } from
{xk } such that
lim xkl = x̄ ∈ D.
l→∞
Then together with (2.24), we obtain
lim f (xkl ) = f (x̄) = f ∗ ,

l→∞
which is completes the proof. 2

2.4 Discrete Utility Maximization Problem 33
2.4 Discrete Utility Maximization Problem

In previous sections we have considered the utility maximization problem
with continuous variables or goods. However, if goods are discrete goods
such as cars, televisions computers and so on. In this case the problem is:

U (x) −→ max,Px ∈ D,
(DU M P ) :
D = {x ∈ Rn+ | ni=1 pi xi = m, xi nonnegative integers i = 1, ..., n, }.
(2.25)
We shall call this problem as the discrete utility maximization prob-
lem (DUMP). The problem belongs to combinatorial optimization problem
which is, in general, NP-hard. Solutions of (DUMP), however have not
only theoretical but also practical importance in economics, as they enable
economists to work with demands and supplies for discrete goods.
Consider the discrete utility maximization problem of the type:
n
X
U (x) = ui (xi ) −→ max, (2.26)
i=1
subject to:
n
P
pi xi = m,
i=1 (2.27)
xi nonnegative integers, i = 1, ..., n
where ui : R+ −→ R, i = 1, ..., n, are concave functions, and the positive
parameters pi and m are given. Such separable utility function is justified
in microeconomics by the fact that total preference is the sum of all sepa-
rate preference uj to each good j. This problem is known also ”knapsack”
problem (2.25). For solving (2.26), we use dynamic programming approach
[6]. The general recursive equation for this problem is given by
f1 (y1 ) = maxh i{u1 (x1 )}, y1 = 0, 1, ..., m,

m
x1 =0,1,... p1 (2.28)
p 1 x1 ≤ y 1
fi (yi ) = maxh i{ui (xi ) + fi−1 (yi − pi xi )}
m
xi =0,1,... pi
(2.29)
pi xi ≤ yi
yi = 0, ..., m,
fn (m) = max U (x),
Pn
pi xi = m, (2.30)
i=1
xi ≥ 0
h i h i
m
where pi
is the largest integral amount included in m
pi
.
It has been shown in [23] that formulas (2.28)-(2.30) generate the global
solution to problem (2.26).
Now let’s solve the following problems by dynamic programming:
Problem 1.
U = 0.8 log(x1 − 15) + 0.2 log(x2 − 15) → max

Constraints:
200x1 + 250x2 = 3700
xi : integers xi ∈ [0; 30]
Solution: x∗ = {9, 8}.
Problem 2.
U = 0.5 log(x1 − 15) + 0.2 log(x2 − 15) + 0.3 log(x3 − 11) → max
Constraints:
500x1 + 300x2 + 100x3 = 5000

xi , i = 1, 2, 3, are integers.
Solution: x∗ = {10, 8, 12}
We consider the following utility maximization problem

n
Y
U (x) = xi αi → max
i=1
n
X
p i xi = I
i=1
n
X
xi − nonnegative integers, where αi ≥ 0, αi = 1
i=1
2.4 Discrete Utility Maximization Problem 35
The dynamic programming procedure for this problem is as follows:

α1
f1 (y1 ) = maxh i{x1 }, y1 = 0, 1, ..., m.
m
x1 =0,1,... p1 (2.31)
p 1 x1 ≤ y 1
αi
fi (yi ) = maxh i{xi fi−1 (yi − pi xi )}
m
xi =0,1,... pi
(2.32)
p i xi ≤ y i
yi = 0, ..., m,
fn (m) = n
max U (x),
P
pi xi =m (2.33)
i=1
xi ≥ 0
Problem 3.
U (x) = x0.1 0.2 0.3 0.2 0.1 0.1

1 x2 x3 x4 x5 x6 → max,
2x1 + 7x2 + 3x3 + 4x4 + 5x5 + 6x6 = 58.

Formulas (2.31)-(2.33) give the following integer solutions and value:
x∗ = (3, 2, 5, 3, 1, 1), Umax = 2.5884
Problem 4.
U = 0.1 log(x1 − 10) + 0.2 log(x2 − 8) + 0.3 log(x3 − 2)+
+0.1 log(x4 − 12) + 0.1 log(x5 − 8) → max
20x1 + 30x2 + 40x3 + 50x4 + 60x5 = 10000
xi , (i = 1, 5) are nonnegative integers.
Optimal solutions are : x∗ = {0, 19, 0, 19, 19} and Umax = 2.5015
Bibliography
[1] Arrow,K.J. and Enthoven, A.C., Quasiconcave Programming.,

Econometrica 29, 779-800, 1961.
[2] Avreil,M., Diewert, W.E., Schaible, S. and Zang, I. Generalized Con-

cavity, (forthcoming)
[3] Avreil,M., Diewert, W.E., Schaible, S., and Ziemba, W.T. Introduc-
tion to Concave and Generalized Concave Functions This volume,
21-50, 1981.
[4] Afriat, S.N.The Construction of Utility Function from Expenditure

Data ., International Economic Review 8, 67-77, 1967.
[5] Bazaraa,M.S. and Shetty, C.M. Foundations of Optimization.

Springer-Verlag, Berlin, 1976.
[6] Bertsekas,D.P., Nolinear Programming., 2nd edition, Athena Scien-

tific, Belmont, MA,1999.
[7] Debreu, G. Theory of Value. John Wiley and Sons, New York, 1959.
[8] Deaton and Muellbauer., Understanding consumption., London.,

1980a.
[9] Deaton and Muellbauer., An Almost Ideal Demand System., The

American Economic Review, Vol 70, 1980b.
[10] Diewert, W.E. Applications of Duality Theory. In Frontiers of Quan-

titative Economics, Vol.II. (M.D. Intriligator and D.A. Kendrick,
eds.) North-Holland Publishing Company, Amsterdam, 1974.
[11] Diewert, W.E. Hicks Aggregation Theorem and the Existence of a

Real Value Added Function. In Production Economics: A Dual Ap-
proach to Theory and Applications, Vol.2. (M. Fuss and D. McFad-
den, eds.) North-Holland Publishing Company, Amsterdam, 1978a.
BIBLIOGRAPHY 37
[12] Duvobitzki A.Y. and Milutin A.A. Problem in extremum., Moscow,

1965.
[13] Ferland, J.A. Mathematical Programming Problems with Quasi-

convex Objective Functions. Mathematical Programming 3, 296-301,
1972.
[14] Guddat.J., Guerra Vasquet.F.,and Jongen.H.TH., Parametric Opti-

mization: Singularities, Parthtollowing and Jumps, Jonhn Wiley and
Sons, New-York,1990.
[15] Hicks, J.R. Consumer’s Surplus ans Index-Numbers. Review of Eco-

nomic Studies 9, 126-137, 1941-2.
[16] Katzner, D.W. Static Demand Theory. MacMillan, New York, N.Y.
[17] Karlin, S. Mathematical Methods and Theory in Games, Program-

ming and Economics. Vol. I. Addison-Wesley, Palo Alto, California,
1959.
[18] Nesterov Yurii., Introduction lectures on Convex Optimization,

Kluwer Academic Publishers, London, 2004.
[19] Mangasarian, O.L. Pseudo-Convex Functions. SIAM Journal Control

3, 281-290, 1965.
[20] Mangasarian, O.L. Nonlinear Programming. McGraw-Hill Book Co.,

New York, N.Y., 1969.
[21] Shephard, R.W. Theory of Cost and Production Functions. Princeton

University Press, Princeton, N.J., 1970.
[22] Surish, P.A. Optimal control theory. New-York, 2009.
[23] Taha., A.H,, Operations research. New Jersy, 2005.
[24] Vasiliev,O.V., Optimization Methods., World Federation Publisher.

1995.
3
Game Theory
3.1 The Curvilinear Search Algorithm for Solving Three-Person Game 39
3.1 The Curvilinear Search Algorithm for Solv-

ing Three-Person Game
Introduction
It is well known that game theory plays an important role in economics, op-
timization and operations research. Game theory is found to be a powerful
mathematical tool for modeling of firm competitions at oligopoly markets
where each firm maximizes its own profit using the same price which de-
pends on the sum of quantities produced by the firms. Existence of Nash
equilibrium points have been proved in [12, 15]. In recent years, compu-
tational methods of game theory or equivalently, finding Nash equilibrium
points in various games have been intensively studying in the literature
[1, 3, 4, 5, 7, 8, 9, 10, 11, 13, 14, 16]. As usual, finding Nash equilibrium
points in zero-sum games leads to linear programming while in non-zero sum
game it requires solving nonconvex optimization problems. Global search
for Nash equilibrium points have been mainly studied for polymatrix [28]
and hexamatrix games by global optimization techniques [1, 13, 16].
But it seems to us that a little attention has been paid to computational
aspects of non-zero sum three-person game.
3.1.1 Non-zero sum three-person game

Consider the three-person game in mixed strategies with payoff matrices
(A, B, C) for players 1,2 and 3.
A = (aijk ), B = (bijk ), C = (cijk ),
i = 1, . . . , m, j = 1, . . . , n, k = 1, . . . , s.
Denote by Dq the set
q
( )
X
q
Dq = u∈R ui = 1, ui ≥ 0, i = 1, . . . , q .
i=1
A mixed strategy for player 1 is a vector x = (x1 , x2 , . . . , xm ) ∈ Dm rep-

resenting the probability that player 1 uses a strategy i. Similarly, the
mixed strategies for players 2 and 3 are y = (y1 , y2 , . . . , yn ) ∈ Dn and
z = (z1 , z2 , . . . , zs ) ∈ Ds . Their expected payoffs are given by:
m X
X n X
s
f1 (x, y, z) = aijk xi yj zk ,
i=1 j=1 k=1
40 Game Theory
m X
X n X
s
f2 (x, y, z) = bijk xi yj zk ,
i=1 j=1 k=1
m X
X n X
s
f3 (x, y, z) = cijk xi yj zk .
i=1 j=1 k=1
Definition 3.1. A triple of mixed strategies x∗ ∈ Dm , y ∗ ∈ Dn , z ∗ ∈ Ds ,

is a Nash equilibrium if

 f1 (x∗ , y ∗ , z ∗ ) ≥ f1 (x, y ∗ , z ∗ ), ∀x ∈ Dm ,
∗ ∗ ∗ ∗ ∗
f2 (x , y , z ) ≥ f2 (x , y, z ), ∀y ∈ Dn ,
f3 (x∗ , y ∗ , z ∗ ) ≥ f3 (x∗ , y ∗ , z), ∀z ∈ Ds .

It is clear that
f1 (x∗ , y ∗ , z ∗ ) = max f1 (x, y ∗ , z ∗ ),
x∈Dm
f2 (x , y , z ) = max f2 (x∗ , y, z ∗ ),
∗ ∗ ∗
y∈Dn
f3 (x∗ , y ∗ , z ∗ ) = max f3 (x∗ , y ∗ , z).

z∈Ds
For further purpose, it is useful to formulate the following statement.

Theorem 3.1. A triple strategy (x∗ , y ∗ , z ∗ ) is a Nash equilibrium if and
only if
 Pm Pn Ps ∗ ∗ ∗
Pn Ps ∗ ∗
 Pi=1 Pj=1 Pk=1 aijk xi yj zk ≥ Pj=1 Pk=1 aijk yj zk , i = 1, 2, . . . , m,
m n s ∗ ∗ ∗ n s ∗ ∗
i=1 Pj=1 Pk=1 bijk xi yj zk ≥ j=1 Pk=1 bijk xi zk , j = 1, 2, . . . , n,
 Pm n s ∗ ∗ ∗
Pn s ∗ ∗
i=1 j=1 k=1 cijk xi yj zk ≥ j=1 k=1 cijk xi yj , k = 1, 2, . . . , s.
(3.1)
Proof. Necessity: Assume that (x∗ , y ∗ , z ∗ ) is a Nash equilibrium. Then by
definition 3.1, we have
m X
X n X
s n X
X s
aijk x∗i yj∗ zk∗ ≥ aijk xi yj∗ zk∗ , ∀x ∈ Dm , (3.2)
i=1 j=1 k=1 j=1 k=1
m X
X n X
s m X
X s
aijk x∗i yj∗ zk∗ ≥ aijk x∗i yj zk∗ , ∀y ∈ Dn , (3.3)
i=1 j=1 k=1 i=1 k=1
m X
X n X
s m X
X n
aijk x∗i yj∗ zk∗ ≥ aijk x∗i yj∗ zk , ∀z ∈ Ds . (3.4)
i=1 j=1 k=1 i=1 j=1
In the first inequality (3.16), successively choose x = (0, 0, . . . , 1, . . . , 0) with

1 in each of the m spots, in (3.17) choose y = (0, 0, . . . , 1, . . . , 0) with 1 in
each of the n spots, and in (3.18) choose z = (0, 0, . . . , 1, . . . , 0) with 1 in

each of the s spots. We can easily see that
n X
X s
f1 (x∗ , y ∗ , z ∗ ) ≥ aijk yj∗ zk∗ , i = 1, . . . , m,
j=1 k=1
m X
X s
∗ ∗ ∗
f2 (x , y , z ) ≥ bijk x∗i zk∗ , j = 1, . . . , n,
i=1 k=1
Xm X n
f3 (x∗ , y ∗ , z ∗ ) ≥ cijk x∗i yj∗ , k = 1, . . . , s.
i=1 j=1
Sufficiency: Suppose that for a triple (x∗ , y ∗ , z ∗ ) ∈ Dm ×Dn ×Ds , conditions

(3.15) are satisfied. We choose x ∈ Dm , y ∈ Dn and z ∈ Ds and multiply
(3.15) by xi , yj and zk respectively. We obtain
m
" m n s # m X n Xs
X XXX X
∗ ∗ ∗
xe aijk xi yj zk ≥ aijk xi yj∗ zk∗ ,
e=1 i=1 j=1 k=1 i=1 j=1 k=1
n
" m n s # m X
n X
s
X XXX X
ye bijk x∗i yj∗ zk∗ ≥ bijk x∗i yj zk∗ ,
e=1 i=1 j=1 k=1 i=1 j=1 k=1
m
" m X
n X
s
# m X
n X
s
X X X
ze cijk x∗i yj∗ zk∗ ≥ cijk x∗i yj∗ zk .
e=1 i=1 j=1 k=1 i=1 j=1 k=1
Pm Pn Ps
Taking into account that i=1 xi = j=1 yj = k=1 zk = 1 we have
f1 (x∗ , y ∗ , z ∗ ) ≥ f1 (x, y ∗ , z ∗ ), ∀x ∈ Dm ,
f2 (x∗ , y ∗ , z ∗ ) ≥ f2 (x∗ , y, z ∗ ), ∀y ∈ Dn ,
f3 (x∗ , y ∗ , z ∗ ) ≥ f3 (x∗ , y ∗ , z), ∀z ∈ Ds ,
which shows that (x∗ , y ∗ , z ∗ ) is a Nash equilibrium. The proof is complete.
Now we are ready to generalize Mills’s theorem [8] formulated originally
for the bimatrix game of two players for three-person matrix game as follows.
Theorem 3.2. A triple strategy (x∗ , y ∗ , z ∗ ) is a Nash equilibrium for the
non-zero sum three-person game if and only if there exist scalars (p∗ , q ∗ , t∗ )
such that (x∗ , y ∗ , z ∗ , p∗ , q ∗ , t∗ ) is a solution to the following nonconvex opti-
mization problem:
m X
X n X
s
max F (x, y, z, p, q, t) = (aijk + bijk + cijk )xi yj zk − p − q − t
(x,y,z,p,q,t)
i=1 j=1 k=1
(3.5)
42 Game Theory
subject to:
n X
X s
aijk yj zk ≤ p, i = 1, . . . , m, (3.6)
j=1 k=1
m X
X s
bijk xi zk ≤ q, j = 1, . . . , n, (3.7)
i=1 k=1
m X
X n
cijk xi yj ≤ t, k = 1, . . . , s, (3.8)
i=1 j=1
m
X
xi = 1, xi ≥ 0, i = 1, . . . , m,
i=1
n
X
yj = 1, yj ≥ 0, j = 1, . . . , n, (3.9)
j=1
s
X
zk = 1, zk ≥ 0, k = 1, . . . , s.
k=1
Proof. Necessity: Now suppose that (x∗ , y ∗ , z ∗ ) is a Nash equilib-

rium point. Choose scalars p∗ , q ∗ and t∗ as: p∗ = f1 (x∗ , y ∗ , z ∗ ), q ∗ =
f2 (x∗ , y ∗ , z ∗ ), and t∗ = f3 (x∗ , y ∗ , z ∗ ).
We show that (x∗ , y ∗ , z ∗ , p∗ , q ∗ , t∗ ) is a solution to problem (3.5)-(3.9). First,
we show that (x∗ , y ∗ , z ∗ , p∗ , q ∗ , t∗ ) is a feasible point for problem (3.5). By
Theorem 3.1, the equivalent characterization of a Nash equilibrium point,
we have  Pn Ps
 Pj=1 P k=1 aijk yj∗ zk∗ ≥ f1 (x∗ , y ∗ , z ∗ ),
m s ∗ ∗
i=1 Pk=1 bijk xi zk ≥ f2 (x∗ , y ∗ , z ∗ ),
 Pm n ∗ ∗
i=1 j=1 cijk xi yj ≥ f3 (x∗ , y ∗ , z ∗ ).
The rest of the constraints are satisfied because of x ∈ Dm , y ∈ Dn and
z ∈ Ds . It meant that (x∗ , y ∗ , z ∗ , p∗ , q ∗ , t∗ ) is a feasible point. Choose any
x ∈ Dm , y ∈ Dn , z ∈ Ds and multiply (3.19)-(3.21) by xi , yj and zk
respectively. If we have sum up these inequalities, we obtain
m X
X n X
s
f1 (x, y, z) = aijk xi yj zk ≤ p,
i=1 j=1 k=1
m X
X n X
s
f2 (x, y, z) = bijk xi yj zk ≤ q,
i=1 j=1 k=1
m X
X n X
s
f3 (x, y, z) = cijk xi yj zk ≤ t.
i=1 j=1 k=1
Hence, we get
m X
X n X
s
F (x, y, z, p, q, t) = (aijk + bijk + cijk )xi yj zk − p − q − t ≤ 0
i=1 j=1 k=1
for all x ∈ Dm , y ∈ Dn and z ∈ Ds .

