Journal of Mathematics

and System Science

Volume 3, Number 2, February 2013 (Serial Number 12)

David Publishing

David Publishing Company

www.davidpublishing.com
Publication Information:
Journal of Mathematics and System Science is published monthly in hard copy and online (Print: ISSN 2159-5291;
Online: ISSN 2159-5305) by David Publishing Company located at 9460 Telstar Ave Suite 5, EL Monte, CA 91731,
USA.
Aims and Scope:
Journal of Mathematics and System Science, a monthly professional academic and peer-reviewed journal, particularly
emphasizes new research results in theory and methodology, in realm of pure mathematics, applied mathematics,
computational mathematics, system theory, system control, system engineering, system biology, operations research
and management, probability theory, statistics, information processing, etc.. Articles interpreting practical application
of up-to-date technology are also welcome.
Editorial Board Members:
William P. Fox (USA)
Wattanavadee Sriwattanapongse (Thailand)
Zoubir Dahmani (Algeria)
Claudio Cuevas (Brazil)
Elisa Francomano (Italy)
Baha ŞEN (Turkey)
Assia Guezane-Lakoud (Algeria)
Shelly SHEN (China)
Manuscripts and correspondence are invited for publication. You can submit your papers via web submission, or
E-mail to mathematics@davidpublishing.org, jmss.mathematics@hotmail.com, jmss.mathematics@yahoo.com.
Submission guidelines and web submission system are available at http://www.davidpublishing.org or
http://www.davidpublishing.com.
Editorial Office:
9460 Telstar Ave Suite 5, EL Monte, CA 91731, USA
Tel: 1-323-984-7526, 323-410-1082
Fax: 1-323-984-7374, 323-908-0457
E-mail: mathematics@davidpublishing.org, jmss.mathematics@hotmail.com, jmss.mathematics@yahoo.com
Copyright©2013 by David Publishing Company and individual contributors. All rights reserved. David Publishing
Company holds the exclusive copyright of all the contents of this journal. In accordance with the international
convention, no part of this journal may be reproduced or transmitted by any media or publishing organs (including
various websites) without the written permission of the copyright holder. Otherwise, any conduct would be
considered as the violation of the copyright. The contents of this journal are available for any citation. However, all
the citations should be clearly indicated with the title of this journal, serial number and the name of the author.
Abstracted / Indexed in:
Database of EBSCO, Massachusetts, USA
Index Copernicus, Poland
Ulrich’s Periodicals Directory
CSA Technology Research Database
Norwegian Social Science Data Services (NSD), Database for Statistics on Higher Education (DBH), Norway
Universe Digital Library S/B, Proquest
Summon Serials Solutions
Chinese Database of CEPS, Airiti Inc. & OCLC
Chinese Scientific Journals Database, VIP Corporation, Chongqing, P. R. China
Subscription Information:
Price (per year):
Print $420; Online $300; Print and Online $560
David Publishing Company
9460 Telstar Ave Suite 5, EL Monte, CA 91731, USA
Tel: 1-323-984-7526, 323-410-1082; Fax: 1-323-984-7374, 323-908-0457
E-mail: order@davidpublishing.com

D DAVID PUBLISHING

David Publishing Company

www.davidpublishing.com
 Journal of Mathematics
and System Science
Volume 3, Number 2, February 2013 (Serial Number 12)

Contents
Applied Mathematics

55 Interpolation in 2 with Minimum Value of the Uniform Norm of the Laplace Operator
Sergey I. Novikov

62 Model for the Vaccination Against the Poliomyelitis

Irene Duarte Gandica, Lina Marcela Ocampo, Edwin Fernando Duque Marín, Maria Mercedes
González

68 Bayesian Lasso for Screening Experiments

Hidehisa Noguchi, Yoshikazu Ojima and Seiichi Yasui

75 Relationship Between Classical and Quantum Physics

Skender Ahmetaj, Skender Kabashi, Gazmend Kabashi, Valon Veliu

83 Oscillation Theorems for Second Order Differential Equations and Their Applications
Xhevair Beqiri, Elizabeta Koci

System Science

89 Ranking Independent Variables in Linear Models

Deng-Yuan Huang and Ren-Fen Lee

96 On the Isolation Property Over a Database Domain

Bektur S. Baizhanov and Beibut Sh. Kulpeshov

101 A Parallel Finite Element Algorithm for Navier-Stokes Equations

Mohamed Abdelwahed

110 Calculation of Infinite Series Through Polygamma Functions’ Algorithms and Laplace Transforms
Héctor Luna-García, Luz María García Cruz and R. Mares
Journal of Mathematics and System Science 3 (2013) 55-61 D DAVID PUBLISHING

Interpolation in 2 with Minimum Value of the Uniform

Norm of the Laplace Operator

Sergey I. Novikov
Institute of Mathematics and Mechanics, Ural Branch of RAS, Yekaterinburg 620990, Russia

Received: September 11, 2012 / Accepted: December 11, 2013 / Published: February 25, 2013.

Abstract: Problems, which are studied in the paper, concern to theoretical aspects of interpolation theory. As is known, interpolation is
one of the methods for approximate representation or recovery of functions on the basis of their given values at points of a grid.
Interpolating functions can be chosen by many various ways. In the paper the authors are interested in interpolating functions, for which
the Laplace operator, applied to them, has a minimal norm. The authors interpolate infinite bounded sequences at the knots of the
square grid in Euclidian space. The considered problem is formulated as an extremal one. The main result of the paper is the theorem, in
which certain estimates for the uniform norm of the Laplace operator applied to smooth interpolating functions of two real variables are
established for the class of all bounded (in the corresponding discrete norm) interpolated sequences. Also connections of the considered
interpolation problem with other problems and with embeddings of the Sobolev classes into the space of continuous functions are
discussed. In the final part of the main section of the paper, the authors formulate some open problems in this area and sketch possible
approaches to the search of solutions. In order to prove the main results, the authors use methods of classical mathematical analysis and
the theory of polynomial splines of one variable with equidistant knots.

Key words: Interpolation, Laplace operator, minimal norm, B-spline.

1. Introduction L( x) is constructed in the form L( x) = ∑ αν ϕν ( x) ,

ν

where {ϕν } is a certain system of linearly

This paper is devoted to one of the problems in
interpolation theory.
independent functions, chosen in advance. For example,
As is known, interpolation is one of the methods of ν
if n = 1 and ϕ= ν ( x) x= , ν 0,1,..., N , then the
approximation. Interpolating of a function f ( x) ,
authors have the interpolation by algebraic
x = ( x1 , x2 ,..., xn ) on a domain D ⊆  n (n ≥ 1)
polynomials of degree N and if ϕ0 ( x) = 1,
by the values of f at knots { x }k∈A of a grid is the
(k )

ϕ2 k −1 ( x) sin kx=
= , ϕ2 k ( x) cos
= kx (k 1,..., m) ,
process of constructing another function
then the authors arrive at the trigonometric
L( x) ≡ L( x; f ) such that L( x ) = f ( x ( k ) ) ,
(k )
interpolation. The choice of the system of functions
k ∈ A. Function L( x) is called the interpolating
{ϕν } is determined by the properties of the function
function or the interpolant. Here A is the set of all
class for which an interpolation process is studied and
non-negative integers or a finite subset of it. Values of
(or) an interpolation algorithm is constructed. The
f can be not given at the points lying between the
interpolation by algebraic and trigonometric
knots of the grid. In that case, the authors deal with
polynomials, splines and many other systems of
interpolation of discrete data { f ( x )}
(k )
k∈A
. Usually functions of one variable was investigated in sufficient
detail.
For many applications, the interpolants that change
Corresponding author: Sergey I. Novikov, senior
researcher, research field: approximation theory. E-mail: regularly upon transition from one knot of a grid to
Sergey.Novikov@imm.uran.ru.
56
Interpolation in 
2
with Minimum Value of the Uniform Norm of the Laplace Operator
other knots are most important. This leads to problems
of choosing of interpolants on the basis of
minimization of a norm of the derivative of order
m (m ≥ 1) for fixed interpolated data. Originally,
such a minimization problem was considered by
J.Favard [1] for functions of one variable in the case of
the uniform norm. Actually, this problem as well as
some more general problems of interpolation (so-called
multiple interpolation) was studied in Ref. [1]. Later
Boor [2], Karlin [3], Tikhomirov and Bojanov [4],
Fisher and Jerome [5], and many others found solutions
of the Favard type interpolation problems for various
interpolation conditions and spaces of interpolating
functions. The number that is equal to the norm of
m -th derivative of the “minimal” interpolant is called
the value of the Favard interpolation problem.
A key role in studying of the minimization problems
with interpolation restrictions played the splines, i.e.
the smooth piecewise-polynomial functions. In 1957,
J.Holladay [6] found an interesting property of the
interpolating cubic splines. It turned out that they
provide the least L2 -norm of the second derivative
over the corresponding set of interpolating functions.
This is so-called the minimum norm property. It was
transferred on interpolating splines of an arbitrary odd
degree and L-splines [7]. The minimum norm property
served as a starting point for many subsequent
generalizations (see Refs. [8, 9] and other papers).
Often it is required to interpolate not one set of
numerical data, but some collection. On this way, the
authors arrive at problems of extremal interpolation.
The class of interpolated data whose m -th finite
differences are bounded, is chosen. The problem of
extremal interpolation is to find a minimum value for
L∞ -norm of m -th derivative of interpolants under
interpolation of the class of data by smooth functions of
one variable at the equidistant knots. This means that
the authors take the supremum of the values of the
Favard interpolation problems over the chosen class of
interpolated data. First, this problem was solved by
Subbotin in the paper [10]. Later, various
modifications of the extremal interpolation problem
were researched (see Refs. [11, 12], and many other
papers). For the history of these problems and
contributions of many authors, the authors refer to the
reader to Refs. [13, 14], and references therein.
Interpolation by functions of several variables was
studied much less. Here the problem of finding of
optimal interpolants has certain specific details. Since
on  ( n ≥ 2) there are partial derivatives of the
n
same order with respect to the different variables, the
authors have to minimize either the Sobolev semi-norm
of order m or a norm of a linear differential operator
(of order m ) applied to interpolants. Also
spline-similar functions arose here ([7, 15], and other
papers). In  ( n ≥ 2) , the extremal interpolation
n
problems were considered for operators of the mixed
derivatives ∂ / ∂x1 1 ∂x2 2 ...∂xn n , m1 + m2 + … +
m m m m
mn = m [11], and for linear differential operators that
were represented as a superposition of n
one-dimensional linear differential operators with
constant real coefficients [16]. The corresponding
difference operators were constructed by a
superposition of n appropriate one-dimensional
linear difference operators, and classes of interpolated
data were defined by these operators. Every
one-dimensional difference operator was chosen so
that it vanish the grid values of all functions from the
null space of the differential operator. Solutions of
these problems were obtained in the above quoted
papers. However, the most of the partial linear
differential operators cannot be represented as a
superposition of one-dimensional ones. One of such
differential operators is the well-known Laplace
operator ∆ =∂ / ∂x1 + ... + ∂ / ∂xn . In the case
2 2 2 2
when the Laplace operator in  n , n ≥ 2 is
considered, there is no appropriate statement of the
extremal interpolation problem, since there is no finite
difference operator that annihilates all harmonic
functions on a rectangular lattice. More information
about this and other results in this area can be found in
the paper [14].
 57
Interpolation in 
2
with Minimum Value of the Uniform Norm of the Laplace Operator

Now the authors pass to statement of the problem This value is the norm of the Laplace operator applied
considered in the present paper. to the "best" function from the class Y∞ ( z ) when the
=
Let z {z j ,ν : ( j ,ν ) ∈  2 } be an arbitrary real "worst" sequence from set M ∞ is interpolated.
The problem || ∆u ||∞ → u∈inf Y ( z ) is the Favard type
=
sequence, {
|| z ||l∞ sup | z j ,ν |: ( j ,ν ) ∈  2 } . The
∞

interpolation problem for an arbitrary fixed

authors introduce the class of interpolated sequences as interpolated sequence z ∈ M ∞ . Therefore Eq. (1) can
follows: be interpreted as the Favard type interpolation problem
=
M∞ {=
z: z {z j ,ν }( j ,ν )∈2 , || z ||l∞ ≤ 1 } for the whole class M ∞ of the interpolated sequences.
For functions of one variable, the Laplace operator is
By D ( ) the authors denote the space of infinite
2
turned into the second derivative. In this case, a sharp
differentiable compactly supported functions defined
value of the analog of (1) is a simple consequence from
on  2 . Let C (  2 ) be the space of continuous Subbotin’s results [10].
functions on  2 , and C 1 ( 2 ) be the space of For the ball of radius R (0 < R < +∞) under
continuously differentiable functions on  2 . These finite number N of the interpolation conditions, an
spaces are equipped with the uniform norms, i.e. analog of Eq. (1) was studied by the author in Refs. [17,
18] where were found orders in N and R under
|| f || X = sup | f ( x, y ) | , where X is C ( 2 ) or
( x , y )∈
2
some additional constraints on the arrangement of
C1 ( 2 ) correspondently. The authors admit that the interpolation points. For cubes, similar problems were
considered in author’s papers [19, 20].
norms can be equal to the infinity. Let L∞ ( ) be the
2
Note also that Fisher and Jerome [5] obtained,
space of measurable essentially bounded functions among many other interesting results, certain important
with norm || f ||L 2
( )
= ess sup| f ( x, y ) | . For short properties of extremal functions in minimization
∞
( x , y )∈
2

problems for the uniform norm of the Laplace operator

the authors shall write || f ||∞ in these cases.
applied to interpolating functions when the number of
As usual, ∆ denotes the Laplace operator. Now the interpolation conditions, given inside of a bounded
authors put domain, was finite.
U= {u ∈ C ( ) : ∆u ∈ L
1 2
∞ ( 2 )} .
2. Results and Discussion
here ∆u is understood in the Sobolev generalized
The aim of the paper is to prove the following result.
sense, namely v = ∆u if the equality
Theorem:
∫∫ v( x, y ) ϕ ( x,=
y ) dx dy ∫∫ u ( x, y ) ∆ϕ ( x, y ) dx dy
≤ A∞ (  2 ) ≤ 48
2 2
1
 

is true for every function ϕ ∈ D ( 2 ) . 8

In order to prove the Theorem the authors need some
The authors shall interpolate at points ( j ,ν ) ∈  2 ,
additional tools.
and let
{u ∈ U : || u || < +∞, }
One of them is inequalities of the
Y∞ ( z= u ( j , ν=
) z j ,ν ∀( j , ν ) ∈ 
2
) ∞
Landau-Kolmogorov type between the uniform norms
be a class of functions that interpolate the fixed
of a function, its partial derivatives of the first order
sequence z ∈ M ∞ .
and the Laplace operator. Such inequalities were
The main problem, which the authors study in this
obtained by V.G.Timofeev [21] in the following form
paper, is to investigate the following value:
∂f
A∞ (  2 ) sup inf ∆u ( )
1/ 2

= (1) ≤ 2 f ∞
∆f ∞
(2)
z∈M ∞ u∈Y∞ ( z ) ∞ ∂x ∞
58
Interpolation in 
2
with Minimum Value of the Uniform Norm of the Laplace Operator

∂f otherwise. Since z
j ,ν ∈ M ∞ the authors have
( )
1/ 2
≤ 2 f ∆f
A∞ (  2 ) ≥ inf  || ∆u ||∞
∞ ∞
∂y ∞ (3)
u∈Y∞ ( z )
The same inequalities for functions defined on
Let u = u ( x, y ) be an arbitrary function from the
 n , n ≥ 3 were established in Ref. [22].
class Y ( z) .
Another tool is the parabolic B-spline of one variable ∞
Now the authors decompose  2 into closed
with equidistant knots. B-splines are well known in
squares of the unit measure
mathematics. Detailed information about B-splines and
 2 =  Q j ,ν ,
their applications can be found, for example, in Refs.
where
[23-25], and references therein. The authors shall give
an explicit representation and list some properties of
Q=
j ,ν {( x, y ) ∈  2
: j ≤ x ≤ j + 1, ν ≤ y ≤ ν + 1 }
and j , ν run over all integers. It is evident that the
the parabolic B-spline with equidistant knots.
neighboring squares are intersected in sets of zero
Let a ∈  be an arbitrary fixed point, h > 0 be
plane measure. In every square Q j ,ν the authors
any number, and δ =− {a h / 2, a − h / 6, a + h / 6, a + h / 2}
represent the differentiable function u by Taylor’s
be a partition of the interval [ a − h / 2, a + h / 2] .
formula (see Ref. [26])
Define the following function: B (t ) = 0 if
=
u ( x , y ) u ( x0 , y 0 ) +
t < a − h / 2 and t ≥ a + h / 2 ,
∂u ( x0 + t ( x − x0 ), y0 + t ( y − y0 ))
1

9 h + ( x − x0 ) ∫ dt + (4)
2

B (t=
) t − a +  ∂x
2 2 0

∂u ( x0 + t ( x − x0 ), y0 + t ( y − y0 ))
1
if a − h / 2 ≤ t < a − h / 6 , + ( y − y0 ) ∫ ∂y
dt
h 2
 0
B ( t= − 3(t − a )
2
) 3  where ( x0 , y0 ) is an arbitrary fixed point from Q .
4  j ,ν

if a − h / 6 ≤ t < a + h / 6 , and Choosing x0 = j if j ≠ 0 and x0= j + 1 if j = 0 ,

9 h
2 y0 =ν if ν ≠ 0 and y0= ν + 1 if ν = 0 , the authors
B (t=
) t − a − 
2 2 have u ( x0 , y0 ) = 0 . From Eq. (4) it follows the
if a + h / 6 ≤ t < a + h / 2 . inequality
∂u ( x, y )
Function B (t ) is the parabolic B-spline with knot | u ( x, y ) | ≤ | x − x0 | max +
( x , y )∈Q ν ∂x
set δ and has the following properties:
j,

B (t ) is continuously differentiable on the whole ∂u ( x, y )

+ | y − y0 | max
real line; ∂y ( x , y )∈Q j ,ν

[ a − h / 2, a + h / 2] is the support of B (t ) ; Since | x − x0 | ≤ 1 , | y − y0 | ≤ 1 and any point

B (t ) is positive throughout the interior of its support; ( x, y ) ∈  2 belongs to a certain square Q j ,ν , the
B (t ) has the symmetry about the point a , i.e. authors obtain
B ( a − τ ) = B ( a + τ ) ∀τ ∈  ; | u ( x, y ) | ≤
B (t ) has the greatest value at the unique point ∂u ( x, y ) ∂u ( x, y )
t =a. ≤ max + max ≤
( x , y )∈Q j ,ν ∂x ( x , y )∈Q j ,ν ∂y
Proof of theorem. Firstly, the authors obtain the
∂u ∂u
lower estimate for Eq. (1). ≤ +
∂x ∂y
Choose an interpolated sequence z = z
j ,ν { }( j ,ν )∈ 2
∞ ∞

Now the authors take the least upper bound over all
so that z
j ,ν
= 1 if ( j ,ν ) = (0, 0) and z
j ,ν = 0 points ( x, y ) ∈  and get the inequality
2
 59
Interpolation in 
2
with Minimum Value of the Uniform Norm of the Laplace Operator

∂u ∂u lines, passing through these knots in parallel to the

u ∞
≤ + coordinate axes, break up every square V j ,ν into nine
∂x ∞ ∂y ∞
small squares, i.e.
In order to estimate the norms of the partial 3
derivatives in the right-hand side, the authors use V j ,ν =  I q , s
q , s =1
inequalities Eq. (2). Applying them, the authors get
u ∞ ≤ 8 ∆u ∞ . From u (0, 0) = 1 the authors have =
where I q ,s [ aq( j ) , aq( +j )1 ] × [ as(ν ) , as(ν+1) ], ( q, s = 1, 2,3 )
=1 u (0, 0) ≤ u ∞ ≤ 8 ∆u ∞ , i.e. ∆u ≥
1
8 ∞ are the small squares. In every I q , s , function
∀u ∈ Y∞ ( z) . Taking to the greatest lower bound over Fj ,ν ( x, y ) is written by a single formula in
Y∞ ( z) and returning to Eq. (3), the authors complete corresponding to the explicit representations of the
the proof of the lower estimate. parabolic B-splines B j ( x) and Bν ( y ) .
Now the authors obtain the upper estimate. From Eq.
Fix any V j ,ν and consider the centric small square
(3) the authors have
A∞ (  2 ) ≤ sup || ∆u ||∞ (5) ( j ,ν )
I 2,2 =
 j − 1 , j + 1  × ν − 1 , ν + 1 
z∈M ∞  6 6   6 6 
where u is any function from Y∞ ( z ) . Applying the Laplace operator to function Fj ,ν ( x, y ) ,
The authors shall construct an interpolating function performing simple calculations and taking the modulus,
from the class Y∞ ( z ) for any sequence z ∈ M ∞ by the authors get
using of the parabolic B-splines. The authors represent 1
96 | z ν | 3 ( x − j ) + 3 ( y − ν ) −
| ∆Fj ,ν ( x, y ) | =
2 2

 2 as the union of a countable set of the closed

j,
2
Now the authors put x1 = y −ν . It is
x − j , y1 =
 1 1  1 1
x1 , y1 ∈ [−1/ 6, 1/ 6] . Finding the
squares V j ,ν = j
 2− , j + × ν − , ν + , i.e.
2   2 
evident that
2
maximum of the function
 =2
 V j ,ν . On every square V j ,ν the authors
1
( j ,ν )∈
2
ϕ ( x1 , y1 ) = 3( x12 + y12 ) − by standard methods of
2
build the function Fj ,ν ( x, y ) = c j ,ν B j ( x ) Bν ( y ) where
c j ,ν is a constant, B j ( x) is the parabolic B-spline of the differential calculus in two variables, the authors
one variable x with knots obtain
j − 1/ 2, j − 1/ 6, j + 1/ 6, j + 1/ 2 , and Bν ( y ) is the   1 1  1
max ϕ ( x1 , y1 ) : x1 , y1 ∈  − ,   =
parabolic B-spline of one variable y with knots   6 6  2
( j ,ν )
Thus ∆F j ,ν ( x, y ) ≤ 48 z j ,ν , ( x, y ) ∈ I 2,2 . For
ν − 1 / 2, ν − 1 / 6, ν + 1 / 6, ν + 1 / 2 . The constant
all other small squares the authors perform the same
c j ,ν is determined from the interpolation condition operations. As results the authors have inequalities
Fj ,ν ( j ,ν ) = z j ,ν . As a result the authors have ∆F j ,ν ( x, y ) ≤ Cq ,s z j ,ν , ( x, y ) ∈ I q( ,js,ν )
c j ,ν = C z j ,ν where C = 16 / 9 . where the Cq , s ( q, s = 1, 2,3) are certain explicitly
Now the authors denote the knots of the B-splines in computed constants. It turns out that the constant,
1
the horizontal direction by a1( j )= j − , which has been found for the centric small square, is
2 greatest. Therefore
ak( j ) =ak( −j )1 + h, k =2,3, 4 and in the vertical
{
max ∆Fj ,ν ( x, y ) : ( x, y ) ∈ V j ,ν =
48 z j ,ν }
1 (ν )
direction by a = ν− , 1
Since z l∞
≤ 1 , the authors have the estimate
2
ak(ν ) =ak(ν−1) + h, k =2,3, 4 , where h = 1/ 3 . The {
max ∆Fj ,ν ( x, y ) : ( x, y ) ∈ V j ,ν ≤ 48 }
60
Interpolation in 
2
with Minimum Value of the Uniform Norm of the Laplace Operator

for every point ( j ,ν ) ∈  2 . Now the authors put inequalities Eq. (2) are known, and the products of the
G j ,ν ( x, y ) = Fj ,ν ( x, y ) if the point ( x, y ) belongs to parabolic B-splines can be written for any their number
in corresponding to dimension of the space.
the fixed square V j ,ν , and G j ,ν ( x, y ) = 0 otherwise.
The problem considered in this paper can be
The authors define a function F = F ( x, y ) by
generalized by means of replacement of the uniform
F ( x, y ) = ∑G j ,ν
( x, y ) . Function F lies in the norms l∞ and L∞ by the norms l p and
j ,ν ∈
Lp (1 ≤ p < ∞) respectively. In that case, the set of
class Y∞ ( z ) and
interpolated data is defined as follows
∆F = ∑ ∆G ν ( x, y ) ≤  
1/ p
 
 ∑ z j
∞ j, p
ν
j , ∈ ∞
=
M p = z : z {z j } j∈n ,  ≤ 1 , the
  j∈n  
≤ ∑
j ,ν ∈
∆G j ,ν ( x, y ) ≤ 48
class of interpolating functions as
∞

Applying Eq. (5), the authors complete the proof of Yp ( z=

) {u ∈ U : || u || < +∞, u ( j=)
p
z j ∀j ∈ 
n
} where
the upper estimate. Theorem is proved. U= {u ∈ C ( ) : ∆u ∈ L ( )} , and an analog of
1 n
p
n

The proved theorem gives only an interval, in which the value Eq. (1) is
1/ p
Eq. (1) lies. A sharp value of Eq. (1) is unknown. It is  
possible the upper estimate can be decreased by means = Ap  n ( )
sup inf  ∆u ( x) dx  .
z∈M p u∈Yp ( z )  n  ∫
p

of application any finite elements or box splines of an  

appropriate smoothness [27] instead of the products of
the parabolic B-splines. However this approach can be
realized only by using of computer aids for symbolic
computations. It is most probable that both the lower
and upper estimates in the theorem are not final, i.e.
they are not equal to the sharp value.
If the restriction || u ||∞ < +∞ is deleted from the
definition of class Y∞ ( z ) then the value Eq. (1) is
equal to zero, since there is a harmonic function which
interpolates any sequence z ∈ M ∞ (see, for example,
[14]). Therefore this restriction ensures the
non-degeneration of Eq. (1). Simple arguments enable
to understand why this occurs. It is known that every
bounded harmonic function is a constant (the Liouville
theorem). However it cannot interpolate an arbitrary
bounded sequence. Therefore, imposing restriction
|| u ||∞ < +∞ on the class of the interpolants, the authors
delete all harmonic functions, except constants, from
the class.
The proved theorem can be transferred on smooth
functions defined on  n , n ≥ 3 . The tools, which
have been used for the proof of the theorem, have the
needed analogs when the authors deal with  n , n ≥ 3 :
There is a connection between this value and the
known Sobolev theorem on embeddings of the Sobolev
classes into the space of continuous on 
n
functions
(see Refs. [26, 28], and references therein). According
to this theorem, the Sobolev class W p2  n ( ) is
embedded into the space C  ( )n
if 2 p > n , and
there is no such embedding if 2 p < n [28, Chapter 5].
( )
It turns out that Ap  n = 0 if 2 p < n . For all
other p and n , Ap  n ( ) is finite and not equal to
zero. These facts can be proved by means of the
products of the parabolic B-splines of one variable by
the analogy of the proof of the upper estimate in the
theorem. If 2 p ≥ n then sharp values of Ap  n ( )
are unknown. For p = 2 , some estimates for them can
be established by methods of the harmonic analysis.
Here only the cases of n = 2,3, 4 are interesting.
These results will be published later. In our opinion, the
problem to find sharp values of Ap (  n ) for
2 p ≥ n , is very difficult.
3. Conclusions
In the paper, the two-sided estimate is obtained in
one of the series of the problems about minimization of

Interpolation in 
2
with Minimum Value of the Uniform Norm of the Laplace Operator
the norm of the Laplace operator applied to the smooth
bounded interpolating functions, on the class of the
interpolated bounded sequences.
Acknowledgments
This research was supported by the Russian
Foundation of Basic Researches (project no.11-01-00347)
and by the Program of Integration Projects between the
Ural and Siberian Branches of RAS.
References
[1] J. Favard, Sur l’interpolation, Journal de Mathematiques
Pures et Appliquèes 19 (9) (1940) 281-306.
[2] C. de Boor, On the best interpolation, Journal of
Approximation Theory 16 (1) (1976) 28-42.
[3] S. Karlin, Interpolation properties of generalized perfect
splines and solutions of certain extremal problems,
Transactions of the American Mathe-matical Society 206
(1975) 25-66.
[4] V.M. Tikhomirov, B.D. Bojanov, Some convex problems
of approximation theory, Serdica 5 (1) (1979) 83-96. (in
Russian)
[5] S. Fisher, J. Jerome, Minimum norm extremals in function
spaces, Lecture Notes in Mathematics 479 (1975) 1-209.
[6] J.G. Holladay, A smoothest curve approximation,
Mathematical Tables and other Aids to Computation 11
(1957) 233-243.
[7] J.H. Ahlberg, E.N. Nilson, J.L. Walsh, The Theory of
Splines and Their Applications, Academic Press, New
York and London, 1967.
[8] M. Atteia, Generalisation de la definition et des proprietes
des “spline functions”, Comptes Rendus des Seances de
L’Académie des Sciences 260 (13) (1965) 3550-3553.
[9] P.-J. Laurent, Approximation et optimisation, Hermann,
Paris, 1972.
[10] Yu.N. Subbotin, On the relationship between finite
differences and corresponding derivatives, Proceedings of
the Steklov Institute of Mathematics 78 (1965) 24-42. (in
Russian)
[11] Yu.N. Subbotin, Functional interpolation in the mean with
the least n-th derivative, Proceedings of the Steklov
Institute of Mathematics 88 (1967) 30-60. (in Russian)
[12] V.T. Shevaldin, A problem of extremal interpolation,
Mathematical Notes 29 (4) (1981) 310-320.
[13] Yu.N. Subbotin, Variations on a spline theme,
Fundamental and Applied Mathematics 3 (4) (1997)
1043-1058. (in Russian)
61
[14] S.I. Novikov, Problems of extremal functional
interpolation, in: Proceedings of the International Summer
Mathematical Stechkin Workshop on Function Theory,
Aleksin Russia, 2007, pp.100-109. (in Russian)
[15] J. Duchon, Splines minimizing rotation-invariant semi
norms in Sobolev spaces, Lecture Notes in Mathematics
571 (1977) 85-100.
[16] S.I. Novikov, V.T. Shevaldin, A problem of extremal
interpolation for multivariate functions, Proceedings of the
Steklov Institute of Mathematics (Supplementary issue) 1
(2001) 150-166.
[17] S.I. Novikov, Interpolation with minimal value of the
norm of the Laplace operator in a ball, Proceedings of
Institute of Mathematics of NAS of Ukraine 5 (1) (2008)
248-262. (in Russian)
[18] S.I. Novikov, Interpolation in a ball with minimum value
of L p -norm of the Laplace operator, Proceedings of the
Steklov Institute of Mathematics (Supplementary issue)
277 (1) (2012) 152-160.
[19] S.I. Novikov, Interpolation in a square with minimum
value of the norm of the Laplace operator, in: Abstracts of
International Conference “Theory of Approximation of
functions and its Applications” dedicated to the 70th
anniversary of prof. A.I. Stepanets. Kyev, Institute of
Mathematics of NAS Ukraine, 2012, p.78. (in Russian)
[20] S.I. Novikov, Interpolation in a square with minimum
value of the uniform norm of the Laplace operator,
Proceedings of the Steklov Institute of Mathematics 18 (4)
249-257. (in Russian)
[21] G. Timofeev, Kolmogorov’s type inequality with the Laplace
operator, in: Theory of Functions and Approximations, 1983,
Saratov, Saratov State University, pp.84-92. (in Russian)
[22] V.G. Timofeev, An inequality of Landau type for
functions of several variables, Mathematical Notes 37 (5-6)
(1985) 369-377.
[23] I.J. Schoenberg, Cardinal interpolation and spline functions,
Journal of Approximation Theory 2 (1969) 167-206.
[24] C. de Boor, Splines as linear combinations of B-splines, in:
Approximation Theory II, Proceedings of International
Symposium, Austin, Texas, USA, 1976, pp.1-47.
[25] S.B. Stechkin, Yu.N. Subbotin, Splines in numerical
Mathematics, Nauka, Moscow, 1976 (in Russian).
[26] V.I. Burenkov, Sobolev spaces on domains, Teubner Texts
in Mathe-matics, Vol. 137, G.Teubner Verlag GmbH,
Stuttgart, 1998.
[27] C. de Boor, K. Höllig, S. Riemenschneider, Box Splines,
Applied Mathematical Science, Vol. 98, Springer Verlag,
New York and Berlin etc., 1993.
[28] R. Adams, Sobolev spaces, Academic Press, New York,
San Francisco and London, 1975.
Journal of Mathematics and System Science 3 (2013) 62-67 D DAVID PUBLISHING

Model for the Vaccination against the Poliomyelitis

Irene Duarte Gandica1, Lina Marcela Ocampo1, Edwin Fernando Duque Marín1, Maria Mercedes González2
1. Escuela de Investigación en Biomatemática,Universidaddel Quindío, Armenia, Quindío 630001, Colombia
2. Grupo GYMOL, Universidaddel Quindío, Armenia, Quindío 630001, Colombia

Received: January 8, 2012 / Accepted: February 11, 2013 / Published: February 25, 2013.

Abstract: The poliomyelitis is an acute infection produced by the polio virus that affects the human central nervous system. It is
transmitted by fecal-oral and respiratory contact.There are two types of vaccine, OPV (live attenuated virus) and IPV (inactivated
polio virus). Currently, there is a plan of vaccination until the age of 5 with OPV. The children vaccinated expel a virus (derived from
the vaccine) to the environment, and some of the people that have oral contact with them, get vaccinated by the herd behavior.
Nevertheless, taking into account the lately observed facts about the reversion to virulence of the oral polio vaccine during its
circulation in the environment, a change in the current vaccination schedule is being contemplated, where the oral polio vaccine can
be replaced by the inactivated vaccine. Nowadays, In Colombia the inactivated oral polio vaccine is recommended for children
presenting immune deficiency who are vaccinated with IPV. These children do not expel poliovirus to the environment. This work
presents a mathematical model that describes the dynamics of the infection in a population where the two types of vaccination are
carried out. The population is divided into two groups of age and Michaelis-Menten interactions. Different strategies of vaccination
are simulated and analyzed.

Key words: Poliomyelitis, vaccine OPV, vaccine IPV, vaccination, mathematical model.

1. Introduction The disease transmission by fecal-oral contact is

very common in developing countries with poor
The poliomyelitis is an acute infection produced by
sanitary conditions. While the oropharyngeal
the polio virus that affects the human central nervous
transmission is frequent in industrialized nations and
system. In its acute form, it causes inflammation in the
during the outbreaks. A week after the start of the
motor neurons of the spinal cord and the brainstem and
disease few viruses remain in the throat. However, they
leads to the paralysis, muscle atrophy and very often,
continue being excreted in the feces during 6 to 8
deformity. In the worst cases, it can cause permanent
weeks. The cases are probably more infectious in the
paralysis or even death by paralyzing the diaphragm.
first days, before and after the start of the symptoms.
The human being is the only reservoir and the infection
On average, the incubation period, since the moment of
spreads from person to person. Due to the large number
the exposure to the onset of the paralysis, is from 7 to
of sub-clinical infections, it is sometimes difficult to
21 days (with a minimum of 4 days and a maximum of
find the source of a case. It has not been shown the
40). Some relatively asymptomatic days preceded the
existence of long-term carriers, except in the rare
initial disease, prior to the paralysis [1, 2].
circumstance where the virus has been isolated
There are two types of vaccine, OPV (live attenuated
repeatedly and by long periods in immunodeficient
viruses) and IPV (inactivated viruses). Currently, there
people. Those cases have not been associated with
is a plan of vaccination until the age of 5 with OPV. All
polio outbreaks.
people who have not been immunized can be
susceptible to get the poliomyelitis. The
Corresponding author: Irene Duarte Gandica, asociate
professor, research field: differential partial equations, epidemiological tests show that the newborns born to
mathematical modelling. E-mail: iduarte@uniquindio.edu.co.
 Model for the Vaccination Against the Poliomyelitis
mothers with antibodies are protected in natural form
against the paralytic disease during some weeks. The
immunity is acquired after an infection by the wild
virus or by vaccination. The immunity acquired by the
natural infection (that includes mild and subclinical
infections) or by the complete series of the live
poliovirus vaccine orally administered, produces an
antibody response as well as a local response in the
intestinal cells. It is estimated that this is a life-long
immunity and that it can block the infection by
subsequent wild viruses, interrupting the chain of
transmission. The application of the IPV confers
humoral immunity. But the mucosal immunity is
relatively smaller. Consequently, the IPV does not
offer resistance to the carrying and spread of the wild
virus in the community. It is believed that there is little
or no cross-protective immunity among different types
of poliovirus.
The poliovirus is an enterovirus that is divided into
three antigenic types: 1, 2 and 3. All 3 types can cause
paralysis. Type 1 more frequently, type 3 in a smaller
measure, and type 2 rarely causes it. Most of the
epidemics are caused by the poliovirus type 1. Vaccine
related cases, containing the three types of virus, are
generally produced by type 2 or 3. The poliovirus
derived from the Sabin vaccine, which has caused
outbreaks in Egypt, Philippines, Haiti, Madagascar and
the Dominican Republic, is a virus that has mutated in
more than 1% regarding the VDPV (original Sabin
strain) and that has reverted to the neurovirulence. Two
types of vaccine-derived poliovirus are known: The
iVDPV (i of immunodeficiency), isolated in immune
deficient people, and the cVDPV (c of circulating), that
has been isolated in outbreaks and has shown to have
the same biological and epidemiological characteristics
of the wild virus. The initiative of polio world
eradication has significantly diminished the quantity of
cases in the world: From around 350,000 estimated
cases in 1988 to just 784 notified in 2003. Otherwise,
the disease keeps behaving, from an epidemiological
point of view, in the same way it used to happen in
63
highly incident times, that is, it mostly affects the
poorest non-immunized populations and preserves the
same epidemiological characteristics. On the other
hand, it is important to mention that the appearance of
outbreaks caused by vaccine-derived virus is a
phenomenon that has raised the need to reach the world
eradication.
In 1994 the ICCPE (International Certification
Commission for poliomyelitis Eradication) concluded
that the natural circulation of the wild poliovirus in the
Americas had been interrupted. One of the risks to be
faced in the poliomyelitis post-eradication stage is the
appearance of polio cases due to an oral
vaccine-derived virus [3-6]. Once the world
certification of polio eradication was achieved, the
World Health Organization recommends to interrupt
the use of the OPV, leaving the IPV as the only
alternative. Before presenting the scenarios in this
stage, all the risks representing the interruption of the
vaccination with VPO should be analyzed, since the
costs generated when having the IPV as the only option
are high and many countries can refuse to use it.
This work presents a model in differential equations
that describes the dynamics of the infection in a
population where the two types of vaccination are
carried out. It intends to contribute, through the
simulation of different hypothetical scenarios, to the
design of a plan to minimize and/or avoid the risks of
the replacement of the OPV by the IPV.
2. Mathematical Model
The model considers the following assumptions:
- A closed human population is considered (Global
death rate = birth rate).
- No migrations are taken into account.
- It is considered that an individual has acquired the
infection, only, if he develops a flaccid paralysis and it
is demonstrated that it was caused by the poliovirus.
- From a year, the circulating vaccine-derived
poliovirus can recover virulence.
- The people who have never been vaccinated and
64 Model for the Vaccination Against the Poliomyelitis

have contact with the vaccine virus (that has mutated variables is:
between a 1 and 15%), can develop infection. 1
1 1 1 1 1 1
- A proportion of the people having oral contact
1
with the Sabin virus (that has mutated less than 1%)
2
circulating in the environment, develop immunity. 2 2 2 2 1

- A proportion of the vaccinated population does

1 1 2
not develop antibodies.
- A child is considered to have been vaccinated, 1

only if he/she has received 5 doses until the age of 5

(Complete immunization schedule).
The human population is considered to be divided
into susceptible, vaccinated and those who acquire the
infection .
At the same time, the susceptible are divided into the
ones in age of vaccination (0 to 5 years of age),
and the ones over 5 years old, ; and the
vaccinated into vaccinated with OPV, , and
vaccinated with IPV, . Each of these variables
denotes an individual proportion in the time
measured in days.
The population of virus is considered to be divided
into Sabin circulating virus , and circulating
vaccine virus . These variables are measured in
Plaque Forming Units (PFU).
: Vaccination coverage rate with OPV;
: Vaccination coverage rate with IPV;
: Probability to be infected with the circulating
vaccine virus;
: Death rate of individuals with or below 5 years
old;
: Death rate of individuals under 5 years old;
: Global death rate;
: Sabin poliovirus Rate thrown to the environment
per individual;
: Rate in which the circulating Sabin virus recovers
virulence;
: Polio degradation rate (death) in the environment;
: Probability to acquire herd immunity;
: Susceptible population transition rate with or
below 5 years old, to children under 5 years old.
The system corresponding to these assumptions and
1 2
1 1 2
The parameters and are considered to be
, dependent from the virus populations, type
Michaelis-Menten, so that:
,
1 1
, where represents the vaccine virus infectiousness
rate, and the Sabin virus immunization rate. Besides
is estimated from the direct proportion of the virus
quantity contained in a plaque forming unit and the
Sabin virus quantity thrown to the environment by a
vaccinated person per day [7- 12].
2.1 Simulations
The fourth-order Runge-Kutta method [13-15] to
numerically approach the solutions of the system,
implemented in the mathematical software Matlab, was
utilized.
The behavior of the human populations (Susceptible,
Vaccinated and Infected) and of the virus (Sabin virus
and vaccine virus) was simulated for 2 years. The
values of the parameters used in all the simulations are
estimated from the direct proportion of the virus
quantity contained in a plaque forming unit, and the
Sabin virus quantity thrown to the environment by a
vaccinated person per day during 15 days [16- 18]. The
annual mortality rates in children and adults are the
ones estimated by DANE in Quindío Department
according to the projections from the Census carried
 Model for the Vaccination Against the Poliomyelitis 65

out in 2005, 24.9/6236 , 16.8/552703;

7.13/552703 ; 0.05 ; 1; 1 ,
1/365 since from a year, the vaccine virus begins to
recover virulence.
With the objective of generating different scenarios,
the vaccination coverage and the initial populations
were varied.
Figs. 1-4 show some human population
behaviors, , e , and the virus populations
and .
Figs. 1 and 2 have the same initial populations, with Fig. 1 Human populations and virus behavior when the
different vaccination coverage. It can be observed that initial populations are: . , . ,
when the vaccination coverage is diminished with OPV, . , . , , , and the
vaccination coverage is: . , . .
and it is enlarged with IPV, the quantity of infected
people keeps constant, close to 1.5% of the total
population, and the peak of the Sabin virus diminishes
from 300 to 200 PFU.
Fig. 3 has the same vaccination coverage than Figure
1. But the initial populations were varied increasing the
susceptible, below and over 5 years old, and
diminishing the quantity of vaccinated people with oral
vaccine. The infected population diminishes
considerably and the peak of the Sabin virus rises
twice.
Finally, Fig. 4 corresponds to a scenario where no Fig. 2 Human populations and virus behavior when the
oral vaccination is presented and there is coverage of initial populations are: . , . ,
0.01% with an inactivated vaccine. This with an initial . , . , , , and the
vaccination coverage is: . , . .
susceptible population of children over 5 years old of
50%, of a susceptible population of children below 5
years old of 10% and of an oral vaccinated population
of 39%.

3. Results and Discussion

3.1 Results

From the results of the simulations, it is observed

that the vaccine virus quantity presents an increase
during the first 6 months (200 days) and subsequently
tends to zero, being stabilized after around 2 years and
Fig. 3 Human populations and virus behavior when the
a half (900 days) (Fig. 1, Fig. 2 and Fig. 3).
initial populations are: . , . , . ,
The infected population remains in all of the cases, . , , , and the vaccination coverage
except in the scenario in Fig. 4, around 10% of the is: . , . .
66 Model for the Vaccination Against the Poliomyelitis

Acknowledgment
We would like to thank Departamento
Administrativo de Ciencia, Tecnología e Innovación,
COLCIENCIAS, and Universidad del Quindío.