But with p∗ = f1 (x∗ , y ∗ , z ∗ ), q ∗ = f2 (x∗ , y ∗ , z ∗ ), and t∗ = f3 (x∗ , y ∗ , z ∗ ), we
have F (x∗ , y ∗ , z ∗ , p∗ , q ∗ , t∗ ) = 0. Hence, the point (x∗ , y ∗ , z ∗ , p∗ , q ∗ , t∗ ) is a
solution to the problem (3.5)-(3.9).
Sufficiency: Now we have to show reverse, namely, that any solution of
problem (3.5)-(3.9) must be a Nash equilibrium point. Let (x̄, ȳ, z̄, p̄, q̄, t̄) be
any solution of problem (3.5)-(3.9). Let (x∗ , y ∗ , z ∗ ) be a Nash equilibrium
point for the game, and set p∗ = f1 (x∗ , y ∗ , z ∗ ), q ∗ = f2 (x∗ , y ∗ , z ∗ ), and
t∗ = f3 (x∗ , y ∗ , z ∗ ). We will show that (x̄, ȳ, z̄) must be a Nash equilibrium
of the game. Since (x̄, ȳ, z̄, p̄, q̄, t̄) is a feasible point, we have
n X
X s
aijk ȳj z̄k ≤ p̄, i = 1, . . . , m, (3.10)
j=1 k=1
m X
X s
bijk x̄i z̄k ≤ q̄, j = 1, . . . , n, (3.11)
i=1 k=1
m X
X n
cijk x̄i ȳj ≤ t̄, k = 1, . . . , s. (3.12)
i=1 j=1
Hence, we receive
m X
X n X
s
aijk x̄i ȳj z̄k ≤ p̄,
i=1 j=1 k=1
m X
X n X
s
bijk x̄i ȳj z̄k ≤ q̄,
i=1 j=1 k=1
m X
X n X
s
cijk x̄i ȳj z̄k ≤ t̄.
i=1 j=1 k=1
Adding these inequalities, we obtain

m X
X n X
s
F (x̄, ȳ, z̄, p̄, q̄, t̄) = [aijk + bijk + cijk ] x̄i ȳj z̄k − p̄ − q̄ − t̄ ≤ 0.
i=1 j=1 k=1
(3.13)
44 Game Theory
We know that at a Nash equilibrium F (x∗ , y ∗ , z ∗ , p∗ , q ∗ , t∗ ) = 0. Since

(x̄, ȳ, z̄, p̄, q̄, t̄) is also a solution, F (x̄, ȳ, z̄, p̄, q̄, t̄) be equal to zero:
m X
X n X
s
F (x̄, ȳ, z̄, p̄, q̄, t̄) = ( aijk x̄i ȳj z̄k − p̄) +
i=1 j=1 k=1
Xm X n X s
+( bijk x̄i ȳj z̄k − q̄) + (3.14)
i=1 j=1 k=1
Xm X n X s
+( cijk x̄i ȳj z̄k − t̄) = 0.
i=1 j=1 k=1
Consequently,
 Pm Pn Ps
 Pi=1 Pj=1 Pk=1 aijk x̄i ȳj z̄k = p̄,
m n s
i=1 Pj=1 Pk=1 bijk x̄i ȳj z̄k = q̄,
 m
P n s
i=1 j=1 k=1 cijk x̄i ȳj z̄k = t̄.
Since a point (x̄, ȳ, z̄, p̄, q̄, t̄) is feasible, we can write the constraints (3.10)-
(3.12) as follows:
 Pm Pn Ps Pn Ps
 Pi=1 Pj=1 Pk=1 aijk x̄i ȳj z̄k ≤ Pj=1 P k=1 aijk ȳj z̄k , i = 1, . . . , m,
m n s m s
i=1 Pj=1 Pk=1 bijk x̄i ȳj z̄k ≤ Pi=1 Pk=1 bijk x̄i z̄k , j = 1, . . . , n,
 Pm n s m n
i=1 j=1 k=1 cijk x̄i ȳj z̄k ≤ i=1 j=1 cijk x̄i ȳj , k = 1, . . . , s.
Now taking into account the above results, by Theorem 3.1 we conclude
that (x̄, ȳ, z̄) is a Nash equilibrium point which completes the proof.
3.1.2 The curvilinear global search algorithm

In order to solve problem (3.5)-(3.9), we use curvilinear search algorithm
introduced in [2]. This algorithm allows us to search for the minimum
value of the function along the scanning domain curve. The algorithm
was originally developed for solving box-constrained optimization problems,
therefore, we convert our problem from the constrained to unconstrained
form using penalty function techniques. For each equality constraint g(x) =
0, we construct a simple penalty function ĝ(x) = g 2 (x). For each inequality
constraint q(x) ≤ 0, we also construct the corresponding penalty function
as follows:
0, if q(x) ≤ 0,
q̂(x) = 2
q (x), if q(x) > 0.
Thus, we have the following box-constrained optimization problem:
γX γX
fˆ(x) = f (x) + ĝi (x) + q̂j (x) → min,
2 i 2 j X
X = x ∈ Rn |xi ≤ xi ≤ xi , i = 1, ..., n .

where γ is a penalty parameter, x and x - are upper and below bounds. For
original x, y and z variables the constraint is the box [0, 1]; for p, q and t
box constraints are [0, p], [0, q] and [0, t]. Values of p, q and t are chosen
from some intervals. An initial value of a penalty parameter γ is chosen not
too large (something about 1000) and after finding some local minimums
we increase it for searching another local minimum.
The proposed algorithm starts from some initial point x1 ∈ X. At each
k−th iteration the algorithm performs randomly “drop” of two auxiliary
points x̃1 and x̃2 and generating a curve (parabola) which passes through
all three points xk , x̃1 and x̃2 . Then we solve one-dimensional minimization
problem along this curve. If a solution to this problem is better than xk ,
we use it as a new minimum point, otherwise, we start a new iteration from
the previous point. Details are presented in Algorithm 1:
46 Game Theory
Algorithm 1 The Curvilinear Multistart Algorithm

Input: x1 ∈ X – initial (start) point; K > 0 – iterations count; δ > 0;
N > 0; εα > 0 — algorithm parameters.
Output: Global minimum point x∗ and f ∗ = f (x∗ )
1: for k ← 1, K do f k ← f (xk )
2: generate stochastic point x̃1 ∈ X
3: generate stochastic point x̃2 ∈ X
4: generate stochastic α-grid:
−1 = α1 ≤ .... ≤ αi ≤ −δ ≤ 0 ≤ δ ≤ αi+1 ≤ ... ≤ αN = 1

5: Let x̂(α) = ProjX α2 (x̃1 + x̃2 )/2 − xk + α/2 (x̃2 − x̃1 ) + xk
where ProjX (z) - projection of point z onto set X.
//note that x̂(−1) = x̂1 , x̂(1) = x̂2 , x̂(0) = xk .
6: f∗k ← f k
7: α∗k ← 0
8: for i ← 1, (N − 2) do
//Convex triplet
9: if f (x̂(αi )) > f (x̂(αi+1 )) and f (x̂(αi+1 )) < f (x̂(αi+2 )) then
//Refining the value of minima using
//Golden-Section search method with accuracy εα
10: α∗k ← GoldenSectionSearch(f, αi , αi+1 , αi+2 , εα )
11: if f (x̂(α∗k )) < f∗k then
12: f∗k ← f (x̂(α∗k ))
13: α∗k ← α∗k
14: end if
15: end if
16: end for
//Start local optimization algorithm
17: xk+1 ← LOptim(x̂(α∗k ))
18: end for
19: x∗ ← xk
20: f ∗ ← f (xk )
3.1.3 Computational results

The proposed method was implemented in C language and tested on some
examples.
Three problems of type (3.5) − (3.9) have been solved numerically for
dimensions 2 × 2 × 2 (Problems 1–3) and 5 × 6 × 4 (Problem 4). In all cases,
Nash equilibrium points were found successfully. These problems were:
Problem 1. Let a111 = 2, a112 = 3, a121 = −1, a122 = 0, a211 = 1,

a212 = −2, a221 = 4, a222 = 3, b111 = 1, b112 = 2, b121 = 0, b122 = −1,
b211 = −1, b212 = 0, b221 = 2, b222 = 1, and c111 = 3, c112 = 2, c121 = 1,
c122 = −3, c211 = 0, c212 = 2, c221 = −1, c222 = 2.
Then problem (3.5)–(3.9) can be written as:
F (x, y, z, p, q, t) = 6x1 y1 z1 + 7x1 y1 z2 − 4x1 y2 z2 + 5x2 y2 z1 +
+6x2 y2 z2 − p − q − t → max .


 2y1 z1 + 3y1 z2 − y2 z1 − p ≤ 0,
y1 z1 − 2y1 z2 + 4y2 z1 + 3y2 z2 − p ≤ 0,




x1 z1 + 2x1 z2 − x2 z1 − q ≤ 0,




−x1 z2 + 2x2 z1 + x2 z2 − q ≤ 0,




3x1 y1 + x1 y2 − x2 y2 − t ≤ 0,



2x1 y1 − 3x1 y2 + 2x2 y1 + 2x2 y2 − t ≤ 0,
x1 + x2 = 1,




y1 + y2 = 1,




z1 + z2 = 1,




x1 ≥ 0, x2 ≥ 0, y1 ≥ 0, y2 ≥ 0,




z1 ≥ 0, z2 ≥ 0, p ≥ 0, q ≥ 0, t ≥ 0.

Nash equilibrium points are:

F∗ x∗ y∗ z∗ p q t
0 (0; 1) (0; 1) (0; 1) 3 1 2
0 (1; 0) (1; 0) (1; 0) 2 1 3
2.08 · 10−8 (0.5191; 0.4809) (0.5888; 0.4112) (0.5382; 0.4618) 1.2281 0.5 0.9327
3.37 · 10−8 (0.75; 0.25) (0.8333; 0.1667) (1.0; 0.0) 1.5 0.5 1.9583
Problem 2. Let a111 = 5, a112 = 3, a121 = 6, a122 = 7, a211 = 0, a212 = 8,
a221 = 2, a222 = 1, b111 = 2, b112 = 4, b121 = −1, b122 = 0, b211 = 3, b212 = 5,
b221 = 4, b222 = 9, and c111 = 2, c112 = 0, c121 = −4, c122 = −1, c211 = −2,
c212 = 6, c221 = 8, c222 = 9.

0 (1; 0) (1; 0) (1; 0) 5 2 2
2.6 · 10−8 (0.5; 0.5) (0.5454; 0.4545) (0; 1) 4.8181 4.5 3.4545
9.9 · 10−8 (0.8; 0.2) (1; 0) (0.5; 0.5) 4.0 3.2 1.2
Problem 3. Let a111 = 3, a112 = 2, a121 = 1, a122 = 5, a211 = 8, a212 = 4,
a221 = 1, a222 = 3, b111 = 3, b112 = 2, b121 = 4, b122 = 0, b211 = 1, b212 = 8,
48 Game Theory
b221 = 6, b222 = 6, and c111 = 3, c112 = 1, c121 = 9, c122 = 2, c211 = 4,

c212 = 7, c221 = 2, c222 = 3.
0 (1; 0) (0; 1) (1; 0) 1 4 9
0 (0; 1) (1; 0) (0; 1) 4 8 7
0 (0.5; 0.5) (0; 1) (1; 0) 1 5 5.5
−1.33 · 10−15 (0.7; 0.3) (0; 1) (1; 0) 1 4.6 6.9
Problem 4. We have considered competitions of 3 companies sharing the

bread market of city Ulaanbataar where each company maximizes own profit
depending on its manufacturing strategies. The problem was formulated
as the three-person game with profit matrices A = {aijk }, B = {bijk },
C = {cijk }, i = 1, 5, j = 1, 6, k = 1, 4. The matrix data can be downloaded
from [17]. In this case the problem had 18 variables with 18 constraints.
The solution of the problem found by the proposed algorithm was:
 ∗

 F = 0,
x∗ = (0, 0, 0, 0, 1),




 y ∗ = (0, 0, 0, 0, 0, 1),


z ∗ = (1, 0, 0, 0),
p∗ = 65,




q ∗ = 160,




 ∗
t = 53.
It means that first and second companies must follow their 5-th and 6-th
production strategies while third company applies for its 1-st production
strategy. Companies’s maximum profits were 65, 160 and 53 respectively.
3.2 A Computational Method for Solving N -Person Game 49
3.2 A Computational Method for Solving N -

Person Game
3.2.1 Nonzero sum N -person game
Consider the n-person game in mixed strategies with matrices (Aq , q =
1, 2, . . . , n) for players 1, 2, . . . , n.
Aq = aqi1 i2 ...in , q = 1, 2, . . . , n

i1 = 1, 2, . . . , k1 , . . . , in = 1, 2, . . . , kn ,
Denote by Dp the set
p
X
q
Dp = {u ∈ R | ui = 1, ui ≥ 0, i = 1, . . . , p}, p = k1 , k2 , . . . , kn
i=1
A mixed strategy for player 1 is a vector x1 = (x11 , x12 , . . . , x1k1 ) ∈ Dk1 , where
x1i represents the probability that player 1 uses a strategy i. Similarly,
the mixed strategies for q-th player is xq = (xq1 , xq2 , . . . , xqkq ) ∈ Dkq , q =
1, 2, . . . , n. Their expected payoffs are given by for 1-th person :
k1 X
X k2 kn
X
f1 (x1 , x2 , . . . , xn ) = ... a1i1 i2 ...in x1i1 x2i2 . . . xnin .
i1 =1 i2 =1 in =1
and for q-th person
k1 X
X k2 kn
X
1 2 n
fq (x , x , . . . , x ) = ... aqi1 i2 ...in x1i1 x2i2 . . . xnin ,
i1 =1 i2 =1 in =1
q = 1, 2, . . . , n.
Definition 3.2. A vector of mixed strategies x̃q ∈ Dkq , q = 1, 2, . . . , n is a

Nash equilibrium if
 1 2 n 1 2 n 1
 f1 (x̃ , x̃ , . . . , x̃ ) ≥ f1 (x , x̃ , . . . , x̃ ), ∀x ∈ Dk1