References
[1] Restrepo A, Robledo J, Leiderman E, Restrepo M, Botero
D, Bedoya V., Enfermedades Infecciosas.
Corporaciónpara Investigaciones Biológicas, Medellín,
2003.
[2] A. Roy, M. Robert, Infectious Diseases of Humans,
Oxford University Press Inc., New York, 1991.
[3] M. Eichner, K. Dietz, Eradication of poliomyelitis: When
Fig. 4 Human populations and virus behavior when the
can one be sure that polio virus transmission has been
initial populations are: . , . , . ,
terminated?, American Journal of Epidemiology 143 (8)
. , , , and the vaccination coverage
(1996) 16-22.
is: , . .
[4] P. Fine, G. Oblapenko, R. Sutter, Polio control after
total population. Comparing the results of the certification: Major issues outstanding, Bulletin of the
World Health Organization 82 (2004) 47-52.
simulations, it is concluded that a single unit of Sabin
[5] Organización Panamericana de la Salud, Erradicación de
virus in the environment leads to infection, and given la Poliomielitis, Guía Práctica,Publicación Científica y
the strong resistance of the virus, it is necessary to be Técnica No. 607, 2005.
very attentive to the circulating virus quantity in the [6] N. Alvis, F. De la Hoz, J. Narváez, Impacto económico
de la introducción de la vacuna inactivada inyectable
environment, since this represents one of the most
contra la poliomielitis en Colombia,Revista
impacting factors on the increase of the infected Panamericana de Salud Pública 27(5) (2010) 352–359.
population. [7] R.J.T. Radboud, M.A. Pallansch, O.M. Kew,V. M.
It is also shown that the increase of the initial Cáceres, R.W. Sutter, K.M. Thompson, A dynamic model
of poliomyelitis outbreaks: Learning from the past to help
susceptible population causes an increase of the Inform the future, American Journal of Epidemiology
infected population. 162 (2005) 358-372.
[8] T.J. Bailey, The Mathematical Theory of Infectious
4. Conclusions Diseases, Hafner Press, New York, 1957.
[9] F. Brauer, C. Castillo-Chavez, Mathematical Models in
This leads to conclude that before suspending the use Population Biology and Epidemiology, Springer-Verlag,
of the oral vaccine, the percentage of the vaccinated New York, 2001.
people with OPV should be very close to 100% of the [10] K.L. Sutton, H.T. Banks, C. Castillo-Chavez, Estimation
of invasive pneumococcal disease dynamics parameters
population, since the probability of infection is
and the impact of conjugate vaccination in Australia,
considerably low. Mathematical Biosciences and Engineering 5 (1) (2008)
Finally, this work constitutes a first attempt of 175-204.
mathematical modeling of the poliomyelitis in the [11] C. Castillo-Chavez, A. Rivas, N. Hengartner, J. Hyman, G.
Chowell, Critical response to post-outbreak vaccination
post-eradication stage. The simulations have permitted against foot-and-mouth disease, Contemporary
to evaluate the human population and virus behavior Mathematics, American Mathematical Society, 2005.
corresponding to different vaccination scenarios. This [12] J.D. Murray, Mathematical Biology, Springer–Verlag,
New York, 1993.
allows to an approximation to real settings that lead to
[13] B. Carnahan, H.A. Luther, J.O. Wilkes, Applied
the evaluation of the risks involving the replacement of Numerical Methods, John Wiley and Sons, New York,
the OPV vaccine by the IPV vaccine. 1969.
 Model for the Vaccination Against the Poliomyelitis 67

[14] R.J. Burden, J.D. Faires, Numerical Analysis, Thomson
Brooks, USA, 2005.
[15] C. Johnson, Numerical Solution of Partial Differential
Equations by the Finite Element Method, Cambridge
University Press, New York, 1987.
[16] B. Cvjetanovic, B. Grab, H. Dixon, Epidemiological
models of poliomielitis and measles and their application
in the planning of immunization programmes, Bulletin of
the World Health Organization 60 (3) (1982) 405-422.
[17] W. Dowdle, C. Wolff, Post-eradication poliovirus
facility-associated community risks, Biologicals 34 (2006)
127-132.
[18] T. Hovi, L. Shulman, H. Van der Avoort, J. Deshpande,
M. Roivainen, E. Gourville, Role on environmental
poliovirus surveillance in global polio eradication and
beyond, Epidemiology and Infection 140 (1) (2011) 1-13.
Journal of Mathematics and System Science 3 (2013) 68-74 D DAVID PUBLISHING

Bayesian Lasso for Screening Experiments

Hidehisa Noguchi, Yoshikazu Ojima and Seiichi Yasui

Department of Industrial Science and Technology, Tokyo University of Science, Chiba 2788510, Japan

Received: January 22, 2012 / Accepted: February 12, 2013 / Published: February 25, 2013.

Abstract: The purpose of variable selection is to identify important predictors for response variables. Although there are many
varieties of variable selection methods, almost all of them have a problem of not accounting for the relationship between predictors.
Therefore it would well happen that the selected subset of identified predictors leads to hard-to-interpret model consisted of only
interaction terms. In design of experiments, the analysis is driven by the effect heredity principle which governs the relationship
between an interaction and its corresponding main effects. In this paper, the authors extend the variable selection method the Lasso
with effect heredity principle to its Bayesian version. In the example, the authors analyze the data obtained from typical screening
design Plackett-Burman design and compare the result from the ordinary Bayesian Lasso and proposed method.

Key words: Variable selection, Bayesian lasso, effect heredity principle.

1. Introduction than higher order main effects or interaction effects

and the effects of the same order are equally likely to
In the initial stage of experiments, factor screening
be important. Using this principle, the authors can
should be performed to eliminate unimportant factors.
focus main effects and two-factor interactions. It is
Through a screening experiment, the authors would
natural to accept this principle because higher order
like to identify main effects of important factors and
interactions are not easy justified physically. The
to obtain some insight about which factors may be
effect heredity principle governs the relationship
involved in two-factor interactions. Therefore, in this
between predictors that is an interaction and its
paper, the authors consider main and interaction
corresponding main effects. Nelder called it the
effects as candidate factors. In such situation, variable
marginality principle [2]. According to the heredity
selection plays an important part to select a subset of
principle, in order for an interaction to be active, at
active factors which exhibit the strongest effects. The
least one of its parent factors should be active. This
purpose of variable selection in regression analysis is
principle can be interpreted as two types [3]. One is
to identify the smallest and interpretable subset that
the strong heredity principle which allows an
explains the data well.
interaction to be active only if both corresponding
In the analysis of screening experiments, there are
main effects are active, for example, if the interaction
three principles of effect sparsity, effect hierarchy and
AB is active, both main effects A and B must be active.
effect heredity to select good subset [1]. The effect
The other is weak heredity principle which allows an
sparsity principle states that the number of active
interaction to be active if one or more of its parent’s
effects in factorial experiments is relatively small.
factors are active, for example, if the interaction AB is
This principle focuses on the vital few and not the
active, only A or B, or both of them must be active.
trivial many. The effect hierarchy principle indicates
Weak heredity is useful because of its flexibility
that lower order effects are more likely to be active
however, the model following only the weak heredity
Corresponding author: Hidehisa Noguchi, Ph.D., research
can’t be invariant to linear transformations of
field: statistical quality control. E-mail: ngc1120@gmail.com.
 Bayesian Lasso for Screening Experiments
predictors.
Although there are many variable selection methods,
almost all of them do not account for the effect
heredity principle. Because of that it might well
happen the selected subset of active factors leads to
hard-to-interpret model consisted of only interactions.
There are some methods tackling this problem.
Hamada and Wu extended stepwise regression to
consider effect heredity principle [4]. Chipman
proposed the Bayesian variable selection following the
principle [3] and analyzed the experiments with
complex aliasing patterns [5]. Yuan et al. extended the
least angle regression [6] to follow the principle [7].
Choi et al. and Noguchi et al. addressed the problem
applying the Lasso [8] with effect heredity principle
[9-10].
When the authors take into account not only main
effects but interactions as candidate factors in the
analysis of designed experiment, the number of factors
can be larger than observations. In this paper, the
authors have use of the variable selection method the
Bayesian Lasso [11] which is Bayesian version of the
Lasso because of their merit of being able to apply
even when the number of variables exceeds the
number of observations. And the authors extend the
Bayesian Lasso to follow the effect heredity principle.
Extension to Bayesian approach let us obtain credible
intervals of estimates.
The rest of paper is organized as follows. In Section
2, the authors explain the brief overview of the Lasso
and its Bayesian version. In Section 3, the authors
introduce our models and shows parameter structure
in the hierarchical model. The authors demonstrate the
proposed method through two examples in Section 4,
and conclude with a discussion in Section 5.
2. Lasso and Bayesian Lasso
Usually, experimental data can be modeled in the
form of the following general linear model
Y  Xβ  ε (1)
where the authors have n observations on a response Y
69
and X is an n×f matrix of predictors and  is a vector
of n error variables = (n) each of which is
assumed independent and identically distributed
normal errors with zero mean and unknown variance
2 and = (1, …, f)T. Throughout this paper, the
authors assume Y is centered so that the observed
mean is 0. Since the authors focus on regression
model with main effects and two-factor interactions,
p p
f ( x )    i xi    ij xi x j
(2)
i 1 i j j 2
where p is the number of main effects.
The Lasso is proposed for variable selection and
parameter estimation and the estimates can be
interpreted as constrained version of ordinary least
squares. It fits a linear model by solving the
optimization problem
p p
 |  |   | 
min(Y  Xβ ) (Y  Xβ )   (
T
i ij
|) (3)

i 1 i j j 2
where > 0 is the tuning parameter controlling the
amount of shrinkage for the fitted coefficients.
Sufficiently large value of  produces shrunken
estimates of regression coefficients to be exactly zero.
It may improve prediction accuracy and interpretation.
In screening experiments, this characteristic
corresponds to the effect sparsity principle.
Tibshirani suggests that estimates of the Lasso can
be interpreted as a Bayesian posterior mode using
normal likelihood and an independent Laplace prior to
each regression coefficient [8]. The first explicit
application of the Bayesian Lasso was proposed by
Park and Casella [11].
3. Models
3.1 Strong and Weak Heredity Model
In this paper, the authors extend the Bayesian Lasso
for selecting subset following effect heredity principle.
Choi et al. [9] considered the strong heredity model
with following reparameterization of the coefficient
for the interaction terms xixj.
p p
f s ( x)    i xi    ij  i  j xi x j (4)
i 1 i j j 2
70 Bayesian Lasso for Screening Experiments

Instead of a single parameter ij, the coefficient for

f

 (β)   exp  |  h |
the interaction term xixj is expressed as the product of h 1 2 (7)
ij, i and j. If one of i and j is equal to 0,
f  1   2  2  2 h2 
  exp  h2   exp  d
interaction term for xixj is removed. This expression h 1
0
2 h2  2 h  2  2 
Eq. (4) let the model follow the strong heredity This motivates us to represent the following
principle. Whereas, Noguchi et al. [10] proposed the hierarchical model Eq. (8).
weak heredity model expressed in Eq. (5). Y | X, β heredity ,  2 ~ N ( Xβ heredity ,  2 I)
β heredity |  2 , 12 ,  ,  2f ~ N (0,  2 D τ )
p p
f s ( x)    i xi    ij (  i2   j2 ) xi x j (5)
(8)
2
 
i 1 i j j 2 f

With above expression, the authors can consider the  ,,  ~ 

weak heredity principle as a model. This situation
where the interaction term for xixj is removed only
when both i and j are equal to 0 that represents the
weak heredity principle.
The Lasso estimators for Eqs. (4) and (5) denoted
by Eq. (6) instead of Eq. (3).
min(Y  Xβ heredity )T (Y  Xβ heredity )
 gamma prior Gamma (a, b), a > 0, b > 0, to the
(6)
p p
 main  |  i |  inter  |  ij |
i 1 i j j 2
here, heredity is the reparameterized regression
coefficient vector. There are two tuning parameters,
main and inter. The first tuning parameter main
controls the estimates for main effects. The second
tuning parameter inter controls the estimates for
interactions. To estimate the parameters, the authors
need to use an iterative approach similar to shooting
algorithm and the result of variable selection might be
different depending on initial value of the algorithm.
Compared to the Lasso with the heredity principle,
it is more straightforward to estimate parameters with
the Bayesian Lasso. Only one tuning parameter is
required to control main and interaction effects and
the authors don’t need to use the iterative algorithm to
estimate parameters.
3.2 Hierarchical Model
In the Bayesian Lasso, each element of regression
coefficient parameter  has independent and identical
Laplace prior which can be represented as a scale
mixture of normals with an exponential mixing
density in Eq. (6) [12].
2
1
2
f exp   
2 2
h 2 d 2
h
h 1 2
 2 ~  ( 2 )d 2
where D= diag(12,…,f2). Park and Casella
suggested to use the noninformative scale-invariant
marginal prior (2)=1/2 [11]. Instead of , the
authors regard 2/2 as the parameter of interest.
Following Ref. [11], the authors assign a conjugate
parameter 2/2. To consider the prior for 2/2 as a
noninformative, the authors set a and b as small value
(e.g., a = 0.1 and b = 0.1).
The joint posterior distribution of the parameters
given data can be simulated using the MCMC
algorithm Gibbs sampler which proceeds to draw from
each fully conditional posterior distribution given the
current values of all other unknown and the observed
data. Updating each h from the normal distribution
(h|Y, h, 2, 2), where h represents all elements
of  without h, updating each h2 from the inverse
Gaussian, and updating 2/2 from the gamma
distribution (2|2). In the example, the updating is
performed using the Bayesian software WinBUGS14
[13].
3.3 Further Discussion
To incorporate the effect heredity principle in the
linear model Eq. (1), the authors use the
reparameterized regression coefficient heredity. In
doing so, the authors can identify active factors
considering with the relationship between an
interaction and its corresponding main effects.
However it may lead us to ignoring the models which
 Bayesian Lasso for Screening Experiments 71

do not obey the heredity principle. In Bayesian Table 1 Plackett-Burman 12-run design with response data.
Design Response Y
variable selection, the heredity principle is adopted by No. A B C D E F G H I J K M1 M2
setting conditional probability for an interaction given 1            1.058 6.058
2            1.004 4.733
its corresponding main effects [3]. For example, if the 3            -5.200 4.625
4            5.320 5.899
authors consider two factors A and B and their 5            1.022 7.000
6            -2.471 5.752
interaction AB, the conditional probability for the 7            2.809 5.682
8            -1.272 6.607
interaction being active is as follows 9            -0.955 5.818
 p 0 if ( A ,  B )  (0, 0) 10            0.644 5.917
 11            -5.025 5.863
Pr( AB  1)   p1 if one of  A ,  B  1 12            3.060 4.809
 p if ( A ,  B )  (1, 1)
 2 Bayesian Lasso

where  is binary indicator variable which represent

1.5
whether the effect is active or not. That is, A can take

1.0
0.5
two values 0 or 1. When A = 0, main effect A is

0.0
inactive and A = 1 means A has active effect. By

-0.5
introducing the above indicator parameter  into the
-1.0

hierarchical model Eq. (8), it would be possible to A D F H J AC AF AI BC BF BI CD CH CK DH DK EH EK FI GH GK HK JK

consider the model selected from every possible Strong Heredity Model

combination of factors.
2
1

4. Examples
0

In this section, the authors apply the models Eq. (4)

-1

and Eq. (5) to the Bayesian Lasso and analyze the data
-2

obtained from 12-run Plackett-Burman design. A D F H J AC AF AI BC BF BI CD CH CK DH DK EH EK FI GH GK HK JK

Weak Heredity Model

4.1 Simulated Screening Experiment
3

Hamada and Wu [4] constructed the example of a

2

simulated screening experiment using the 12-run

1

Plackett-Burman. Design which can accommodate up

0

to 11 factors labeled A-K and response data M1 are

-1

shown in Table 1. The data are constructed based on A D F H J AC AF AI BC BF BI CD CH CK DH DK EH EK FI GH GK HK JK

the true model Y = A + 2AB + 2AC +  where ~ N(0,
=0.25), that is, factor A has an active effect and
interactions AB and AC are active while the remaining
factors D-K are inactive. The authors consider a total
of 66 candidate effects consisted of 11 main effects
and 55 two-factor interactions. Through this example,
the authors confirm our proposed model can identify
the active effects and compare the results of the usual
Bayesian Lasso.
Fig. 1 shows boxplots of 50,000 samples from each
66 factor’s posterior distribution obtained by the usual
Lasso and proposed methods. To identify the factors
Fig. 1 Boxplots for samples from posterior distributions in
the simulated screening experiment M1.
which have strong effect, posterior median estimates
are compared. Note that the median estimates won’t
be exactly zero therefore the authors need to specify
threshold value to identify effects. Add to using
turning parameter, specifying the amount of a
threshold value also can be interpreted as effect
sparsity principle.
If the authors ignore the factors whose absolute
value of median estimate under 0.25, identified factors
are only interactions AB and AC. Active main effect A
72 Bayesian Lasso for Screening Experiments

can’t be identified with the ordinary application of the Table 2 Factors and levels for cast fatigue experiment.
Bayesian Lasso and this is the same result of the Level
Lasso and least angle regression. The strong heredity Factor － ＋
A ：Initial structure As received  treat
model identify the subset including main effects A, B,
B ：Bead size Small Large
C, E, H, I, K and the interaction IK. The selected C ：Presure treat None HIP
subset does not conform to the true model but follows D ：Heat treat Anneal Solution treat/age
strong heredity principle. When the authors apply E ：Cooling rate Slow Rapid
F ：Polish Chemical Mechanical
weak heredity model, active main effect A and
G ：Final treat None Peen
interactions AB and AC can be identified. The authors
Bayesian Lasso

note that extended stepwise selection proposed by

0.6
Hamada and Wu could not identify any of the

0.4
0.2
important effects [4]. It can be said our proposed

0.0
method have the advantages.

-0.2
-0.4
4.2 Cast Fatigue Experiment
-0.6 A B C D E F G AC AE AG BD BF CD CF DE DG EG

Hunter et al. [14] used the 12-run design to study Strong Heredity Model

the effects of seven factors on the fatigue life of

weld-repaired castings. The factors and levels are
0.5

listed in Table 2. The seven factors A-G were assigned

0.0

using the first seven columns of the 12-run design

shown in Table 1 and the response M2 is the logged
-0.5

lifetime of the casting. This example has analyzed by A B C D E F G AC AE AG BD BF CD CF DE DG EG

Hamada and Wu [4] and the main effect F and Weak Heredity Model

interaction FG are identified. Using the model

1

consisted of these factors, one can obtain predicted

0

fatigue lifetimes by Eq. (8).

-1

Yˆ  5.7  0.458F  0.459FG (8)

-2

To increase the predicted life, setting of factor F

-3

and G should be “” and “” respectively which A B C D E F G AC AE AG BD BF CD CF DE DG EG

means changing chemical to mechanical on polish and
no peening probably is able to make fatigue lifetimes
longer. Therefore, the goal of this experiment would
be to identify the subset (F, FG) which has adjR2 of
0.871. Note that the procedure of the stepwise type
variable selection method ignores stochastic errors
inherited in the stages of variable selection. It makes
their theoretical properties rather hard to understand.
The authors applied our proposed models and the
ordinary Bayesian Lasso to the data M2 in Table 1. Fig.
2 shows the results. From the result of the ordinary
application of the Bayesian Lasso, main effect F and
Fig. 2 Boxplots for samples from posterior distributions in
the simulated screening experiment M2
interactions AE and FG are identified by applying
the threshold value 0.1. This result is same as the
Lasso. The interaction AE is identified but its
corresponding main effects A and E are not. And the
subset consisted of these factors has adjR2 of 0.553.
When the authors apply the strong heredity model and
set the threshold value 0.3 for absolute median
estimates, only one main effect F is identified and the
subset has adjR2 of 0.395. On the other hand, the
authors can identify the important factors F and FG
 Bayesian Lasso for Screening Experiments 73

Bayesian Lasso
with the weak heredity model and the combination of

3
the factors is following the weak heredity principle.

2
Here, the authors can see the variance of samples from

1
posterior distributions for inactive main effects is

0
smaller than inactive interactions when proposed

-1
models are applied. This might be interpreted that the

-2
A C E G I K AD AG AJ BD BG BJ CE CH CK DG DJ EG EJ FH FK GJ HJ IK
model can consider effect hierarchy principle.
Strong Heredity Model
However, in Fig. 1, there are not so large difference of

5
variance of the samples between main and interaction

4
3
effects. Therefore, it might depend on the number and

2
size of effects.

1
0
5. Discussion

-1
-2
A C E G I K AD AG AJ BD BG BJ CE CH CK DG DJ EG EJ FH FK GJ HJ IK

The authors have extended the Lasso with the effect

Weak Heredity Model

heredity principle to its Bayesian version that can 2

simultaneously fit and estimate main and interaction

1

effects. The effect sparsity principle is adopted by

0

tuning parameter of the Bayesian Lasso and the effect

-1

hierarchy principle is adopted by our models when

-2

there is relatively small number of considered factors.

A C E G I K AD AG AJ BD BG BJ CE CH CK DG DJ EG EJ FH FK GJ HJ IK

In the examples, the Bayesian Lasso identified same Fig. 3 Boxplots for samples from posterior distributions in
factors as the Lasso and the Bayesian Lasso with the the example from Hamada and Wu [4].

effect heredity principle could select the subset
following the principle. The weak heredity model
works well where there are a few active interactions
that are partially aliased with the main effects. Here,
the authors illustrate some limitations with the brief
example. The authors used the proposed methods on
Hamada and Wu’s [4] constructed example, the same
design as Table 1 with true model Y = 2A 4C 2BC
2CD , where ~ N(0, =0.5). Fig. 3 shows the
result. When the authors set threshold value 0.25 for
absolute value of posterior samples, the Bayesian
Lasso selects the subset (C, AK, AF, DG, BH) and its
adjR2 is 0.870. The strong heredity model selects the
subset (C, G, D, J, H) which has adjR2 of 0.913. The
weak heredity model also can't identify active factors
even true model is following the weak heredity
principle. The selected subset is (C, G, BH, DG, GH)
with adjR2 of 0.872. The interaction BH is contained
without its main effects B and H. This illustrates a
limitation of the design that may require additional
experimentation to resolve. The authors have two
choices of the model, strong and weak heredity, based
on the Bayesian Lasso. However, in practice, the
authors don't know which version to use. Thus, the
authors should try both of them and consider the result
with physical background. Limitation for variable
selection with the methods might depend on the effect
size and number of considered factors. Choosing
adequate design might be helpful to identify active
factors in screening experiment.
Acknowledgments
This work was supported by grants from the Tokyo
University of Science of Japan.
References
[1] C.F.J. Wu, M. Hamada, Experiment: Planning, Analysis,
and Parameter Design Optimization, 2nd ed., New York,
Wiley, 2000, pp. 172-173.
74 Bayesian Lasso for Screening Experiments

[2] J.A. Nelder, The selection of terms in response-surface 267-288.

models: How strong is the weak-heredity principle?, The
American Statistician 52 (1998) 315-318.
[3] H. Chipman, Bayesian variable selection with related
predictors, Canadian Journal of Statistics 24 (1996)
17-36.
[4] M. Hamada, C.F.J. Wu, Analysis of designed
experiments with complex aliasing, Journal of Quality
Technology 24 (1992) 130-137.
[5] H. Chipman, M. Hamada, C.F.J. Wu, A bayesian variable
selection approach for analyzing designed experiments
with complex aliasing, Technometrics 39 (1997) 372-381.
[6] B. Efron, T. Hastie, I. Johnstone, R. Tibshirani, Least
angle regression (with discussion), The Annals of
Statistics 32 (2004) 407-499.
[7] M. Yuan, V.R. Joseph, Y. Lin, An efficient variable
selection approach for analyzing designed experiments,
Technometrics 49 (2007) 430-439.
[8] R. Tibshirani, Regression shrinkage and selection via the
lasso, Journal of the Royal Statistical Society B 57 (1996)
[9] N.H. Choi, W. Li, J. Zhu, Variable selection with the
strong heredity constraint and its oracle property, Journal
of the American Statistical Association 105 (2010)
354-364.
[10] H. Noguchi, Y. Ojima, S. Yasui, A practical variable
selection for linear models, Frontiers in Statistical Quality
Control 10 (2012) 349-360.
[11] T. Park, G. Casella, The bayesian lasso, Journal of the
American Statistical Association 103 (2008) 681-686.
[12] D.F. Andrews, C.L. Mallows, Scale mixtures of normal
distributions, Journal of the Royal Statistical Society B 36
(1974) 99-102.
[13] D. Spiegelhalter, A. Thomas, N. Best, D. Lunn, Bugs:
Bayesian inference using gibbs sampling, version 1.4.
MRC, Biostatistics unit, Cambridge, UK,
http://www.mrc-bsu.cam.ac.uk/bugs/.
[14] G.B. Hunter, F.S. Hodi, T.W. Eager, High-cycle fatigue
of weld repaired cast Ti-6A1-4V, Metallurgical
Transactions 13A (1982) 1589-1594.
Journal of Mathematics and System Science 3 (2013) 75-82 D DAVID PUBLISHING

Relationship Between Classical and Quantum Physics

Skender Ahmetaj, Skender Kabashi, Gazmend Kabashi, Valon Veliu

Department of Physics, Faculty of Sciences, University of Prishtina, Theresa nr 4, Prishtina 10000 Kosovo.

Received: January 22, 2012 / Accepted: February 12, 2013 / Published: February 25, 2013.

Abstract: Schrödinger’s equation is one the equations that mark the beginnings of the systematic quantum physics. It was shown that
it follows from the Dirac’s equation and the relationship with classical physics, i.e. with classical field theory was established. The
subject of this work is the relationship between classical relativistic physics and the quantum physics. Investigation carried out in this
work, shows that the free electromagnetic field, spinor Dirac’s field without mass, spinor Dirac’s field with mass, and some other
fields are described by the same vibrational formulation. The conditions that a field be described by Maxwell’s equations of motion
are given in this work, and some solutions of these conditions are also given. Non-relativistic approximation of the equations of the
non-quantified field are the Schrödinger’s equations. Dirac’s equation as a special case, contains Maxwell’s equations and the
Schrödinger’s equation.

Key words: Relationship, quantum physics, classical physics.

1. Introduction variational description of the law of motion of the

physical objects is given, Lagrange’s equations are
Schrödinger’s equation is one the equations that
second order differential equations, whereas canonical
marks the beginnings of the systematic quantum
equations are first order differential equations. Dirac’s
physics. It was shown that it follows from the Dirac’s
equation is a first order differential equation. It is in
equation and the relationship with classical physics,
conformity with Maxwell’s equations of
i.e. with classical field theory was established. The
electrodynamics, which are first order differential
aim of this work is to show the relationship between
equations, as well. This question is not only a formal
the classical relativistic physics and the quantum
mathematical problem. It is profoundly physical
physics. For this purpose, in the second part of this
because it contains the well-founded language of
work, the authors have obtained Maxwell’s equations
physics. In this language generalized momenta, energy
from the Dirac’s equation, whereas in the third part we
etc., are defined. Therefore, if Dirac’s equation
brought the Maxwell’s equations into the form of the
represents and describes a physical object, then its
Dirac’s equation for leptons. In the fourth part of this
character is the character of canonical equations of that
work, it is shown that Schrödinger’s equation for the
object. According to this fact and taking into account
free particle is obtained from the Dirac’s equation. In
the analogy with electromagnetic field, “potentials” of
this part, the case when the particle is affected by the
the Dirac’s field as Lagrange’s variables of that field
external electromagnetic field is analyzed as well
are introduced, despite the conventional approach
2. Relationship between Electromagnetic where the field itself plays the role of the Lagrange’s
and Dirac’s Field variables. A correct and consistent formulation of the
Dirac’s field is built on this basis [1].
According to the classical mechanics where the
Lagrange’s density of the linear theory is:
Corresponding author: Skender Ahmetaj, professor,   K [(i    )(i    )] (1)
research field: Field Theory, modeling. E-mail:
skenderahmetaj@ yahoo.com.
76 Relationship Between Classical and Quantum Physics

where K is a constant and     0 . or

Lagrange’s equations, given by Ex  Ex , Bx  Bx
  
 0 (2) E y  E y ch  Bz sh , By  By ch  Ez sh (8)
 x   
   Ez  Ez ch  By sh , Bz  Bz ch  E y sh
 x 
are then G  G, F   F
     0,       0 (3) On the other hand, if they describe spinor field, then
and the corresponding canonical equations are during the transition from one to another Lorentz’s

  i    , 
i     0 (4) system, they must be transformed according to the law
   
Second of the canonical equations is the Dirac’s Ax  Ax ch   sh , bx  bx ch  fsh
equation. It is shown, then, that the canonical equation
2 2 2 2
 
of the (free) Dirac’s field is not the Lagrange’s Ay  Ay ch  bz sh
equation, but the Dirac’s equation. Lagrange’s
2 2
 
equation Eq. (3) is mathematically equivalent to the by  by ch  Az sh (9)
Lagrange’s equations for potentials of the free
2 2
   
electromagnetic field Az  Az ch  by sh , bz  bz ch  Ay sh
2 2 2 2
    A  0 (5)
   
or     ch  Ax sh , f   fch  bx sh
 2 2 2 2
 1 2 A 1  2  
A  2 2  0 ,   2 2  0 (6)
Ex'  Ex ch  Gsh ,
c t c t 2 2
Correct canonical formalism for the Maxwell’s field
 
must give all four Maxwell’s equations that contain Bx  Bx ch  Fsh
2 2
first space and time derivatives of the components
   
E , B of the field, whereas the correct Lagrange’s E y  E y ch  Bz sh
2 2
formalism must lead to the partial differential Eq. (5)
 
without introducing additional conditions, such as By  By ch  Ez sh (10)
2 2
Lorentz’s condition. Equivalence of the dynamical    
equations of these two fields, Dirac’s field and the Ez  Ez ch  By sh Bz  Bz ch  E y sh
2 2 2 2
two-vector or tensor-scalar field (that contains    
standard electromagnetic field) can be shown also by G  Gch  Bx sh F   Fch  Ex sh
2 2 2 2
observation of the transformation properties of the Relativistic invariant Lagrangean Eq. (1), through
Lagrangean Eq. (1). the variational principle, leads to the correct equations
Components Ax , Ay , Az ,  , C x , C y , C z , f of the of the electromagnetic field. Two four-quantities
potential , which are Lagrange’s variables, if they  A , C   appear.If they are four-vectors (one of
describe tensor-scalar field, must be transformed them pseudo-four-vector), then F  is an
according to the law [2] antisymmetric tensor of the second range, and G and F
Ax  Ax ch   sh , bx  bx ch  fsh are scalars (in relation to the Lorentz’s transforms
    ch  Ax sh , f   fch  bx sh (7)  ( F   g  G )  0
Ay  Ay , by  by  ( F   g  F )  0
v 1
Az  Az , bz  bz , th  F    A    A      C   C 
c 2
 Relationship Between Classical and Quantum Physics 77

1
F    C     C       A   A 
After separating the real and imaginary parts of
2 each four components of the Dirac’s field function,
G   A , F   C 

(11) the authors have
Taking F = G = 0, these equations become: 
 F   g  G   2 A  0 
 F 
 0,  F 
0 (12)   
 F   g  F   2C   0 (18)
1
i.e., equations of the free Maxwell’s field. The authors
see that equations of the free Maxwell’s field are, in
F    A    A     C    C 
2
 
1
F    C     C       A   A 
fact, a special case of the canonical Eq. (11). If the

authors chose C  0, then we obtain equations of 2
the standard electromagnetic theory of the field, which G    A , F   C 


are the equations of the conventional theory of the free For F  G  0,   0 The authors obtain
electromagnetic field. Therefore, it follows that the canonical equation for the real Proca’s field,
same dynamic equations describe free Dirac’s field if A , C  are four-vectors,
  
and the electromagnetic field [2]. In the classical rotB   0 E   2 A
theory of the electromagnetic field, Lagrange’s 
divB   2 f (19)
variables are the potentials A of the electromagnetic   
field, whereas the tensor of the electromagnetic field rotE   0 B   2C

forms the four-vector  and the potential A . The divE   2
authors now discuss the field with the mass term. In For the spinor field, introducing linear combination
that case the density of the Lagrangean is of spinor  and 
  K   ,   i    
   2   (13) I 
1 1
  i  ,  II    i  (20)
where  is a scalar constant. In order that this 2 2
1 1
 must be a
Lagrangean be a relativistic scalar,  or    I  II  ,    I  II 
scalar as well. Using the Euler-Lagrange’s equations,  2 i 2
the authors obtain Lagrange’s equations of motion Eq. (19) become,

  
    2   0,     2    0 (14)
    i 



   I  0,  i 



   II  0 (21)
According to the definition of the conjugated These are Dirac equations with eight -
momentum, the authors have: component functions and with positive and negative
   iK  ,     iK (15) mass terms. According to this, the authors conclude
Now, the density of the Hamiltonian, is that we come to the Dirac’s field described by the
H     0    0       Lagrangean Eq. (13) and both Eq. (21), if we remain
in the scope of the classical relativistic theory [2].
1
   0   i  i    i i   K 2
 (16)
K 3. Conditions of the Maxwell’s Field
Canonical equations in covariant notation follow:
In the contemporary physics of the electromagnetic
      K 2  0
fields, Maxwell’s operator plays a very important role.
1
         0 (17) The question arises, which physical objects are
K formulated by the same or analogous equations of
1  0
  0     physics. We shall discuss this question in this section.
K Maxwell’s system of equations of the free field, we
     0   K  2 
 0
can write in the matrix form [3]
78 Relationship Between Classical and Quantum Physics

D  0 (22) only field that satisfies the Maxwell’s Eqs. (22) and
  D (23) (23). The authors are interested in whether any other
where field exists that can be described by these equations.
D    (24) For this purpose, the authors will seek for the
is called Maxwell’s operator, and  
are matrices conditions of the relativistic invariance of the field
that satisfy conditions of the antisymmetry equations. The authors start from the Lorentz’s

 

,   2g  (25)
transform given by Ref. [3]

ch -sh 0 0
1 0 0 ai
  0
, i 
x  a x , a   -sh ch
(26) 0 0 (30)
0 1 88 ai  0 
0 0 1 0
0 0 0 1 0 0 1 0
0 0 0 1
0 0 1 0 0 0 0 1
a 1
, a 2
,
0 1 0 0 1 0 0 0 Corresponding field transform, the authors will
1 0 0 0 0 1 0 0 write in the form:
0 1 0 0   R ,   S ,    1, R  1, S  1 (31)
1 0 0 0 (27) where R and S are operators of the transform for
a3 
0 0 0 1  and  , respectively.
0 0 1 0 Requirement for the relativistic invariance of the
Where  and Φ are given in components: Lagrangean,
 Ex   x   K  R  0 R  K 
  K     (32)
E  -A y gives R  R  
 0 0
 y (33)
 Ez  -A z whereas relativistic invariance of the Eq. (23) gives
 
F f R 1a  S   
,    
 0
   ,   (28) (34)
 Bx  Cx
  Therefore, the conditions of the relativistic
 By  Cy invariance of the “Maxwellian” Eqs. (22) and (23) are
B  Cz given by Eqs. (33) and (3). One of the solutions of the
 z
 G  - set of Eqs. (33) and (34) is
and F and G are scalars. We operate with the operator a b c d
D on the field functions  , and obtain
R , S  (35)
  b a 88 d c 88
 0 E  rotB  G  0 where
 
 0 B  rotE  F  0 (29) ch      0 0 0

divE   0G 0 ch 0 0
 a
divB   0 F 0 0 ch 0
Chosing scalars F, G such that F = G = 0, the 0 0 0 ch     
authors obtain Maxwell’s equations for the free field. 0 0 0 sh     
Thus, the authors have shown that we can describe
0 0 sh 0
electromagnetic field by Eq. (22) and Eq. (23). These b
0  sh 0 0
equations are called “Maxwell’s equations”. We
make the question: Is the electromagnetic field the sh      0 0 0
 Relationship Between Classical and Quantum Physics 79

ch 0 0 0 1 0 0 0 0 0 0 0
0 ch      0 0 0 ch 0 0 0 0 sh 0
c
0 0 ch      0 0 0 ch 0 0  sh 0 0
0 0 0 ch 0 0 0 1 0 0 0 0
R
0 0 0 sh 0 0 0 0 1 0 0 0
0 0 sh      0 0 0  sh 0 0 ch 0 0
d (36)
0  sh      0 0 0 sh 0 0 0 0 ch 0
sh 0 0 0 0 0 0 0 0 0 0 1
we obtain
ε is a parameter confined by the condition:
R→1, S→1 for Ф→0 i.e.       ,   0   0 Ex  Ex , Bx  Bx
E y  E y ch  Bzsh, By  By ch  Ezsh
4. Some of the Solutions and Their
Ez  Ezch  By sh, Bz  Bzch  E y sh
Interpretation
F  F , G  G ' (38)
The authors will seek for the list of solutions of the In this case  is consisted from the antisymmetric
Eqs. (33) and (34) for the values of the parameters tensor F  and the scalars G and F . That means
1 1 that for the choice   1,   , we have two-vector
  ,   1, , 0,  , 1.
fiel  A , C  , or tensor-scalar field  F , G , F  .
  
2 2
If we chose λ = 1,   , operating with the For F  G  0, C   0 this field is reduced into the
operator S on  classic standard electromagnetic field.

ch 0 sh 1 
0 0 0 0 0 If the authors chose   ,  , than R  S .
2 2
0 0 0 0 0 0 0 0
Therefore,  and  are of the same magnitude
0 0 0 0 0 0 0 0 in relation to the Lorentz’s transform. Operating with
0 0 0 ch sh 0 0 0 operator R on   , the authors obtain:
S
0 0 0 sh ch 0 0 0  
Ex  Ex ch  Gsh
0 0 0 0 0 0 0 0 2 2
0 0 0 0 0 0 0 0  
Ez  Ezch  By sh
sh 0 0 0 0 0 0 ch 2 2
 
G  Gch  Exsh
the authors obtain 2 2
Ax  Ax ch   sh  
Bz  By ch  Ezsh (39)
C x  C x ch  f sh 2 2
 
E y  E y ch  Bzsh
Ay  Ay , C y  C y (37) 2 2
 
Az  Az , C z  Cz F  F ch  Bxsh
2 2
   ch  Axsh, f  f ch  C xsh  
Bz  Bzch  E y sh
In this case  is consisted by four-vectors 2 2
 
( , A) and ( f , C ).  
Operating with the operator R on  ' ,
Bx  Bx ch  F sh
2 2
80 Relationship Between Classical and Quantum Physics

The authors see that we have to deal here with the  c 

i    k k   mc 2 (41)
real spinors’ transform ( Ex , G ), ( E y , Bz ), t i k x
( Ez ,  By ), ( F , Bx ).    
Therefore, Φ and Ψ are spinors, and the field is a for m  0 i  c    k  i  k   (42)
t  k x 
spinor field, and the equations of motion Eqs. (22) and
  
(23) satisfy the spinor field as well. This field is or i  c P (43)
t
without the rest mass.
Eq. (43) represents the Dirac’s equation for leptons.
For the case   0,  ( )  ,   0, we
The authors begin with the expressions
apply R on  ' and S on  ' ; we have:   
Ex  Ex ch  Gsh, Bx  Bx ch  F sh   E 
j xk
El   jkl (44)
E y  E y , By  By   
Ez  Ez , Bz  By

  B   jkl
j xk

Bl (45)

F  F ch  Bxsh, G  G 'ch  Exsh and the components j  1, 2,3 of the Maxwell’s
equations will be
Ax  Ax , C x  Cx (40)
 1 B j
 jkl El  0 (46)
Ay  Ay ch  C zsh xk c t
C y  C y ch  Azsh  1 E j
 jkl Bl  0 (47)
Az  Azch  C y sh xk c t
B j
C z  Czch  Ay sh 0 (48)
x j
f  f ,    
  E j
In this case A and C form antisymmetric tensor 0 (49)
x j
of the second rank,  and f are scalars, and
E x , E y, Ez,G  and B x , By , Bz, F  form Substitute the Crammers’ complex vector
four-vectors analogous electrodynamics where  j  E j  iB j in these equations, multiplying Eqs.
 
 E,B 
form antisymmetric tensor of the (48) and (49) by i jkl and thereafter, adding Eq. (46)
electromagnetic field. Dirac’s equation with Eq. (47), and Eq. (48) with Eq. (49), the authors
 0, without mass term is   0  ,
 iD    obtain:
where  is a field function with components  l i  j
 jkl  0 (50)
 Ex , E y , Ez , F , Bx , By , Bz , G  xk c t
 j
Operating with the operator D on the field function 0 (51)
Eq. (28), we obtain Eq. (29). For the values x j
F  G  0 Eq. (29) are the Maxwell’s equations for Differentiate the Eq. (50) according to x j
the free field.
 2 l i    j 
 jkl     0 (52)
5. Description of the Maxwell’s Equations in x j xk c t  x j 
the Form of the Dirac’s Equation We will now try to bring the expression
We will show that Maxwell’s equations in the  l i  j
 jkl  0 (53)
vaccum can be written in the form of the Dirac’s xk c t
equation for massless particles [4] into the form of the Dirac’s equation for leptons (43)
 Relationship Between Classical and Quantum Physics 81

 j     iD    I  0,  iD    II 0 (62)
i  c  i jkl   i  l (54)
t  xk  The authors will pass to the non-relativistic
The authors use the matrix Schrödinger’s field.
S  k  jl  i jkl (55) The authors begin with II  0,  I  0, properties
to obtain of the matrices Eqs. (25), (26) and (27), and the
authors will take that [5, 6]
 j
i  c  Sk  jl Pk  l (56) 
t    I  eikx 0
(63)
i.e., Maxwell’s equations in the form of the Dirac’s 
equation. Matrix  S k  jl is 3  3, k  1, 2,3 Then the authors will have:

matrix i 0   i j a j ikx0 
e 0 (64)
 111 112 113  i j a j  i 0   
 S1  jl  i j1l  i   211  212  213  i 0   i j a j  0 (65)
 
 311  312  313   2  i 0   i j a j    0 (66)
0 0 0  0 0 0  In the non-relativistic limit:
   
 S x  i  0 0  1    0 0 i  i 0  2  (67)
0 1 0  0 i 0   that
    From the equation (66), we take for
 0 0 i i
   j a j  (68)
 S2  jl  S y   0 0 0  2
 i 0 0  Substitute  in Eq. (33) and obtain
 
 0 i 0   i 
i 0   i j a j    j a j  
 S3  jl  S z   i 0 0  (57)  2 
0 0 0 1 1
  
2

 j i a j ai   ai a j  
2

These matrices have the properties
(69)
 Si , S j   i ijk S k (58)
From the assymmetrical condition Eq. (25) the
S 2  S12  S22  S32  2 (59) authors have:
so that the authors have: 0 aj 0 ai 0 ai 0 aj
  2 g ji
S 2  S  S  1  2 (60) a j
0 a i
0 a i
0 a j
0
 j a a a a
j i i j
 2 ij
(70)
i.e. S  1 i  c  S k  jl Pk l (61)
t
a j  a i  a i  a j  2 ij
and the Eq. (61) describes the particle with a spin
S  1. But, this particle is the quant of the Using Eq. (70), the Eq. (69) takes the form
electromagnetic field―a photon. 1
i 0     (71)
2
6. Non-Relativistic Boundary and the
which is the Schrödinger’s equation for the free
Schrödinger’s Equation
particle. Now we take the case when the particle is
From the Dirac’s equations with the positive and under the influence of the external electromagnetic
negative mass term field. The motion of a free relativistic particle of mass
82 Relationship Between Classical and Quantum Physics

m and charge q is described by the Dirac’s equation where E   E  mc 2 ,  i ,  j   2i ijk k and
[7-9]   2
   p  p   p I .
i    ic   mc 2  (72)
t
where the 3 matrices  i and the matrix  are given 7. Conclusions
by
Investigation carried out in this work shows that the
 0 i  I 0 
i   i  and     (73) free electromagnetic field, spinor Dirac’s field without
 0  0 I  mass, spinor Dirac’s field with mass, and some other
 are the Pauli spin matrices, and I is the unit 2  2
i,
fields are described by the same variational
matrix. Describe how the Dirac’s equation should be formulation. The conditions that a field be described
modified if particle comes under the influence of an by Maxwell’s equations of motion are given in this
external electromagnetic field described by the work, and some solutions of these conditions are also

four-potential A   A , A  . Hence deduce that the
 0
given. Non-relativistic approximation of the equations
Dirac equation for a particle of charge q in the presence of the non-quantified field are the Schrödinger’s
of an electrostatic central potential V  r   qA , with
0
 equations. Dirac’s equation as a special case, contains
A  0, may be written as Maxwell’s equations and the Schrödinger’s equation.
 
  Hˆ   where Hˆ  ic   mc  V  r  (74)
2
i The authors conclude that the basis of the quantum
t
By seeking energy eigenstates of the form physics lie on the classical physics, including special
   iEt  theory of relativity.
  r , t     r  exp    (75)
    References
and writing the Dirac wave function   r  in terms
of two-component spinors [1] K. Ljolje, S. Vobornik, Canonical formulations of fields

   r 
satisfying the D'Alambert's equation, Fizika 11 (4) 171
 r      (76) (1979) 171-180.
 r   [2] K. Ljolje, Some remarks on variational formulations of
show that the Dirac equation may be written in the physical fields, Fortschr. Phys. 36 (1) 1988 9-32.
[3] S. Ahmetaj, Physical Fields Described by Maxwell`s
form

 E  V  mc    r   c  p   r   0
 Equations © 2007: American Institute of Physics.
2
[4] D. Bjorken, S.D. Drell, Relativistic Quantum Field,

 E  V  mc   r   c  p    r   0 (77)
 McGraw-Hill, New York, 1965.
2
[5] E.D Hoker, D. Gieser, Quantum Field Theory, Los
Angeles, 2004
Show that for a weak potential in the [6] P.J. Mulders, Introduction to Quantum Field Theory,
non-relativistic limit, these equations are reduced to an Amsterdam, 2004
energy eigenvalue equation of Schrodinger form [7] R. Casalbuoni, Advanced Quantum Field Theory, Firenze,
1999.
 ˆ  2

  p [8] H. Hees, Introduction to Relativistic Quantum Field

 V   r 

E   r   (78)
 2m  Theory, Texas University, 2004.
  [9] J. Norbury, Quantum Field Theory, Milwankee, 2000.
Journal of Mathematics and System Science 3 (2013) 83-88 D DAVID PUBLISHING

Oscillation Theorems for Second Order Differential

Equations and Their Applications

Xhevair Beqiri1, Elizabeta Koci2

1. Departament of mathematics, State University of Tetova Macedonia, Tetovo 389 044, Republic of Macedonia
2. Departament of mathematics, University of Tirana, Tirana 355 4, Albania

Received: January 14, 2012 / Accepted: February 12, 2013 / Published: February 25, 2013.