 ··· ··· ··· ··· ··· ···


fq (x̃ , x̃ , . . . , x̃ ) ≥ fq (x̃ , . . . , x̃ , x , x̃q+1 , . . . , x̃n ), ∀xq ∈ Dkq
1 2 n 1 q−1 q
··· ··· ··· ··· ··· ···





fn (x̃ , x̃ , . . . , x̃ ) ≥ fn (x̃ , x̃ , . . . , x ), ∀xn ∈ Dkn .
1 2 n 1 2 n

50 Game Theory
It is clear that
f1 (x̃1 , x̃2 , . . . , x̃n ) = maxx1 ∈Dk1 f1 (x1 , x̃2 , . . . . . . , x̃n ),
··· ··· ··· ···
fq (x̃1 , x̃2 , . . . , x̃n ) = maxxq ∈Dkq fq (x̃1 , x̃2 , . . . , x̃q−1 , xq , x̃q+1 , . . . , x̃n ),
··· ··· ··· ···
fn (x̃ , x̃ , . . . , x̃ ) = maxxn ∈Dkn fn (x̃1 , x̃2 , . . . , x̃n−1 , xn ).
1 2 n
Denote by
kq−1
k1 X
X k2 X kX
q+1 kn
X
... ... aqi1 i2 ...in x1i1 x2i2 . . . xq−1 q+1 n
iq−1 xiq+1 . . . xin ,
i1 =1 i2 =1 iq−1 =1 iq+1 =1 in =1
1 2 q−1 q+1
, ϕiq (x , x , . . . , x ,x , . . . , xn ) = ϕiq (x|xq ), iq = 1, 2, . . . , kq , q = 1, 2, . . . , n.
For further purpose, it is useful to formulate the following statement.
Theorem 3.3. A vector strategy (x̃1 , x̃2 , . . . , x̃n ) is a Nash equilibrium if

and only if
fq (x̃) = ϕiq (x̃|x̃q ) (3.15)
for
x̃ = (x̃1 , x̃2 , . . . , x̃n )
iq = 1, 2, . . . , kq ,
q = 1, 2, . . . , n.
Proof. Necessity : Assume that x̃ is a Nash equilibrium. Then by the

definition, we have
k1 X
X k2 kn
X
fq (x̃) = ... aqi1 i2 ...in x̃1i1 . . . x̃nin ≥
i1 =1 i2 =1 in =1
k1 X
k2 kq−1
kq kq+1 kn
X X X X X
≥ ... ... aqi1 i2 ...in x̃1i1 . . . x̃iq−1 xq x̃q+1 . . . x̃nin
q−1 iq iq+1
i1 =1 i2 =1 iq−1 =1 iq =1 iq+1 =1 in =1
= fq (x̃1 , x̃2 , . . . , x̃q−1 , xq , x̃q+1 , . . . , x̃n ), (3.16)
q = 1, 2, . . . , n.
In the inequality 3.16, successively choose xqiq = 1, iq = 1, 2, . . . , kq . We
can easily see that fq (x̃) = ϕiq (x̃|x̃q ), for iq = 1, 2, . . . , kq ; q = 1, 2, . . . , n.
Sufficiency : Suppose that for a vector x̃ ∈ Dk1 ×Dk2 ×. . .×Dkn , conditions
3.15 are satisfied. We choose xq ∈ Dkq , q = 1, 2, . . . , n and multiply 3.15 by

xqiq respectively. We obtain
kq k1 kq kn
X X X X
xqiq fq (x̃) ≥ ... ... aqi1 i2 ...in x̃1i1 . . . x̃iq−1 xq x̃q+1 . . . x̃nin ,
q−1 iq iq+1
iq =1 i1 =1 iq =1 in =1
q = 1, 2, . . . , n.
Pkq
Taking into account that iq =1 xqiq = 1, q = 1, 2, . . . , n, we have
fq (x̃1 , x̃2 , . . . , x̃n ) ≥ fq (x̃1 , . . . , x̃q−1 , xq , x̃q+1 . . . x̃n ), ∀xq ∈ Dkq , q = 1, 2, . . . , n
which shows that x̃ is a Nash equilibrium. The proof is complete.
Theorem 3.4. A mixed strategy x̃ is a Nash equilibrium for the nonzero

sum n-person game if and only if
(x̃, p̃) is a solution to the following bilinear programming problem :
n
X n
X
1 2 n
max F (x, p) = fq (x , x , . . . , x ) − pq (3.17)
(x,p)
q=1 q=1
subject to :
ϕiq (x|xq ) ≤ pq , iq = 1, 2, . . . , kq , (3.18)
Proof. Necessity : Now suppose that x̃ is a Nash point. Choose vector

p̃ as : p̃q = fq (x̃), q = 1, 2, . . . , n. We show that (x̃, p̃) is a solution to
problem (3.17)-(3.18). First, we show that (x̃, p̃) is a feasible point for the
problem. By Theorem 3.3, the equivalent characterization of a Nash point,
we have
ϕiq (x̃|x̃q ) ≥ fq (x̃1 , . . . , x̃n ), q = 1, 2, . . . , n.
The rest of the constraints are satisfied because x̃q ∈ Dkq , q = 1, 2, . . . , n. It
meant that (x̃, p̃) is a feasible point. Choose any xq ∈ Dkq , q = 1, 2, . . . , n.
Multiply (3.18) by xqiq , q = 1, 2, . . . , n. respectively. If we have sum up these
inequalities, we obtain
fq (x) ≤ pq , q = 1, 2, . . . , n.
Hence, we get
k1 X
k2 kn n
! n
X X X X
F (x, p) = ... aqi1 i2 ...in x1i1 x2i2 . . . xnin − pq ≤ 0
i1 =1 i2 =1 in =1 q=1 q=1
52 Game Theory
for all xq ∈ Dq , q = 1, 2, . . . , n.
But with p˜q = fq (x̃), we have F (x̃, p̃) = 0 Hence, the point (x̃, p̃) is a
solution to the problem (3.17)-(3.18).
Sufficiency : Now we have to show reverse, namely, that any solution of
problem (3.17)-(3.18) must be a Nash point. Let (x̄, p̄) be any solution
of problem (3.17)-(3.18). Let x̃ be a Nash point for the game, and set
p̄q = fq (x̄).
We will show that x̄ must be a Nash equilibrium of the game. Since (x̄, p̄)
is a feasible point, we have
ϕiq (x̄|x̄q ) ≤ p̄q , iq = 1, 2, . . . , kq , q = 1, 2, . . . , n. (3.19)
Hence, we have
fq (x̄) ≤ p̄q , q = 1, 2, . . . , n.
Adding these inequalities, we obtain

k1 X
k2 kn n
! n
X X X X
F (x̄, p̄) = ... aqi1 i2 ...in 1 2 n
x̄ x̄ . . . x̄ − p̄q ≤ 0 (3.20)
i1 =1 i2 =1 in =1 q=1 q=1
We know that at a Nash equilibrium F (x̃, p̃) = 0. Since (x̄, p̄) is also a
solution, F (x̄, p̄) be equal to zero :
k1 X k2 kn n
! n
X X X X
F (x̄, p̄) = ... aqi1 i2 ...in x̄1 x̄2 . . . x̄n − p̄q = 0 (3.21)
i1 =1 i2 =1 in =1 q=1 q=1
Consequently,
fq (x̄) = p̄q , q = 1, 2, . . . , n.
Since a point (x̄, p̄) feasible, we can write the constraints (3.19) as follows :
k1 X
X k2 kn
X
... aqi1 i2 ...in x̄1i1 x̄2i2 . . . x̄nin ≥
i1 =1 i2 =1 in =1
kq−1
k1 X
X k2 X kX
q+1 kn
X
≥ ... ... aqi1 i2 ...in x̄1i1 x̄2i2 . . . x̄q−1 q+1 n
iq−1 x̄iq+1 . . . x̄in ,
i1 =1 i2 =1 iq−1 =1 iq+1 =1 in =1
for
iq = 1, 2, . . . , kq , q = 1, 2, . . . , n.
Now taking into account the above results, by Theorem 3.3 we conclude
that x̄ is a Nash point which a completes the proof.
3.2.2 Computational results

The following 4 person game has been considered. The problem was solved
numerically by the Cirvilinear Global Search Algorithm, described in 3.2.1.
Problem 3.1 Let A1 is a11111 = 1, a11211 = 0, a11121 = 0, a11112 = 0, a12111 =

1, a12112 = 0, a12121 = 0, a12211 = 0, a11122 = 1, a11212 = 0, a11221 = 0, a12221 =
0, a12212 = 0, a12122 = 0, a11222 = 0, a12222 = 1,
A2 is a21111 = 0, a21211 = 1, a21121 = 0, a21112 = 0, a22111 = 0, a22112 = 1, a22121 =
0, a22211 = 0, a21122 = 0, a21212 = 0, a21221 = 1, a22221 = 0, a22212 = 1, a22122 =
1, a21222 = 0, a22222 = 1.
and A3 is a31111 = 0, a31211 = 1, a31121 = 0, a31112 = 0, a32111 = 0, a32112 =
1, a32121 = 1, a32211 = 1, a31122 = 0, a31212 = 1, a31221 = 0, a32221 = 1, a32212 =
0, a32122 = 1, a31222 = 1, a32222 = 0,
A4 is a41111 = 0, a41211 = 0, a41121 = 0, a41112 = 1, a4 2111 = 0, a4 2112 =
0, a42121 = 1, a42211 = 0, a41122 = 1, a41212 = 2, a41221 = 0, a42221 = 1, a42212 =
1, a42122 = 0, a41222 = −1, a42222 = 0.
Solution of this problem is also not unique and consist of several sets,
such as:
∗ ∗ ∗ ∗
1. F ∗ = 0, x1 = (0, 1)T , x2 = (0, 1)T , x3 = (t, 1 − t)T , x4 = (1, 0)T ,
p∗1 = 0, p∗2 = 0, p∗3 = 1, and p∗4 = 1 − t, where t ∈ [0, 0.5].
∗ ∗ ∗ ∗
2. F ∗ = 0, x1 = (0, 1)T , x2 = (u, 1 − u)T , x3 = (0, 1)T , x4 = (1, 0)T ,
p∗1 = 0, p∗2 = 0, p∗3 = 1, and p∗4 = 1, where u ∈ [0, 1].
∗ ∗ ∗ ∗
3. F ∗ = 0, x1 = (0, 1)T , x2 = (0, 1)T , x3 = (v, 1 − v)T , x4 = (0, 1)T ,
p∗1 = 1 − z1∗ , p∗2 = 1, p∗3 = 0, and p∗4 = v, where v ∈ [0.5, 1].
∗ ∗ ∗
Solution (1) and (2) meets in point x1 = (0, 1)T , x2 = (0, 1)T , x3 =
∗
(0, 1)T , x4 = (1, 0)T .
Some single points of Nash equilibrium also are:
∗ ∗ ∗ ∗
F∗ x1 x2 x3 x4 p∗1 p∗2 p∗3 p∗4
0 (1, 0) (0, 1) (1, 0) (0, 1) 0 0 1 2
0 (1, 0) (1, 0) (1, 0) (0, 1) 0 0 0 1
Bibliography
[1] Strekalovsky, A. S and Orlov, A. V.,Bimatrix Game and Bilinear Pro-

gramming, 2007, Nauka.
[2] Gornov, A. Yu and Zarodnyuk, T. S., Optimization, Simulation, and

Control, pages 289-299, 2013, Springer, New York.
[3] Lemke, C. E. and Howson, T. T., Equilibrium Points of Bimatrix Games,

SIAM J. Appl. Math.,pages 413-423, 1961,.
[4] Dickhaut, J. and Kaplan, T., A Program for Finding Nash Equilibria,
Mathematica J., pages 87-93, 1991.
[5] McKelvey, Richard D. and McLennan, Andrew, Computation of equi-

libria in finite games, Handbook of Computational Economics, pages 87-
142, 1996, Elsevier.
[6] Howson, J. T.,Equilibria of Polymatrix Games, Management Sci., vol.18,

pages 312-318, 1972.
[7] Yanovskaya, E. B., Equilibrium Points in Polymatrix Games, Latvian

Math. Col lect., vol.8, pages 381-384, 1968.
[8] Mills, H., Equilibrium Points in Finite Games,J. Soc. Indust. Appl.
Mathemat.,vol.8(2), pages 397-402, 1960.
[9] Mangasarian, O. L., Equilibrium Points in Bimatrix Games, J. Soc.

Indust. Appl. Mathemat., vol.12, pages 778-780, 1964.
[10] Mangasarian, O. L. and Stone, H., Two-Person Nonzero Games and

Quadratic Programming, J. Mathemat. Anal. Appl.,No9, pages 348-355,
1964.
[11] Vasilyev, I. L. and Klimentova, K. B. and Orlov, A. V., A parallel

search of equilibrium points in bimatrix games, Numerical methods and
programming, vol.8, pages 233-243, 2007.
BIBLIOGRAPHY 55
[12] Nash, John F., Equilibrium points in n-person games, Proc. of the Nat.
Acad. of Sci. USA, vol.36, pages 48-49, 1950.
[13] Orlov, Andrei V. and Strekalovsky, Alexander S. and Batbileg, S.,On

computational search for Nash equilibrium in hexamatrix games, Opti-
mization Letters, vol.10(2), pages 369-381, 2014.
[14] Batbileg, S. and Enkhbat, R., Global Optimization Approach To Game

Theory, Journal of Mongolian Mathematical Society, vol.14, pages 2-10,
2010.
[15] Nash, John F.,Non-Cooperative Games, Annals of Mathematics,

vol.54(2), pages 286-295, 1951.
[16] Strekalovsky, A. S. and Enkhbat, R., Polymatrix games and optimiza-

tion problems, Automation and Remote Control, vol.75(4), pages 632-645,
2014.
[17] https://www.dropbox.com/s/137aaahvbniau9q/problem-4-data.txt
[18] Strekalovsky A.S. (1998) Global Optimality Conditions for Nonconvex

Optimization. Journal of Global Optimization 12: 415–434.
[19] Enkhbat R, Tungalag N., Gornov A., Anikin A. The Curvilin-

ear Search Algorithm for Solving Three-Person Game. Proc. DOOR
2016. CEUR-WS. 2016. Vol. 1623. P. 574–583. http://ceur-ws.org/Vol-
1623/paperme4.pdf.
[20] Gornov A., Zarodnyuk T. (2014) Computing technology for estimation

of convexity degree of the multiextremal function. Machine learning and
data analysis 10 (1): 1345–1353. http://jmlda.org
[21] Von Neumann, J., Morgenstern, O. (1944) Theory of games and eco-
nomic behavior. Princeton University Press
[22] Melvin Dresher (1981) The Mathematics of Games of Strategy. Dover

Publications.
[23] Vorobyev N.N. (1984) Noncooperative games. Nauka.
[24] Germeyer YU.B. (1976) Introduction to Operation Reseach. Nauka.
[25] Owen,G. (1971) Game Theory. Saunders, Philadelphia.
[26] Gibbons R. (1992) Game Theory for Applied Economists. Princeton

University, Press.
56 BIBLIOGRAPHY
[27] Horst R. and Tuy H. (1993) Global Optimization. Springer-Verlag.
[28] Howson J. T. (1972) Equilibria of Polymatrix Games. Management Sci

18: 312–318.
4
Cost Minimization Problem
58 Cost Minimization Problem
4.1 Pseudoconvexity Properties of Average

Cost Function
Introduction
Cost functions play important role in the firm theory. Cost function is
defined as a solution of the following cost minimization problem subject to
a given level of production.
min(r1 y1 + r2 y2 + ... + rn yn ), (4.1)
subject to :
f (y1 , y2 , ..., yn ) = x, (4.2)
where f : R+ → R is a continuously differentiable strictly concave pro-
duction function, ri (i = 1, 2, ..., n) are factor prices, x is output, yi (i =
1, 2, ..., n) are factors.
The solutions yi∗ (r1 , r2 , ..., rn , x), i = 1, 2, ..., n, to problem (4.1)-(4.2)

constitute the cost function C as follows:
X n
C(x, r1 , r2 , ..., rn ) = ri yi∗ (r1 , r2 , ..., rn , x)
i=1
If factor prices are fixed, i.e, ri = ri0 , i = 1, 2, ..., n then one variable function
C(x, r1 , r2 , ..., rn ) is a called as a short run cost function defined as:
C(x) = C(x, r10 , r20 , ..., rn0 ).
For example, if the productionPfunction is of type
f (y1 , y2 , ..., yn ) = y1α1 y2α2 ...ynαn , ( ni=1 αi 6 1, αi > 0) then the cost func-
tiom C(x) can be easily found as:
" n
# n1 n
Y r0 i
X
C(x) = x · · αi
i=1
αi i=1
4.1.1 Average cost function

We consider some properties of the average cost function ϕ(x) defined as
ϕ(x) = C(x)
x
, x ∈ R+ , where R+ = {x ∈ R|x > 0}.
Definition 4.1. [9] A differentiable function h : R+ → R is pseudoconvex
at y ∈ R+ if h(x) − h(y) < 0 implies h0 (y)(x − y) < 0, ∀x ∈ R+ . A function
h(·) is pseudoconvex on R+ if it is pseudoconvex at each point y ∈ R+ .
4.1 Pseudoconvexity Properties of Average Cost Function 59
Lemma 4.1. The average cost function ϕ(x) is pseudoconvex on R+ .
Proof. We foloow up a proof given in [5] for the general concave-convex

fractional function. It is clear that C(x) > 0 for all x ∈ R+ . Take any point
y ∈ R+ . Introduce the function ϕ(x) : R+ → R as follows:
ψ(x) = C(x)y − xC(y).

Clearly, C(x) is convex and differentiable and ψ(x) = 0. On the other hand,
it is obvious that the inequality ϕ(y) > ϕ(x) is equivalent to ψ(y) > ψ(x).
Since ψ(·) is convex and the differentiable, then we have]:
0 > ψ(x) − ψ(y) > ψ 0 (x)(x − y).

Taking into account
C 0 (y)y − C(y) ψ(y)(x − y)

ϕ0 (y)(x − y) = 2
= < 0,
y y2
we conclude that ϕ(y) > ϕ(x) implies ϕ0 (y)(x − y) < 0 which prove the
assertion.
4.1.2 Average cost minimization problem.