Abstract: The authors present several oscillation theorems for differential equation of second order
(r (t ) g ( ( x(t )) x' (t ))' q (t ) f ( x(t ))  0 (1)
and for differential equation with damping term
Mx' ' (t )  p (t ) x' (t )  q (t ) x(t )  0 (2)
where M > 0, r(t) is positive continuous function. The conclusion is based also on building function where coefficients are involved in
the equation and positive functions used by Philo H (t , s ) and averaging techniques. Our results generalized, extend to some already
known oscillation criteria in the literature. Also, here we give some applications of oscillation solution of . (1) and (2), where p(t) and
q(t) are positive. The original purposes of differential equation are the mathematical formulation of the vibration frequency and the
amplitude profile of a vibrating string with friction which the mass may have to encounter air resistance in its motion and in electric
circuit containing an ac voltage source, an indicator, a capacitor, and a resistor in series is analyzed mathematically, the equation that
results is a second order linear differential equation with oscillatory solution.

Key words: Differential, second order, oscillation, equation, applications etc.

1. Introduction there exists a sequence {n }n 1 of points in the

interval [t 0 , } , such that nlim n   and x( )  0 ,
 n
In this paper the authors consider the oscillation
behavior, of solutions in the second order differential
n N , otherwise it is said to be nonoscillatory. An
equation is said to be oscillatory if all its solutions are
equation :
oscillatory, otherwise it is considered that is
(r (t ) g ( ( x(t )) x' (t ))' q (t ) f ( x(t ))  0 (1)
nonoscillatory solution.
and the differential equation of second order with
In the following with respect to (1) and (2), the
damping term
authors shall assumed that:
Mx' ' (t )  p (t ) x' (t )  q (t ) x(t )  0 (2)
A1) r (t )  C ([t 0 , )), r (t )  0 , for t  t 0 ;
By a solution of (1) or (2) the authors consider a
A2) f ( x)  C (), xf ( x)  0 , for x  0 ;
function x(t ), t  [t x , )  [t 0 , } , which is twice
A3) g ( x)  C () , 0  m  g ( x)  M , for x   ;
continuously differentiable and satisfies (1) or (2)
A4) q(t )  C ([t 0 , )), q (t )  0 for t  t 0 .
on the given interval. The number depends on that
It is well known that Hille for equation
particular solution x(t) under consideration.
x' ' (t )  p (t ) x(t )  0
The authors consider only non-trivial solutions. A

solution x(t) of (1) or (2) is said to be oscillatory if 1 
proved that is oscillatory if  p(s)ds  4t
, if  is
t
Corresponding author: Xhevair Beqiri, Ph.D., research
field: differential analisys. E-mail: xhevairb03@yahoo.com. any small positive number.
84 Oscillation Theorems for Second Order Differential Equations and Their Applications

If in the (2) substituted Put

1 p(s) t
 
F (t )   f (l ) g ( (l ))dl
ds
2 M (*)
x(t )  u (t )e x ( t1 )
the authors get Multiplying (1) by r (t ) g ( ( x(t )) x' (t ) , the authors
p ' (t ) p 2 (t ) have
u ' ' (t )  (q(t )   )u (t )  0 (3)
2 4 1
In this paper is presented theorems, to use ((r (t ) g ( ( x(t )) x' (t )) 2 )' r (t )q(t ) g ( ( x(t )) x' (t )q (t ) f ( x(t ))  0
2
generalized Riccati-type transformations, averaging Integrating from t1 to t, the authors obtain
technique and positive function H(t,s), which explain 1 1
(r (t ) g ( ( x(t )) x' (t )) 2  (r (t1 ) g ( ( x(t1 )) x' (t1 )) 2 
results for oscillatory nature of differential equations 2 2
t
and their aplications.Our results extend and improve a
 r (t )q (t ) F ( x (t ))   ( r ( s ) q ( s ))' F ( x ( s ))ds  0 (7)
number of existing results (see [2],[7],[9]). t1

Denote
2. Main Result
1
(r (t1 ) g ( ( x(t1 )) x' (t1 )) 2
c
The nature of solutions of second order differential 2
equations is studied by many authors. Here, the authors Now by (7) it follows that:
give some conditions for coefficients where the (2) t
r (t )q(t ) F ( x(t )  c   (r ( s )q( s ))' F ( x( s ))ds
have bounded solutions. Taking in account the result
t1
the authors have obtained from previous researches , t
(r ( s )q ( s ))'
here the authors present more generalized criteria that c r ( s )q( s )F ( x( s ))ds (8)
t1
r ( s)q( s)
define oscillation solution of (1) and (2) that can be
applied in dynamics , electricity etc. Hence, by Granwalls-it inequality, the authors get
t
( r ( s ) q ( s ))'
Before the authors prove the theorems, the authors  r (s)q(s)
ds

state the following well-known Gronwall’s inequality r (t )q(t ) F ( x(t )  c  e t1

(9)
Lemma 2.1 Let I  [t 0 , T ) be an interval of real from this
c
numbers, and suppose that F ( x(t ))  , t1  t 0  0 (10)
t r (t1 )q(t1 )
u (t )  c   q ( s )u ( s )ds , for t  I (4) so, F ( x (t )) is bounded and from (*) the solution
t0
x(t ) is bounded .
where c is a nonnegative constant, and u, q  C ( I ,   ) .
Now the authors present some sufficient conditions
Then
for (1) to be oscillatory.
t
u (t )  c exp  q ( s )ds , for t  I (5) For convenience the class of functions H (t , s ) which
t0 comes from Philos, where H (t , s )  0 , for t  s  t 0 ,
The authors give the bounded solution of (1) with: H (t , t )  0 , are continuous and have partial derivatives
Theorem 2.2 Let r (t ), q(t )  C ([t 0 , )) and H and H satisfying: H (t , s )  h (t , s ) H (t , s )
1

(r (t )q(t ))'  0 , for t  t 0 and dt ds s

2


H (t , s )
 f (l ) g (l )dl   , (6) and
s
 h1 (t , s ) H (t , s ) .
then the solution x(t ) of (1) such that The authors define certain conditions for the
x(t1 )  0, t1  t 0 for some t1  [t 0 , ) is bounded. coefficients of (1) using positive function of Philo,
Proof Let x(t ) be a arbitrary solution of (1) such
that x(t1 )  0, t1  t 0 . when the solution is oscillatory.
 Oscillation Theorems for Second Order Differential Equations and Their Applications 85

Theorem 2.3 Let exist constant k, m, M, such that: exists and c  (a, b) , such that
f ' ,  '  k  0 , 0  m  g (t )  M (11) w' (t )   A0 (t )  A1 (t ) w 2 (t ) (16)
and then
t
1 h2 (t , s ) r ( s ) M
 2 H  1 ( s, a )h1 ( s, a )
c 2
lim sup  ( q ( s ) H (t , s )  ) ds = (12) 1
 ( H (s, a) A0 (s) 

)ds 
H  (c , a ) a
2
t  H (t , t ) 0 t0
4k 4 A1 ( s )
 2 H  1 (b, s )h2 (b, s )
c 2
then (1) is oscillatory . 1 (17)
 
H (b, c ) a
( H  (b, s ) A0 ( s ) 
4 A1 ( s )
)ds  0
Proof Assumed that x(t ) is a nonoscillatory

solution of (1). Then exist t 0 , such that x(t )  0 , for Proof If multiplying (16) with H (t , s)  0 and
t  t0 . integrate from t to c, the authors have
c

H

The authors define ( s, t ) w' ( s )ds  
r (t ) g ( ( x(t )) x' (t ) t
w(t )  , for t  t 0 (13)
f ( ( x(t ))
c c

 H (s, t ) A0 (s)ds   H (s, t )w (s)ds

  2

Thus by (13), to consider (11) and (1), the authors t t

obtain hence
c
w 2 (t )k 2
H

q(t )   w' (t ) 
(14) ( s, t ) A0 ( s )ds   w(c) H  (c, t ) 
rM t
Multiplying (14) on both side with positive function  1 2
 
H (t , s) and integrate from t 0 to t , the authors get c   H 2 ( s, t )h1 ( s, t ) 
  H (s, t ) A1 (s) w  2

t t
 ds 
A ( s ) 
 q ( s ) H (t , s )ds  H (t , t0 )w(t0 )   w( s ) h2 (t , s ) H (t , s )ds t 

1

t0 t0

 2 H  1 ( s, t )h1 ( s, t )
c 2

t 2 2
w (t , s )k H (t , s ) ds
 ds , 4 A1 ( s )
t
rM t
0

If t  a  and dividing it with H  , the authors get

t

 q( s) H (t , s)ds  H (t , t0 )w(t0 )   2 H  1 ( s, a ) h1 ( s, a ) 
t0 c
1
H

( s, a ) A0 ( s )  ds   w(c ) (18)
2 H 
(c, a )  4 
 wk H (t , s ) h2 (t , s ) rM 
t a  A1 ( s ) 
    ds Similar to this, if (16) multiplies with H (t , s )  0 

t0  r ( s) M 2k  and integrate from c to t , the authors obtain

t 2 t
h2 (t , s )r ( s ) M
 ds H

( s, t ) w' ( s )ds  
t0 4k 2 c
t t
hence
 H (s, t ) A0 (s)ds   H (s, t )w (s)ds
  2
t 2
1 h (t , s ) r ( s ) M
H (t , t 0 ) t
( q ( s ) H (t , s )  2 )ds  w(t 0 ) (15) c c

0
4k 2 hence
t
thus by (15) it follows that
H

t 2
( s, t ) A0 ( s )ds  w(c) H  (t , c) 
1 h (t , s )r ( s ) M
lim sup 
t  H (t , t 0 ) t
(q( s ) H (t , s )  2
4k 2
)ds

c
 1 
2
0 t   H 2 (t , s ) h (t , s ) 
  H (t , s) A1 (s) w(s)  2

is bounded , that is the contradiction with condition 2
 ds 
c A1 ( s ) 
given in the theorem.  
Lemma 2.4 Let A0 , A1  C ([t 0 , ), ) , where c 2  1
 H (t , s )h2 (t , s )
2
 ds
A1  0 and w  C 1 ([t 0 , ), ) . If (a, b)  [t 0 , ) t
4 A1 ( s )
86 Oscillation Theorems for Second Order Differential Equations and Their Applications

p 2 (t ) p ' (t )
if t  b and dividing with H  , the authors get w' (t )  w 2 (t )  Kq(t )   0
4M 2M
b  1  the rest of the proof is similar.
 H  (b, c) A ( s )   H (b, c)h1 (b, c) ds  w(c)
2
1
  
H (b, c) c 
0
4 A1 ( s )  (19) Application For instance, the mass may have to

If (18) and (19) are added, the authors obtain (17). encounter air resistance in its motion. A typical
Based on the relation of Hile, for (2) can be defined physical assumption is that the frictional force points
oscillation nature of solutions like as follows: opposite to the direction of motion and has magnitude
f ( x) proportional to the speed. If there is no external forces,
Theorem 2.5 Let q(t )  0 , and  K  0 , for
x this leads to the differential equation
x  0 , where K is constant. If for large t   , mu ' 'u ' ku  0 (21)
p (t )  O(1) and where u ' is magnitude of frictional force, with
 characteristic equation
K p 2 ( s) p' ( s ) 
[ M
q( s ) 
4M 2

2M
]ds  0
t
(20)
mr 2  r  k  0
t

1 which have the roots

where  0  , then (2) is oscillatory.
4     2  4km
r1, 2 
Proof Let x(t) be a nonoscillatory solution of (2) , 2m
which without loss of generality, can be assumed to be From the last equation the authors have three cases
x(t )  0 , from that f ( x(t ))  0 , for t  t 0  0 . By for the general solution, but oscillation solution the
(16-20) it is easy to conclude that there exists t1  t 0 , authors have for  2  4km and the general solution
such that has the form

K p 2 ( s) p' ( s )  u (t )  Ae r1t  Be r2t
M[ q ( s ) 
4M 2

2M
]ds  0 , for t  t1 ,
t For   4km   2 , the authors have the general
t
which yields that there exists an integer n(t )  t such solution and it has the following form
 
that  t  t
u (t )  e ( A cos(t )  B sin(t ))  Re 2 m cos(t   )
2m

t'
K p 2 (s) p' (s)
 [ M q(s)  4M 2 
2M
]ds  1
t
and where R  A 2  B 2 , and   tan( ) .
B
t A
12 12 2 In (2) if M = m, p(t) =  , q(t) = k, and f(x) = x, the
  , for t '  n(t ) ,
t t' t authors have (21) which satisfies the theorems.
1 In the figure is demostrate a motion of a mass on a
where  0   1    .
4 spring without friction.
x' (t ) From Hook’s Law, which claims that if the spring is
Defining z (t )  , for t  t 0 . By (2) z(t)
x(t ) stretched (or compressed) x units from its natural length
satisfies the equation (see fig.1), it exerts a force that is proportional to x:
p (t ) f ( x)
z ' (t )  z 2 (t )  z (t )  q(t ) 0 equilibrium position
M Mx(t )
f ( x)
from  K  0 , the authors get
x
m
p (t ) K
z ' (t )  z 2 (t )  z (t )  q(t ) 0
M M
p(t ) 0 x
Let w(t )  z (t )  , the authors have Fig. 1 Motion of a mass on a spring without friction.
2M
 Oscillation Theorems for Second Order Differential Equations and Their Applications 87

restoring force  kx d 2x dx x(t )

2
LR   V cos t (22)
where k is a positive constant (called the spring dt dt C
constant). If the authors ignore any external resisting The first step in solving this equation is to obtain the
forces (due to air resistance or friction) then, by general solution of the corresponding homogeneous
Newton’s Second Law equation
Force equals mass time acceleration, the authors have d 2x dx x(t )
2
R 
L 0
dx 2 k dt dt C
m 2
 kx , or x' ' (t )  x(t )  0
dt m but notice that this differential equation has exactly the
This is a case given on (21) where   0 , or satisfied same mathematical form as the equation for damped

k 0 oscillator,
 M ds  t
, for t  t1  0
d 2x dx
t M
2
K  kx(t )  0 (23)
When an electric circuit containing an ac voltage dt dt
source, an indicator, a capacitor, and a resistor in series The general solution of (1) was found to be
Kt
is analyzed mathematically, the equation that results is  k K2 k K2
xe 2M (c1 cos  t  c 2 sin  t)
a second order linear differential equations. The M 4M 2 M M2
voltage v(t) produced by the ac source will be if K  2 Mk .
expressed by the equation v  V sin t , where V is the The particular solution of (1) will have the form
maximum voltage generated. An inductor is a circuit x(t )  A sin t  B cos t , where from (22) the authors
element that opposes changes in current, causing a get A and B.
dx It is easy to see that (2) satisfies the condition of the
voltage drop of L( ) , where x is the instantaneous
dt 
K K2 
current and L is a proportionality constant known as the theorems given above, because ( M
k
4M 2
) 0 ,
t
t
inductance. A capacitor stores charge, and when each
plate carries a magnitude of charge q , the voltage drop
for t  t1 , from where the solution is oscillatory.
q
across the capacitor is , where C is a constant called
C
References
the capacitance. Finally, resistor opposes the flow of
current, creating a voltage drop equal to xR, when the [1] E. Hille , Nonoscillation theorems, Amer. Math. Soc. 64
(1948) 234-252.
constant R is the resistance. Kirchhoff’s Loop Rule
[2] D. Cakmak, Oscillation criteria for nonlinear second order
states that the algebraic sum of the voltage differences differential equations with damping, Ukrainian
as one goes around any closed loop in circuit is equal to Mathematical Journal 5 (60) 2008 694-700.
zero. Therefore, if the voltage source, inductor, [3] A. Tiryaki , A. Zafer, Oscillation criteria for second order
nonlinear differential equations with damping, Turk J.
capacitor, and resistor are all in series, then
Math. 24 (2000) 185-196.
dx q
V sin t  L   xR  0 [4] M.M.A. El-Sheikh, R.A. Sallam, D.I. Elimy, Oscillation
dt C criteria for nonlinear second order damped differential
Now, if an expression for x(t) the current in the equations, Int. Jour. of nonl. sc. 2010, 297-307.
circuit as a function of time , then the equation to be [5] Xh. Beqiri, Elizabeta Koci, Interval oscillation criteria for
second order nonlinear differential equations with
solved must be written in terms of x. To this and,
damping term, int. jur. of sc. inov. new tech. 1 (2) (2011)
differentiate the previous equation directly, and use the 85-91.
dq [6] Xh. Beqiri, Oscillation criteria for second order nonlinear
definition x(t )  , the authors have perturbed differential equations, Hieritage 6 (2011) 83-90.
dt
88 Oscillation Theorems for Second Order Differential Equations and Their Applications

[7] J. Ohriska, A. Zulova, Oscillation critweria for second
order non-linear differential equation, Institute of
mathematics, Safarik University, Jesenna, 10 (2004) 1-10.
[8] R. Feynman, The Feynman Lectures on Physics, Pearson,
California, 1989.
[9] Ravi P. Agarval, Said R. Grace, Donal O’Regan,
Oscillation Theory for second order linear, half-linear
superlinear and sublinear dynamic equation (2002) Florida
institute of technology.
Journal of Mathematics and System Science 3 (2013) 89-95 D DAVID PUBLISHING

Ranking Independent Variables in Linear Models

Deng-Yuan Huang1 and Ren-Fen Lee2

1. Fu-Jen Catholic University, New Taipei City 24205, Taiwan
2. Kaohsiung University of Applied Science, Kaohsiung 807, Taiwan

Received: December 19, 2012 / Accepted: January 11, 2013 / Published: February 25, 2013.

Abstract: In this article, the true model of size p is considered. The importance of the independent variables will be studied. The model
by dropping independent variable one at a time is called a reduced model, and the size of the model is p-1. The non-centrality quantity
is employed to measure the strength of influence on the true model for the independent variable dropped. The larger influence is defined
to be more important in the true model. In this article, the goal is ranking the importance of the independent variables to study the
structure of the models. It is important in the practical application. The confidence interval approach is used to rank the degrees of the
influence of the independent variables in linear models and achieve the goal. An illustrative example is given, and the modeling
procedure is studied to check the assumptions step by step in this example to make sure the assumptions satisfied in the true model. As
a result, the proposed method can be used efficiently.

Key words: Non-central F-distribution, non-centrality, confidence interval, bonferroni inequality.

1. Introduction squares for the reduced models is denoted by SS p 1,m .

Accordingly, the residual sums of squares for the true
The linear model is considered as follows:
model is denoted by SS p . It is known that:
Y  X   (1)
' '
SS p  Y QY and SS p 1, m  Y Q( m) Y , where,

where Y = Y ,Y ,...,Y
1 2 n  ' is an n 1 vector of
Q  I n  X [ X ' X ]1 X ' and

responses, X  1, X 1 ,..., X p 1  '
is an n  p , (n>p),
1
Q( m )  I n  X ( m ) [ X ( m ) X ( m ) ] X ( m ) . Under the true
' '

SS p
matrix of independent variables of rank p, model assumption, it is shown that and
2
   0 ,.1,...,  p 1  ' is a p 1 SS p1,m
 n
2
p and  n ( p 1),
vector of are distributed as '2
,
 2 p 1, m

parameters.,  is distributed N 0,  I n , and I n  2

 respectively, where,  p 1, m  X  ' Q ( m ) X   / 2 ,
2

m = 1, 2, …, p-1,   , denotes the non-central (central)

'

denotes the n  n identity matrix. Model Eq. (1) is chi-square distribution on  degrees of freedom with
non-centrality  (  =0). Thus, E SS p   ( n  p )
referred as the true model of size p. From this model, 2

the reduced model, each of size p-1, by dropping the

independent variable at a time is obtained. Let X (m ) and E SS p 1, m   ( n  p  1) 2  2 2  p 1, m , m = 1,
denote the “X matrix” of the reduced model obtained 2, …, p-1.
by dropping the independent variable X m , m = 1,
Let a reduced model be Y  X (m)  ( m )   ( m ) . The
2, …, p-1. Correspondingly, the residual sums of SS p 1, m
associate mean square error is . The
n p
Corresponding author: Ren-Fen Lee, Ph.D., professor, expectation is
research field: multiple decision theory. E-mail:
candy@cc.kuas.edu.tw.
90 Ranking Independent Variables in Linear Models

independent variables.
 SS   1  2  2 2 
E  p 1, m   1      p 1, m . It Lemma 1.1 see Ref. [1]: Let Y have the
 n p   n p n p non-central F distribution with u, v degrees of freedom,
shows that the noncentrality  p 1, m will affect the and non-centrality  , denoted by Fu ,v ,  . Then
prediction error in reduced model. PY  y  [1  e 1.794148 x ]1 (2)
Note that Q (m ) , m = 1, 2, …, p-1, are idempotent where
and symmetric, thus they are positive semi-definite. 1

Hence  p 1, m , m = 1, 2, …, p-1, are nonnegative.  2  uy  3  2(u  2 ) 

1     1  2 
Obviously,  1   2  ...   p 1  0 implies that  9v  u     9(u   ) 
x  1
 p 1, m  0 for m = 1, 2, …, p-1. When   0 ,  
2 2

 2(u  2 )  2  uy  
3
 p 1, m can be interpreted as the contribution of X m
in making significant given that other variables are
 9(u   ) 2
9v  u    
 
already in the model. the maximum absolute approximation error is 0.05
By Taylor’s expansion, for u  1, v  5 and   0 .
 SS p 1,m  E  SS p 1,m  (n  p  1) 2  2 2 p 1,m Now, consider the problem of the confidence
 
interval estimates of  p 1 ,  p 1   p 1,1 ,...,  p 1, p 1  .
E
 SS p  E  SS p  (n  p) 2
Let  p 1,[1]   p 1,[ 2 ]  ...   p 1,[ p 1] denote the
(n  p  1)  2 p 1,m
 ordered  p  1, m s . Require that, for each  p 1, m with
'

(n  p)
the confidence coefficient 1   , 0    1
thus
satisfying:
 n  p  SS p 1, m  n  p  1 inf P c2 ,m  ˆ p 1, m   p 1, m  c1, m , m  1,2,., , , p  1 1   0 (3)
E      p 1, m
 2  SS p  2  
Where  0  are given.
p 1
n  p  SS p 1,m  n  p  1
Use ˆ p 1,m    to Theorem 1.2 For each  p 1, m with the confidence
2  SS p  2 coefficient 1   , 0    1 , satisfying
estimate the parameter  p 1, m .

inf P ˆp 1, m  c1m   p 1, m  ˆp 1, m  c2 m  1   
Rewrite ˆ p 1,m as follows:
where
      16C 2  2  4C 2  
3

1  SS  SS p 1 1 1    2A    A  4C (1  A)  1 
2


ˆp 1, m   p 1, m    ˆm  1  u  H
  
  9u   9u   
  1 
2  SS p  2 2 2 2  u
c1m  max   
2( A2  C 2 (1  A))
 
 
 (n  p)   



 
 
SS p 1.m  SS p 1  u  H 1
3
 
where ˆ m  .  
2  u
    1
 A 



SS p
(n  p)  3

1   E  E2 4DF   where
 p 1   p 1,1 ,  p 1, 2 ,...,  p 1, p 1  T . The
c2m  1   
Let 2   2D 
 

statistic ̂m has the non-central F distribution on 1 and  2  2
u  1 ,   n  p , 1    A , 1  B
n-p degrees of freedom with non-centrality parameter  9  9u
 p 1, m , denoted by F1, n  p ,  p 1, m . Note that the ̂ m is     ,

C
1
 ln  D  A2  C 2 1  A
the statistic used to evaluate the importance of the  1.794148   2   
 Ranking Independent Variables in Linear Models 91

E  2 AB , F  B 2  C 2 1  B  Thus from Eq. (8)

1
  3
    3
 1    E  E 2  4 DF 
1
  2u    c 
   1  2c2m   H  2   p 1, m 2 2 m   
 1 
2(u ) 

 1  
 2   2D     9v 

u   9(u ) 2 

  1   
Proof Require that, for each  p 1, m with the 1
 ln 
1.794148  2    (9)
 
2 2

confidence coefficient 1   , 0    1 , satisfying   

 2u 
1
 c  
  3

 
 2(u ) 2   p 1, m 2 2 m   

inf P ˆp 1, m  c1m   p 1, m ˆp 1, m  c2 m  1   (4) 
 9(u )
2
9v  u  
  
   
Let  


P ˆp 1, m   p 1, m  c1m  1 

2
 (5) or

 2  p  1 , m  1  2 c 2 m 3   1 
 2  1
 2 
and 1  
 9v   9u 

P ˆp 1, m   p 1, m  c2 m 

 (6)
1
(10)
 9 u  9 v 2  p  1, m  1  2 c 2 m 3 
 2 2 2 2

2
From Eqs. (5) and (6), it will be satisfied Eq. (4) as
 1    
follows:   ln  
 1 . 794148  2 
inf P  c  ˆ   2m, c p  1, m p 1 m 1m  C2
Let u 1 ,   n p  , where

 inf P ˆ p  1, m   p 1 m ,  c ,1 m  36
 1    

 sup P ˆ p  1, m   ,  c2 m
p 1 m  C  ln 
 1.794148   2   


1 First, find the confidence interval of  p 1, m .

From Eq. (5), let From Eq. (10), for the convenience, let
  1   1 
  A , 2 p 1, m  1  2c2 m 3  Y
P ˆ p 1, m  2   p 1, m   c1m    P ˆ p 1, m  2   c1m    2  1
  2   2   (7) 1 
(1)
1.794148 x , 
 9 
1
 [1  e p 1 m
]  1 2
2 1 B
First, find a range of the non-centrality  p 1, m . 9u
By Lemma 1, for fixed y, ˆ ( p 1) is stochastically Thus
decreasing in terms of  p 1,m  0 . Then from Eq. (6), AY  B
C

P ˆ p1,m  y  P 0 ˆ p1,m  y . p 1,m
  1  B   1  AY 
1
2 2
(11)

Thus, the following result is obtained.

Since C  0 , then AY  B  0 ,
  1 
P ˆ p  1, m  2   p  1, m   c 2 m  
  2  
then  AY  B   C 2 (1  B )  (1  A )Y 2 .
2

  
D  A2  C 2 1  A , E  2 AB , and
1
 P p 1 ,m  0 ˆ p  1, m  2   p  1, m   c 2 m   Let
  2 
1.794148 x (p2)1,m 1  F  B 2  C 2 1  B  .
 [1  e ]  (8)
2 Then
where DY 2  EY  F  0 (12)
71.794748 1
1
 2  u ( 2 p 1, m  1  2c2 m )   2( u ) 
3 1   
1     1  2
Let 0    1  e 3 2  and
 9 v  u   9( u )  2  
x (2)
p 1, m  1
2
 
2 2 C , then
 2 ( u ) 2  u ( 2    1  2 c )  3
   n p 
  p 1, m 2 m
 36
 9(u ) 2 9v   

D  A2  C 2 1  A  0
u
 
92 Ranking Independent Variables in Linear Models

F  B 2  C 2 1  B   0   1 
P ˆ p 1, m  2  p 1, m   c1m 
  2 
Thus E  4 DF  0 .
2

 1 
From Eq. (11)  P p1,m H ˆ p 1, m  2  c1m 
 2 
 E  E 2  4 DF  E  E 2  4 DF
Y or Y  1.794148 x (p1)1 , m 1
2D 2D  [1  e ]
 
1 1
And AYB0 implies A 2 p1,m  1  2c2m 3 B
 2  u 1  2c1m    2(u  2 H ) 
3
The following results are obtained. 1     1  2 

 9v  u  H   9(u  H ) 
 3

1   E  E 2  4 DF   1
 p 1, m    1  2c2 m   2

 2(u  2 H )  2  u 1  2c1m  
2
2  2D 
  3


 3
  9(u  H ) 2 9v  u  H  
or  1   E  E 2  4 DF    
   1  2 c2m 
p 1, m
2  2D  1

 
 2  u 1  2c1m   3
1B 1
3
1   1 
and 
p 1, m      c2 m  9v  uH 
2  A 2  1
Let  2

 2(u  2 H )  2  u 1  2c1m  
2
3
   3

 1    E  E 2  4DF   1  B  1
3 
 
c2m  max 1  ,     c2 m 
2
  
 2 D

 2  A 2
  
(13)  9(u  H ) 2 9v  u  H  
 
 3

1    E  E 2  4DF   1
 1 
2   2D    2  u 1  2c1m   3
   1    1
be guaranteed the positive of  p 1, m . Since  9v  u  H 
 1
 E  E 2  4 DF  
2 2

 4  2  u 1  2c1m   
3
2D
 E  E 2  4 DF
 9u 9v  u  H  
  
2 A2 1  2  
 ln  (15)
E E 2
1.794148   

2 A2

Let 1    A, , C  
2 1    
 2E  E 2B B  ln  
  2    9v   1.794148   2   
2 A2 A A A
Thus u(1  2c1m )
Let X 
 E  E  4 DF 2
B uH
0 From Eq. (15)
2D A
Thus 1
 2  u 1  2c1m   3
  
3
 1    1
1   E 
 p  1, m    
E 2  4 DF   1  2 c 2 m   (14)  9v  u  H  1  2  
 ln 
 2 
  2D 
  
1
1. 794148   
 
2 2

 4  2  u 1  2c1m   
3

 H  9u 9v  u  H  
From Eq. (7)  
 Ranking Independent Variables in Linear Models 93

1
AX 3  1 1  u  H 1
3

 C and c1m      1
2  u 
1
4 
2 2  A 
  1  A X 
3

 9u  Thus
  
3

1 1    16C 2  2  4C 2     (16)
 2A    A  4C (1  A) 1  
1
2
1  u  H 
Since C  0 and AX  1  0 , thus X   9u 
3 3   9u  
.      1 
2  u 2( A2  C 2 (1  A)) 
A c1m  max    
 
 
    
1  
4  1  u  H 
3
1 2 2 1 
AX  1  C   1  A X 3   2  u     1 
3
 A 
   
 9u 
From Eqs. (13) and (16), the confidence interval of
2 1
4 2
  p 1,m is as follows:
A2 X 3  2 AX 3  1  C 2   1  AX 3  the parameter

2
 9u 
1

inf P ˆp 1, m  c1, m   p 1, m  ˆp 1, m  c2 m  1  
 p 1,m  

A 
2
4C
2
 C 2 (1  A) X  2 AX  1 
3 3
0 as an interval estimate of  p 1,m with confidence
9u
coefficient 1   .
 4C 2 
1
2 A  4 A 2
 4 A 2
 C 2
1 A  1 
9
  2. Example
 u 
X3  2 2

2 A  C 1  A  In the following example, the data were originally
3 reported in Ref. [2] for assessing the effects of air
 2 
 2 A  4 A2  4A2  C 2 1  A1  4C   pollution on the median housing value (medv).
  9u  
 The dependent variable is Median value of owner
X  
 2A2  C 2 1  A  homes. The considered independent variables are as
  follows:
 
3 Crim: Per capita crime rate;
 2 
 2 A   16C  A2  4C 2 1  A1  4C  
2
Zn: Proportion of residential land zoned 25K+;
  9u   9u  
   Indus: Proportion of non-retail bus acres per town;
 
 2A2  C 2 1  A  Chas: Charles River connection dummy;
  nox: Nitric oxides concentration pp 10M;
 
1 3 rm: Average # of rooms per dwelling;
and X  ( ) . age: Proportion of owner-occ dwellings before
A
1940;
1
 71.794748
 dis: Wtd dist to five Boston employment ctrs;
when 0    1 1  e 3 2 
 2
and   n  p  C
2   36 rad: Access index to radial hwys;
tax: Full-value prop-tax rate per $10K;
then A  C 1  A  0
2 2
ptratio: Pupil-teacher ratio by town;
 16C 2  2  4C 2  b: Proportion of blacks per town;
then   A  4C 2 1  A1  0
 9u   9u  lstat: % lower status of the population.
There are 506 observations. The data was used by
 3

 
 2A   16C 2  2  4C 2    the classification and regression trees (CART) methods
   A  4C 2 (1  A)1   
1u H   9u   9u   to integrate the interaction into the linear regression
c1m     1
2 u  2( A  C (1  A))
2 2
 
    models. There are eight branches in the CART tree.
   
  The correspondence eight levels are defined as follows:
94 Ranking Independent Variables in Linear Models

Level 1: rm ≤ 6.941, lstat ≤ 14.4, dis ≤ 1.38;
Level 2: rm ≤ 6.941, lstat ≤ 14.4, dis > 1.38
Level 3: rm ≤ 6.941, lstat > 14.4, crim ≤ 6.99
Level 4: rm ≤ 6.941, lstat > 14.4, crim > 6.99
Level 5: rm > 6.941, rm ≤ 7.437, nox ≤ 0.66
Level 6: rm > 6.941, rm ≤ 7.437, nox > 0.66
Level 7: rm > 6.941, rm > 7.437, nox ≤ 0.68
(4) Testing the equality of the variances of the
residuals.
H 0 :  12   22  ......   k2   2
The systematic sampling method is used to separate
residuals into five groups. Here, the size of these
groups is n1  102, n2  n3  n4  n5  101.
Let X i 1 , X i 2 ,  , X in , i  1, 2,  , k , be k
i

Level 8: rm > 6.941, rm > 7.437, nox > 0.68 independent random samples from k normal;
The modeling procedures are as follows: populations 
N  i ,  i2  , i  1, 2,  , k ,
(1) Transform the dependent variable  2

The dependent variable medv was transformed to respectively.. Let i  k

i
1 be as the distance
stabilize the variances and achieve a symmetric (  ) 2
j
k

j 1
distribution under each fixed combination of ,
between the variance  2
i and all other  2j s , and
independent variables. Let y  medv 0.134912 . Use
si2
CART (the classification and regression trees) to find ˆi  k 1
as an estimator of i , where
the interaction between the explanatory variables. Each ( s ) 2
j
k

class is consisted as an interaction. For larger the size j 1

1 ni 1 ni
of sub-sample, class 5 and class 6 are merged into one
class. Class 7 and class 8 are also merged into one class.
si2  
ni  1 j 1
( X ij  X i ) 2 and X i 
ni
X
j 1
ij ,

The linear model is considered as follows: i  1, 2,  , k .

y   0  1crim   2 zn   3indus   4chas   5nox
Lemma 2.1 see Ref. [3]. Given 0    1,   0,
  6 rm   7 age   8dis   9 rad  10tax  11 ptratio (17)
and sample sizes n1 ,  , n k , the critical value C for
 12b  13lstat  14CA2  15CA3  16CA4  17CA5
the decision rule R satisfying
 18CA6  
 
The interaction terms by dummy variables are added Pr  max ˆi  C | H 0    is approximately given
1 i  k 
in model Eq. (17).
 a[ k ]  k  
(2) Use the least squares method to estimate the by C  exp ln  1  b[ k ] , where
parameters
 1 .5   
 
481 2

inf  Yi   0  1 X i ,1  ...  18 X i ,18  (18) 2(k 2  2k ) 1 2

k
a[ k ]  
k 2 ( n[1]  1) k 2
n
j 1
i 1
j 1

(3) Normal Testing

2 1 k 1
After computing, the Durbin-Watson statistic = b[ k ]    
2 n[ k ]  1 k j 1 n j  1
1.338, R = 86.5%. Use the Kolmogorov-Smirnov
(KS) test to test the following hypothesis: The decision rule R is as follows:
H 0 : The distribution of ei is normal distribution; Reject H 0 whenever ̂ i  C for some
H 1 : The distribution of ei is not normal i  {1, 2,  , k} , P{max ˆi  C | H 0 }   .
1i  k
distribution. And the probability of correct decision
The p-value = 0.045 of the KS test is obtained. inf P {CD | R }  P * , where
Under the significant level  =0.01, the null  2 

hypotheses is accepted. 
    ( 12 ,  ,  k2 )  [ k ]  
2

 Ranking Independent Variables in Linear Models 95

1 Table 1 Sample sizes, sample variances and statistic of

P*  five groups.
1.5(ln C  ln   b[1] ) 
group 1 2 3 4 5
1  exp 
 a[ k ]  ni 102 101 101 101 101
si2 0.0010 0.0009 0.0008 0.0007 0.0012
1 1 k 1 ̂ i
bi     1.0997 0.9574 0.8831 0.8121 1.3243
ni  1 k j 1 n j  1
Table 2 Confidence intervals of  p1,m .
(k  1) 2 2 1 k
2
ai  2   , b[1]  min{bi } Omitted
SSE ̂ m ˆ p 1, m Lower Upper
k (ni  1) k 2 j 1
n i
j 1 variable limit limit
j i
TRUE 0.4649
a[ k ]  max{ai } , n1  n2   ...  nk  crim 0.5276 65.70 32.35 26.22 32.85
i
zn 0.4687 4.03 1.52 0.00 2.02
Here, k = 5, n[1]  101 , n[ 5 ]  102 . Let the indus 0.4658 1.00 0.00 0.00 0.50
significant level  = 0.05, then, a [ 5 ] =0.015992, chas 0.4678 3.09 1.04 0.00 1.55
nox 0.4960 32.63 15.82 9.69 16.32
b[ 5] = -0.00982 and C = 1.458746. Since all ˆ i
rm 0.4654 0.58 0.00 0.00 0.50
are smaller than C, then the null hypotheses is accepted. age 0.4666 1.78 0.39 0.00 0.89
That is all of the variances of the residuals are equal. dis 0.4998 36.55 17.77 11.65 18.28
(5) Ranking independent variables rad 0.5128 50.20 24.60 18.47 25.10
As mentioned, in regression model, the noncentrality tax 0.4909 27.22 13.11 6.99 13.62
ptratio 0.5137 51.14 25.07 18.94 25.57
 p 1, m can be interpreted as the contribution of
b 0.4703 5.66 2.33 0.00 2.83
independent variable X m in making significant given lstat 0.5490 88.08 43.54 37.41 44.04
that other variables are already in the model. To create
a confidence interval to estimate  p 1, m with coefficient of the simultaneous confidence intervals of
 p 1 is 0.995.
confidence coefficient 0.95, model Eq. (18) is rewritten
as follows: 3. Conclusions
medv0.134912  0  1crim  2 zn  3indus  4chas  5nox
(19)
 6rm  7 age  8dis  9rad  10tax  11 ptratio  12b The structure of the linear models is very important
 13lstat  14CA2  15CA3  16CA4  17CA5  18CA6   in practical application. The structure levels are the
Then, u = 1, p = 19, n = 506,  = 487, A = main parts of the model. This article is to rank and
0.9995, B = 0.7778. Let the confidence coefficient handle the most important levels. It will be helpful for
be 0.95, then  = 0.05. And C = -2.04195, D = making decision. An example for the modeling
0.9972, E = -1.5549, F = -0.32163. By equation procedure is also given.
(13), c 2 m = 0.5032. By equation (14), H = 0. Use
References
equation Eq. (16), c1m = 6.1248. After computing, the
[1] D.Y. Huang, Selection procedures in linear models,
results are listed in table 2.
Journal of Statistical Planning and Inference 54 (1996)
From table 2, there are five groups of variables: (zn, 271-273.
indus, chas, rm, age, b), (nox, dis, tax), (rad, ptratio), [2] D. Harrison, D.L. Rubinfeld, Hedonic prices and the
(crim), (lstat). That is, the contribution of independent demand for clean air, Journal of Environmental
Economics & Management 5 (1978) 81-102.
variable “lstat” is larger than other variables. The
[3] C.C. Lin, D.Y. Huang, Multiple decision procedures for
second important independent variable is “crim”. Since normal variances, Communication in Statistics Series B, --
 =0.05, using Bonferroni inequality, the confidence Simulation and Computation 36 (2) (2007) 265-275.
Journal of Mathematics and System Science 3 (2013) 96-100 D DAVID PUBLISHING

On the Isolation Property over a Database Domain

Bektur S. Baizhanov1 and Beibut Sh. Kulpeshov2

1. Institute of Mathematics and Mathematical Modeling, Almaty 050000, Kazakhstan
2. Department if Information Systems and Mathematical Modeling, International Information Technology University, Almaty 050000,
Kazakhstan

Received: January 3, 2013 / Accepted: February 11, 2013 / Published: February 25, 2013.

Abstract: The authors consider relational databases organized over an ordered domain with some additional relations – a typical
example is the ordered domain of rational numbers together with the operation of addition. In the focus of our study are the FO
(first-order) queries that are invariant under order-preserving permutations–such queries are called order-generic. It was discovered that
for some domains order-generic FO queries fail to express more than pure order queries. The collapse result theorem was proved for
locally generic queries over a linearly ordered domain with the Pseudo finite Homogeneity Property (or / and the Isolation Property) by
Belegradek et al.. Here the authors consider a circularly ordered domain and prove the collapse result theorem over a quasi circularly
minimal domain.

Key words: O-minimality, database query, circularly ordered domain.