Consider the average cost minimization problem:

C(x)
min ϕ(x) = (4.3)
x∈R+ x
Lemma 4.2. Any local minimizer x∗ of ϕ(x) on R+ is also a global mini-

mizer.
Proof. On the contrary, assume that x∗ is not a global minimizer. Then

there exists a point u ∈ R+ such that
ϕ(x∗ ) > ϕ(u). (4.4)
Since R+ is a convex set, x∗ + α(u − x∗ ) = αu + (1 − α)x∗ ∈ R+ , ∀α : 0 <

α < 1.
By Taylor’s expansion, we have
ϕ(x∗ + α(u − x∗ )) = ϕ(x∗ ) + αϕ0 (x∗ )(u − x∗ ) + 0(αku − x∗ k),

∗
k)
where limα→0+ 0(αku−x
α
= 0.
Since x is a local minimizer of ϕ(·) on R+ , there exists 0 < α∗ < 1, so that
∗
ϕ(x∗ + α(u − x∗ )) − ϕ(x∗ ) > 0, ∀α : 0 < α < α∗ .

which implies
ϕ0 (x∗ )(u − x∗ ) > 0
Since ϕ(·) is pseudoconvex, ϕ0 (x∗ )(u − x∗ ) > 0 implies that ϕ(u) > ϕ(x∗ )
contradicting ϕ(x∗ ) > ϕ(u) in (4.2). This completes the proof.
Lemma 4.3. The average cost minimization problem (4.1) has a unique
solution.
Proof. On the contrary assume that there exist x1 , x2 ∈ R+ such that
x1 < x2 and minx∈R+ ϕ(x) = ϕ(x1 ) = ϕ(x2 )
Take any point x ∈ (x1 , x2 ). Since ϕ(x) is pseudoconvex ϕ(x1 ) < ϕ(x) im-
plies that ϕ0 (x)(x1 − x) < 0 or equivalently ϕ0 (x) > 0 (sincex1 − x < 0
On the other hand, ϕ(x2 ) < ϕ(x) implies also ϕ0 (x)(x2 − x) < 0 or equiva-
lently ϕ0 (x) < 0 which contradicts ϕ0 (x) > 0. The proof is complete.
Lemma 4.2 allows us to apply the bisection methods for solving problem
(4.1) numerically.
4.1.3 Numerical results

A unique global minimizer of the problem (4.1) can be computed as:
C 0 (x)x − C(x)
ϕ0 (x) = =0
x2
Then a solution x∗ satisfies the following equation C 0 (x∗ )x∗ − C(x∗ ) = 0
equivalently,
M C 0 (x) = AC(x)
Example, if C(x) = ax2 + bx + c
then
ϕ(x) = ax + b + xc ϕ0 (x) = a − xc2 x∗ = ac
p
If
C(x) = ax3 + bx2 + dx + e
then the average cost function is
C(x) e
= ax2 + bx + d +
x x
Clearly, the following equation
e
ϕ0 (x) = 2ax + b − 2
= 0 or 2ax3 + bx2 − e = 0
x
4.1 Pseudoconvexity Properties of Average Cost Function 61
has a unique solution.

Using a mongolian carpet manufacturing company’s data we obtained the
following total cost function.
C(x) = 0.0001x3 + 5.51x2 − 14611x + 3 · 107 ,

where x is output (carpet in thousand sq.m). C is the total cost of the
company. Solving by the bisection method, we found an optimal solution
x∗ = 2243.7
4.2 Economic Order Quantity Model

Introduction
The economic order quantity (EOQ) formula plays an important role in
inventory management. EOQ model has had on a century of researchers
and practitioners in the fields of operations management and operations
research. EOQ model appeared in Harris (1913) describes a very simple de-
terministic inventory planning model with a tradeoff between fixed ordering
cost and inventory carrying cost [12]. EOQ lays the foundation for all kinds
of extensions and real world management applications [12, 13, 14, 15, 16].
Both the deterministic and the stochastic EOQ models were developed in
[12, 16, 17, 18, 19]. The Economic Order Quantity (EOQ) is the number of
units that a company should add to inventory with each order to minimize
the total costs of inventorysuch as holding costs, and shortage costs. The
EOQ is used as part of a continuous review inventory system in which the
level of inventory monitored at all times and a fixed quantity is ordered each
time the inventory level reaches a specific reorder point. The EOQ provides
a model for calculating the appropriate reorder point and the optimal re-
order quantity to ensure the instantaneous replenishment of inventory with
no shortages. It can be a valuable tool for small business owners who need
to make decisions about how much inventory to keep on hand, how many
items to order each time, and how often to reorder to incur the lowest pos-
sible costs. The EOQ model assumes that demand is constant, and that
inventory is depleted at a fixed rate until it reaches zero. At that point,
a specific number of items arrive to return the inventory to its beginning
level. Since the model assumes instantaneous replenishment, there are no
inventory shortages or associated costs. Therefore, the cost of inventory un-
der the EOQ model involves a tradeoff between inventory holding costs (the
cost of shortage, as well the cost of tying up capital in inventory rather than
investing or using it other purposes) and order costs (any fees associated
with placing orders, such as delivery charges). Ordering a large amount at
one time will increase a small businesss holding costs, while making more
frequent orders of fewer items will reduce holding costs but increase order
costs. The EOQ model finds the quantity that minimizes the sum of these
costs. q
2Dc
The EOQ model computed by the following formula EOQ = Ch
,
where D is the demand per time unit, c is the ordering cost, ch is the
holding cost a per unit and time unit.
Research on the EOQ can be primarily classified into three areas of
interest:
4.2 Economic Order Quantity Model 63
1) The performance of the EOQ against other lot-sizing rules: Lot size
refers to the quantity of an item ordered for delivery on a specific date or
manufactured in a single production run. Choi et al. [20] examined the EOQ
versus eight other rules in multi-echelon MRP systems using FORTRAN,
with the EOQ performing in the lower third. Evan L. Porteus [21] has
introduced a model that shows a significant relationship between quality
and lot size. Melnyk and Piper [22] examined the effects of lead time errors
on different lot sizing rules.
2) Extensions of the EOQ model: EOQ extensions include an applica-
tion to retail cycle stock inventories [23], the addition of cost changes under
a finite or infinite time horizon [24], the inclusion of storage size considera-
tions [17], and the addition of damage costs [16, 25] provided an exhaustive
summary of the research on EOQ models that handle partial backordering,
and some even more current models with partial or full backordering include
those by [18, 26, 27] and [28].
3) The role of the EOQ in logistics: Research that addressed the use
of EOQ models in transportation and logistics first appeared in the 1980s.
Eppen [29] was the first to discuss the ”impact of inventories” on locations
by exploiting risk pooling effects. Tanchoco et al. [30] considered the im-
pact of material handling and the transportation of unit loads on lot sizing.
Ng et al. [31] studied a type of EOQ problem where the (maximum) ware-
house capacity is a decision variable. Furthermore, they assumed that the
warehouse cost dominates all the other inventory holding costs. Keskin et
al. [32] study generalized vendor selection models aimed at optimizing the
total logistical costs including not only the vendor-specific fixed manage-
ment and purchasing costs considered in traditional models, but also the
transportation, inventory replenishment, and holding costs.
In their models, the authors mainly used convex optimization as a math-
ematical tool. However, dynamic models for EOQ have been less considered.
To fulfill this gape, we formulate in this paper EOQ model as a calculus
of variations. The new problem is a simple optimal control which can be
solved by Euler-Lagrange equation as optimality conditions. In particular,
we obtain the well-known EOQ formula [12].
4.2.1 Calculus Variations Approach to EOQ

Assume that holding cost per unit depends on its rate and quantity of items.
In other words,
C = C(q, q 0 , t) (4.5)
where C : R ∗ R ∗ R → R, q : R → R, cq = q(t), C is the cost function at
moment t per each inventory. And q is the quantity inventory.
Then total cost is

ZT
TC = C(q, q 0 , t)q(t)dt.
0
Economic order quantity problem can be formulated as a problem of calculus

of variations with a free left side constraint q0 .

RT 0
 min F (q, q , t)dt


0 (4.6)

 q(0) = q0
q(T ) = 0

where, q(t): order quantity function, q0 : economic order quantity, T :

final time, F (q, q 0 , t) = C(q, q 0 , t)q(t).
In order to solve problem (4.6), we need to solve Euler-Lagrange equation
[33] as an optimality condition for the calculus of variations.
δF (q, q 0 , t) d δF (q.q 0 , t)

− =0 (4.7)
δq dt δq 0
Let q∗ = q ∗ (c1 , c2 , t) be a solution to problem (4.7). The constants c1

and c2 can be found from the boundary conditions [26] solving a system of
nonlinear equations.
0
(
δF (q ∗ q ∗ 0)
δq ∗ (4.8)
∗
q (T ) = 0
0
For example, if we have F (q, q 0 , t) = 12qt + qq 0 + q 2 , q(0) = q0 and q(1) = 0,
then Euler-Lagrange equation is:
00
12t + q 0 − q 0 − 2q = 0
The solution is q(t, c1 , c2 ) = t3 + c1 t + c2 .

Clearly, q(0) = c2 = q0 , and the system (4) gives
= q + 2q 0 = (t3 + c1 t + c2 ) + 2(3t2 + c1 )|t=0 = c2 + 2c1

δF
δq 0
q(T ) = T 3 + c1 T + c2 = 0
Now we have condition (4) as

c2 + 2c1 = 0
T 3 + c1 T + c2
4.2 Economic Order Quantity Model 65
T 3 2T 3 2T 3
Hence, we have c1 = 2−T and c2 = T −2
, consequently, q0 = T −2
.
Then optimal solution is
T3 2T 3
q ∗ (t) = t3 + t+
2−T T −2
In practice, usually type of order quantity function is given. For instance,
q(T ) = bT + d, and holding cost per unit is constant ch . Then total cost
minimum problem is:

RT
 minT C = Dc + (bt + d)ch dt


q0
0

 q(0) = d = q00
q(t) = bT + d = 0,

where, D is demand for the period T , and c ordering cost. b can be

expressed as
d q0
b=− =−
T T
Then the cost function T C has the form.
ZT
Dc q0
min T C = + − t + q0 ch dt
q0 T
0
or equivalently,
Dc q0 ch T
TC = − + q0 ch T.
q0 2
Taking derivative of T C, we have
dT C Dc ch T
= 2 − + qch T = 0.
dq0 q0 2
Now we obtain
Dc ch T
− .
q02 2
We find optimal order quantity q0 as
r
2Dc
q0 = .
ch T
If we take T =1, then we have well known economic order quantity formula
(EOQ) in the literature [12].
r
2Dc
q0 = .
ch
Bibliography
[1] Akira Takayama, Analytical Methods in Economics, Harvester

Wheatsheaf, 1994
[2] Basskin.L, Using Cost-Minimization Analysis to Select From Equally

Effective Alternatives. Formulary 33(12):1209-1214,1998.
[3] Briggs AH, O’Brien BJ, The death of Cost-Minimization Analysis?

Health Economics, 10(2):179-184, 2001.
[4] Enkhbat, R.: Quasiconvex progarmming, Lambert Publisher, Ger-

many, 2009.
[5] Enkhbat, R; Ya.Bazarsad and J.Enkhbayar A Methofd for Fractional

Programming, International Journal of Pure and Applied Mathemat-
ics,Volume 73 No. 1 2011, 93-99
[6] J.Henderson and R.Quandt, Microeconomic Theory: A Mathemati-

cal Approach, McGraw-Hill Book Company, 1980.
[7] M.Mahmud Khan, Disha Ali, Zohra Ferdousy and Abdullah Ali- Ma-
mun, A Cost-Minimization Approach to PlanningJ The Geograph-
ical Distribution of Health Facilites, Oxford Journals, Medicine,
Health Policy and Planning, Volume 16, Issue 3, pp.264-272.
[8] Michael D.Intriligator, Mathematical Optimization and Economic

Theory, SIAM, 2002
[9] Mishra.S.K, Lai.K.K, Wong.S, Generalized Convexity and Vector

Optimization, Springer, 2009
[10] Newby D, Hill S, Use of Pharmacoeconomics In Prescribing Re-

search. Part 2: Cost Minimization Analysis-When Are Two Ther-
apies Equal? Journal of Clinical Pharmacy and Therapeutics 28(2):
145-150,2003.
BIBLIOGRAPHY 67
[11] C.Rose and R.Yates, Minimizing The Average Cost of Paging Under
Delay Constraints, Wireless Networks1 (1995) 211-219.
[12] Drake, M.J. and Marley, K.A. (2014) A Century of the EOQ.
In: Tsan-Ming, C., Ed., Handbook of EOQ Inventory problems,
Springer, New York, 3-22.
[13] Axs¨ter, S. (1996) Using the Deterministic EOQ Formula in Stochas-

tic Inventory Control. Management Science, 42, 830-834.
[14] Huang, W., Kulkarni, V.G. and Swaminathan, J.M. (2003) Optimal
EOQ for Announced Price Increases in Infinite Horizon. Operations
Research, 51, 336-339.
[15] Khan, M., Jaber, M.Y., Guiffrida, A.L. and Zolfaghari, S. (2011) A
Review of the Extensions of a Modified EOQ Model for Imperfect
Quality Items. International Journal of Production Economics, 132,
1-12.
[16] Pentico, D.W. and Drake, M.J. (2011) A Survey of Deterministic

Models for the EOQ and EPQ with Partial Backordering. European
Journal of Operational Research, 214, 179-198.
[17] Rao, S.S. and Bahari-Kashani, H. (1990) Economic Order Quantity

and Storage Size Some Considerations. Engineering Costs and Pro-
duction Economics, 19, 201-204.
[18] Zhang, R.-Q., Kaku, I. and Xiao, Y.-Y. (2011) Deterministic EOQ
with Partial Backordering and Correlated Demand Caused by Cross-
Selling. European Journal of Operational Research, 210, 537-551.
[19] Brill, P.H. and Chaouch, B.A. (1995) An EOQ Model with Random
Variations in Demand. Management Science, 41, 937-936.
[20] Choi, H., Malstrom, E.M. and Classen, R.J. (1984) Computer Simu-
lation of Lot-Sizing Algorithms in Three-Stage Multi-Echelon Inven-
tory Systems. Journal of Operations Management, 4, 259-277.
[21] Evan, L.P. (1986) Optimal Lot Sizing, Process Quality Improvement
and Setup Cost Reduction. Operations Research, 34, 137-144.
[22] Melnyk, S.A. and Piper, C.J. (1985) Leadtime Errors in MRP: The
Lot-Sizing Effect. International Journal, 23, 253-264
68 BIBLIOGRAPHY
[23] Bassin, W. (1990) A Technique for Applying EOQ Models to Retail

Cycle Stock Inventories. Journal of Small Business Management, 28,
48-55.
[24] Lev, B. and Weiss, H.J. (1990) Inventory Models with Cost Changes.
Operations Research, 38, 53-63.
[25] Chyr, F., Tsong, M.L. and Chin-Fu, H. (1990) An Extension of the
EOQ Production Model Based on Damage Costs. International Jour-
nal of Operations and Production Management, 10, 71-76.
[26] Toews, C., Pentico, D.W. and Drake, M.J. (2011) The Deterministic
EOQ and EPQ with Partial Backordering at a Rate That Is Lin-
early Dependent on the Time to Delivery. International Journal of
Production Economics, 131, 643-649.
[27] Chung, K.-J. and Cardenas-Barrn, L.E. (2012) The Complete Solu-
tion Procedure for the EOQ and EPQ Inventory Models with Linear
and Fixed Backorder Costs. Mathematical and Computer Modellling,
55, 2151-2156.
[28] Taleizadeh, A.A., Pentico, D.W., Jabalameli, M.S. and Aryanezhad,
M. (2013) An EOQ Model with Partial Delayed Payment and Partial
Backordering. Omega, 41, 354-368.
[29] Eppen, G. (1979) Effects of Centralization on Expected Costs in
a Multi-Location Newsboy Problem. Management Science, 25, 498-
501.
[30] Tanchoco, J.M.A., Davis, R.P. and Wysk, R.A. (1980) Economic
Order Quantities Based on Unit-Load and Material Handling Con-
siderations. Decision Sciences, 11, 514-521.
[31] Ng, C.T., Chenga, T.C.E., Kotov, V. and Kovalyov, M. (2009) The
EOQ Problem with Decidable Warehouse Capacity: Analysis, Solu-
tion Approaches and Applications. Discrete Applied Mathematics,
157, 1806-1824.
[32] Keskin, B.B.H. and ster, S.. (2010) Integration of Strategic and
Tactical Decisions for Vendor Selection under Capacity Constraints.
Computers and Operations Research, 37, 2182-2191.
[33] Enkhbat, R. (2006) Optimization-2 (Calculus of Variations with Eco-
nomic Applications). NUM Press, Ulaanbaatar.
5
Investment Model
70 Investment Model
5.1 The Basic Valuation Model

The value of any asset is the present value of all future cash flows it is
expected to provide over the relevant time period. The value of an asset is
therefore determined by discounting the expected cash flows back to their
present value, using the required return commensurate with the asset’s risk
as the appropriate discount rate.
We can express the value of any asset at time zero,V0 - net present value, as
the valuation equation
n
X ak
V0 = k
(5.1)
k=1
(1 + i)
where V0 - net present value and value of the asset at time zero, ak - cash
flow expected at the end of year k, i - appropriate required return (discount
rate), n - relevant time period.
If we know ak as function ak = ϕ(k) of k then formula (5.1) becomes
n
X ϕ(k)
V0 = (5.2)
k=1
(1 + i)k
If we replace (5.2) with an integral sum then we have

Z n
ϕ(x)
V0 = x
dx.
1 (1 + i)
When ak = a, k = 1, n then formula (5.2) has the form


1

 1 − (1 + i)n (1 + i) 

a
V0 =  
1+i  i 
We can find n as
V0 · i
ln 1 −
a
n=
1
ln
1+i
for given ai and V0 and compute i from the equation:
n
X ak
V0 = , ak > 0 (5.3)
k=1
(1 + i)k
Introduce the function

ak
ϕk (i) = .
(1 + i)k
5.1 The Basic Valuation Model 71
Lemma 5.1. The equation (5.3) has the unique solution.