1. Introduction first-order logic FO is used as the query language, its

formulas may use database relations as well as the
In the relational model of databases introduced by
domain relations, while variables range over the entire
E.F. Codd [1, 2] a database state is thought as a finite
domain.
collection of relations between elements. For example,
The expressive power of database queries has been
a father-son relation can be represented in the form of
studied in Refs. [3-7].
one binary relation (or a two-column table). The names
of the relations and their arities are fixed and are called 2. Definitions and Preliminary Results
a database scheme. Particular information stored in the
Let M   M , ,  be a linear ordering. If the
relations of a given scheme is called a database state.
authors connect two endpoints of the linearly ordered
As the authors acquire more and more information
set M (possibly, these are –  and + ) then the
about fathers and sons, the database states change, but
authors obtain a circular ordering.
the scheme (one binary relation) does not. Database
Definition 2.1 A circular ordering is described by a
relations (tables) are always going to be finite.
ternary relation K which satisfies to the following
However, it is often convenient to assume that there is
conditions:
an infinite domain–for example, the integer or rational
(co1)  x  y  z (K(x, y, z)  K(y, z, x));
numbers or the strings–such that the data elements are
(co2)  x  y  z (K(x, y, z)  K(y, x, z)  x = y 
chosen from this domain. Functions and relations
y = z  z = x);
defined over the entire domain like  and , may also
(co3)  x  y  z (K(x, y, z)   t [K(x, y, t)  K(t,
be used in querying, for example, if the language of
y, z)]);
(co4)  x  y  z (K(x, y, z)  K(y, x, z)).
Corresponding author: Beibut Sh. Kulpeshov, professor,
research field: model theory and its applications in computer
The signature of a relational structure L is a
science. E-mail: kulpesh@mail.ru.
 On the Isolation Property Over a Database Domain
non-empty set with the mapping assigning to each
relational symbol in L the relation of the same arity
over this set. Let M be an infinite structure of
signature L. Here the authors consider circularly
ordered structures. This means that L includes a ternary
relational symbol K of which the interpretation in M
satisfies to axioms of the circular order. The authors fix
the circuit of database SC and enter the following
notations:
L 0 = {K}, L’ = L 0  SC, L’’ = L  SC
A query of a database can be formally determined as
a mapping which is accepted by the state of a database
and it makes a new relation of fixed arity over M .
The authors consider two languages for querying.
Queries of the first language are formulas of signature
L’–the authors name them by limited. Queries of the
second language are formulas of signature L’’ – the
authors name them by expanded. In Ref. [7] the authors
have proved reducibility of expanded queries to limited
ones over an 0-categorical quite o-minimal domain.
Here the authors prove reducibility of expanded
queries to limited ones over a circularly ordered
domain.
Recall the necessary definitions.
Definition 2.2 A k-аry query  is locally generic
over finite states if a   if and only if ( a ) 
((s)) for every partial < -isomorphism : X  M,
where X  M, for any finite state s over X and for any
k-tuple a in X.
A state s expands an universe M of the signature
L to a L  -structure which the authors denote by
( M , s ).
Definition 2.3 A  -state for a L -structure W
is pseudo finite in W if (W , s ) is a model of the
first order L -theory of all (W , r ) where r is a
finite state over W .
A pseudo finite set is a special case of a pseudo
finite state. A signature consisting of one unary
relation and some other relations, and such systems of
this signature in which all closed formulas true in all
finite systems of the signature hold are considered.
97
Then an interpretation of this unary relation in such a
system is called a pseudo finite set.
Definition 2.4 The authors say that a complete
theory Т has the Isolation Property if there is a cardinal
 such that for any pseudo-finite set A and for any
element a of a model of the theory Т there exists A 0
 A such that |A 0 | <  and tp ( a / A 0) isolates tp ( a /
A).
Definition 2.5 Let A  M where M is a circularly
ordered structure. The set A is convex if for any
elements a, b  A the following holds: for any
element c  M such that K(a, c, b) the authors have
that c  A or for any element c  M such that
K(b, c, a) the authors have c  A.
Definition 2.6 Let A, B be convex disjoint subsets of
a circularly ordered structure M . The authors say that
A is the immediate predecessor of the set B (denote this
by A = IP(B)) or В is the immediate successor of the set
А (denote this by В = IS(А)) if for any a  A, b  B and
c  M such that K(a, c, b) the authors have that c  A 
B.
Notation 2.7 (1) Let M = M, =, <, … be a linearly
ordered structure. The authors denote by c(M) the
structure M* = M, =, K3, … in which the linear order
< is replaced by the ternary relation K3 as follows: for
every a, b, c  M K(a, b, c)  a  b  c  b  c  a 
cab
(2) Let M1, M2, …, Mm be linear orderings. If the
authors replace < by K3 as above, the authors denote
by c(M1, …, Mm) the circularly ordered sum M1, …,
Mm such that for every 1  i  m–1 Mi = IP(Mi + 1) and
Mm = IP(M1).
(3) K0(x, y, z):= K(x, y, z)  y  x  y  z x  z .
Definition 2.8 Let I  M where M is a circularly
ordered structure. The set I is called an open interval if
I = {c  M: M  K0(a, c, b)} for some elements a, b
 M.
If I is an open interval, the authors will sometimes
write I = (a, b) if the authors wish to show the
endpoints of I. Similarly, the authors can define closed,
semi-open-semi-closed and etc. intervals in M . By an
98 On the Isolation Property Over a Database Domain
interval in M the authors will understand any of the
above called types of intervals in M . Obviously, both
an interval and a singleton are convex sets.
The following notion has been originally introduced
by Dugald Macpherson and Charles Steinhorn in Ref.
[8]. They have described circularly ordered groups
that are circularly minimal.
Definition 2.9 [8] A circularly ordered structure M
=  M, =, K3, … is called circularly minimal if any
definable (with parameters) subset of the structure
M is an union of finitely many intervals and points
in M .
The notion of circular minimality is a variant of
o-minimality. O-minimality has been developed in
many works [9-13].
Example 2.10 Let G = {z  C | |z| = 1}, where C is
the set of complex numbers. Consider the structure G
= G, =, *, K3, where * is a binary operation being
the multiplication on complex numbers.
Obviously, G, * is a group. It isn’t difficult to
prove that G is circularly minimal.
Example 2.11 Let M = c( + *, Q,  + *, Q),
=, K3, where  is the order of natural numbers, * is
the inverse order on natural numbers, and Q is the
order of rational numbers.
Obviously that M is a circularly minimal
structure. Denote the first elements of these two
orderings that are equivalent to  + * by a and c,
and the last elements of these orderings by b and d.
Consider the following formula:
(x) := y  z [K0(y, x, z)   t1  t2 (K0(y, t1, x) 
K0(x, t2, z))]
Obviously that (x)  (K0(b, x, c)  K0(d, x, a)), i.e.
(M) is an union of two open intervals, but the authors
see that the endpoints of these intervals are not
definable over : a, b, c, d  acl() \ dcl(). This is
in contrast with the o-minimal case where acl(A) =
dcl(A) for all A  M.
The following notion has been introduced jointly
with D. Macpherson:
Definition 2.12 A circularly ordered structure M =
M, =, K3, … is called weakly circularly minimal if
every definable (with parameters) subset of the
structure M is an union of finitely many convex sets.
Weakly circularly minimal structures have been
originally studied in Ref. [14].
Example 2.13 Let M = c(Q1, Q2, Q3, Q4), =, K3,
P1, where Qi is a copy of the ordering of rational
numbers for each i  4, and P(M) = Q1  Q3.
It can be proved that M is a weakly circularly
minimal structure. It is easy to see that P(M) is an union
of two convex sets that are neither intervals nor points.
Therefore, M is not circularly minimal. The authors
can also observe that P(M) consists of two convex sets
each of which is not -definable. This contrasts with
the weakly o-minimal case, where every parametrically
definable set is a union of finitely many definable with
the same parameters convex sets.
3. Results
Generalizing the notion of circular minimality, the
authors call a circularly ordered structure M quasi
circularly minimal if every definable subset of the
structure is a finite union of points, intervals and
-definable sets. A complete theory T is said to be
quasi circularly minimal if every its model is quasi
circularly minimal.
There are quasi circularly minimal theories that are
not circularly minimal. The simplest example is the
theory T of densely ordered sets without endpoints
with a distinguished subset that is dense in the universe,
and its complement is also dense in the universe. It can
be showed that T is the theory of the structure
 R, , K 3 , Q is quasi circularly minimal. It is not
circularly minimal since the distinguished subset is not
a finite union of intervals and points.
The authors prove the following theorem:
Theorem 3.1 Any quasi circularly minimal theory
has the Isolation Property.
Proof Let A be a pseudo finite set in a model M
of a quasi circularly minimal theory T , and a  M .
Show that there exists A0  A such that
 On the Isolation Property Over a Database Domain
| A0 |  | T | and tp(a / A0 ) isolates tp(a / A) .
Consider an arbitrary L -formula  ( x, y ) . Since
T is quasi circularly minimal, the formula  ( x, y )
is equivalent to the formula  ( x, y ) being a
disjunction of formulas of the form
 ( x)   ( y )   ( x, y ) where  ( x, y ) has one of
the following forms for some terms t and t  in
variables of the tuple y : x  x, x  t, x  t,
K 0 (t , x, t ) where
K 0 ( x, y, z ) :  K ( x, y, z )  y  x  y  z  x  z .
Let S be the set of all terms t containing in
 ( x, y ) on all disjuncts of the formula  ( x, y ) .
Obviously, S is finite.
Let F be an arbitrary finite subset of the structure
M . Then the set D :  {t | t  t (c ) for some t  S
and c  F } is finite, i.e. there exists s   such
that |D|  s and D  {t1 ,..., t s } with
K 0 (t1 , t 2 ,..., t s } .
If D   then  ( x, y ) contains only the
formula x  x . If | D |  1 then  ( x, y ) contains
the formula x  t or x  t . If | D |  2 then for
any d  M there exists i  s such that
M | K (t1 ,..., t i , d , t i 1 ,..., t s ) . Then the authors
introduce the following notations:
m ( F , d ) :  ti , m ( F , d ) : ti 1
Since the set A is pseudo finite in M , then the
same is true for A (that is for F ), i.e. there exist
m  m ( A, a) and m  m ( A, a) .
 
Consequently, there are t , t  S , c , c  A
  
such that m  t (c ), m  t (c ) .

Clearly that m  m if and only if a  t (c ) for

some t ( y )  S and c  A . If m  m then
M | K 0 (m , a, m ) . If  ( x, y ) contains the
formula x  x then a  A , i.e. tp ( a / A) is
algebraic and as the set A0 the authors can take {a} .
If  ( x, y ) contains the formula x  t then
a  t (c ) for some t  S and c  A and as the set
A0 the authors can take {c } . If each from  ( x, y )
contains only one formula x  t (c ) for some t  S
and c  A then as the set A0 the authors can also
99
take {c } .
Let now the authors have K 0 (m , a, m ) .
Suppose that the formula  ( x, c )  tp (a / A) for
some c  A , i.e. M |  (a, c ) , whereas
 ( x)  tp (a / ) ,
 ( y )  tp ( A) , where tp( A) is the set of all
L( A) -propositions that are true in M . Show that
tp ( A)  tp (a / ) | K 0 (m , x, m )   ( x, c )
It is sufficiently to show that
tp ( A) | K 0 (m , x, m )   ( x, c )
The formula  ( x, y ) has the following form
K 0 (t ( y ), x, t ( y )) . Consequently,
M | K 0 (t (c ), a, t (c )) . Then by the definition of
t (c ) and t  (c  ) the authors have:
M | K (t (c ), t (c ), t  (c  ), t (c )) , i.e.
tp ( A) | K 0 (m , x, m )  K 0 (t (c ), x, t (c ))
Then as the set A0 the authors take the set of all
c , c  .
Corollary 3.2 Let Т be a quasi circularly minimal
theory. Then every extended query being locally
generic over finite states is equivalent to a limited
query.
4. Conclusions
Thus, the authors conclude that if the authors
consider a quasi circularly minimal domain of
databases then locally generic queries in an expanded
language can be expressed by queries of databases in
limited language.
The condition of quasi circular minimality is
essential for holding the Isolation Property. For
example, if the authors consider the structures of the
3
form ( M , K 3 , E ) where ( M , K ) is a dense
circularly ordered set without endpoints, and E is an
equivalence relation on M with infinitely
many classes all of which are dense, the authors
see that the theory of these structures is not quasi
circularly minimal, and it does not have the Isolation
Property.
100 On the Isolation Property Over a Database Domain
Acknowledgments
This work was supported by Kazakhstan grant
“Development and studying logical methods of
modeling a behavior of complex systems” (priority
“Information and telecommunication technologies”;
sub priority “Information technologies”) for
2012-2014 years.
References
[1] E.F. Codd, A relational model for large shared data banks,
Communications of ACM 13 (6) (1970) 377-387.
[2] E.F. Codd, Relational completeness of database
sublanguages, Database systems, Prentice-Hall (1972)
33-64.
[3] M. Benedikt, G. Dong, L. Libkin, L. Wong, Relational
expressive power of constraint query languages, Journal
of ACM 45 (1) (1998) 1-34.
[4] O.V. Belegradek, A.P. Stolboushkin, and M.A. Taitslin,
Extended order-generic queries, Annals of Pure and
Applied Logic 97 (1999) 85-125.
[5] M.A. Taitslin, A general condition for collapse results,
Annals of Pure and Applied Logic 113 (1-3) (2002)
323-330.
[6] S.M. Dudakov, M. A. Taitslin, Translation results for
query languages in theory of databases, Uspekhi
Mathematicheskich Nauk 61 (2) (2006) 3-66. (in Russian).
[7] B.Sh. Kulpeshov, On problem of expressiveness of
database queries, International Journal of Mathematics,
Computer Sciences and Information Technology 3 (2)
(2010) 123-128.
[8] H.D. Macpherson, Ch. Steinhorn, On variants of
o-minimality, Annals of Pure and Applied Logic 79 (1996)
165-209.
[9] A. Pillay, Ch. Steinhorn, Definable sets in ordered
structures I, Transactions of the American Mathematical
Society 295 (1986) 565-592.
[10] H.D. Macpherson, D. Marker, Ch. Steinhorn, Weakly
o-minimal structures and real closed fields, Transactions
of the American Mathematical Society 352 (2000)
5435-5483.
[11] B.Sh. Kulpeshov, Weakly o-minimal structures and some
of their properties, The Journal of Symbolic Logic 63
(1998) 1511-1528.
[12] B. Herwig, H.D. Macpherson, G. Martin, A. Nurtazin, and
J.K. Truss, On 0-categorical weakly o-minimal
structures, The Annals of Pure and Applied Logic 101
(2000) 65-93.
[13] B.S. Baizhanov, Expansion of a model of a weakly
o-minimal theory by a family of unary predicates, The
Journal of Symbolic Logic 66 (2001) 1382-1414.
[14] B.Sh. Kulpeshov, H.D. Macpherson, Minimality
conditions on circularly ordered structures, Mathematical
Logic Quarterly 51 (2005) 377-399.
Journal of Mathematics and System Science 3 (2013) 101-109 D DAVID PUBLISHING

A Parallel Finite Element Algorithm for Navier-Stokes

Equations

Mohamed Abdelwahed
Department of Mathematics, College of Science, King Saud University, Riyadh 11451, Kingdom of Saudi Arabia

Received: January 21, 2013 / Accepted: February 11, 2013 / Published: February 25, 2013.

Abstract: The authors propose a numerical algorithm for the two-dimensional Navier-Stokes equations written in stream
function-vorticity formulation. The total time derivative term is treated with a first order characteristics method. The space
approximation is based on a piecewise continuous finite element method. The proposed algorithm is used to simulate the mechanical
aeration process in lakes. Such process is used to combat the degradation of the water quality due to the eutrophication phenomena. For
this application high computing facilities and capacities are required. In order to optimize the computing time and make possible the
simulation of real applications, the authors propose a parallel implementation of the numerical algorithm. The parallelization technique
is performed using the Message Passing Interface. The efficiency of the proposed numerical algorithm is illustrated by some numerical
results.

Key words: Navier-Stokes equations, stream function, vorticity, high performance computing.

1. Introduction eutrophication phenomena [4, 5]. It consists in

pumping a source of compressed air into the reservoir
Over the past several decades, computational
bottom via injectors in order to create a dynamic, and
methods of solving the Navier-Stokes equations for
aerate the water by bringing it into contact with the
fluid flow problems have proved to be very useful for
surface air (Fig. 1).
better understanding the physics of various flows [1-3].
Using the fact that the water phase is dominant, the
The increase in computer power and the advent of
flow can be described by a simplified model based on
inexpensive parallel computing have driven the
the Navier-Stokes equations. The aeration effects are
advances in algorithm development, allowing for more
taken into account through local boundary conditions
and more flows of increased complexity to be
for the velocity [6]. Then, the water velocity u(x,t) and
investigated. Motions and regions contain large
the pressure p(x,t) solve the following system:
gradients, such as shocks and interfaces, near-wall
 u
regions of complex boundaries, and transitional and  t  uu  vu  p  f     in   [0,T ]
turbulent flows, remain a humbling limitation of these 
computational methods. divu = 0                                     in   [0,T ] (1)
 u = u                                          in   [0,T ]
In this paper the authors deal with a problem  d

requiring a high performance processor for practical 

implementations. It is the mechanical aeration process
in the eutrophication lakes. Such a process is used to
combat the water quality degradation caused by the
Corresponding author: Mohamed Abdelwahed, Ph.D.,
associate professor, research fields: numerical analysis and
optimization. E-mail: mabdelwahed@ksu.edu.sa.
where Ω is the lake domain, v is the fluid viscosity, f is
the gravitational force, T is the final time of simulation,
Γ = Γs ∪ Γf ∪ Γinj the boundary of Ω (Fig. 2) and ud is
a given boundary velocity equal to uwind on Γs (the
surface in contact with atmosphere), 0 on Γf (the
bottom like boundary) and uinj on Γinj (the injector
102 A Parallel Finite Element Algorithm for Navier-Stokes Equations
wind
air bubbles
injector
Fig. 1 Aeration process.
Γs
Ω   =  d                                       on   [0,T ]
Γinj
Γf
Fig. 2 Domain.
boundary). Where uwind is the wind velocity and uinj is
the injector velocity.
For such application high computing facilities and
capacities are required [7]. In fact, let us consider, for
example, only a real two dimensional simulation. The
domain is a two dimensional cross section of a lake
whose average dimensions can reach 500 m × 20 m,
the injector length is about 10m with more than 100
holes of 1 cm for its diameter. In order to detect the
bubbles effect, it is necessary to make a mesh around
the injector area with elements having a size of the
order of bubbles volume: this should generate a mesh
of the order of 10 million nodes. To have an idea of the
difficulty, a simulation for T = 10 mn with a mesh of
5.105 nodes, requires approximately two days on a
good computer.
In order to reduce the computing time and make
possible the numerical simulation of the dynamic
aeration process with more fine meshes, the authors
propose two alternatives.
The first one consists in reducing the number of
equations by using the the steam function―vorticity
formulation [8-10]. The stream function ψ and the
vorticity ω are linked to the velocity field u by the
relations:
   u  and  u   (2)
The    formulation is more suitable compared
with primitive variables (velocity and pressure) for
representing the kinematic nature of the flow. Such a
formulation has two main advantages. The first one is
related to the automatic satisfaction of the divergence
free condition. The second one concerns the reduction
of the equations number. Using Eq. (2), from Eq. (1)
the authors obtain the following system:
 
 t  u   v     f     in    [0,T ]

     = 0                               in    [0,T ] (3)

 
  g                                     on   [0,T ]
 n
Here the boundary data ψd and g are calculated using
ud and the relations Eq. (2).
The second alternative concerns the reduction of the
computing time through the parallel implementation of
the numerical algorithm using MPI (Message Passing
Interface) [11].
The paper is organized as follows. In Section 2, the
authors introduce the numerical analysis of the used
model. The time discretization scheme is presented in
Section 2.1. The used finite element method for the
space approximation is described in Section 2.2. In
Section 3, the authors illustrate the parallel algorithm
and the authors present some numerical results
showing the efficiency of the proposed method. In
Section 4 the authors deal with a real application. The
authors treat the aeration process using the obtained
parallelized finite element algorithm. The paper ends
with some concluding comments in Section 5.
2. Numerical Analysis
The authors present in this section a numerical
analysis of the system Eq. (4). The time discretization
scheme is based on the characteristics method [12, 13].
The space approximation is performed using the
regularization-stabilization technique introduced in
Refs. [14, 15] for the biharmonic problem.
2.1 Time Discretization
Let ψn+1 and ωn+1 being the approximations of ψ and
ω at time tn+1 = (n + 1)  t , where  t is the time step.
The time discretization of the problem Eq. (4) leads to
 A Parallel Finite Element Algorithm for Navier-Stokes Equations 103

finding ψn+1 and ωn+1 solutions to 1  n oX n 

F n 1     f n 1   (8)
  
n 1
v t 
 +u   v n 1    f n 1    in 
 t  The system Eq. (7) is computed iteratively in time.
 n 1
 +  = 0                                     in  (4)
n 1 At each time step, the authors have to solve a steady
  n 1
state problem of quasi-Stokes type having the
 n 1 =  d，
     g                              on  following generic form:
 n

The authors use the characteristics scheme [12, 13]     0                in 
which consists in approximating the total derivative of 
    = F        in  (9)
the function ω as follow:  
   =  d，
     g      on 
 +u   ( x , t )
n 1
 n
 t  where ψd, g, λ and F are given.
 ( x, t )   ( X ( x, t n 1 ; t n ), t n )
n 1
 (5) 2.2 Finite Element Approximation
t
   oX
n 1 n n
The variational formulation of problem Eq. (9) is:
  
t Given ( d , g  H 3/2 ()  H 1/2 () and
where Xn(x) = X(x, tn+1; tn) is the position at time tn of F  L2 () , find ( , g  H 1 (, )  H 1 ()
particle of fluid which is at point x at time tn+1. solutions to
  d    d   g d 
   
The system Eq. (4) is then equivalent to the
resolution of the following system :
  H (, )
1
find ψn+1 and ωn+1 solutions to 
  n 1 n 1  oX
n n
   ,   1,1     d  (10)
 t  v  n 1
   f +       in  
 t
 F ,   1,1   H 0 ()
1
 n 1
 +  = 0                                    in  (6)
n 1

 such that    d  on 
 n 1 =  d，  n 1
     g                             on  where
 n
H 1 (, )  {  L2 (),   H 1 ()}
 n 1  n 1 1
with H () is the dual space of H 0 () and
1
Replacing the term by  in the
t t <.,.>-1,1 is the duality product between H 1 () and
H 01 () .
first equation of Eq. (6), the authors obtain the
The existence and uniqueness results for the solution
following equivalent problem:
of Eq. (10) are detailed in Ref. [16]. For the spatial
find ψn+1 and ωn+1 solutions to:
approximation, the authors used the C0 optimal finite

 n 1   n 1  0                          in  element method [15].
 Let (Th)h be a regular family of decomposition of 
 1
   n 1 = F n 1       in 
n 1
(7) in triangles K. For each triangle K∈(Th)h, the authors
 v  t denote by hK its diameter and mesK its area. The
 n 1  n 1 authors set h  max K Th hK .
  = 
     
d，  g                 on 
 n The authors associate to each decomposition Th, the
where set Ch of the internal edges. Then, for every edge T of
104 A Parallel Finite Element Algorithm for Navier-Stokes Equations

Ch, there exist two triangles K and K’ of Th such that K represents the jump of the normal derivative on each
edge T of Ch, where vK = v|K and  n vK is the normal
K
≠ K’ and T = ∂K ∩ ∂K’.
The authors consider the discrete spaces: derivative of vK on ∂K.
Yh  {  C 0 (); K  Th ,  |K  P1 ( K )} The authors obtain the following discrete
formulation:
Yh0  Yh  H 01 ()
a (h ,  h )  b( h , h )  g h d   
Where P1(K) is the space of piecewise linear functions
 
defined on K . ( h )  Yh
The authors denote by ωh, ψh and Fh respectively the 
approximations of ω, ψ and F in Yh. The classical b(h ,h )  d ( h ,h )   Fh h d  (13)

 h  Yh
0
discrete variational formulation of the problem Eq. (10)
is written as: 
 h  d  Yh ,    h  Yh
0

 a (h ,  h )  b( h , h )   g h d         h  Yh

  This problem is known well posed, for more detail
 b(h , h )  d ( h , h )   Fh h d    h  Y (11)
0
one can consult Refs. [16].
h

 The obtained form ah is then uniformly coercive on

 h   d  Y ,  h  Yh
0
h
the discrete space Yh. The discrete formulation remains
where a, b and d are the following three bilinear forms: consistent and leads to a stabilized finite element
a( h ,  h )    h h d ,             ( h ,  h )  Yh  Yh method. Using this stabilization-regularization

technique, one can obtain with the element P1, an
b( h , h )     h  h d    ( h ,  h )  Yh  Yh unconditionally convergent method. The theoretical


d ( h ,h )    h h d   ( h ,h )  Yh  Yh0 results concerning this technique are described in Ref.

[15] for the bi-harmonic problem and in Ref. [16] for
The authors remark that the compatibility constant
0 the quasi-Stokes equations.
between spaces Yh and Yh is independent on h but the
associated bilinear form a(.,.) to Yh is elliptic for the 2.3. Linear System
norm of L2(Ω) and consequently the coercivity Let n the number of mesh nodes, (θi) i = 1, n be a base
constant depends on h. To make up for such 0
of Yh and (ηi) i = 1, n a base of Yh . The system Eq. (13)
inconvenient, the authors use an approximation method is written for all j = 1, …, n:
based on a technique of stabilization described in Refs.  n n

[8, 14], which consists in changing this principal form  i i j

 a ( ,  )    i Ah (i , j )
 i 1 i 1
a(.,.) into another form ah(.,.) by adding a stabilizing  n

term βAh:   i b(i , j )  g j d 

 
ah ( h ,  h a( h ,  h  Ah ( h ,  h (12)  i 1 (14)
 n n

where β is a positive real parameter and Ah is the  i b(i , j )   i d (i , j )

discrete bilinear form defined by:  i 1 i 1

 F  d 
Ah ( h ,  h )   mes(T ) [  h h ][ h h ]dT     h j
T
T Ch which is equivalent to the following linear system :
( h ,  h )  Yh  Yh MX = N (15)
with where

 n vK   n vK '  if  T  Ch , K , K '  Th  A   S    BT   

K K'
G 
[ h v]T   M   ,  X   and  N  
0                      if  T  K  , K  Th     B           D    F
 A Parallel Finite Element Algorithm for Navier-Stokes Equations
*The matrix A = (ai, j)1 ≤ i, j ≤n with ai, j = a (θi, θj) is
computed from the bilinear form a(.,.).
*The matrix S = (si, j)1 ≤ i, j ≤n with si, j = Ah (θi, θj)
represents the bilinear form Ah (.,.).
* The matrix B = (bi, j)1 ≤ i, j ≤n
with bi, j = b(ηi, ηj).
* D = -λB.
* G and F are associated to the second member with
G j   g j d  and Fj   Fh j d  .
 
For the storage of these matrices, the authors used
the Morse compact method keeping only non-zero
terms. For the resolution, the authors used the
BICGSTAB (bi-conjugate gradient stabilized)
algorithm [17.
3. Parallelization
The goal of the parallel implementation is to make
possible the simulation of real application cases,
needing high performance computing, in reasonable
computing time. The authors can ideally hope to divide
the real time (Elapsed time) by the number of used
processors. The authors present in the following a
parallelization of the solver. The implementation was
done with Fortran and MPI library for communications
[11].
3.1 Algorithmic Analysis
Some carried out numerical tests on one-processor
machine showed that the resolution part of the linear
system Eq. (15) represents the major part of the total
CPU (computing) time of the code. Consequently the
authors focused on the parallel treatment of this part of
the solver which uses the BICGSTAB algorithm. In
fact at each iteration step, the following operations are
down.
3.1.1 Matrix-Vector Product
This operation is the most significant one in term of
CPU time; then naturally the authors are going to share
the different tasks on different processors in order to
minimize computing time. The strategy used consists
105
to distribute the lines of the matrix between the
processors. Each processor contains the total vector X
and a block Mi of lines of the matrix (Fig. 3).
Each processor carries out locally the product of a
block Mi (compound of a some number of lines of the
matrix M) and of the vector X. The result is a vector
(MX)i of dimension the number of lines of the block Mi.
After that, a synchronous communication is used for
the update of the total vectors after each product
matrix-vector. No communications are needed in the
local computing of each processor.
3.1.2 Scalar Product
Each processor computes a scalar partial product of
two corresponding local vectors by multiplying these
two vectors and by adding the components of the
partial result. Then each processor sends to all
processors its partial result and receives from each the
partial resulting product which is added to its own
result. Consequently, a total communication is
necessary in this scalar product.
3.1.3 Linear Operations
Each processor contains one part of each total vector
considered as a local vector. It is possible to make a
linear combination of two local vectors; the obtained
vector result is a local vector. Consequently, no
communication is needed for this operation.
3.2 Overhead Time
The most important part of overhead is generally
resulting from the exchanges of messages between
processors (send/receive). For that, one starts with the
distribution of all the necessary data to be used by
M X MX
M1 MX 1
M2 MX 2
M3 MX 3
Mi MX i
Mn MX n
Fig. 3 Matrix-vector product.
106 A Parallel Finite Element Algorithm for Navier-Stokes Equations

processors before beginning any calculation. Each Table 1 Meshes.

processor will receive a package of data containing the Mesh nodes elements unknowns
useful part of the matrix M and the vector X. All the 1 2304 4418 4608
2 9216 18050 18432
exchanges carried out after correspond only to the
3 36864 72962 73728
update of the directions vectors in the BICGSTAB 4 147456 293378 294912
algorithm which means also that the registers addresses
2200
are remained unchanged, this fact will increases the 1 Proc
2 Procs
2000
performance. 4 Procs
8 Procs
1800 16 Procs

3.3 Performance Analysis 1600

The authors present in this section some numerical

1400

Execution time (s)

results in order to show the performance of the 1200

parallelized solver and the importance of the 1000

parallelization for such application. The simulations 800

are performed on a HP V class computer which is a 600

MIMD (multiple instructions multiple data machines) 400

of 16 processors with a shared memory cross bar 200

architecture. 0
1 2 3 4

The parallel performance has been investigated Mesh

Fig. 4 Elapsed time.

through different meshes described in the Table 1, and
combined with partitions of 2, 4, 8 and 16 processors,
respectively
The authors denote that passing from the mesh i to
the mesh i + 1 (i = 1, …, 3), the authors resolved a
problem with four times more data. The results will be
analyzed through the two following types of curves:
* Curve of execution time which informs us on the
evolution of the elapsed time (computing and
communication times) of the program according to the
number of processors.
* Curves of speed-up and efficiency which permit to
evaluate the parallel algorithm program and the
influence of the communication time between
processors.
Let’s note that all measures presented in this section
are measured for one iteration time and are given by the
average of five tests at least.
3.3.1 Execution time
The authors show in Fig. 4 the simulation time for
the four meshes according to the number of processors.
The authors deduce that the execution time increases
when the authors raise the number of unknowns. The
authors mention also that, for the mesh 1, the execution
time on 16 processors (0.37 s) is near to the one on 8
processors (0.41 s). The authors explain this by the fact
that for this case, the number of operations done by
every processor is too small in comparison with the
communication time spent for exchanging information
between the 16 processors.
Thus, it is necessary to raise the number of
unknowns when the authors increase the number of
processors so that the execution time dominates the
communication time. In fact, for the mesh 2, the
execution time on 8 processors is 4.02 s whereas on 16
processors it is 2.07 s.
Finally the authors conclude that more the mesh is
finer or more the number of unknown is important,
more the gain of time becomes significant. This is
proportional to the treated case and is often polluted by
the communication time between processors which
becomes more significant when the number of
processors increases.
 A Parallel Finite Element Algorithm for Navier-Stokes Equations 107

3.3.2 Speed-up and Efficiency 16

2 Procs
4 Procs
The authors succeed to reach good speed-up. Indeed, 14 8 Procs
16 Procs

the authors show in Fig. 5 that the speed-up rises with

12
the mesh parameter. Particularly, for the mesh 4, the
speed-up (1.99) reaches its optimal value Eq. (2) on 10

Speed−up
two processors. Concerning speed-up gotten with four 8

processors, the authors notice that for a small mesh

6
(mesh 1), the authors cannot reach a good speed-up
(2.95). Indeed the calculation volume is less than the 4

communication volume between the four processors. 2

Therefore, the communication cost degrades the

0
performance of the parallel algorithm. But, when the 1 2
Mesh
3 4

authors increase the number of unknowns (mesh 4), the Fig. 5 Speed-up.
authors can reach speed-up near to its optimal value on 100

four processors (3.85). 90

After the study about the speed-up, the authors

80 11
00
00
11
2 Procs
11
00
00
11 00 4 Procs
11
analyze the efficiency. The histogram of Fig. 6 shows
70
00
11 00
11
00 8 Procs
11
11
00 00
11 00
11
60
00
11 00
11
00
11 00 10 16 Procs
11
that when the authors fix the number of processors and 00
11
Efficiency

00
11 00
11 00
11
00
11
00
11
50
00
11
00
11 00
11 00
11
vary the mesh, the efficiency raises. Indeed, when the 40
00
11 00
11 00
11
00
11 00
11 00
11
00
11
authors raise the problem size, the speed-up improves 11
00 00
11 00
11 00
11
00
11
30
00
11 00
11 00
11 00
11
00
11
00
11 00
11 00
11 00
11
whatever is the number of processors, what entails an 00
11
20

00
11 00
11 00
11 00
11
00
11 00
11 00
11 00
11
00
11
00
11
10
improvement of the efficiency. But, when the authors 00
11
00
11 00
11 00
11 00
11
0 00
11 00
11 00
11
fix the mesh (for example the mesh 1) and continue
1 2 3 4
Mesh

increasing the number of processors, the efficiency is
reduced. Indeed, if the authors increase the number of
processors, the communication cost rises. Therefore,
the speed-up and efficiency decrease. Thus, the
increase of the number of processors requires an
increase of the problem size in order to guarantee the
efficiency of the parallelization, which is called
scalability.
4. Numerical Simulation of the Aeration
Process
This section is devoted to a real application. The
authors give a numerical simulation of the aeration
process using the obtained parallelized stabilized finite
element algorithm.
The authors present in this section a numerical
simulation of the dynamic created using the parallel
solver. The authors consider a two dimensional cross
Fig. 6 Efficiency.
section of the lake containing one injector. The
computational domain is a two dimensional domain
with length 150 m and maximum height 25 m
containing one injector with 5m length. The used mesh
is about 497,635 nodes. The authors used the following
boundary conditions: the no slip condition on the
bottom, uwind = (0.01m/s, 0,0), and uinj = (0,0,0.1
m/s).The numerical simulation is done using 16
processors for T = 10 min. The execution time is about
4.26 h. The authors remark that the execution time
using one processor is about two days.
Fig. 8 shows a zoom near the injector (see Fig. 7) of
the velocity field due to the wind effect. The authors
remark that this effect concerns only the top of the
domain which is not sufficient to aerate the lake
domain. The authors present in Fig. 9 a zoom near the
injector of the created dynamic by the mechanical
108 A Parallel Finite Element Algorithm for Navier-Stokes Equations

Fig. 7 Zoom near the injector.

After 1 min

Fig. 8 The initial motion (the wind effect).

aeration process for different time steps. The generated

vertical motion mixes the water of the bottom with that
in the top, thus oxygenating the lower part by bringing
it in contact with the surface air.

5. Conclusions
In this work, the authors have proposed a parallel
After 4 min
algorithm for the two dimensional Navier-Stokes
equations written in stream function-vorticity
formulation. In order to optimize the computing time,
the authors have proposed a parallel implementation
for the obtained solver. The parallelization process is
performed using MPI. The obtained solver is used to
simulate the mechanical aeration process in an
eutrophic lake. The proposed approach will be
extended in a forthcoming work to a three dimensional
model taking into account the two-phases air bubbles
and water. The authors denote finally that the proposed
algorithm is general and can be adapted for many other
important engineering problems related to the
Navier-Stokes equations. After 7 min
 A Parallel Finite Element Algorithm for Navier-Stokes Equations
After 10 min
Fig. 9 The created motion (the aeration effect).
References
[1] M. Abdelwahed, M. Hassine, M. Masmoudi, Control of
mechanical aeration via topological sensitivity analysis,
Journal of computational and applied Mathematics 228
(2009) 480-485.
[2] M.D. Gunzburger, Perspectives in flow control and
optimization, Advances in Design and Control, SIAM,
Philadelphia, 2002.
[3] B. Mohammadi, O. Pironneau, Applied shape
optimization for fluids, Numerical Mathematics and
Scientific Computation, Oxford University Press, New
York, 2001.
[4] Z. Gurkan, J. Zhang, S.E. Jorgensen, Development of a
structurally dynamic model for forecasting the effects of
restoration of lake Fire, Ecological modeling 197 (2006)
89-102.
[5] R. Lawson, M.A. Anderson, Stratification and mixing in
Lake Elsinore, California: An assessment of axial flow
pumps for improving water quality in a shallow eutrophic
lake, Water research 41 (2007) 4457-4467.
109
[6] F. Dabaghi, Numerical Aspects for aeration process
modeling in eutrophised water basins, Journal of system
analysis modeling simulation 30 (2000) 1-23.
[7] M. Abdelwahed, M. Amara, Numerical analysis of a two
phase flow model, Journal of computational Methods 9 (3)
(2012) 28.
[8] M. Amara, C. Bernardi, Convergence of a finite element
discretization of the Navier-Stokes equations in vorticity
and stream function formulation, Mathematical Modeling
and Numerical analysis 33 (5) (1999) 1033-1056.
[9] C. Bernardi, V. Girault, Y. Maday, Mixed spectral element
approximation of the Navier Stokes equations in the
stream function and vorticity formulation, IMA Journal of
Numerical Analysis 12 (1992) 565-608.
[10] V. Girault, J.Giroire, A. Sequeira, A stream
function-vorticity variational formulation for the exterior
Stokes problem in weighted Sobolev spaces, Math. Meth.
in the Applied Sciences 15 (5) (1992) 345-363.
[11] HP MPI User’s Guide, Fourth edition, Helwet Packard,
1999.
[12] J. Douglas, T.F. Russel, Numerical methods for
convection dominated diffusion problems based on
combining the method of characteristics with finite
element methods or finite difference method, SIAM, J.
Numeric. Anal. 19 (1982) 871-885
[13] O. Pironneau, On the transport diffusion algorithm and its
applications to the Navier-Stokes equations. Numer Math.
38 (1982) 309–332.
[14] M. Amara, Une méthode optimale de classe C0
d’approximation du bilaplacien, Comptes Rendus de
l’Academie des Sciences 319 (1994) 1327-1330.
[15] M. Amara, F. Dabaghi, An optimal C0 finite element
method for the 2D biharmonic problem. Numerisch
Mathematik 90(1) (2001) 19–46.
[16] M. Amara, F. Dabaghi, M. Hassine, Une formulation
mixte utilisant un élément fini C0 optimal pour la
résolution des équations de Quasi-Stokes incompressible :
Analyse théorique et validation numérique. Research
report 4142: INRIA, 2001.
[17] Y. Saad, Iterative methods for sparse linear systems. PWS
series in computer science, 1996.
Journal of Mathematics and System Science 3 (2013) 110-113 D DAVID PUBLISHING

Calculation of Infinite Series Through Polygamma

Functions’ Algorithms and Laplace Transforms
Héctor Luna-García1, Luz María García Cruz1 and R. Mares2
1. Área de Física Atómica Molecular Aplicada, CBI,UAM-A, Azcapotzalco, México 02200, México
2. Departamento de Física, ESFM, IPN, Edificio 9. U.P. Zacatenco, México 07738, México

Received: January 22, 2012 / Accepted: February 12, 2013 / Published: February 25, 2013.

Abstract: Here is introduced some novel algorithms which made use of polygamma functions to get the exact limits of a broad class of
infinite series. Moreover, Laplace transform is used to find the sum of many convergent infinite series. These exact limits are found in
different branches of physics for some special cases series and are in complete agreement with the values found by other authors.
Moreover, the methods presented here are generalized and applied to other wide variety of sums, including alternating series. Finally,
these methods are simple and quite powerful to calculate the limits of many convergent series as you can see from the examples
included.

Keywords: Polygammafunction, Laplace transform infinite series.

1. Introduction criteria to a series does not provide its limit. In this

Infinite series appear in most fields of physics such article, the authors present some algorithms to calculate

the exact limit of a convergent series through Eulerian
1
as the series:   n  a  n  b 
n 1
mainly used in the functions applied to a broad class of infinite series. This
work is divided into five parts: This introduction,
theory of quantized fields theory at finite temperature Section 2 illustrates the method for some special cases;
in Section 3 the method is generalized; Section 4 shows
 1
n

[1] and the series   2n  1 which is very useful to the Laplace transform used to get the infinite series
n 1 sum; Finally, Section 5 presents conclusions and future
the determine the neutrons average distance generated projects.
under isotropy. Thus, infinite series are an important
2. Special Cases of Infinite Series
part of fundamental analysis in course and applications.

1

 1
n
2.1 The Infinite Series   n  a  n  b 
The series   2n  1
n 1
is the Leibnitz definition for n 1

This series is denoted as:

the irrational π number. To treat the infinite series, in 
1
classical literature, it is usual to introduce some S  c, d    (1)
n 1  n  c  n  d 
convergence criteriadue to D’Alembert, Cauchy,
Where c and d are not equal to negative integer
Raabe, Kummer and Gauss authors [2], which ensure
and c  d . The S  c, d  term is decomposed into
the convergence of infinite series or in worse cases the
partial fractions to obtain:
divergence. However, the application of a convergence 
 1 1  1
S  c, d       (2)
Corresponding author: Héctor Luna-García, Ph.D., n 1  n  d n  c  c  d 
professor, research fields: molecular physics and applied
mathematics. E-mail: lghm@correo.azc.uam.mx. and using the Feynman identity
 Calculation of Infinite Series Through Polygamma Functions’ Algorithms and Laplace Transforms 111

  0
1 1 z c ln z
A 0
 e  Ax dx (3)   dz
n  c 1 z
2
n 1 1

Where A  0 , the infinite series in Eq. (2) becomes      c  1

1
(9)
 
1
S  c, d     e nx  e dx  e cx  dx (4) where 
1
 .
c  d n 1 0
The Eq. (9) is now expressed in terms of  n  z 
Introducing the sum into the integral in Eq. (4), one has
polygamma function [3].
S  c, d  
1

e  dx
 e cx  e x Applying Eq. (9) the already reported results are
cd 
0 c  d 

obtained:
2
 e  e  dx

1
  
x  Nx 1

  1 
lim  (5) 2
6
N 
1  e 
n 1
x

n 1 n

1 2
Considering now N   , the result is 
n 1  n  1
2
   2 
1
6
1

S  c, d   

e  dx
 e cx  e x
dx (6) This first series appears in the black body theory [4].
0  c  d  1  e x 2.3 Generalization of Infinite Series Limits
x
Changing z  e variables and using the Eulerian
To generalize the infinite series so far treated, let us
functions properties [3], the Eq. (6) finally is expressed

1
as: consider the infinite series  , and apply
n  a
1
m
1 z z d c n 1
S  c, d    dz
c  d  0 1 z the most general Feynman identity
  d  1   c  1 1 1


 m  1! 0
  x m 1e Ax dx (10)
c  d  A m

  d    c   c  d  And following the same procedure of Section 2.1,

  2 2 (7)
c  d  cd the series can be written as:
where   z  is the called digamma function. 
1  1   m1 a  1
m2

Eq. (7) is the most general expression result of this


n 1  n  a 
m

 m  1!
  (11)

section. As an application, the authors present the

following examples: 3. Polynomial Infiniteseries

1 1
S 1, 2     Let us consider the following infinite series, where
n 1  n  1 n  2  2 QN 1  n  y PN  n  are polynomials depending on n
 1  1
S  0,     4 1  ln 2  of degree (N-1) and N, respectively:
 2  n 1 n n  1  2  

QN 1  n 
(12)
2.2 Series of the section 2.1 type, for the case c  d n 1 PN  n 

The polynomial PN  n  is defined as:

The 1 2 Feynman identity applied to this case is:
A k
PN  n     n  ai 
mi
 (13)
1
A2 0
 xe  Ax dx (8) i 1

where ai  i  1, 2,..., k  are distinct, real and

Following the same procedure as was used in section different to any negative integers. Then, for every
2.1, leads to the following result: polynomial
112 Calculation of Infinite Series Through Polygamma Functions’ Algorithms and Laplace Transforms

QN 1  n   c0  c1n  ...  cN 1n N 1 (14) commute. Now, let us present two applications of the
the infinite series is expressed inpartial fractions as Laplace transform for the evaluations of some infinite
 k Aj series.
S  a0 ,..., ak ; c0 ,..., cN 1    (15) For the first example, the following expression is
n  a 
mi
n 1 j 1
j considered
where A j are the coefficients in the fractions. s  ae at  be bt 
 L
And applying the Feynman identity done in Eq. (10),  s  a  s  b   a  b 
the series can be written as: Then, the infinite series associated with this
S  a0 ,..., ak ; c0 ,..., cN 1  expression is given by
 
 nx  ae 
 ax
 k Aj   (16) 
n  be bx
   x
m j 1  n  a j x
e dx   n   a  b
e  dx
n 1 j 1 m j  1 0 n 1  n  a  n  b  0  
Following similar procedures to those already made, which becomes
1

n 1 at a  bt b
it finally reaches the value of the infinite series in terms
  n  a  n  b   a  b  1 t
dt
of polygamma function. n 1 0

S  a0 ,..., ak ; c0 ,..., cN 1  where t  e  x , the variable change is used. Applying

the digamma function properties [3], the infinite series
 1
mj

 m j 1 (17)
a  1
k
   Aj limit is
n 1 j 1 m  1 !
j

n b  b  1  a  a  1

j

Applying Eq. (17) and the properties of the Eulerian n 1  n  a  n  b  a b
functions [3], the authors have b  b   a  a  1
  (20)
 a b ab
 8n  18n  1  2  16  

1 1
2 1 For the second example, the following expression of
n 1
the infinite series is used
This series is found in Ref [5]. 
 k 

As another example, the authors find n  a
k
n 1

1 1
  2n  1 2n  1  2
n 1
Since, the term in the summation is the Laplace
transform of t k 1e  at , applying the method directly,
the results is
4. Infinite Series and Laplace Transform 
 k   

    e  nt t k 1e  at dt
The Feynman identities given by Eqs. (3), (8) and n  a
k
n 1 n 1 0
(10) correspond to the Laplace transforms of 1, x, xm-1,
x a  ln x 
1 k 1
f  s  is the Laplace   1
k 1
respectively. In general, if
transform of g  x  , one has

0
1 x
dx (21)


where x  e  t , variable change is used.
f  s    e  sx g  x  dx (18)
Finally, the limit of infinite series is
 k 
0

Where the sum  f  n  can be written as    1    a  1
k k 1
(22)
n  a
k
n n 1

 
 f  n    g  x    e  nx  dx
 n 
(19) To finish the algorithms exposure, it looks important
to give some information on an alternating series
0

considering that the addition and integration operations example, which have not yet been considered in this
 Calculation of Infinite Series Through Polygamma Functions’ Algorithms and Laplace Transforms 113

work. Due to these series are important in physics, they applications. The algorithms developed are also useful
will be extensively treated in another article applying for solving the alternating series; another important
our general algorithms used here. As a small example, convergent series class.
the following alternating series is solved by the Section
Acknowledgments
2 methods.
Considering the series given by Leibnitz expression Thanks to E. Poulain for fruitful discussions. One of
 1
n


 2n  1 .
n 1
the authors (HLG) thanks CONACYT for his
fellowship as member of SNI for financial support.
Using standard methods of Section 2,
 1  References
n

 
   e    1 e 2 nx  dx (23)
x n

n 1 2n  1
[1] P.Ramond, Field Theory: A Modern Primer, Frontiers in
0  n 1 
 Physics, Vol 74, 2nd ed.,Westview Press, United States,
  1 e 2 nx can be
n
and since the infinite series 2001.
n 1 [2] G.B. Arfken, H.J. Weber, F.E. Harris, Mathematical
also evaluated, finally one has Methods for Physicists: A Comprehensive Guide, 7th ed.,
 1  e  x 
n 1

dt
     arctan t 0  Elsevier Academic Press, USA, 2012, pp. 321-344
1
 2 x
dx 
n 1 2 n  1 1 e 1 t 2
0 0
4 (Chapter 5).
[3] M. Abramowitz, I.A. Stegun, Handbook of Mathematical
5. Conclusions Functions: With Formulas, Graphs, and Mathematical
The limits of a broad class of convergent infinite Tables, Dover Publications, USA, 1964.
series were obtained using some novel algorithms. [4] L.I. Schiff, Quantum Mechanics, 3rd ed., McGraw Hill,
New York, 1968.
These algorithms are very flexible and can be used to
[5] I.S. Gradshteyn, I.M. Ryzhik, A. Jeffrey, D. Zwillinger,
evaluate many series that are important in different Table of Integrals, Series and Products, 6th ed., Academic
physics disciplines, as well as in other relevant sciences Press, USA, 2000, p. 11.