0 1
Pn
Proof . Since ϕk = −k · ak · (1+i)k the function f (i) = k=1 ϕk (i) is
a monotonically decreasing function. On the other hand, the function f (i)
is positive and bonded below. It is clear that
lim f (i) = 0.
i→∞
Then by the mean value theorem the equation

n
X
F (i) = ϕk (i) − V0 = 0
k=1
has a unique solution i∗ which can be computed by the Newton method as
F (im )
im+1 = im − , m = 0, 1, . . .
F 0 (im )
where i0 is an initial approximation and
lim im = i∗ .
m→∞
The proof is complete.
5.1.1 Optimization Approach

Now consider the following minimization problem:
ψ(i, n) → min (5.4)
subject to
n
X ak
= V0 ,
k=1
(1 + i)k
where ψ(i, n) is a risk function.
For ak = a, k = 1, 2, . . . , problem (5.4) can be reduced to one dimensional
minimization problem.
ψ(i, n) → min, i ∈ [iA , iB ] (5.5)
ln 1 − N PaV ·i

n= 1

ln 1+i
72 Investment Model
For example, if
ψ(i, n) = (i − i0 )2 + (n − n0 )2
then the problem is a one dimensional global minimization problem
!2
V0 ·i

ln 1 −
f (i) = (i − i0 )2 + 1
a
− n0 → min (5.6)
ln 1+i iA ≤i≤iB
In general, problem
min f (i)
iA ≤i≤iB
is a global optimization problem.

In order to solve globally this problem, we can use the method of piecewise
linear function [7] described in Chapter 1. Let us consider a continuous
version of problem (5.4):
ψ(i, n) → min (5.7)
Z n
ϕ(x)dx
x
− V0 = 0
1 (1 + i)
We can solve this problem by the Lagrange method.
Z n
ϕ(x)dx
L(i, n, λ) = ψ(i, n) + λ x
− V0
1 (1 + i)
Write down optimality condition:

∂L ∂ψ(i, n) ∂Φ(i, n)


 = +λ =0



 ∂i ∂i ∂i


 ∂L

∂ψ(i, n) ϕ(n)
= +λ =0

 ∂n ∂n (1 + i)n


 Z n
∂L ϕ(x)dx


− V0 = 0


 = x
∂λ 1 (1 + i)
Here Φ(i, n) is an integral:

Z n
ϕ(x)dx
Φ(i, n) =
1 (1 + i)x
We calculate this integral.
∂Φ(i, n) i − i0 ϕ(n)
a. = · ⇒
∂i n − n0 (1 + i)n

ϕ(n) 1+i 1 + i0
Φ(i, n) = −
(n − n0 )(1 + i)n−1 2 − n 1 − n
5.1 The Basic Valuation Model 73
b. If ϕ(n) is an arithmetic progression ϕ(n) = a + bn then the integral

has the following form:
(a + b) ln(1 + i) + b (a + bn) ln(1 + i) + b
Φ(i, n) = −
(1 + i)(ln(1 + i))2 (1 + i)n (ln(1 + i))2
and we can compute problem (5.7) numerically.
5.1.2 Numerical Results

The following problems have been solved numerically on MATLAB.
1. We used a data for efficiency of investment V0 and cash flow of Mongo-
lian National company APU. We assume that cash flow was constant since
2010 year. This company invested 8275,2 million tugrugs in 2006. Using a
cash flow for 2007-2015 and V0 , we can find i∗ from the equation:
9
339, 3 526 1520, 7 X 2967
+ + + = 8275, 2.
1+i (1 + i)2 (1 + i)3 k=4 (1 + i)k
The optimal solution found by by the Newton method has
i∗ = 0.1281.
2. We write problem (5.6) in the following form:
 2
V0 · i
2
 ln 1 − a 
f (i) = (i − i0 ) + 
 ln(1 + i) + n 0
 → min (5.8)
 iA ≤i≤iB
a
The condition > i has an efficient condition of financial investment.
V0
Solutions are found for different data in the following table.
Inputs Solutions
i0 n0 i∗ n∗
0.04 5 0.0412 4.8806
0.06 10 0.0593 7.1982
0.08 15 0.0819 12.5661
0.1 20 0.0984 16.0002
0.12 25 0.1209 18.4395
0.14 5 0.1387 6.3412
0.16 10 0.1621 8.1497
0.18 15 0.1846 13.2106
0.2 20 0.2113 17,4974
0.22 25 0.2351 19.1312
74 Investment Model
5.2 Optimal Control Problem in Financial

Modeling
Introduction
Financial modeling has experienced dramatic improvements since the be-
ginning of the 20th century. Especially stochastic financial modeling was
developed significantly. In 1970 Davis [4] stated a deterministic model for
optimal corporate finance in form of a next non-linear optimal control prob-
lem:  
 ZT 
max P (T )e−ρT + e−ρt [1 − ur (t)]rE(t)dt (5.9)
 
0
Ṗ (t) = c[{1 − ur (t)} rE(t) − ρP (t)] (5.10)

E(t)
Ė(t) = rE(t) ur (t) + us (t) 1 − (5.11)
(1 − δ)P (t)
u(t) ∈ U = u ∈ R2 |ur + us ≤ k/r < 1, ur ≥ 0, us ≥ 0

(5.12)
where P (0) = P0 , E(0) = E0 and the terminal condition is given by planning
horizon T 1 .
Then in joint paper with Elzinga [3] Davis provided a possible analytical
solution to his problem (5.9)-(5.12). In 2005 Chen and Sardar [11] suggested
a computational method for solving Davis and Elzinga model and provided
computer software SCOM2 . However, regardless of the nonconvex struc-
ture3 of the problem, they applied Pontryagin’s maximum principle to the
problem finding a local solution.
In this chapter we suggest the conditional gradient method [1] with
appropriate algorithm for finding local optimums of the original problem.
1
P (t) - market price of a share of stock, E(t) - equity per share of outstanding common
stock (net worth of utility divided by the number of shares outstanding), ur - retention
rate which describes the fraction of earnings retained for increasing the capital assets,
us - stock financing rate concerning new money invested in the company, ρ - market
capitalization rate (or investor discount rate), k - maximum investment rate, r - rate of
return to equity (maximum return allowed by government), δ - discount on share price
resulting from flotation cost, c - a positive constant denoting the responsiveness of the
price to changes in earnings and dividends, T - planning horizon of the optimal financing
program.
2
In 2001 Craven and Islam first developed this package and used for solving optimal
control problems.
3
In (5.11) right side of the differential equation has a state variable in quadratic form
2
(E (t)) which results in nonconvex form of the reachable set.
5.2 Optimal Control Problem in Financial Modeling 75
5.2.1 Conditional gradient method

Let us rewrite state variables as x1 (t) = P (t), x2 (t) = E(t) and restate the
model as follows:
 
 ZT 
−ρT −ρT
max x1 (T )e + e [1 − ur (t)]rx2 (t)dt (5.13)
 
0
x˙1 (t) = c[{1 − ur (t)} rx2 (t) − ρx1 (t)] (5.14)

x2 (t)
x˙2 (t) = rx2 (t) ur (t) + us (t) 1 − (5.15)
(1 − δ)x1 (t)
x1 (0) = x01 , x2 (0) = x02 (5.16)

(5.17)
We introduce a new variable x3 as follows:
Rτ
x3 (τ ) = e−ρt [1 − ur ]rx2 (t)dt,
0
x˙3 (t) = e−ρt [1 − ur ]rx2 (t),
x3 (0) = 0.
Then we can restate (5.13)-(5.17) in a terminal functional form:
min ϕ(x(T )) = −x1 (T )e−ρT − x3 (T )

(5.18)
x˙1 (t) = c[{1 − ur (t)} rx2 (t) − ρx1 (t)] (5.19)

x2 (t)
x˙2 (t) = rx2 (t) ur (t) + us (t) 1 − (5.20)
(1 − δ)x1 (t)
x˙3 (t) = e−ρt [1 − ur ]rx2 (t) (5.21)
x1 (0) = x01 , x2 (0) = x02 (5.22)

(5.23)
Problem (5.18)-(5.23) is nonconvex optimal control problem, and therefore
applying Pontryagin’s maximum principle cannot always guarantee a global
solution. As control constraint set U is compact, we can use conditional
gradient method in infinite dimensional space.
Let us construct Pontryagin’s function as follows:

x2
H(Ψ, x, u, t) = Ψ1 (t)c[(1 − ur )rx2 − ρx1 ]+Ψ2 (t)rx2 ur + us 1 − +
(1 − δ)x1
+ Ψ3 e−ρt [1 − ur ]rx2 (5.24)
76 Investment Model
where Ψ1 , Ψ2 , Ψ3 are the solutions of the next system of adjoint differential

equations:
x22 rΨ2 us

∂H

 Ψ̇1 (t) = − ∂x 1
= cΨ 1 ρ + x21 (1−δ)

∂H
n − ∂x2 = r(cΨ
Ψ̇2 (t) = o 1 +n Ψ3 e−ρt ) o − r[Ψ2 − cΨ1 − Ψ3 e−ρt ]ur −





 −rΨ2 1 − (1−δ)x
 x2 Ψ2 rx2

1
us + (1−δ)x 1
us
∂H (5.25)

 Ψ̇3 (t) = − ∂x3 = 0
∂ϕ
Ψ1 (T ) = − ∂x = e−ρT



 1

 Ψ (T ) = 0
 2


Ψ3 (T ) = 1
The objective functional of (5.13)-(5.17) is J(u) = ϕ(x(T )). It’s gradient

can be calculated as follows:

˙ ∂H ∂H ∂H
J(u) =− =− , ,
∂u ∂ur ∂us
where,
∂H −ρT
(
− ∂u r
= Ψ1hcrx2 −
Ψ2 rx2 + Ψ3 ei rx2
∂H x2 (5.26)
− ∂u s
= − Ψ2 rx2 1 − (1−δ)x1
Then (5.13)-(5.17) can be restated as follows:
min J(u), U ⊂ H,
u∈U (5.27)
J(u) ∈ C 1 (H)
where, U is weak compact, H - Hilbert space, C 1 (H) - space of functionals

continuously differentiable on H.
5.2.2 Computational algorithm

Algorithm 1 (Conditional gradient)
∂H(Ψk , xk , uk , t) ∂H(Ψk , xk , uk , t)

˙ k
J(u ) = − ,− ,
∂ur ∂us
where H(Ψk , xk , uk , t) is computed by (5.24).

˙ k ∂H ∂H
min hJ(u ), ui = − ur − us .
u∈U ∂ur ∂us
Let ūk = (ūkr , ūks ) be a solution of the above problem, i.e.:
˙ k ), u¯k i = minhJ(u
hJ(u ˙ k ), ui, t ∈ [0, T ].
u∈U
5.2 Optimal Control Problem in Financial Modeling 77
uk (α) = uk + α(ūk − uk ), 0 ≤ α ≤ 1.
uk (αk ) : min J(uk (α)) = J(uk (αk )).
0≤α≤1
uk+1 := uk (αk ) = uk + αk (ūk − uk ).

Next theorem provides the convergence of the above algorithm:

Theorem 5.1. [20] Sequence uk generated by Algorithm 1 according to
differential maximum principle of Pontryagin converges to stationary point
of problem (5)-(9) as follows:
˙ k ), ūk − uk i = 0
lim hJ(u
k→∞
5.2.3 Numerical Results

As mentioned above Ping Cheng and Islam Sardar [11] provided their nu-
merical solution. They used next values of system parameters ρ = 0.1, k =
0.15, r = 0.2, δ = 0.1, c = 1, P (0) = 0.5, E(0) = 1 and achieved maxi-
mum value of objective functional f ∗ = 2.04 with terminal values of states
P (T ) = x1 (T ) = 1.99, E(T ) = x2 (T ) = 1.96. To find the best solution of
the problem, as U is a triangular set, we chose 4 initial controls as follows:
u0a = {0, 0} , u0b = {0.375, 0.375} , u0c = {0.75, 0} , u0d = {0, 0.75}.
Table 5.1: ur (0) = 0, us (0) = 0

t x1 (T ) x2 (T ) f∗
10 1.363990024 0.005175294 2.12754
20 1.360725818 0.002854255 2.12284
30 1.359562271 0.002308791 2.12111
50 1.358626698 0.001938102 2.12111
70 1.358227107 0.001795339 2.12111
100 1.357928621 0.001694181 2.11866
Table 5.2: ur (0) = 0.375, us (0) = 0.375

t x1 (T ) x2 (T ) f∗
10 1.353424465 0.00353208 2.11288
20 1.355974687 0.00233097 2.11705
30 1.356695453 0.002011626 2.11796
50 1.357103472 0.001781852 2.11826
70 1.357210419 0.001689978 2.11823
100 1.357257493 0.001623573 2.11813
78 Investment Model
Table 5.3: ur (0) = 0.75, us (0) = 0

t x1 (T ) x2 (T ) f∗
10 1.353424465 0.00353208 2.11288
20 1.355974687 0.00233097 2.11705
30 1.356695453 0.002011626 2.11796
50 1.357103472 0.001781852 2.11826
70 1.357210419 0.001689978 2.11823
100 1.357257493 0.001623573 2.11813
Table 5.4: ur (0) = 0, us (0) = 0.75

t x1 (T ) x2 (T ) f∗
10 1.354151332 0.002330051 2.11096
20 1.352959975 0.001681294 2.10937
30 1.352536452 0.001508424 2.10878
50 1.352196923 0.001383136 2.1083
70 1.35205227 0.001332709 2.1081
100 1.35194438 0.001296109 2.10794
From this we can see that the solution we found is f ∗ = 2.12754 > 2.04,
which is better than the previous solutions proposed by Chen and Sardar.
Bibliography
[1] Bredies, K., Lorenz, D. A. and Maass P., (2007), A generalized con-
ditional gradient method and its connection to an iterative shrinkage
method, Computational Optimization and Applications, Volume 42,
Issue 2, pp 173-193
[2] Craven, B. D., HAAS, K. De. and Wettenhall, J. M. (1998), Comput-

ing optimal control, Dynamics of Continious, Discrete and Impulsive
Systems, pp. 601-615.
[3] Davis, B. E. and Elzinga, D. J. (1970), The solution of an optimal

control problem in financial modeling, Operations research: the journal
of the Operations Research Society of America, 19(6).
[4] Davis, B. E., Investment and Rate of Return for the Regulated Firm,
The Bell Journal of Economics and Management Science, Volume 1,
No. 2, 245-270, 1970.
[5] Elton, E., Gruber, M., Finance as a Dynamic Process, Prentice-Hall,

New Jersey, 1975.
[6] Goh, G. J. and Teo, K. L. (1987), MISER, an Optimal Control Soft-

ware, Department of Industrial and Systems Engineering, National
University of Singapore.
[7] Krouse, C. G., Lee, W. Y., Optimal equity financing of the corporation,
Journal of Financial and Quantitative Analysis, Voulme 8, 539-563,
1973.
[8] Lee, H. W. J., Teo, K. L., Rehbock, V. and Jennings, L. S. (1997),

Control parameterization enhancing technique for time optimal control
problems, Dynamic Systems and Applications, 6: 243-262.
[9] Modigliani, F. and Miller, M. H., The cost of capital, corporation fi-
nance, and the theory of investment, American Economic Review, Vol-
ume 48, No.3 :261-297, 1958.
80 BIBLIOGRAPHY
[10] Modigliani, F. and Miller, M. H., Corporation income taxes and the
cost of capital, American Economic Review, Volume 53, No.3 :433-443,
1963.
[11] Ping Chen and Islam, S. M. N. (2005), Optimal Control Models in

Finance: A New Computational Approach, Volume 95, Applied Opti-
mization series, Springer.
[12] Ping Chen and Islam, S. M. N. (2008), Advances in Financial Planning

and Forecasting, Volume 3, 37-66, Center for PBBEFR and Airiti Press
Inc.
[13] Pontryagin, L. S., Boltyanskii, V. G., Gamkrelidze, R. V., Mis-

chenko, E. F., The mathematical theory of optimal processes, ”Nauka”,
Moscow, 1969.
[14] Sethi, P. S., Thompson, G. L., Optimal Control Theory: Applications

to Management Science and Economics, Springer, 2000.
[15] Sethi, P.S., Optimal equity financing model of Krouse and Lee: correc-
tions and extensions, Journal of Financial and Quantitative Analysis,
Volume 13, 487-505, 1978.
[16] Jennings, L. S., Fisher, M. E., Teo, K. L. and Goh G. J. (1991),

MISER3.0: Solving optimal control problems - an update, Advances in
Engineering Software, 13.
[17] Tjatjushkin, A. I., Numerical methods for optimization of controlled

systems, Stability and control: Theory and Applications, Volume 3,
No. 2, 150-174, 2000.
[18] Tjatjushkin, A. I., Numerical solution of multiextreme optimal control

problem, Russian Academy of Sciences, No.5, 59-67, 2004.
[19] Vasyliev, O. V., Methods of solving extreme problems: A course book,

”Nauka”, Moscow, 1981.
[20] Vasyliev, O. V., Numerical methods of solving extreme problems: A

course book, ”Nauka”, Moscow, 1988.
6
Population Model
82 Population Model
6.1 Application of Stochastic Differential Equa-

tions in Population growth
Introduction
Population as a function of time is usually described by dynamic models
based on differential equations. The most common practical methods are
component, exponential and logistic models. On the other hand, the pop-
ulation growth can be considered as stochastic variables since a number of
population depends on social and economic policies, political stability, epi-
demic decease, plague, disaster and so on. Due to the unpredictable nature
of population growth, we introduce uncertainty in the population models.
We choose the linear multiplicative noise, that is a random element to the
exponential and logistic growth equations.
We applied the deterministic and stochastic models to the Mongolian
population growth. In recent years, Mongolian population has grown rapidly
and reached 3.1 million people in 2016.
According to estimation of United Nations Population Fund, the popu-
lation of Mongolia will reach 4 million around 2038 under the high scenario
and 2040 under medium scenario. Under the low scenario the population
would not reach 4 million within 2040. Also W.Patrick using a different
approach carried out Mongolian long term population projection as follows:
3065.5 thousand in 2020, 3223.9 thousand in 2025, 3353.1 thousand in 2030,
3466.3 thousand in 2035.
In this chapter, we first examine the existing population models such as
exponential and logistic model. Then we consider the stochastic models for
population growth, compare the models using Monte Carlo simulation and
predict the dynamics of Mongolian population up to 2035 year using the
models.
6.1.1 Population models

Deterministic models
So far different mathematical models have been developed to model popu-
lation dynamics. In this section, we introduce two common deterministic
population models such as exponential and logistic.
1. Exponential model
One of the models for population growth is based on the assumption
that the population grows at a rate proportional to the number of
6.1 Application of Stochastic Differential Equations in Population growth83
population. This yields the following initial value problem (IVP)

(
dY
dt
= rY,
(6.1)
Y (0) = Y0
where r is a measure of the growth rate, t is a period, Y is equal to

the number of individuals at time t and Y0 is the population for the
base year. The equation is called the exponential growth.
The solution of the initial value problem is
Y (t) = Y0 ert . (6.2)
2. Logistic model
A population often increases exponentially in its early stages but levels
off eventually and approaches its carrying capacity because of limited
resources. One of the first to model this limitation was the logistic
growth model (
dY Y
dt
= rY (1 − M ),
(6.3)
Y (0) = Y0
where M is the carrying capacity or the maximum number of popu-
lation for a given period.
By solving the initial value problem we get
M Y0
Y (t) = . (6.4)
Y0 + (M − Y0 )e−rt
6.1.2 Stochastic models

Before we consider Stochastic models, we introduce the stochastic process
or the Wiener process and the stochastic differential equations.
A stochastic process is a family of random variables that depend on time.
The stochastic process can either be a discrete-time process or a continues-
time process. An example of the latter one is Wiener process or Brownian
motion. The Wiener process, W (t), t ≥ 0, satisfies the following properties
[6].
(i) [Independence of increments] For any 0 ≤ s ≤ t, the increment W (t)−

W (s) is independent of the past.
(ii) [Normal increment] For any 0 ≤ s ≤ t the increment W (t) − W (s) has
the normally distribution with mean 0 and variance t − s.
84 Population Model
(iii) [Continuity of paths] W (t) is continuous function of t.

(iv) W (0) = 0.
√
An implication of (ii) and (iv) is that W (t) N (0, t).
By introducing noise into an ordinary differential equation (ODE) a
stochastic differential equation (SDE) is obtained. Let Y (t) be an
unknown stochastic process, W (t) a Wiener process, µ(Y, t) and σ(Y, t)
known functions, then
dY (t) = µ(Y (t), t)dt + σ(Y (t), t)dW (t) (6.5)
is an SDE driven by a Wiener process. The function µ(Y, t) is called the
drift coefficient and σ(Y, t) is called the diffusion coefficient.
An initial value problem for SDE is
(
dY (t) = µ(Y (t), t)dt + σ(Y (t), t)dW (t),
(6.6)
Y (0) = Y0 .)
As the case of deterministic differential equations, SDEs may not be explic-
itly solvable. When this is case, numerical methods can be used to obtain
an approximation of the solution if a solution exists. The noise is called ad-
ditive if the diffusion coefficient, σ, doesnt depend on the stochastic process,
Y (t), or multiplicative if it does depend on Y (t).
1. Stochastic exponential growth model
For simplicity, we consider that the drift coefficient is a proportional
to the stochastic process and the diffusion coefficients is a linear mul-
tiplicative. Then we obtain the stochastic exponential growth model
dY (t) = rY (t)dt + σY (t)dW (6.7)
where r is the growth rate.

For simulation, let the number of population for base year be Y0 at
time t = 0 and ti = iδt, so the numbers of population are to be
determined at discrete points ti . Then our discrete-time model is
√
Y (ti+1 ) = Y (ti ) + rδtY (ti ) + σ δtNi Y (ti ), (6.8)
where Xi (i = 0, 1, 2, . . .) are i.i.d N (0, 1).
Based on the Central Limit theorem and Xi (i = 0, 1, 2, . . .) are i.i.d
N (0, 1), we take a limit δt → 0 to get continuous population model
[1]. Continuous time expression for the model is
1 2 )t+σ
√
Y (t) = Y0 e(r− 2 σ tX
, (6.9)
where X ∼ N (0, 1).

A random variable Y (t) of the form has a so-called log normal distri-
bution, that is, its log is normally distributed. So we can describe
the evolution of the population over any sequence of time points
0 = t0 < t1 < t2 < · · · < tM by
1 2 )(t √
i+1 −ti )+σ ti+1 −ti Xi
Y (ti+1 ) = Y (ti )e(r− 2 σ , (6.10)
for i.i.d Xi ∼ N (0, 1).
2. Stochastic logistic growth model

We choose the linear multiplicative noise, σX(t), a random element
to the logistic growth equation
Y (t)
dY (t) = rY (t) 1 − dt + σY (t)dW (6.11)
M
where r is a measure of the growth rate and M is a carriyng capacity
of the population. The solution can be shown to be [3]
c2
Y0 M e(r− 2 )t+cW (t)
Y (t) = Rt c2
(6.12)
M + Y0 r 0 e(r− 2 )s+σW (t) ds
If there is no noise in the system, i.e. when σ = 0, (6.12) reduces to

the solution of the deterministic logistic growth model. The integral in
the denominator of equation (6.12) is not readily obtainable. However,
expected value of it can be calculated and found to be [4]
er t
E(I(t)) = ,
r
Rt c2
where I(t) = 0
e(r− 2 )s+σW (s) ds.
Since equation (6.12) is a quotient of two dependent random variable
its expected value is not readily obtainable. However, applying the
Delta method the expected value can be approximated by the ex-
pected value of the numerator divided by the expected value of the
denominator, i.e.
Y E(Y )
E ≈ .
Z E(Z)
So this yields
Y0 M e r t
E(Y (t)) = (6.13)
M + Y0 (er t − 1)
86 Population Model
which is the solution of deterministic logistic growth model. Thus,

the expected value of the solution to stochastic logistic growth model
can be approximated with the solution of deterministic model.
For discretization, let the number of population for base year be Y0
at time t = 0 and ti = iδt, so the numbers of population are to be
determined at discrete points ti . Then our discrete-time model for
logistic growth is
Y (ti ) √
Y (ti+1 ) = Y (ti ) + rY (ti ) 1 − δt + σ δtNi Y (ti ), (6.14)
M
where Xi (i = 0, 1, 2, . . .) are i.i.d N (0, 1).
Simulation results
We applied the deterministic and stochastic models to the Mongolian popu-
lation growth. In recent years, Mongolian population has grown rapidly and
reached 3.1 million people in 2016. In our work, we estimated the growth
rate using statistical data from 2000 to 2016 obtained by the National Sta-
tistical Office of Mongolia. According to the least square estimation, the
average population growth rate was 1.59 during that periods. We use
Figure 6.1: Mongolian population/thous.people/
Sourse:National Statistical Office of Mongolia
Monte Carlo simulation to generate the stochastic process. Since we have

already estimated population growth rate as r = 0.0159 and we choose
M = 35000 and different σ, we can generate sample paths starting from
2403.1 thousands in 2000 using different pseudo random number generators
for each path. In the following graphs show some sample paths and mean
for the stochastic exponential and logistic models. We compared our mod-
els such us deterministic and stochastic models with real data from 2000
to 2016 and predicted Mongolian population during 2017-2035. Results are
Figure 6.2: Stochastic exponential model
Figure 6.3: Stochastic logistic model
shown on the graphs and Table 6.1. The graphs show that stochastic models
are more suitable for prediction of the population.
Table 6.1 Population up to the year 2035 (thousand)

Year Exponential Stoc. exp. Logistic Stoc. log.
model model model model
2017 3098.3 3097.7 3038.0 3041.6
2019 3198.3 3204.5 3127.2 3131.5
2021 3301.5 3309.8 3218.9 3226.5
2023 3408.1 3413.8 3312.9 3320.7
2025 3518.0 3524.3 3409.4 3418.6
2027 3631.6 3638.3 3508.4 3520.9
2029 3748.8 3761.6 3600.0 3623.2
2031 3869.7 3880.3 3714.1 3730.3
2033 3994.6 3999.7 3820.9 3835.5
2035 4123.5 4121.1 3930.3 3946.8
88 Population Model
Figure 6.4: Comparision of the models
We can conclude that
• In this chapter we only considered the linear diffusion coefficient in

exponential and logistic growth models. Other diffusion coefficient are
plausible. For example, the logistic growth model can be perturbed
with white noise.
• According to statistical data, stochastic models are more suitable for

prediction of the population.
• According to the stochastic exponential prediction, our population

will reach 3309.8 thousand in 2021, 3524.3 in 2025, 3880.3 in 2031
and 4121.1 thousand in 2035 where as the number of population will
be 3226.5 thousand in 2021, 3418.6 in 2025 and 3946.8 thousand in
2035 by the stochastic logistic model.
Bibliography
[1] Desmond J.Higham, (2004)An introduction to Finantial option val-

uation, Cambridge University Press.
[2] R.Enkhbat, N.Tungalag, D.Bayanjargal Mathematical model for pre-

diction of population growth, First International conference on Fore-
sight and Forecasting 2010, Proceeding, pp.32-37, 2010.
[3] Kloeden, Peter E. and Platen, Eckhard (1999). Numerical Solution

of Stochastic Differential Equations. Berlin:Springer.
[4] Kristina Lundquist.,(2011) A comparative study of Stochastic and

deterministic population models, Umea Universitet.
[5] Paul Wilmott.,(2007)Introduces Quantitave Finance, 2nd edition,

John Wiet and Sons, Ltd.
[6] Zszislaw Brzezniak and Tomas Zastawniak.,(2005) Basic Stochastic

Processes, Springer.
[7] Richard L.Burden, J.Douglas Faires (2010), Numerical Analysis, 9th

edition, Cengage Learning.
7
Optimization Applications in Finance
7.1 A Mathematical Model of Marketing Advertisement for Insurance
Companies 91
7.1 A Mathematical Model of Marketing Ad-
vertisement for Insurance Companies
Introduction
After a certain period of time, the spending on advertising can increase
the sales and provide an opportunity to growth market extent (Bain [2],
Vidal-Wolfe [18], Nelson [8]). Erickson [5] confirmed that the enterprises
cost-effectiveness reduced and expected consequences couldnt be reached
when insufficient attention was brought to the customer behavior. Fruchter
[6] and Sethi [12] explained that ineffective advertising spend is the reason,
Tsurumi [17] and Bensoussan [3] explained that the insufficient amount
of advertising is the cause. Sasieni [11] described the optimal control with
respect to mathematical modeling, Neumann [9] dynamic process, Sethi [12]
evolutionary behavior change, Tapiero [15] and Sethi scholastic form, Deal
[4] - the direct impact of competition, Teng Thompson [16], and Jorgensen
[7] an indirect effect, Wells [19] - the behavior of the consumer.
7.1.1 Extension of advertising model

If π(0) is preliminary profit, π(t) = mx(t)−cu2 (t) is an operating profit, and
−rπ(T ) expr(T −t) = γ is non-operating profit, then the objective function is
described by the following formula [10].
Z ∞
−rt 2
J (x(t), t) = max E exp (γ + mx(t) − cu (t))dt
u(t)≥0 0
Extension of advertising model is based on a stochastic differential game,

which is formulated as follows:
Z ∞
−r1 t 2
J1 (x(t), t) = max E exp (γ1 + m1 x(t) − c1 u1 (t))dt
u1 (t)≥0 Z0 ∞
−r2 t 2
J2 (y(t), t) = max E exp (γ + m2 y(t) − c2 u2 (t))dt
√u2 (t)≥0 0 √
dx = [ρ11 u1 (x, y)√ 1 − x − ρ21 u2 (x, y) x − δ1 (2x − 1)]dt + θ1 σ1 (x, y)dw1
√
dy = [ρ22 u2 (x, y) 1 − y − ρ12 u1 (x, y) y − δ2 (2y − 1)]dt + θ2 σ2 (x, y)dw2 ,
(7.1)
here, x, y- market share of the first and second company’s sales, u1 , u2 -
advertising spending of the first and second company’s sales, ρ11 - effective-
ness of the first company’s advertising spending, ρ12 - reaction coefficient of
the first companys market share to impact of the second company’s mar-
ket share, ρ22 - effectiveness of the second company’s advertising spending,
92 Optimization Applications in Finance
ρ21 - reaction coefficient of the second company’s market share to impact

of the first company’s market share, m1 , m2 - marginal profit of the first
and second company’s sales, c1 , c2 - marginal cost of advertising of the first
and second company’s sales, δ1 , δ2 - loss of the first and second company’s
market, γ1 , γ2 - coefficients of the first and second company’s non-operating
profit, θ1 , θ2 -coefficients of the first and second company’s customer behav-
ior, σ1 dw1 , σ2 dw2 - white noises of the first and second company’s market
share.
As some characteristics of the coefficients, the extensions of advertising
model were described in Vidale Wolfe model [18], if γi = 0, ci = 1, θ1 =
θ2 = 0, δ1 = δ2 , ρ21 = 0, ρ12 = 0, Sethi model [13], if γi = 0, ci =
1, θ1 = θ2 = 1, δ1 = δ2 , ρ21 = 0, ρ12 = 0, Sorger model [14], if γi =
0, θ1 = θ2 = 0, δ1 = δ2 = 0, ρ11 = ρ12 , ρ21 = ρ22 , Prasad model [10], if
γi = 0, θ1 = θ2 = 1, δ1 = δ2 = δ, ρ11 = ρ12 , ρ21 = ρ22 .
We examine case when γi 6= 0, ci 6= 0, θ1 6= θ2 , δ1 6= δ2 , ρ11 6= ρ12 , ρ21 6=
ρ22 .
7.1.2 Problem solution

To find the closed-loop Nash Equilibrium strategies, we use the classical
model forming the Hamilton-Jacobi-Bellman equation for each company
[10]:
 (
2 0 √ )
γ1 + m 1 x − c 1 u + J (ρ
1 11 1 u 1 − x−
√ 1

 r1 J1 = umax

θ2 σ2
 00
−ρ21 u2 (x, y) x − δ1 (2x − 1)) + 12 1 J1

1 (t)≥0
(
2 0 √ ) (7.2)
 γ 2 + m 2 x − c u
2 2 + J (ρ
2 22 2u 1 − y−
 r2 J2 = max √

θ2 σ2
 00
 u2 (t)≥0 −ρ12 u1 y − δ2 (2y − 1)) + 22 2 J2
0 0
where J1 , J2 and u∗1 , u∗2 denote the competitors advertising policies in (7.2),
respectively. We obtain the optimal feedback advertising decisions
√ √
 0 0
 u∗ (x) = J1 ρ11 1−x J1 ρ11 y
1 2c√
= 2c1
0
1
J2 ρ22 1−y
0 √
J2 ρ22 x
(7.3)
 u∗ (x) = =
2 2c2 2c2
Following Sethi [13], we find solutions for value functions J1 = α1 + β1 x and

J2 = α2 + β2 y. From equations (7.2) and (7.3) we determine the unknown
Companies 93
coefficients of x0 , x, y 0 , y.
β12 ρ211

0


 x : r1 α1 = γ1 + 4c1
+ β1 δ1
 x : r β = m − β12 ρ211 − β1 β2 ρ21 ρ21 − 2β δ

1 1 1 4c1 2c2 1 1
0 β22 ρ222


 y : r2 α2 = γ2 + 4c2 + β2 δ2
 y : r β = m − β22 ρ222 − β1 β2 ρ12 ρ11 − 2β δ

2 2 1 4c2 2c1 2 2
Substituting W1 = r1 + 2δ1 , W2 = r2 + 2δ2 , R2 = ρ22 /2c2 , R1 = ρ11 /2c1

into above system we get
W1 β1 = m1 − ρ211 R1 β12 − ρ21 β1 β2

W2 β2 = m2 − ρ222 R2 β22 − ρ12 β1 β2
The system reduces to the following four power equation:
β14 + k1 β13 + k2 β12 − k3 β1 − k4 = 0 (7.4)

where r1 > 0, r2 > 0, ρ11 > 0, ρ22 > 0, ρ12 > 0, ρ21 > 0, m1 > 0, m2 >
0, c1 > 0, c2 > 0, δ1 > 0, δ2 > 0, W1 = r1 + 2δ1 > 0, W2 = r2 + 2δ2 >
0, R2 = ρ22 /2c2 > 0, R2 = ρ11 /2c1 > 0.
From equation (7.3), we pick β̄1 as the only real positive solution out of
the four roots, β̄1 = β1 (i∗ ), where i∗ = {i ∈ {1, 2, 3, 4}|β1 (i) > 0}. While i∗
may depend on the data, there will only be one i∗ in every case. Then, β̄2
can be found
m1 − ρ11 R1 β̄12 − W1 β̄1
β̄2 = .
ρ21 R2 β̄1
Then optimal advertising is
 √
 ∗ β̄1 ρ11 1 − x
 u1 (x) =

2c
√1
∗ β̄2 ρ22 1−y
 u2 (y) = .


2c2
The state equation (7.1) becomes as:

dx = [b11 − (b11 + b21 )x]dt + θ1 σ1 (x, y)dw1
(7.5)
dy = [b22 − (b22 + b12 )x]dt + θ2 σ2 (x, y)dw2 ,
where
ρ211 ¯ ρ21 ρ22 ¯
b11 = β 1 + δ1 , b21 = β 2 + δ1 ,
2c1 2c2
ρ222 ¯ ρ12 ρ11 ¯
b22 = β 2 + δ2 , b12 = β 1 + δ2 ,
2c2 2c1
If we search for solutions of system (7.5), the long run equilibrium market
share (x̄, ȳ) is obtained by taking the limits as t → ∞ and are given by
x̄(t) = b11b+b
11
21
(7.6)
ȳ(t) = b22b+b
22
12
.
Computational results
We tested our proposed model on selected insurance companies of Mongo-

lia. Computational results are given in the following tables:
Discount rate Regression model
r1 = 0.022 π(t) = 0.143e0.022t , R2 = 0.41, DW = 0.42
top 5 companies
r2 = 0.004 π(t) = 0.106e0.004t , R2 = 0.22, DW = 2.36
other companies
Value Function Regression model
m1 = 0.45 π1t = 0.45 · xt−1 − 56.53 · u21t
c1 = 56.53 (9.83) (4.75)
2
γ1 = 0 R = 0.43, DW = 0.62
Sorger and Prasad model top 5 companies
Value Function Regression model
m2 = 0.23 π2t = 0.07 + 0.23 · yt−1 − 1.80 · u22t
c2 = 39.79 (4.32) (2.76) (2.23)
γ2 = 0.07 R2 = 0.40, DW = 2.31
Extented model other companies
State equation GARCH Model
√ √
ρ11 = 2.71 xt = 0.93 · xt−1 + 2.71 · u1t 1 − xt−1 − 1.80 · u2t xt−1
ρ21 = 1.80 (24.5) (1.91) (2.03)
1−0.93
δ1 = 2 = 0.03 σ1t = 0.00006 + 0.6 · σ1,(t−1) + 0.15 · e21,(t−1)
2 2
θ1 = 0.6 (0.08) (0.15) (0.17)

2
Extention model R = 0.40, DW = 2.31 P rob.F (1.15) = 0.23
√ √
ρ22 = 2.04 yt = 0.06 + 0.92yt−1 + 2.04u2t 1 − yt−1 − 2.62u1t yt−1
ρ12 = 2.62 (3.66) (11.08) (2.81) (4.56)
1−0.92
δ2 = 2 = 0.04 σ1t = 0.00002 + 1.20 · σ1,(t−1) + 0.27 · e21,(t−1)
2 2
θ2 = 1.2 (2.38) (8.77) (−2.49)

Extention model R2 = 0.98, DW = 1.73 P rob.F (1.15) = 0.98
Other parameters are calculated by the model.
Companies 95
The parameters in the model Computed parameters
R1 = ρ2c111 > 0 2.71
R1 = 2·56.532 = 0.0240
R2 = ρ2c222 > 0 2.046
R1 = 2·39.789 = 0.0257
W1 = r1 + 2σ1 > 0 W1 = 0.0226 + 2 · 0.0342 = 0.0909
W2 = r2 + 2σ2 > 0 W2 = 0.0037 + 2 · 0.0420 = 0.0876
2
ρ11 ρ12 ρ21 − ρ22 (ρ211 )/4 > 0 2.71 · 2.62 · 1.80 − 2.04 · 2.71
4
= 9.0356
(2ρ21 ρ12 − ρ22 ρ11 )W1 + ρ11 ρ12 W2
k1 =
2ρ ρ ρ 2
>0 k1 = 3.6120 > 0
R1 11 12 21 − ρ22 (ρ11 )/4
(2ρ21 W1 −ρ22 W1 )W1 +m1 R1 (ρ11 ρ22 −2ρ12 ρ21 )+2m2 ρ2
k2 =
R1 2 ρ ρ ρ −ρ (ρ2 )/4
11 12 21 22 11
21 R2 >0 k2 = 1.4682 > 0
2m1 (ρ21 W1 −ρ22 W1 )
k3 = R12 ρ11 ρ12 ρ21 −ρ22 (ρ211 )/4
<0 k3 = −4.8322
ρ22 m21 2.0426·0.45032
k4 = R12 ρ11 ρ12 ρ21 −ρ22 (ρ211 )/4
>0 k4 = 0.0242 ·9.0356
We rewrite four power equation using the above parameters
β14 + 3.612 · β13 + 1.4682 · β12 + 4.8322β1 − 79.789 = 0.
The equation has two complex solutions. One negative and one positive.
One positive solution is β¯1 = 2.2016 and β¯2 = 0.911. Therefore, the market
equilibrium levels are b11 = 0.1772, b21 = 0.0763, b22 = 0.0897, and b12 =
0.1801.
The large insurance company occupies 69.9% of the market equilibrium,
and a small insurance company has been stabilized at 33.3% percent of the
market equilibrium. Optimal advertising spending of the large insurance
company is
√ √
∗ β¯1 ρ11 1 − x 2.0216 · 2.71 1 − 0.699
u1 (x) = = = 0.027.
2c1 2 · 56.532
The optimal advertising spending of the small insurance company is
√ √
β¯2 ρ22 1 − y 0.911 · 2.0426 1 − 0.3325
u∗2 (y) = = = 0.019.
2c2 2 · 39.789

Evaluation
Introduction
The assessment of credit risk has always been important to banks and other
financial institutions. Recently banks have devoted even more resources
than usual to this task. This is because, under the proposals in Basel II,
regulatory credit risk capital may be determined using a bank’s internal
assessments of the probabilities that its counter-parties will default[20].
One popular approach to assessing credit risk involves [24] model. This
model assumes that a company has a certain amount of zero-coupon debt
that will become due at a future time T . The company defaults if the
value of its assets is less than the promised debt repayment at time T . As
inputs, Mertons model requires the current value of the company’s assets,
the volatility of the company’s assets, the outstanding debt, and the debt
maturity. A way of implementing Merton’s model estimates the current
value of the company’s assets and the volatility of the assets from the market
value of the company’s equity. A debt maturity date is chosen and debt
payments are mapped into a single payment on the debt maturity date in
some way[22].
We develop a new way of implementing [24] model. This is based on
household net income to be maximized subject to stochastic differential
equations. We consider objective function of the household net income
introduced by [26] and [25]. Our approach provides a credit risk assessment
and an amount of potential loan.
7.2.1 Model formulation

We consider the following model formulated in [25].
∞ 

(1 − wE )wc
Z
J(x(t)) = max E  e−ρt x(t) − u(t)2 dt , (7.7)
u≥0 r
0
p
dx(t) = ρβu(t) x(t)dt + σx(t)dW (t), (7.8)
rl = E l − e−r(T −t) x(T ) ,

(7.9)
u(t) ≥ 0, σ > 0, (7.10)
where J is value function, ρ: household income growth, wE : household ex-
penses to income ratio, wC : volume of loan to collateral ratio, x: household
Evaluation 97
income to collateral ratio, u: potential loans to collateral ratio, t: maturity
of loan, r: interest rate, β: parameter of continuously function, σ: volatility
of the collaterals, W : Brownian motion, rl : credit risk to collateral ratio, l:
loan payment to collateral ratio, T : future time.
Hamilton-Jacobi-Bellman equation [10] for the problem is the following:
√ ∂J(x) 1 2 2 ∂ 2 J(x)

(1 − wE )wC 2
ρJ(x) = max x − u + ρβu x + σ x
u r ∂x 2 ∂ 2x
(7.11)
Taking a derivative with respect to u, we obtain as
√
∗ ρβ x ∂J(x)
u (x) = .
2 ∂x
Suppose that the value function is linear, that is J(x) = ax. Then the
value of a satisfies the equation:
2 2
ρβ (1 − wE )wC
a2 − ρa + = 0.
4 r
Then the unknown coefficient a is found as:
q
1 ± 1 − β 2 (1−wrE )wC
a1,2 = =λ (7.12)
ρβ 2 /2
Substituting (7.12) into (7.8), we get:
dx(t) ρ2 β 2
= λdt + σdW (t). (7.13)
x(t) 2
Now it is worth to mention well-known [21].
Lemma 7.1. [21] Let f (t, x) be a function for which the partial derivatives
ft (t, x), fx (t, x) and fxx (t, x) are defined and continuous. Then, for every
T > 0, with the Ito process x(t) replacing the Brownian motion W (t), we
have
1
df (t, W (t)) = ft (t, W (t))dt + fx (t, W (t))dW (t) + fxx (t, W (t))dW (t)dW (t)
2
A rigorous proof of this lemma can be found in [20].
Now we rewrite equation (7.13) as
dx(t) ρ2 β 2
= dln(x(t)) = λdt + σdW (t).
x(t) 2
This equation reduces to

dx(t) 1 dx(t) dx(t)
dln(x(t)) = − =
x(t) 2 x(t) x(t)
ρ2 β 2 ρ4 β 4 2 ρ2 β 2

1 2
= λdt+σdW (t)− λ dtdt + 2σ λdtdW (t) + σ dW (t)dW (t) .
2 2 4 2
Applying Ito Lemma 7.1 to the above equation, we obtain
2 2
σ2

ρβ
dln(x(t)) = λ− dt + σdW (t).
2 2
The solution of this equation can be easily found:

ρ2 β 2 2
2
λ− σ2 (T −t)+σ(W (T )−W (t))
x(T ) = x(t)e . (7.14)
−(W (T )−W (t))
If we substitute θ = T − t and y = √
θ
, then it becomes:
√

ρ2 β 2 2
2
λ− σ2 θ−yσ θ
x(T ) = x(t)e .
Substituting this into (7.9), we obtain
rl = E l − e−r(T −t) x(T ) =

ρ2 β 2
= l(1 − Φ(h1 )) − e−rθ xe 2
λ
(1 − Φ(h2 )),
where
ρ2 β 2 σ2 l

2
λθ − ln e−rθ
− 2 x
h1 = √ ,
σ θ
2 2
ρ β σ2 l

2
λ + 2
θ − ln e−rθ x
h2 = √ ,
σ θ
Z∞
1 1 2
Φ(h1 ) = √ e− 2 y dy,
2π
0
Z∞ √ √
1 1
θ)2
Φ(h2 ) = √ e− 2 (y+σ d(y + σ θ).
2π
0
Multiplying (8) by C, we get

ρ2 β 2
R = L(1 − Φ(h1 )) − e−rθ Ie 2
λ
(1 − Φ(h2 )),
Evaluation 99
where R: risk, L: loan payment, I: household, λ: parameter of value
function. We simplify the above expression in the following:
(1+r)t −1
wI L

R r(1+r)t
2 2
−rθ ρ 2β λ
= (1+r)t −1 (1 − Φ(h1 )) − e e (1 − Φ(h2 )) =
I t w I I
r(1+r)

S 2 2
−rθ ρ 2β λ
= (1 − Φ(h1 )) − e e (1 − Φ(h2 )) ,
K
where S: required loans, K: potential loans, C: collateral.
7.2.2 Bank requirements and criteria

We can easily see that
K wC
= (1 − wE ) ≥ 0,
C wI
√
∗ ρβ x
u (x) = λ ≥ 0,
2
where wC > 0, wI > 0, wE ≤ 1, ρ > 0, β > 0, λ ≥ 0, x ≥ 0.
From (7.12), we can choose a positive λ such that
q
1 − 1 − β 2 (1−wrE )wC
λ= ≥ 0,
ρβ 2 /2
(1 − wE )wC
1 − β2 ≥ 0.
r
The bank criteria allow us to formulate the following assertion.
Lemma 7.2. Let N I be household net income, and E be household ex-
penses. If N I = L and K = tL then wI = 1 − wE .
Proof. It can be derived from the equalities:

K I −E E Lt I (1 − wE )K
=t· = 1− · · · .
C C I C L wI C
It follows that wI = 1 − wE which proves the lemma.
Lemma 7.3. Let C be collateral. If the following conditions
a) dI = ρuI
dC
b) √
C
= −σdW
c) x = βdt
are satisfied, then the following stochastic differential equation holds.
p
dx(t) = ρβu(t) x(t)dt + σx(t)dW (t).
Proof. Putting I = Cx, we have
dI = dxC + dCx.
Using condition (a) to (c), we obtain

dC √
dx(t) = ρux − x = ρβu xdt + σxdW,
C
The proof is completed.
Household information: Household income is $1500 per month. It spends
$400 on mortgage and $450 on household expenses. Growth of annual aver-
age income is 10 percent. The household requests loan of $8000 with 2 year
maturity of loan. Now our model should decide whether the bank customer
is eligible for loan or not.
The bank criteria: The bank gives a loan when RI < 0.4 and wC = 0.7,
β = 0.16, r = 0.2, wE = σ. Now we need to do the following computations:
$400 + $450
wE = = 0.57,
$1500
q
1 − 1 − 0.162 (1−0.57)·0.7
0.2
λ= 2
= 15,
0.1 · 0.16 · 0.5
wI = 1 − wE = 1 − 0.57 = 0.43,
(1 + 0.2 )24 − 1

12
K = 0.2 · 0.43 · $1500 = $12673,
12
(1 + 0.2
12
)24
2 2
0.1 ·0.16 0.572 $8000

2
15 − 2
· 2 − ln e−0.2·2 ·$12673
h1 = √ = −0.324
0.57 2
2 2
0.1 ·0.16 0.572 $8000

2
15 + 2
· 2 − ln e−0.2·2 ·$12673
h2 = √ = 0.482,
0.57 2
Φ(h1 ) = Φ(−0.324) = 0.37,
Φ(h2 ) = Φ(0.482) = 0.69,
R $8000 h 0.12 ·0.162
i
= (1 − 0.37) − e−0.2·2 e 2 15 (1 − 0.69) = 0.26.
I $12673
R
Since I
= 0.26 < 0.4 then we conclude that the bank can afford credit.
7.3 Optimization Problem of Foreign Reserves 101
7.3 Optimization Problem of Foreign Reserves

Introduction
There are many works on strategic asset allocations but less attention has
been paid to foreign reserves allocation. Particularly, the notion strate-
gic asset allocation was introduced in Brennan et al. [29] to describe the
portfolio optimization problem with time-varying returns and long-term in-
vestor objectives. In general, the problem of long-term investments is a
well-established research field introduced by Merton [34] and Samuelson
[37], respectively. In the continuous time, using stochastic optimal control
theory, Merton [34] has developed the problem of lifetime portfolio selec-
tion under uncertainty. In [34] important financial economic principles were
established, but there are no explicit results for portfolio choice problems.
Merton’s paper [35] highlights the difficulties in solving complex cases of
assets dynamics with stochastic factors. Advances in numerical techniques
and the growth of computing power led to the development of numerical so-
lutions to multi-period portfolio optimization problems, which can be solved
by a discrete state approximation [29, 30]. However, the use of numerical
dynamic programming is very often restricted to few factors, due to the fact
that the algorithms use excessive computation time and come to be numeri-
cally unreliable for high dimensions. Therefore, closed-form solutions of the
Merton model in continuous time with two or three stochastic factors are
given in [31, 32, 36].
In [28], Bielecki et al. present a closed-form solution to the portfolio opti-
mization problem in continuous time for multiple assets and multiple fac-
tors with an infinite time horizon. Under the assumption of the uncorre-
lated residual of the asset prices and factors, they find the optimal portfolio
allocation decision for many assets and many factors. In [28], general ana-
lytical solutions Merton’s type consumption-investment problem had been
received. Brownian motion with a white noise was examined in [28]. Based
on [33], we suppose non-portfolio problem in the continuous time with one
stochastic factor controls optimal value of foreign reserves when the ex-
change rate fluctuates.
7.3.1 Merton’s type consumption - investment

problem
Assume that the investor lives from 0 to time T ; his wealth at time t is
denoted by Xt . The investor starts with a known initial wealth X0 . At
time t the investor must choose what fraction of his wealth to consume ct
and what fraction to invest in the riskless and risky portfolio. Let’s assume
that F (t, ct ) is the period utility function for consumption at time t, and
Φ(T, XT ) is the boundary condition. Then the optimization problem for
the investor is "Z #
T
max F (t, ct )dt + Φ(T, XT ) , (7.15)
u0 ,u1 ,c 0
where E is the expectation operator, T is the planning horizon. Suppose

there are two assets.
1. The investor can invest money in the bank at the deterministic short
rate of interest r, i.e, the investor has access to the risk-free asset B
with
dB(t) = rB(t)dt. (7.16)
2. The investor can invest in a risky asset with price process S, where
we assume that the S-dynamics are given by a standard Black-Scholes
model
dS(t) = αS(t)dt + σS(t)dW, (7.17)
here α: drift function, σ: diffusion coefficient, W (t): standard Brow-
nian motion.
If we assume that, at time t the investor owns n1 , n2 units of asset, then

the total wealth is determined by
X(t) = n1 (t)B(t) + n2 (t)S(t). (7.18)
The portfolio n1 , n2 remains unchanged over the time interval [t, t + ∆t[,
and assuming the consumption pattern is constant in interval [t, t + ∆t) in
the same way as for portfolio selection, the budget equation becomes
dX(t) = n1 (t)dB(t) + n2 (t)dS(t) − c(t)dt. (7.19)
Let
n1 (t)B(t) n2 (t)S(t)
u0 (t) = , u1 (t) = (7.20)
X(t) X(t)
the share of wealth in assets, with
u0 (t) + u1 (t) = 1. (7.21)
Then the budget equation (7.19) can be written as
dX(t) = X(t)[u0 (t)r + u1 (t)α]dt − c(t)dt + u1 (t)σX(t)dW (t). (7.22)

Now, we may formally state the consumers utility maximization problem

as a stochastic optimal control problem
hZ T i
max
0 1
E F (t, ct )dt + Φ(T, XT )
u ,u ,c 0
dX(t) = X(t)[u (t)r+u1 (t)α]dt − c(t)dt + u1 (t)σX(t)dW (t)
0
(7.23)
X0 =x0
c(t) ≥0, ∀t ≥ 0
u (t) + u1 (t) =1, ∀t ≥ 0,
0
here X is a state process, u0 , u1 and c are control variable.
7.3.2 Foreign reserves problem

In problem (7.23), we assume the following.
1. There’s only one stochastic factor described by the risky asset.
2. Social welfare increases due to foreign reserves.
3. Changes of foreign reserves isn’t always constant.
4. Boundary condition of foreign reserves is zero.

There are u0 (t) = 0, u1 (t) = w(t) for (1), F (t, ct ) = e−δt Rtγ for (2),
c(t) = βR(t) for (3), and Φ(T, XT ) = 0 from (4).
According to the above assumptions, (7.23) can be written as follows.
hZ T i
maxE e−δt Rtγ dt
w,R 0
dX(t) = w(t)αX(t)dt−βR(t)dt + w(t)σX(t)dW (t)
(7.24)
X0 =x0
R(t) ≥0, ∀t ≥ 0
w(t) =1, ∀t ≥ 0,
where the objective function does not contain X(t), shown in (7.23) and
assumption (4).
The Hamilton-Jacobi-Bellman (HJB) equation has the following form.
( )
2
∂V γ ∂V ∂V 1 ∂ X
+ sup e−δt Rt + wαX − βR + w2 σ 2 X 2 = 0 (7.25)
∂t R≥0,w∈R ∂X ∂X 2 ∂X 2
Let R and w be solutions to problem (7.25). Assuming an interior solution,

the first order conditions are
γRγ−1 = βeδt VX , (7.26)
αVX
w=− . (7.27)
Xσ 2 VXX
Taking into account assumption (2), we have
V (t, X) = e−δt hX γ .
The boundary conditions require that
Φ(T, XT ) = 0.
Also, we can get the following equations
∂V
= e−δt hX γ − σe−δt hX γ , (7.28)
∂t
∂V
= γe−δt hX γ−1 , (7.29)
∂X
∂ 2V
2
= γ(γ − 1)e−δt hX γ−2 . (7.30)
∂X
Substituting (7.29) into (7.26), (7.30) and (7.27), we get
R∗ (t, X) = (βh(t))−1/(1−γ) X, (7.31)
α
w∗ (t, X) = . (7.32)
(1 − γ)σ 2
If we substitute the expressions (7.28) and (7.31) - (7.32) into the HJB
equation, we obtain:
X γ ḣ(t) + Ah(t) + Bh(t)−γ/(1−γ) = 0,

(7.33)
where A and B are the constants given by
1 α2 γ
A= −δ
2 (1 − γ)σ 2
B = (1 − γ)β −γ/(1−γ) .
If we search y in the following form y = h1/(1−γ) ), we get

B − 1−γ
A
t B
y(t) = y(0) − e + . (7.34)
A A
We can easily see that
B
lim y(t) = .
t→∞ A
Now we can find the optimal value of foreign reserves using the expression
(7.35).
A
R∗ (t, X) = β −1/(1−γ) X . (7.35)
B
From here we find expected foreign reserves:
A
E(R∗ ) = β −1/(1−γ) E(X). (7.36)
B
7.3.3 Estimation of econometric model

From (7.24), a stochastic differential equation is written as
dX(t) R(t)
= w(t)αdt − β dt + w(t)σdW (t).
X(t) X(t)
In order to construct econometric equation, assume that w(t) = w and

wσdW (t) = u(t). Then we have the following econometric equation:
dX(t) R(t)
= wα − β + u(t).
X(t) X(t)
The first-order autocorrelation is:
u(t) = ρu(t − 1) + e(t).
where ρ can be computed

Cov(wσdW (t), wσdW (s))
ρ= . (7.37)
V ar(wσdW (t))
Corollary 1. if s, t are positive, then Cov(dW (t), dW (s)) = V ar(dW (t))σ(s−

t) for the differential process, where σ(s − t) is the continuous Dirac delta
function. If we set σ as σ(s − t) = 1, then ρ becomes
(
0, f or t 6= s
ρ= 1
wσ
, f or t = s.
Corollary 2.τ Let R(t) be some foreign reserves with an exponential growth
R(t) = R(0)e γ t . If objective function is
"Z #
T
max E e−δt Rtγ dt .
w,R 0
Then, there exists τ so that τ ≈ σ. Proof follows from

"Z #
R
Rγ (0) (τ −σ)T
e−rt Rγ (0)eτ t dt =

max E e − 1 > 0. (7.38)
w,R 0 τ − σ)
If τ − σ > 0, the objective function is positive. Otherwise, the inequality

is valid if 0 < τ < σ. Formula (7.32) can be written:
(αw)w
w∗ = .
(1 − γ)(σw)2
From the above, we have

αw
γ =1− 2
= 1 − αwρ2 .
(σw)
Optimal value of foreign reserves is

1 α2 γ
1
− 1−γ A 1 2 (1−γ)σ 2
−δ
E(R∗ ) = β E(X) = β − 1−γ − γ E(X) =
B (1 − γ)β 1−γ
! !
1 1h α i 1 1
= αγ − δ E(X) = αwγ − τ E(X).
β(1 − γ) 2 (1 − γ)σ 2 β(1 − γ) 2
We made an estimation of the econometric model based on observations
of 132 months for the period from 2006 to 2016 years. The first-order
autocorrelation estimation gives the following values:
dX(t) R(t)
= 0.182 − 0.072 · + 0.957 · AR(1), R2 = 0.908, DW = 1.46
X(t) X(t)
(2.41) (−3.2) (35.7)

It means that all parameters of the equation are statistically significant.
Taking into account Corollary 2, the quarterly average growth of foreign
reserves is
τ /γ = 0.066.
Also, expected exchange rate level is
E(X) = 1465.715.
Thus, all parameters are the following:
wα = 0.182,
β = 0.072,
ρ = 0.957.
If we compute γ and τ then
γ = 1 − 0.182 · 0.9572 = 0.8333,
τ = 0.8333 · 0.066 = 0.055,

and the expected optimal value of foreign reserves is 2.5 billion USD.
∗ 1 1
E(R ) = αwγ − τ E(X) =
β(1 − γ) 2
1465.715 1
= · · 0.182 · 0.8333 − 0.055 ≈ 2544.
0.072 · (1 − 0.8333) 2
Bibliography
[1] Alister, L. and Pessemier, E.A.: Varied consumer behavior, paper no.
774-775, Krannert Graduate School of Management, Purdue Univer-
sity (1981).
[2] Bain, J.S.: Barriers to New Competition. Cambridge, MA: Harvard
University Press (1956).
[3] Bensoussan, A., Butler, A., and Nat. Ph.: Towards a dynamic the-
ory of the multiproduct firm, A generalization of the Nerlove-Arrow
optimality conditions, working paper 35-73, European Institute for
Advanced Studies in Management, Brussel (1973).
[4] Deal, K.: Optimizing advertising expenditures in a dynamic duopoly,
Operations research 22, 297-304 (1979).
[5] Erickson, G.M.: Advertising strategies in a dynamic oligopoly, Jour-
nal of Marketing Research 32, May, 233-237 (1995).
[6] Fruchter, G., and S. Kalish.: Closed-loop advertising strategies in a
duopoly, Management Science 43(1), 54-53 (1997).
[7] Jorgensen, S.: A survey of differential games in advertising, Journal
of Economic Dynamics and Control 4, 341-368 (1982).
[8] Nelson, P.: Advertising as information, Journal of Political Economy
82, 729-754 (1974).
[9] Neumann, C.P.: Dynamic Advertising Strategy, A managerial ap-
proach. Journal of Business Administration 7, 1-21 (1976).
[10] Prasad, A. and Sethi, S.P.: Advertising under uncertainty, A stochas-
tic differential game approach, Forthcoming in the journal of opti-
mization theory and applications (2003).
[11] Sasieni, M.W.: Prising and advertising for profit, Paper presented
at Pennsylvania State University, Oct (1981).
BIBLIOGRAPHY 109
[12] Sethi, S.P.: Optimal advertising for the Nerlove-Arrow model un-
der a budget constraint, Operations research quarterly 28, 683-693
(1977b).
[13] Sethi, S.P.: Deterministic and stochastic optimization of a dynamic

advertising model, Optimal Control Applications and Methods 4,
179-184 (1983).
[14] Sorger, G.: Competitive dynamic advertising. A modification of the

Case game, Journal of Economics Dynamics and Control 13, 55-80
(1989).
[15] Tapiero, C.S.: A generalization of the Nerlove-Arrow Model to multi-

Firm advertising under uncertainty, Working paper 79-6, European
Institute of advertising studies in management, Brussel (1979).
[16] Thompson, G.L. and Teng, J.T.: Optimal pricing and advertising
policies for new product oligopoly models, Marketing Science 3, 148-
168 (1984).
[17] Tsurumi, H.: Simultaneous determination of market share and ad-

vertising expenditure and dynamic conditions, The case of a firm
within the Japanese Pharmaceutical.
[18] Vidal, M.L. and Wolfe, H.B.: An Operations Research study of sales
response to advertising, Operations Research 5, 370-381 (1957).
[19] Wells, W.A. and Sunder, S. and Zoltners, A.A.: A multi-period in-
teger programming approach to the product mix problem, Educa-
tors proceeding, American Marketing Association, Chicago, 493-497
(1976).
[20] Hull, J., White, A.: Valuing Credit Default Swaps I, No Counter-
party Default Risk, Journal of Derivatives, 8, 29-40 (2000).
[21] Ito, K.: Stochastic integral, Proc. Imperial Acad. Tokoy 20, 519-524
(1944).
[22] John, H., Alan, W.: Merton’s model, Credit Risk, and Volatility
Skews (2004).
[23] Marsh, T.A., Rosenfeld, E.R.: Stochastic processes for interest rates
and equilibrium bond prices, J. Financial Economics 38, 635-650
(1975).
110 BIBLIOGRAPHY
[24] Merton, R. C.: On the Pricing of Corporate debt: The Risk Structure
of Interest Rates, Journal of Finance 2, 449-470 (1974).
[25] Sethi, S.P.: Deterministic and stochastic optimization of a dynamic
advertising model, Optimal Control Applications and Methods 4,
179-184 (2003).
[26] Vidale, M.L., Wolfe, H.B.: An Operations Research study of sales
response to advertising, Operations Research 5, 370-381, (1958).
[27] Batkhurel, .: The Optimal level of foreign reserves in Mongolia, Mon-
golbank, Research working paper-6, 99-111 (2010).
[28] Bielecki, T.R., Pliska, S.R., Sherries, M.: Risk-sensitive asset alloca-
tion. J. Econ. Dynam. Control 24, 11451177 (2000).
[29] Brennan, M.J., Schwartz, E.S., Lagnado, R.: Strategic asset alloca-
tion. J. Econ. Dynam. Control 21, 13771403 (1997).
[30] Brennan, M.J., Schwartz, E.S.: The use of Treasury bill futures in
strategic asset allocation programs. In: W.T. Ziemba and J.M. Mul-
vey (Eds.), Worldwide Asset and Liability Modelling, Chapter 10.
Cambridge University Press, Cambridge, UK (1999).
[31] Campbell, J.Y., Chacko, G., Rodriguez, J., Viceira, L.M.: Strategic
asset allocation in a continuous-time var model. J. Econ. Dynam.
Control 28, 21952214 (2003).
[32] Haugh, M.B., Lo, A.W.: Asset allocation and derivatives. Quant.
Finan. 1, 4572 (2001).
[33] Nyamsuren, D., Batsukh, Ts.: Mertons type portfolio optimization
problem in the finite-horizon case with HARA utility function and
proportional transaction costs, explicit solution, International Jour-
nal of Mathematical Archive 3(1), 78-84 (2012).
[34] Merton, R.C.: Lifetime portfolio selection under uncertainty: The
continuous case. Rev. Econ. Stat. 51, 247257 (1969).
[35] Merton, R.C: Optimum consumption and portfolio rules in a
continuous-time model. J. Econ. Theory. 3, 376-413 (1971).
[36] Munk, C., Srensen, C., Vinther, T.N.: Dynamic asset allocation
under mean reverting returns, stochastic interest rates, and inflation
uncertainty. Proceedings of the 30th Annual Meeting of the European
Finance Association, Glasgow (2003).
BIBLIOGRAPHY 111
[37] Samuelson, P.A.: Lifetime portfolio selection by dynamic stochastic

programming. Rev. Econ. Stat. 51, 239246 (1969).

Lambert 2018

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lambert 2018

Uploaded by

Copyright:

Available Formats

Optimization Applications in

Economics and Finance

R.Enkhbat, J.Enkhbayar, N.Tungalag,

1 Solow Growth Theory 1

2 Utility Maximization Problem 19

4 Cost Minimization Problem 57

7 Optimization Applications in Finance 90

1.1 Parametric Optimization Approach to the

Economic equilibria condition is:

k 0 = s(t)ϕ(k) − (µ + η)k = 0. (1.6)

Now we consider per capita consumption function

c = (1 − s)ϕ(k) → max, (1.8)

Then problem (1.8)-(1.9) is equivalent to the following one dimensional

So called the golden rule of level of accumulation is determined from (1.9)

Economic growth with nonconvex production functions

when f (K, L, t) and L(t) are arbitrary continuously differentiable given

Then a per capita consumption function is

C(t) (1 − s(t))f (K(t), L(t), t)

Also, a per capita capital function is

Economic equilibria condition is written as

K 0 (t)L(t) − L0 (t)K(t) = s(t)f (K, L, t) − µ(t)K(t) L(t) − L0 (t)K(t) = 0.

From this equation we find s(t):

Denote by KB the maximum of function K(t) defined by (1.12), i.e.,

and introduce the function φ(K, L, t) as follows

Definition 1.1. If the following condition

holds then the interval [KA , KB ] is called an economic efficient interval.

We can write function f (K, L, t) as

Proof. Since f (K, L, t) is a continuously differentiable function with

which proves the lemma.

Proof. We discretize [tA , tB ] in the following way

Define  > 0 as follows:

To reinforce the theoretical results, we solve the following problems nu-

φ(K, L) = 0.000108K 5 − 0.00596LK 4 + 0.11365K 3 −

and the labor grows at exponential rate η.

Table 1.1: Example 1

Table 1.2: Example 2

Table 1.3: Example 3

1.2 Global Optimization Approach to the Solow

The solution of (1.21) is

The problem (1.22) is equivalent to one variable maximization problem

c = ϕ(k) − (µ + η)k → max, k > 0. (1.23)

We find stationary points of the problem

It can be easily calculated that

If we substitute z = e−M γk and A = M 2 γϕ0 (M − ϕ0 ) then

This reduces to the following quadratic equation

(M − ϕ0 )2 (µ + η)z 2 + 2ϕ0 (M − ϕ0 )(µ + η) − A z + (µ + η)ϕ20 = 0.

The solution z is found as

Capital accumulation equation is

The stability condition is

Then problem (1.22)-(1.23) reduces to one variable parametric maximiza-

Now assume that the production function f (K, L) is differentiable and

The problem (1.8)-(1.9) becomes a nonconvex optimization problem and

F (k) = −ϕ(k) + (µ + η)k → min, k ∈ [0, b]. (1.29)

Problem (1.29) belongs to a class of global optimization and its solution

P (k1 , y) = F (y) − M |k1 − y|, k1 ∈ [0, b].

S1 (k1 ) → min, k1 ∈ [0, b].

Find a point k11 such that

F (k11 ) = min F (k10 ); F (y 1 ) , k11 = k10 ∨ y 1 .

Construct piecewise linear function S2 (k1 ) by

S2 (k1 ) = max S1 (k1 ); P (k1 , y 1 ) .

Let y 2 be the solution of the problem:

S2 (k1 ) → min, k1 ∈ [0, b].

Define k12 so as to satisfy

F (k12 ) = min F (k11 ), F (y 2 ) , k12 = k11 ∨ y 2 .

Define > 0 as follows: