Volume 1 July{December 1995 ISSN 1083-0464

SOUTHWEST JOURNAL
OF
PURE AND APPLIED MATHEMATICS

EDITOR IN CHIEF: IOANNIS ARGYROS

MANAGING EDITOR: MOHAMMAD TABATABAI
TECHNICAL EDITOR: GARY HUCKABAY
EDITORIAL BOARD
G. ANASTASSIOU I. ARGYROS
D. CHEN P. VAN FLEET
D. JANKOVIC T. McKELLIPS
C. RIECKE D. RUCH
F. SZIDAROVSZKY M. TABATABAI

CAMERON UNIVERSITY
DEPARTMENT OF MATHEMATICAL SCIENCES
LAWTON, OK 73505, U.S.A
 Southwest Journal
of
Pure and Applied Mathematics
Editor in Chief: Ioannis Argyros
Managing Editor: Mohammad Tabatabai
Technical Editor: Gary Huckabay
Editorial Board
G. Anastassiou anastasg@hermes.msci.memst.edu
I. Argyros ioannisa@cua.cameron.edu
D. Chen dchen@uafsysb.uark.edu
P. Van Fleet mth pvf@shsu.edu
D. Jankovic draganj@cuok.cameron.edu
T. McKellips terral@dragon.cameron.edu
C. Riecke rick@penguin.cameron.edu
D. Ruch mth dkr@shsu.edu
F. Szidarovszky szidar@sie.arizona.edu
M. Tabatabai mohammad@tiger.cameron.edu
Editorial Oce Publication Program
SWJPAM 2 issues per year
Cameron University July, December
Department of Mathematical Sciences
Lawton, OK 73505, U.S.A.

Information for Authors

Scope. SWJPAM is an electronic journal devoted to all aspects of Pure and Applied mathematics, and
related topics. Authoritative expository and survey articles on subjects of special interest are also welcomed.
Since SWJPAM aims to serve not only the specialists, but also the user of mathematics (promoting, in such
a way, a strong interaction between them), particular emphasis is placed on applications. Thus, SWJPAM
publishes papers which, although not directly contributing to mathematics, contain, however, signicant
results (from any eld) obtained by making an essential and well stressed use of the methods and results of
mathematics.
Submission of Manuscripts. Since SWJPAM is a completely electronic journal, all manuscripts must
be submitted in either AMSTEX, LaTEX, TEX, or post script form. Papers for publication should be sub-
mitted to ftp://rattler.cameron.edu/SWJPAM incoming by anonymous ftp. The articles must be placed
in the `incoming' directory. The submission should include the title of the paper, the author's name and
aliation, and a short abstract which summarizes the paper; a list of keywords and a classication accord-
ing to the Mathematical Reviews' Subject Classication (1991). Authors are informed by e-mail of referee
communication.
Key Words. Key words should be comprehensive and subject specic. It is not necessary to list the subject
area of the journal's coverage as a key word. Four key words should be sucient to cover the major subjects
o a given paper, although more can always be supplied if the author deems it necessary.

c 1995 Cameron University (continued on inside back cover)

 SOUTHWEST JOURNAL
OF
PURE AND APPLIED MATHEMATICS
VOLUME 1, JULY{DECEMBER 1995
CONTENTS
I. Argyros Perturbed Newton-Like Methods and Nondierentiable Operator Equations
on Banach Spaces with a Convergence Structure . . . . . . . . . . 1{12
G. Anastassiou Dierentiated Shift-Invariant Multivariate Integral Operators . . . . . . 13{16
D. Pantic Stochastic Calculus and Initial Value Analysis on the One Dimensional
Diusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 17{26
R. Forslund A Logical Alternative to the Existing Positional Number System . . . . 27{29
I. Argyros Results on Controlling the Residuals of Perturbed Newton-Like Methods
on Banach Spaces with a Convergence Structure . . . . . . . . . . 30{36
M. Tabatabai A Probabilistic Approach to the Estimation of the Distribution of the
Target Location Point . . . . . . . . . . . . . . . . . . . . . . 37{46

ISSN 1083-0464
Electronic Journal: Southwest Journal of Pure and Applied Mathematics
Internet: http://rattler.cameron.edu/swjpam.html
ISSN 1083-0464
Vol. 01, 1995, pp. 1{12.
Submitted: March 8, 1995. Published: June 20, 1995.

PERTURBED NEWTON-LIKE METHODS AND

NONDIFFERENTIABLE OPERATOR EQUATIONS ON
BANACH SPACES WITH A CONVERGENCE STRUCTURE

Ioannis K. Argyros
Abstract. In this study, we use perturbed Newton-likemethods to nd solutions of nonlinear, nondifferentiable
operator equations on Banach spaces with a convergence structure. This technique involves the introduction
of a generalized norm as an operator from a linear space into a partially ordered Banach space. In this way
the metric properties of the examined problem can be analyzed more precisely. Moreover, this approach allows
us to derive from the same theorem, on the one hand, semi-local results of Kantorovich-type, and on the other
hand, global results based on monotonicity considerations. Furthermore, we show that special cases of our
results reduce to the corresponding ones already in the literature. Finally, our results are used to solve integral
equations that cannot be solved with existing methods.

(1991) A.M.S. (MOS) Subject Classication Codes. 65J15, 65B05, 47D15, 47H17, 49D15.
Key Words and Phrases. Banach space, perturbed Newton-like methods, nondierentiable operator.
I. INTRODUCTION
In this study, we are concerned with approximating a solution x of the nonlinear operator equation
F(x) + Q(x) = 0; (1)
where F is a Frechet-dierentiable operator dened on a convex subset D of a Banach space X with values
in X, and Q is a nondierentiable nonlinear operator with the same domain and values in X.
We introduce the perturbed Newton-like method
xn+1 = xn + A(xn) [ (F(xn) + Q(xn))] zn ; x0 = 0 (n  0) (2)
to approximate a solution x of equation (1). Here A(xn ) (n  0) denotes a linear operator which is an
approximation for F 0(xn ) 1 (n  0).
The importance of studying perturbed Newton-like methods comes from the fact that many commonly
used variants of Newton's method can be considered procedures of this type. Indeed, approximation
(2) characterizes any iterative process in which the corrections are taken as approximate solutions of the
Newton equations. In particular for A(x) = F 0(x), Q(x) = 0 and zn = 0 (n  0) we obtain the ordinary
Newton-Kantorovich method [4].
The notion of a Banach space with a convergence structure was used in the elegant paper in [5] (see
also [1], [2], [3], [8]) to solve equation (1), when A(x) = F 0(x), Q(x) = 0 for all x 2 D and zn = 0
for all n  0. However, there are many interesting real life applications already in the literature, where
Department of Mathematics, Cameron University, Lawton, OK 73505, U.S.A.
E-mail Address: ioannisa@cua.cameron.edu
c 1995 Cameron University
Typeset by AMS-TEX
1
 equation (1) contains a nondierentiable term. See for example the applications at the end of this study.
Moreover, since the Newton-like iterates can rarely be computed exactly, suitable choices of the points zn
will \correct" the sequence xn and force it to converge to a solution x of equation (1). The case when
Q(x) = 0 for all x 2 D and F 0(xn) = F 0(xn) 1 (n  0) has already been considered but on a Banach
space without generalized structure [3], [4], [7], [8], [9].
As in [6], we provide semi-local results of Kantorovich-type and global results based on monotonicity
considerations from the same general theorem. Moreover, we show that our results can be reduced to
the one obtained in [6], when Q(x) = 0 for all x 2 D and zn = 0 (n  0), and furthermore to the ones
obtained in [1], [2], [3], [5], [6], [7], [8] by further relaxing the requirements on X.
Finally, our results apply to solve a nonlinear integral equation involving a nondierentiable term that
cannot be solved with existing methods.
II. PRELIMINARIES
We will need the denitions:
Denition 1. The triple (X; V; E) is a Banach space with a convergence structure if
(C1 ) (X; k
k) is a real Banach space;
(C2 ) (V; C; k kv ) is a real Banach space which is partially ordered by the closed convex cone C; the norm
k kv is assumed to be monotone on C;
(C3 ) E is a closed convex cone in X  V satisfying f0g  C  E  X  C;
(C4 ) the operator j j : D ! C is well dened:
jxj = inf fq 2 C j(x; q) 2 E g
for
x 2 D = fx 2 X j9q 2 C : (x; q) 2 E g;
and
(C5 ) for all x 2 D kxk  jxj v .
The set
U(a) = fx 2 X j(x; a) 2 E g
denes a sort of generalized neighborhood of zero.
Denition 2. An operator L 2 C 1(V1 ! V ) dened on an open subset V1 of an ordered Banach space
V is order convex on [a; b]  V1 if
c; d 2 [a; b]; c  d ) L0 (d) L0 (c) 2 L+ (V );
where for m  0
L+ (V m ) = fL 2 L(V m )j0  xi ) 0  L(x1 ; x2; : : : ; xm)g
and L(V m ) denotes the space of m-linear, symmetric, bounded operators on V .
Denition 3. The set of bounds for an operator H 2 L(X m ) is dened to be
B(H) = fL 2 L+ (V m )j(xi; qi) 2 E ) (H(x1; : : : ; xm ); L(q1; : : : ; qm )) 2 E g
Denition 4. Let H 2 L(X) and y 2 X be given, then
H  (y) = z , z = T 1 (0) = nlim n
!1 T (0);
X
1
T(x) = (I H)(x) + y , z = (I H)i y;
i=0
if this limit exists.
We will also need the Lemmas [2], [3], [5], [6]:
2
 Lemma 1. Let L 2 L+ (V ) and a; q 2 C be given such that:
L(q) + a  q and Ln (q) ! 0 as n ! 1: (3)
Then the operator
(I L) : [0; a] ! [0; a]
is well dened and continuous.
The following is a generalization of Banach's lemma [3], [5], [6].
Lemma 2. Let H 2 L(X), L 2 B(H), y 2 D and q 2 C be such that
L(q) + jyj  q and Ln(q) ! 0 as n ! 1:
Then the point x = (I H) (y) is well dened, x 2 S and
jxj  (I L) jyj  q:
Moreover, the sequence
bn+1 = L(bn ) + jyj; b0 = 0
is well dened and
bn+1  q; nlim 
!1 bn = b = (I L) jyj  q:

III. CONVERGENCE ANALYSIS

Let K; L; M; M1 be operators such that L 2 C 1(V1 ! V ), K; M1 2 L+ (V ), M 2 C(V1 ! V ), and xn
be points in D. It is convenient for us to dene the sequences cn ; dn; an; bn (n  0) by
dn+1 = (L + K + M + M1 )(dn ) + L0 (jxnj)cn; d0 = 0; (4)
cn = jxn+1 xnj; (5)
an = (L + K + M + M1 )n (a) for some a 2 C; (6)
bn = (L + K + M + M1)n (0); (7)
and the point b by
b = (L + K + M + M1 )1 (0): (8)
We can now state and prove the main result of this section.
Theorem 1. Let X be a Banach space with convergence structure (X; V; E) with V = (V; C; k kv), an
operator F 2 C 1(D ! X) with D  X, an operator Q 2 C(D ! X), an operator A(x) 2 L(X), operators
L; L1 2 C 1(V1 ! V ) with V1  V , operators K; M1 (V1 ! V ) with K; M1 2 L+ (V ), M 2 C(V1 ! V ), a
point a 2 C, and a null sequence fzn g 2 D such that the following conditions are satised:
(C6 ) the inclusions U(a)  D and [0; a]  V1 are true;
(C7 ) L is order-convex on [0; a] and satises
K + L0 (jxj + jyj) L0 (jxj) 2 B(A(x) F 0(x + y))
for all x; y 2 U(a) with jxj + jyj  a;
(C8 ) L1 is order-convex on [0; a] and satises
L01 (jxj) L01 (0) 2 B(A(0) A(x))
for all x; y 2 U(a) with jxj + jyj  a;
(C9 ) the estimate L01  L0 is true;
3
(C10) M satises the conditions
0  (Q(x) Q(x + y); M(jxj + jyj) M(jxj) 2 E
for all x; y 2 U(a) with jxj + jyj  a,
M(w1) M(w2)  M(w3) M(w4) and M(w)  0
for all w; w1; w2; w3; w4 2 [0; a] with w1  w3 and w2  w4;
(C11) M1 ; xn; zn satisfy the inequality
0  (A(xn)(zn ) A(xn 1)(zn 1); M1 (dn dn 1)) 2 E for all n  1;
(C12) L01 (0) 2 B(I A(0)) and (F(0) + Q(0) + A(0)(z0 )); L(0) + M(0) + M1 (0) 2 E;


(C13) (L + K + M + M1 )(a)  a with 0  L + K + M + M1 ;
and
(C14) (M + M1 + L0 (a) + K)n a ! 0 as n ! 1.
Then
(i) The sequence (xn ; dn) 2 (X  V )N is well dened, remains in E N , is monotone and satises for all
n0
dn  b
where dn; b are given by (4) and (8) respectively and b is the smallest xed point of L+K +M +M1
in [0; a].
(ii) Moreover, the iteration fxng n  0 generated by (2) converges to a solution x 2 U(b) of the equation
F(x) + Q(x) = 0, which is unique in U(a).
(iii) Furthermore, the following estimates are true for all n  0:
bn  dn  b; (9)
bn  an; (10)
jxn+1 xnj  dn+1 dn ; (11)
jxn x j  b dn (12)
and
jxn x j  an bn for M1 = 0 and zn = 0 (n  0) (13)
where dn; an; bn and b are given by (4), (6), (7) and (8) respectively.
Proof. We rst note that b replacing a also satises the conditions of the theorem. Using conditions (C9 )
and (C10) we obtain
jI A(0)j(b) + j (F(0) + Q(0) + A(0)(z0 )j  L01 (0)(b) + L(0) + M(0) + M1(0)
 L01 (0)(b 0) + L(0) + M(0) + M1 (0)
 (L01 (0) + M + M1 )(b 0) + L(0) + M(0) + M1 (0)
 L(b) L(0) + M(b) M(0) + M1 (b) M1 (0) + L(0) + M(0) + M1 (0)
 (L + K + M + M1 )(b) = b:
Hence, by Lemma 2, x1 is well dened and (x1 ; b) 2 E. We also get
x1 = (I A(0))(x1 ) +
 F(0) + Q(0) + A(0)(z )

0
=)
jx1j  L0 (0)jx1j + M(0) + L(0) + M1 (0) = d1
4
and by the order convexity of L
d1 = L0 (0)jx1j + (M + M1 + L)(0)  L0 (0)b + (M + M1 + L)(0)


 L0 (0) + M + M1 (b 0) + (M + M1 + L)(0)
 L(b) L(0) + M(b) M(0) + M1 (b) M1 (0) + M(0) + M1 (0) + L(0)
 (M + M1 + L + K)(b) = b:
That is, we get jx1 x0 j  d1 d0 or
0  (x0 ; d0)  (x1; d1):
We now assume that
0  (xk 1; dk 1)  (xk ; dk) for k = 1; 2; : : : ; n and dk  b:
We need to nd a bound for I A(xn ). We will show that L01(jxnj) 2 B(I A(xn)).
This fact follows from (C7 ), (C12) and the estimate
jI A(xn )j  jI A(0)j + jA(0) A(xn )j
 L01 (0) + L01 (0 + jxnj) L01 (0) = L01 (jxnj):
Using (2) we obtain the approximation
F(x ) + Q(x ) + A(x )(z ) = F(x ) Q(x ) A(x )(z )
n n n n n n n n
+ A(xn 1 )(xn xn 1) + F(xn 1) + Q(xn 1) + A(xn 1)(zn 1):
Hence, by (C7), (C10), (C11) and the induction hypotheses, we obtain in turn
j F(xn) + F(xn 1) + A(xn 1)(xn xn 1)j + jQ(xn 1) Q(xn )j + jA(xn)(zn ) A(xn 1)(zn 1 )j
Z1
 [L0 (jxn 1j + tcn 1) L0 (jxn 1j) + K] cn 1 dt + M(jxn 1j + jxn xn 1j) M(jxn 1j)
0
+ M1 (dn dn 1)
 L(jxn 1j + cn 1) L(jxn 1j) L0 (jxn 1j)cn 1 + Kcn 1 + M(jxn 1j + jxn xn 1j) M(jxn 1j)
+ M1 (dn dn 1)
 L(dn 1 + dn dn 1) L(dn 1) L0 (jxn 1j)cn 1 + K(dn dn 1)
+ M(dn 1 + dn dn 1) M(dn 1) + M1 (dn dn 1) = (M + M1 + L + K)(dn ) dn :
(14)
We can now obtain that
L01 (jxnj)(b dn) + j (F(xn) + Q(xn) + A(xn )(zn ))j + dn
 L0 (dn)(b dn) + L(dn ) + M(dn) + M1 (dn) + K(dn) dn + dn
 (L0 (dn) + M + M1 + K)(b dn ) + (L + K + M + M1 )(dn )
 K(b dn) + L(b) L(dn ) + M(b) M(dn) + M1 (b) M1 (dn) + (L + K + M + M1 )(dn)
= (M + M1 + L + K)(b) = b:
That is, xn+1 is also well dened by Lemma 2 and cn  b dn. Hence dn+1 is well dened too and we
obtain:
dn+1  (M + M1 + L + K)(dn ) + L0 (dn)(b dn)  (M + M1 + L + K)(dn )
+ (L0 (dn) + M + M1 )(b dn)
 (M + M1 + L + K)(dn ) + L(b) L(dn ) + M(b) M(dn) + M1 (b) M1 (dn) + K(b dn)
= (M + M1 + L + K)(b) = b:
5
The monotonicity (xn; dn)  (xn+1 ; dn+1) can be derived from
cn + dn  L0(jxnj)cn + j (F(xn) + Q(xn) + A(xn)(zn ))j + dn
 L0(jxnj)cn + (M + M1 + L + K)dn = dn+1:
The induction has now been completed. We need to show that
bn  dn for all n  1:
For n = 1 and from the denitions of bn; dn
b1 = (L0 (0) + K + M + M1 )1 (0)  d1:
Assume that
bk  dk for k = 1; 2; : : : ; n:
Then, we obtain in turn
bn+1 = (L0 (0) + K + M + M1 )n+1 (0) = (L0 (0) + K + M + M1 )(L0 (0) + K + M + M1)n (0)
 (L0 (0) + K + M + M1 )(dn )
 L0 (0)(dn) + (M + K + M1 )(dn)
 L(dn ) L(0) + (M + K + M1 )(dn)
 (L + K + M + M1)dn  dn  dn+1:
Since dn  b, we have bn  dn  b. We also need to show that
0  an bn  (L0 (a) + K + M + M1 )n (a) (n  1): (15)
For n = 1, we get
a1 b1 = (L + K + M + M1 )(a) (L + K + M + M1 )(0) = (L(a) L(0)) + (M(a) M(0))
Z1
+ K(a 0) + (M1 (a) M1(0)) = [L0 (ta) + K + M + M1 ](a) dt
Z1 0
 (L0 (a) + K + M + M1 )(a) dt = (L0 (a) + K + M + M1 )1(a);
0
which shows (15) for n = 1. Let us now assume that (15) is true for k = 1; 2; : : : ; n, then we can obtain
in turn
an+1 bn+1 = (L + K + M + M1 )n+1 (a) (L + K + M + M1 )n (0)
 
= (L + K + M + M1 ) (L + K + M + M1 )n (a) (L + K + M + M1)


(L + K + M + M1 )n (0)

  
= L (L + K + M + M1 )n(a) L (L + K + M + M1 )n(0)

  
+ (M + K + M1) (L + K + M + M1 )n (a) (M + K + M1 ) (L + K + M + M1 )n (0)

Z1 

= L0 (L + K + M + M1 )n (0) + t (L + K + M + M1 )n (a) (L + K + M + M1 )n(0)
0 

(L + K + M + M1 )n (a) (L + K + M + M1 )n (0) dt

+ (M + K + M1) (L + K + M + M1 )n (a) (L + K + M + M1 )n(0)

Z1

 L0 (L + K + M + M1 )n (a) (L0 (a) + K + M + M1 )n (a) dt
0
+ (M + K + M ) (L0 (a) + K + M + M )n (a)


1 1
 L0 (a)(L0 (a) + K + M + M1 )n (a) + (M + K + M1 )(L0 (a) + K + M + M1 )(a)
= (L0 (a) + K + M + M1 )(a):
6
 By condition (C14) and the above, we deduce that the sequence fbng n  0 is Cauchy in a Banach space
C, and as such, it converges to some b = (L + K
+ M + M1 )1 (0). From (L + K + M + M1 )(b) =
(L + K + M + M1 ) limn!1(L + K + M + M1 )n (0) = limn!1 (L + K + M + M1 )n+1(0) = b, we obtain
(L + K + M + M1 )(b) = b  a;
which makes b smaller than any solution of the inequality
(L + K + M + M1 )(p)  p: (see also [2], [3], [5]).
It also follows that the sequence fxng n  0 is Cauchy in X, and as such, it converges to some x 2 U(b).
By letting n ! 1 in (14) and using the hypotheses that limn!1 zn = 0, we deduce that x is a solution
of the equation F(x) + Q(x) = 0.
To show uniqueness, let us assume that there exists another solution y of the equation F(x)+Q(x) = 0
in U(a). Then, exactly as in [1], [2], [3], [5], [6], by considering the modied Newton-process
xn+1 = xn (F(xn) + Q(xn));
we can show that this sequence converges, under the hypotheses of the theorem. Moreover, as above, we
can easily show (see also [1], [2], [3], [5]) that
jy xnj  an bn for M1 = 0 and zn = 0 (n  0);
from which follows that xn ! y as n ! 1. Finally, estimates (12) and (13) are obtained by using
standard majorization techniques [3], [4], [9].
That completes the proof of the theorem.
We will now introduce results on a posteriori estimates. It is convenient to dene the operator
Rn(q) = (I L0 (jxnj)) Sn (q) + cn
where
Sn (q) = (L + K + M + M1 )(jxnj + q) (L + K + M + M1 )(jxnj) L0 (jxnj)(q)
and the interval
In = [0; a jxnj]:
It can easily be seen that the operators Sn are monotone on In. Moreover, the operators Rn : [0; a dn] !
[0; a dn ] are well dened and monotone. This fact follows from lemma 1 and the scheme
dn +cn  dn+1 ) R(a) dn+1  a dn cn ) Sn (a dn )+L0 (jxnj)(a dn cn )  a dn cn (n  0):
Then, exactly as in [6], we can show:
Proposition. The following implications are true:
(i) if q 2 In satisfy Rn(q)  q, then
cn  Rn(q) = p  q
and
Rn+1(p cn )  p cn for all n  0;
(ii) under the hypotheses of Theorem 1, let qn 2 In be a solution of Rn (q)  q, then
jx xm j  am (m  n)
where
an = qn and am+1 = Rm (am ) cm ;
7
 and
(iii) under the hypotheses of Theorem 1, any solution q 2 In of Rn(q)  q is such that
jx xn j  R1
n (0)  q:

Remark 1.
(a) The results obtained in Theorem 1 and the proposition reduce immediately to the corresponding ones
in [6] (Theorem 5 and Lemmas 10{12) when A(x) = F 0 (xn) 1 , Q(x) = 0 for all x 2 D and zn = 0 for
all n  0 (M; M1 ; K will be 0 then and L = L1).
(b) As in [2], [3], [5], we can also show that if conditions (C6){(C12 ) are satised and there exists t 2 (0; 1)
such that (L + K + M + M1 )(a)  ta, then there exists a1 2 [0; ta] satisfying conditions (C6 ){(C14).
The solution x 2 U(a1 ) is unique in U(a).
(c) From the approximation
A(xn)(zn ) A(xn 1)(zn 1 ) = (A(xn )(zn ) zn ) + ( A(xn 1 )(zn 1) + zn 1 ) + I(zn zn 1)
we observe that (C11 ) will be true if
M1 = 2L00(b) + I with L0 = L1 or L
and zn are chosen so that
jznj + jzn 1j + jzn zn 1j  dn dn 1 (n  1):
(d) Another choice for M1 ; zn can be M1 = j"jI and zn = zn 1 + "n (xn xn 1) (n  1) and j"nj 
j"j (n  0) provided that F 0(x) = I for all x 2 D. It can then easily be seen that (C11) is satised.
The sequence zn must still be chosen to be null. At the end of this paper, in Part V, we have given
examples to cover this case. Several other choices are also possible.
(e) From the proof of the theorem, it can easily be seen that condition (C9) can be replaced by the weaker
condition (which is what we actually need)
L01 (dn)(b dn)  L(b) L(dn) (n  0):
This condition however is not easy to verify in general.
(f) It can easily be seen that the conclusion of Theorem 1 remains valid if in the hypotheses L1 replaces
L everywhere except in condition (C7).
(g) Condition (C7 ) can be replaced by the set of conditions
K + L02 (jxj + jyj) L02(jxj) 2 B(F 0 (x) F 0(x + y))
and
L03 (jxj) L03 (j0j) 2 B(A(x) F 0(x))
for all x; y 2 [0; a] with jxj + jyj  a as long as we can nd L such that
L0 (jxj + jyj) = L03 (jxj) + L02(jxj + jyj) + K and L0 (jxj) = L02 (jxj) + L03 (j0j);
for all x; y 2 [0; a] with jxj + jyj  a. The operators L; L2; L3 2 C 1 (V1 ! V ) dened here must also be
order convex.
(h) It can easily be seen from the proof of theorem 1 that the results obtained in theorem 1 remain valid
if (C11) is replaced by the weaker condition:
0  (A(xn)(zn ) A(xn 1)(zn 1 ); M2 jxn xn 1j) 2 E (n  1)
8
 for some M2 2 L+ (V ). This condition seems to be more helpful since it relates zn with xn and not
an unknown sequence depending on M1 . Under this condition the term M1 (b bn) appearing in the
Proposition can be replaced by M2 jxn xn 1j (n  1). Other choices involving M2 are possible.
(i) Dene the residuals rn = A(xn )(zn ) (n  0) and set n = xn+1 xn (n  0). Then from the
approximation
rn = [(I A(xn )) I ] (zn )
we obtain
jrnj  (L01 (jxnj) + I ) jznj
which shows that rn ! 0 as n ! 1 if zn ! 0 as n ! 1. Consequently, the conclusions of theorem 1
remain true for the system
xn+1 xn = n ; A(xn)n = (F(xn) + Q(xn)) + rn (n  0):
(j) (C11) is equivalent to the condition that (A(xn )(zn ); M1 (dn)) 2 E N (n  1) is an increasing sequence.
Similar observations can be made for (C10).
IV. THE MONOTONE CASE
Let A 2 L(X ! X) be a given operator. Dene the operators P; T (D ! X) by
P(x) = AT(x + u)
T(x) = G(x) + R(x); P(x) = F(x) + Q(x)
and
F(x) = AG(x + u); Q(x) = AR(x + u);
where G; R are as F; Q respectively. We deduce immediately that under the hypotheses of Theorem 1, the
zero x of P is a zero of AT also, if u = 0.
We will now provide a monotonicity result to nd a zero x of AT. The space X is assumed to be
partially ordered and satises the conditions for V given in (C1 ){(C5 ). Moreover, we set X = V; D = C 2
so that j j turns out to be I.
We can now state and prove the main result of this section.
Theorem 2. Let V be a partially ordered Banach space satisfying conditions (C1){(C5), Y be a Banach
space, G 2 C 1(D ! Y ), R 2 C(D ! Y ) with D  V , A; A(x) 2 L(X ! V ), K; M1(D ! V ) with
K; M1 2 L+ (V ), M 2 C(V1 ! V ) and u; v 2 V such that:
(C15) [u; v]  D;
(C16) 2I AG0(u) A(v u) + M + K + M1 2 L+ (V );
(C17) for all w; w1; w2 2 [u; v] : w1  w2 ) AG0 (u + w)  I,
A(w2 )  A(w1)
A(w2)  AG0(w1 + u) + K
A(w1 ) A(w2 ) + AG0 (u + w1 ) AG0(u + w2)  0
and
(Q(0) + A(0)(z0 ))  (M + K + M1 )(0);
(C18)
R
AT(u)+A(u)(z0 )  0, AT(v)+A(v)(z0 )  0 and v u  0v u A(z)dz +AG(v) (M +K +M1 )(v u);
(C19) condition (C10) is satised;
(C20) condition (C11) is satised; and
(C21) (2I AG0(v) A(v u) + M + K + M1 )n(v u) ! 0 as n ! 1.
9
 Then the Newton sequence
y0 = u; yn+1 = yn + (A(yn )) [ AT(yn )] zn (n  0)
is well dened for all n  0, monotone and converges to a unique zero x of AT in [u; v].
R
Proof. The proof follows immediately by setting P(x) = AT(x + u), L1 (e) = e 0e A(z)dz, L2 (e) =
e F(e), L = L1 + L2 and a = v u.
Remark 2.
(a) By setting R(x) = 0 for all x 2 D and zn = 0 (n  0), we immediately deduce that our results reduce
to the ones obtained in [6, thm 13].
(b) If M; R are linear operators, then condition (C19) will certainly be satised
R if we set M(e) = Q(e).
(c) The operators L; L1; L2 can be chosen in general as L1 (e) = z1 +R1 (e) 0e A(z)dz, L2 = z2 +R2(e)
F(e) and L = L1 + L2 with z1 ; z2 2 [u; v] and R1; R2 2 L(D ! Y ). Conditions similar to the ones
in Theorem 2 can then be easily produced for the above choices of L; L1 ; L2. For example, conditions
(C16) and (C21) will become
R1 + R2 + K AG0(u) A(v u) + M1 + M 2 L+ (V )
and
(R1 + R2 AG0(u) A(v u) + M1 + K + M)n(v u) ! 0 as n ! 1
respectively.
V. APPLICATIONS
We will complete this study with two examples that show how to choose L; M; M1; zn in practical
applications. For simplicity, we assume that A(x) = F 0(x) for all x 2 D.
Example 1. We discuss the case of a real Banach space with norm k k. Assume for simplicity that
F 0(0) = I and there exists a monotone operator
f : [0; a] ! IR
such that
kF 00 (x)k  f(kxk) for all x 2 U(a)
and a continuous nondecreasing function g on [0; r], r  a such that
kQ(x) Q(y)k  g(r)kx yk (16)
for all x; y 2 U( 2r ).
We showed in [3], (see also [9]) that (16) implies that
kQ(x + `) Q(x)k  h(r + k`k) h(r); x 2 U(a); k`k  a r; (17)
where Zr
h(r) = g(t) dt:
0
Conversely, it is not hard to see that we may assume, without loss of generality, that the function h
and all functions h(r + t) h(r) are monotone in r. Hence, we may assume that h(r) is convex and hence
dierentiable from the right. Then, as in [9], we show that (17) implies (16) and g(r) = h0(r + 0). Hence,
we can set Zq Zs
L1 (q) = L(q) = kF(0) + Q(0)k + ds f(t) dt (18)
0 0
10
and Zq
M(q) = g(t) dt: (19)
0
In Remark 1 (c) and (d) we have already provided some choices for M1 ; zn . Here, however, for simplicity
let us choose zn = 0 (n  0) and M1 = 0.
Then condition (C13) will be true if
1 f(a)a2 (1 g(a))a + kF(0) + Q(0)k  0: (20)
2
If we set Q = 0 and g = 0, (20) is true if kF(0)kf(a)  21 , which is a well known condition due to
Kantorovich [4, Ch. 18]. If Q 6= 0, condition (20) is the same condition with the one found in [3], [9] for
the Zincenko iteration.
In the example that follows, we show that our results can apply to solve nonlinear integral equations
involving a nondierentiable term, whereas the results obtained in [6] (or [5], [7], [8]) cannot apply.
Example 2. Let X = V = C[0; 1], and consider the integral equation
Z1
x(t) = k(t; s; x(s)) ds on X; (21)
0

where the kernel k(t; s; x(s)) with (t; s) 2 [0; 1]  [0;1] is a nondierentiable operator on X. Consider (21)
in the form (1) where F; Q : X ! X are given by
F(x)(t) = Ix(t)
and Z1
Q(x)(t) = k(t; s; x(s)) ds:
0
The operator j j is dened by considering the sup-norm. We assume that V is equipped with natural
partial ordering, and there exists ; a 2 [0; +1), and a real function (t; s) such that
kk(t; s; x) k(t; s; y)k  (t; s)kx yk
for all t; s 2 [0; 1]; x; y 2 U( a2 ), and
Z1
  sup (t; s) ds:
t2[0;1] 0
Dene the real functions h; f; g on [0; a] by h(r) = r, f(r) = 0 and g(r) =  for all r 2 [0; a]. By choosing
L; M; M1 as in (18) and (19) and Remark 1 (d) respectively, and by setting K = 0, we can easily see that
the conditions (C1 ){(C12) of Theorem 1 are satised. In particular, condition (C13) becomes
(1  jj)a kQ(0)k  0 (22)
which is true in the following cases: if 0   < 1 jj, choose a   = 1kQ(0)jkj ; if  = 1 jj and Q(0) = 0,
choose a  0; if  > 1 jj and Q(0) = 0, choose a = 0. If in (22) strict inequality is valid, then there
exists a solution a of equation (22) satisfying condition (C14). Note that if we choose  2 (0; 1 jj),
a 2 (; +1) and  2 ( 1; 1), condition (22) is valid as a strict inequality. Finally, we remark that the
results obtained in [6] (or [5], [7], [8]) cannot apply here to solve equation (21), since Q is nondierentiable
on X and the zn 's are not necessarily zero.
11
 CONCLUSION
In this study, we used perturbed Newton-like methods to nd solutions of nonlinear, nondierentiable
operator equations on Banach spaces with a generalized structure. This technique involves the introduc-
tion of a generalized norm as an operator from a linear space into a partially ordered Banach space. This
way the metric properties of the examined problem have been shown to be analyzed more precisely. Con-
vergence results and error estimates have been improved compared with the real norm theory. Moreover,
this approach also allowed us to derive, on the one hand, semi-local results of Kantorovich-type and, on
the other hand, global results base on monotonicity considerations, from the same general theorem. All
our convergence theorems were derived from metric estimates. Furthermore, we show that special cases
of our results reduce to the corresponding ones already in the literature [1], [2], [5], [6], [7]. Finally, we
used our results to solve applied problems that cannot be solved with existing methods.
REFERENCES
[1] Argyros, I.K. The secant method in generalized Banach spaces, Appl. Math. Comp. 39, (1990),
111{121.
[2] . Newton-like methods in generalized Banach spaces, Functiones et Approximatio
Commentarii Mathematici, XXII, (1993), 13{20.
[3] Argyros, I.K. and Szidarovszky, F. The Theory and Applications of iteration methods, C.R.C. Press,
Inc. Boca Raton Florida, U.S.A. 1993.
[4] Kantorovich, L.V. and Akilov, G.P. Functional Analysis in normed spaces, Academic Press, New
York, 1978.
[5] Meyer, P.W. Newton's method in generalized Banach spaces, Numer. Funct. Anal. Optimiz. 9 (3
and 4), (1987), 249{259.
[6] . A unifying theorem for Newton's method, Numer. Funct. Anal. Optimiz. 13 (5
and 6), (1992), 463{473.
[7] Pandian, M.C. A convergence test and component-wise error estimates for Newton-type methods,
SIAM J. Numer. Anal. 22, (1985), 779{799.
[8] Vandergraft, J.A. Newton's method for convex operators in partially ordered spaces, SIAM J. Nu-
mer. Anal. 4 (1967), 406{432.
[9] Zabrejko, P.P. and Nguen, D.F. The majorant method in the theory of Newton-Kantorovich approxi-
mations and the Ptak error estimates, Numer. Funct. Anal. Optimiz. 9 (5 and 6), (1987), 671{684.

12
Electronic Journal: Southwest Journal of Pure and Applied Mathematics
Internet: http://rattler.cameron.edu
ISSN 1083-0464
Vol. 01, 1995, pp. 13{16.
Submitted: March 16, 1995. Published: June 30, 1995.

DIFFERENTIATED SHIFT-INVARIANT
MULTIVARIATE INTEGRAL OPERATORS

George A. Anastassiou
Abstract. In a recent paper [1] the author, together with H. Gonska, introduced some wavelet-type, very general
multivariate integral operators over Rd, d  1, and studied their various properties such as shift-invariance, global
smoothness preservation and optimality, convergence to the unit, dierentiation, and preservation of continuous
multivariate probability distribution functions. These operators are introduced through a convolution-like iteration
of another very general multivariate operator with a general multivariate scaling type function. In this paper
the author gives sucient conditions, so that the partial derivatives of the above described multivariate operators
possess most of the above listed nice properties of their originals. In particular, a sucient condition is given so
that a corresponding \global smoothness preservation" sharp multivariate inequality is attained. The corresponding
convergence of dierentiated operators to the dierentiated unit operator is given with rates of convergence.

(1991) A.M.S. (MOS) Subject Classication Codes. 60, 62, 28.
Key Words and Phrases. Wavelet integral operators, Probability distribution functions.
Main Results.
Let X := CU (Rd), d  1 be the space of uniformly continuous functions from Rd into R.
For f 2 X, dene the rst order modulus of continuity of f by
!1 (f; ) := sup d jf(x) f(y)j;  > 0;
x;y2 R
kx yk
where k
k is an arbitrary norm in Rd. It is known [1] that !1 (f; ) < 1, 8 > 0. Here C(Rd) denotes the
Z
space of continuous functions from Rd into R. Let f`k gk2 be a sequence of linear operators from C(Rd) into
itself such that
`k (f; x) = `0 (f(2 k
); x); all x 2 Rd; f 2 C(Rd): (1)
Assume that
`0 : C () (Rd) ! C () (Rd);  2 N be xed; (2)
where C () (Rd) is the space of -times continuously dierentiable 
functions.
In this article we consider only f 2 C () (Rd) such that @xj1@


f@xjm 2 X, where xed i1 ; : : : ; im 2 f1; : : : ; dg
i1 imP
be such that i1 <


< im , and xed j1; : : : ; jm 2 N be such that mr=1 jr = .
From now on we will denote by

f @ := j1 @ f jm ; f 2 C () (Rd); (3)
@xi1


@xim
Department of Mathematics, Memphis State University, Memphis, TN
Email: anastasg@ kermes.msci.memst.edu
Copyright c 1995 by Cameron University
13
where the mixed partial is dened as above.
Let ' be a bounded real valued function of compact support  di=1 [ ai ; ai], ai > 0. We assume that
'  0, ' is Lebesgue measurable and
Z1 Z1



'(x1 u1; x2 u2; : : : ; xd ud ) du1 : : :dud = 1;
| 1 {z 1}
d fold
all (x1; : : : ; xd) 2 Rd (4)
((4) same as Z
'(x u) du = 1; all x 2 R ; d
Rd
where x = (x ; : : : ; x ), u = (u ; : : : ; u )). One can easily prove that
1 d 1 d
Z
Rd '(u)du = 1: (5)
Example. For i = 1; : : : ; d take the characteristic function
 1; xi 2  1 ; 1

'i (xi ) := [ 21 ; 21 ) (xi ) = 2 2
0; else.
Dene
Y
d
' (x) := 'i (xi ); all x = (x1 ; : : : ; xd) 2 Rd:
i=1
 
Note that ' is bounded, supp '  12 ; 12 d , '  0 and ' is Lebesgue measurable. Also
Z 1 Z 1 Yd



'i (xi ui ) du1 : : :dud = 1;
| 1 {z 1} i=1
d fold
i.e., Z
 u)du = 1; all x 2 Rd:
Rd ' (x
Z
Let fLk gk2 be the sequence of linear operators acting on C () (Rd) and dened as follows
Z
L0(f; x) := d `0 (f; u)'(x u) du
R (6)
and
Lk (f; x) := L0 (f(2 k
); 2kx); (7)
all x 2 Rd, k 2 Z. I.e., Z
Lk (f; x) = k

For our convenience we shall use the notation

Rd `k(f; u)'(2 x u) du: (8)

Z ~a Z
= : (9)
~a di=1 [ ai ;ai ]
Furthermore, we obtain that Z ~a
L@k (f; x) = 2k
`k (f; 2k x u)'(u) du
Z ~a
= 2k
@ k
Z Rd `k(f; 2 x u)'(u) du
= 2k
@ k u) du; all x 2 Rd; k 2 Z:
Rd `k (f; u)'(2 x (10)

14
Assumption. Take f 2 C () (Rd),  2 N xed, such that f @ 2 X. For xed  > 0 we assume that
 
sup d j`@0 (f; u) f @ (y)j  !1;1 f @ ; m2r+ n ; (11)
R
u;y2
ku yk1
is true for f as above, m 2 N, n 2 Z+, r 2 Z, where !1;1 is the modulus of continuity !1 dened with
respect to k
k1 .
Here we present our results without proofs due to their length. We intend to show proofs elsewhere.
Denition 1. Let f (
) := f(
+ ),  2 Rd and  be an operator. If (f ) = (f) , then  is called a shift
invariant operator.
Proposition 1. Assume that


`@0 f(2 k
+); 2k u = `@0 f(2 k
); 2k(u + ) ;

(12)
all k 2 Z,  2 Rd xed, all u 2 Rd; any f 2 C () (Rd);  2 N be given. Then L@k , any k 2 Zis a shift
invariant operator.
Next we establish the property of \global smoothness preservation " of operators L@k , k 2 Z.
Theorem 1. For any f 2 C ()(Rd), such that f @ 2 X ,  2 N xed, and any u 2 Rd assume that
j`@0 (f; x u) `@0 (f; y u)j  !1(f @ ; kx yk); all x; y 2 Rd; (13)
where k
k is an arbitrary norm in Rd. Then
!1 ((Lk f)@ ; )  !1(f @ ; ); (14)
any  > 0.
Optimality of (14) follows.
Theorem
Pm j = 2.. Let i1 ; : : : ; im 2 f1; : :j : +1; dg be such that i1 j<


< im , and j1 ; : : : ; jm 2 N be such that
r=1 r Consider gj1 (x) := (xj1 +1)!1
, and gjr (x) := xjrr! , r = 2; : : : ; m; all x 2 R. Denote by f @ :=
@f () d d
@xji11 :::@xjimm , for any f 2 C (R ). Denote also by prir : R 3 (x1 ; : : : ; xd) ! xir , r = 1; : : : ; m the projection
onto the ir coordinate. For all x; y; u 2 Rd assume that
X
m ! !!@
`0 gjr  prir (2 k
) (2k x u)
r=1
Y
m ! !!@
`0 gjr  prir (2 k
) (2k y u) = 2 k
(xi1 yi1 ): (15)
r=1
Then 0 !!@ 1 0 Y !@ 1
Y
m m
!1 @ Lk gjr  prir ;  A = !1 @ gjr  prir ;  A ; (16)
r=1 r=1
for any  > 0; k 2 Z. That is establishing that (14) is sharp!
 Lk
The convergence of @xj@1 :::@x
P
m @
jm ( r=1 jr = ) to @xj1 :::@xjm operator, as k ! +1, with rates given in the
i1 im i1 im
following.
Theorem 3. For f 2 C ()(Rd), d  1, such that f @ 2 CU (Rd), under the assumption (11), it holds
 ma + n 
jLk (f; x) f (x)j  !1;1 f @ ; 2k+r ;
@ @ (17)
where m 2 N, n 2 Z+, k; r 2 Z, and a := max(ai ), i = 1; : : : ; d.
Again, due to their length we intend to present applications of the above results elsewhere.
15
 References
1. G.A. Anastassiou and H.H. Gonska, On some shift-invariant integral operators, multivariate case, Proceedings of Interna-
tional Conference in Approximation Theory and Probability, Santa Barbara, California, May 20-22, 1993 (1994), Plenum,
New York, 41{64.
2. H. Bauer, Ma -und Integrationstheorie (1990), de Gruyter, Berlin.
3. C. Cottin and H.H. Gonska, Simultaneous approximation and global smoothness preservation, Proceedings of the Second
International Conference in Functional Analysis and Approximation Theory Acquafredda di Maratea-Potenza-Italy, 1992.
Supplemento ai Rendiconti del Circolo Matematico di Palermo, Serie II, Numero 33, (1993), 259{279..
4. A.N. Shiryayev, Probability, Springer-Verlag, New York, Berlin, 1984.

16
Electronic Journal: Southwest Journal of Pure and Applied Mathematics
Internet: http://rattler.cameron.edu/swjpam.html
ISSN 1083-0464
Vol. No. 01, 1995, pp. 17{26
Submitted: June 11, 1995. Published: September 5, 1995.

STOCHASTIC CALCULUS AND INITIAL VALUE

ANALYSIS ON THE ONE DIMENSIONAL DIFFUSIONS

Drazen Pantic

Abstract. Stochastic calculus and initial value analysis are used for complete description of distribution type
of diusion processes with Lipschitz coecients. Sucient conditions so that the solutions of stochastic dier-
ential equations posses absolutely continuous one dimensional distribution are given. For stochastic dierential
equations with uniformly elliptic coecients probability density is investigated in detail. Also, the distribution
of inverse process is given.

(1991) A.M.S. Subject Classication Codes. 60H07, 60H10.

Key Words and Phrases. Malliavin calculus, solutions of stochastic dierential equations, point dis-
tributions.
INTRODUCTION
Let fXt;x ; t  0g; x 2 R, be solution of SDE
Zt Zt
Xt;x x = a(Xs;x )dBs + b(Xs;x )ds (1)
0 0

where coecients a and b have bounded and continuous rst order derivatives (a; b 2 C 1b ). Under those
conditions for p > 1 there exists Lp derivative DXt;x of Xt;x w.r.t. initial condition x, so that, as in [2,
x8], Zt Zt
DXt;x 1 = a0(Xs;x )DXs;x dBs + b0 (Xs;x )DXs;x ds: (2)
0 0
Using Ito formula, DXt;x could be computed explicitly as
nZ t Zt o
DXt;x = exp a0 (Xs;x )dBs + (b0(Xs;x ) 21 a0(Xs;x ))2 ds : (3)
0 0
The boundeness of derivatives implies that, for p > 1; DXt;x as well as its inverse (DXt;x ) 1 are in Lp ,
[6], while (DXt;x ) 1 satises following SDE, [10, Lemma (3.7)]:
Zt Zt

(DXt;x ) 1 1= a0 (Xs;x )(DXs;x ) 1 dBs + a0(Xs;x )2 b0(Xs;x ) (DXs;x ) 1 ds:
0 0
Department of Mathematics, University of Belgrade, Studentski trg 16, 11000 Beograd, Yugoslavia
E-mail Address: epantidr@ubbg.etf.bg.ac.yu
c 1995 Cameron University
17
Also, for a Skorohod integrable process u = fus; 0  s  T g a directional derivative Du Xt;x of functional
Xt;x could be expressed in terms of derivatives fDXt;x ; t  0g in the following way, [12],
Zt
Du Xt;x = DXt;x (DXs;x ) 1 a(Xs;x ) us ds: (4)
0
First we shall treat the general case. Dening process u = fus ; 0  s  tg as
us = a(Xs;x )DXs;x
we have Zt
Du Xt;x = DXt;x a(Xs;x)2 ds: (5)
0
Process u is Ito integrable, so for a continuously dierentiable function with compact support f it follows,
by integration by parts formula, [5], and relation (5), that
Zt Zt
E ff 0 (Xt;x )DXt;x a(Xs;x )2 dsg = E fD u (f(Xt;x ))g = E ff(Xt;x ) a(Xs;x )DXs;x dBsg: (6)
0 0
R
Knowing that DXt;x is a.s. positive, we have to nd conditions so that 0t a(Xs;x )2 ds is a.s. positive in
order to apply next Lemma.
Lemma 1. Let X be a random variable such that there exists integrable and almost sure positive random
variable Y and a constant C > 0 so that for each f 2 C 1b we have
j E ff 0 (X)
Y g j C supfj f(x) j: x 2 Rg:
Then probability law generated by the random variable X is absolutely continuous with respect to Lebesgue
measure.
Proof. Let us dene positive bounded linear functional on the space of the continuously dierentiable
functions with compact support with
Z
L(f) = E ff(X)Y g = E ff(X)E fY jX gg = f(x)E fY jX = xgG(dx)
where G is probability measure generated by X. Then, we have
Z
j f 0 (x)E fY jX = xgG(dx) j C supfj f(x) j: x 2 Rg
so using [4, x1] it follows that there exists density g so that
Z Z
f(x)E fY jX = xgdG(x) = f(x)g(x) dx:
Clearly, the previous relation could be extended to the space of bounded measurable functions, so for a
Borel set B we have Z
EfIfX 2 B gY g = Ifx 2 B gg(x) dx:

If B is of Lebesgue measure zero then

E fIfX
2 B gY g = 0:
Knowing that Y > 0 almost sure one easily gets that PfX 2 B g = 0; i.e. probability law generated by
the random variable X is absolutely continuous with respect to Lebesgue measure. 
18
 GENERAL CASE
In the next theorem we shall completely describe the rst moment  from which 0t a(Xs;x )2 ds is a.s.
R
positive, as well as the behavior of diusion process (1) up to that moment.
Theorem 1. Let fXt;x ; t  0g; x 2 R, be solution of SDE (1) with coecients in C b1 , and fYt;x; t  0g
deterministic process dened by
Zt
Yt;x x = b(Ys;x ) ds:
0
Dene
x = inf ft  0 :j a(Xt;x ) j> 0g and Dx = inf ft  0 :j a(Yt;x ) j> 0g:
Then, with probability one, x = Dx and Xt;x = Yt;x for t  Dx .
Proof. Assume rst that 0 < Dx < 1, and let 0 < t < Dx . Then
Zt
Zt

Xt;x Yt;x = a(Xs;x ) a(Ys;x) dBs + b(Xs;x ) b(Ys;x) ds; and
0 0

E f(Xt;x Yt;x)2 g1=2 

nZ t
2o1=2 + En Z t b(X ) b(Y )
ds2o1=2 
E a(Xs;x ) a(Ys;x ) dBs s;x s;x
Z
n t 0

o1=2 n t Z 0

o1=2 
E a(Xs;x ) a(Ys;x ) 2 ds + E t b(Xs;x) b(Ys;x ) 2 ds
Z0 t 1=2  Z0 t 1=2
K E f(Xs;x Ys;x )2gds + t1=2K E f(Xs;x Ys;x )2 gds 
0
Z t 0
1=2
 K(1 + Dx1=2 ) E f(Xs;x Ys;x )2gds ;
0

where K is upper bound for rst derivatives of a and b. Then, using Gronwall inequality, we can conclude
that with probability one X  Y on [0; Dx], and hence Dx  x a.s.
Assume now that 0 < x < 1, and let 0 < t < x . As x is optional time the set
A = f! : t < x (!)g
is Ft measurable, where fFt; t  0g is standard ltration generated by fBt ; t  0g. Then we have
Zt
Ef a(Xs;x )2ds IAg = 0:
0

It is obvious that
Z t 2 Z t^x 2
a(Xs;x ) dBs Ift < x g  a(Xs;x ) dBs :
0 0
Having in mind that t ^ x is optional time we have
nZ t 2 o nZ t^x 2o n Z t^x o
E a(Xs;x ) dBs Ift < x g E a(Xs;x ) dBs =E a(Xs;x )2 ds = 0;
0 0 0
19
 R
and consequently 0t a(Xs;x)dBs = 0 a.s. on set A. Hence for each t > 0 on the set A we have
Zt

Xt;x Yt;x = b(Xs;x ) b(Ys;x) ds; and
0
E f(Xt;x Yt;x )2 Ift < x gg =
nZ t
2 Ift <  go 
=E b(Xs;x) b(Ys;x ) ds x
nZ t
0

o
 tE b(Xs;x ) b(Ys;x ) 2 ds Ift < x g 
n Z0t
o
 tE b(Xs;x ) b(Ys;x ) 2 Ifs < x g ds 
0
Zt 
 tK 2 E (Xs;x Ys;x )2 Ifs < x g ds
0
Let us mark
t = E f(Xt;x Yt;x)2 Ift < x gg. Then previous relation could be written as
Zt

t  tK 2
s ds;
0
from where it follows that
t = 0 for each t > 0. Hence we can conclude that for each t > 0
Pf! : t < x (!); Xt;x 6= Yt;x g = 0:
Using the continuity of paths we have that
Pf! : (9t > 0)(t < x (!); Xt;x 6= Yt;x)g = 0;
so with probability one X  Y on [0; x], i.e. x  Dx a.s. and the proof is complete. 
The following theorem completely describes the type of one dimensional distributions of the solutions
of SDE :
Theorem 2. Under conditions from Theorem 1. the distribution of random variable Xt;x is either
degenerate or absolutely continuous, whether t  Dx or t > Dx .
Proof.
R
The rst part when t  Dx is described in Theorem 1 so let t > Dx . Then 0t a(Xs;x )2ds is
positive a.s. and from relation (6) and Lemma 1 we can reach our conclusion. 
UNIFORMLY ELLIPTIC CASE
Now, we shall treat an important special case when that coecient a is uniformly elliptic in sense that
inf fj a(x) j: x 2 Rg > 0: (EL)
Theorem 3. Let the diusion process fXt;x ; t  0g dened by SDE (1) has continuously dierentiable
coecients a and b with bounded derivatives and let
inf fja(x)j : x 2 Rg > 0:
Then we have :
1 for each t > 0 and x 2 R the probability law of random variable Xt;x is absolutely continuous
w.r.t. Lebesgue measure with transition density pt (x;
);
20
 2 for any bounded, measurable function f and t > 0 the mapping x ! Eff(Xt;x )g is continuously
dierentiable and
@ Zt
@x E ff(Xt;x )g = t
1 Eff(X a(Xs;x ) 1 DXs;x dBs g;
t;x ) (7)
0
3 for each y 2 R and t > 0 transition density pt(x; y) is dierentiable w.r.t. initial condition x and
@ p (x; y) = t Zt
1 Ef a(Xs;x ) 1 DXs;x dBs jXt;x = yg pt (x; y): (8)
@x t 0

Proof. 1 Taking u~s = a(Xs;x ) 1 DXs;x , one has, from (4)

Du~ Xt;x = DXt;x t: (9)
Process fu~g is Ito integrable, so for a continuously dierentiable function with compact support f it
follows by integration by parts formula and relation (9), that
Zt
E ff 0 (Xt;x )DXt;x tg = E fDu~ (f(Xt;x ))g = Eff(Xt;x ) a(Xs;x ) 1 DXs;x dBs g: (10)
0
Now, DXt;x is a.s. positive random variable and its inverse (DXt;x ) 1 is in Lp for
R each p > 1: Also the
t
assumptions we have made on coecient a assure, via a Burgholder theorem, that 0 a(Xs;x ) 1 DXs;x dBs
is in L2 , hence by Lemma 1 assertion 1 holds.
2 First, take f 2 C 1b . Then by [2] and relation (10) one has
@ Zt
0
@x E ff(Xt;x )g = Eff (Xt;x )DXt;x )g = t
1E ff(X a(Xs;x ) 1DXs;x dBsg:
t;x )
0
Now, for x; y 2 R we have
Zy Zt
Eff(Xt;y )g E ff(Xt;x )g = E ff(Xt;z ) a(Xs;z ) 1DXs;z dBs g dzt : (11)
x 0
Both sides of relation (11) are continuous linear functionals on the set Mb of measurable and bounded
functions, so we can conclude that (11) is valid for each f 2 Mb . Also, for f 2 Mb and t > 0 mapping
x ! E ff(Xt;x )g is absolutely continuous.
If we dene semigroup fPt; t  0g of operators on the space Mb of bounded measurable functions as
Ptf(x) = E ff(Xt;x )g, then Markov property of diusion fXt;xg yields, for  > 0
Pt+ f(x) = E fP f(Xt;x )g:
Substituting t with t +  and f with P f, relation (11) reads as
Zy Zt
Pt+ f(y) Pt+ f(x) = E fP f(Xt;z ) a(Xs;z ) 1 DXs;z dBs g dzt :
x 0
Then, from continuity of the function P f(x) it follows that for each x 2 R there exists

lim Pt+ f(y)y Px t+ f(x) ;

y!x
and
21
 @ P f(x) = t 1 E fP (X ) Z t a(X ) 1DX dB g
@x t+  t;x
0
s;x s;x s

hence @x@ Pt+ f(x) is continuous function in x.

Now relation (11) yields that, for each x
@ Zt
1 E ff(X a(Xs;x) 1 DXs;x dBs g:
@x Pt f(x) = t t;x )
0

3 Let us mark Zt
Wt;x = t 1 a(Xs;x) 1 DXs;x dBs : (12)
0
Then relation (10) reads as
E ff 0 (Xt;x )DXt;x g = E ff(Xt;x )Wt;x g;
or, having in mind 1 ,
Z1 Z1
f 0 (u) E fDXt;x jXt;x = ugpt(x; u) du = f(u) E fWt;x jXt;x = ugpt(x; u) du:
1 1
Taking, for  > 0 and y 2 R, Zx
f (x) = e (t y)2 =2 pdt ;
1 2
we get, using partial integration
Z1
e (u y)2=2 EfDXt;x jXt;x = ugpt(x; u) pdu =
1 Z 1 Z u 2
2 dt 
= e (t y) =2 p E fWt;x jXt;x = ugpt(x; u) du =
1 1 2
Z1 Z t  dt
= e (t y)2 =2 E fWt;x jXt;x = ugpt(x; u) du p : (13)
1 1 2
>From relation (11) follows, via [3, Theorem 1.28.], that E fDXt;x jXt;x = ygpt (x; y) is continuous and
bounded function so letting  # 0 one can conclude that
Zy
EfDXt;x jXt;x = ygpt (x; y) = E fWt;x jXt;x = ugpt(x; u)du = E fWt;x IfXt;x  ygg: (14)
1
It is obvious that there exist
lim E fDXt;x jXt;x = ygpt (x; y)
y!1
and as EfDXt;x jXt;x =
gpt(x;
) is nonnegative, integrable function it follows that
lim E fDXt;x jXt;x = ygpt (x; y) = 0:
y!1
Relation (7) implies that
E fWt;x IfXt;x  ygg =
@ @
@x EfIfXt;x  ygg = @x PfXt;x  yg;
so relation (14) read as
E fDXt;x jXt;x
@ PfX  yg:
= ygpt (x; y) = @x (15)
t;x

22
Now, as PfXt;x  yg is continuously dierentiable in x, we have for x1 < x2
Z x2
PfXt;x2  yg PfXt;x1  yg = E fDXt;z jXt;z = ygpt (z; y) dz
Z x2 x1Z
y
= E fWt;z jXt;z = ugpt(z; u) dudz;
x1 1
and by Fubini theorem for almost all u function E fWt;
jXt;
= ugpt(
; u) is integrable and
Z x2 Z y Z y Z x2
E fWt;z jXt;z = ugpt(z; u) dudz = EfWt;z jXt;z = ugpt (z; u) dzdu:
x1 1 1 x1
R
Finally, as PfXt;x  yg = y1 pt (x; u) du we have that, for each y 2 R,
Zy Zy Z y Z x2
pt(x2 ; u) du pt (x1; u) du = EfWt;z jXt;z = ugpt(z; u) dzdu
1 1 1 x1
so for almost all y we have
Z x2
pt(x2 ; y) pt(x1 ; y) = E fWt;z jXt;z = ugpt(z; u)dz
x1
hence pt (x; y) is absolutely continuous function in x and
@
@x pt (x; y) = EfWt;x jXt;x = yg pt (x; y): 
As an appendix to the previous theoremR we shall study in detail partial derivative @x@ pt (x; y). Precisely,
t
we shall show that the process t 1E f 0 a(Xs;x ) 1 DXs;x dBs jXt;xg is a backwards martingale.
As fXt;x ; t  0g is a Markov process, for a f 2 Mb and s < t we have
E x ff(Xt )g = E x fEff(Xt )jFsgg = Ex fE Xs ff(Xt s )gg;

where fFt; t  0g is standard ltration generated by fBt ; t  0g, and E x is expectation operator generated
by initial value X0 = x. Now, 1 of Theorem 3 yields
@ 
@x E x ff(Xt )g = Ex ff(Xt )Wt g = E x f(Xt )E fWt jXt g as well as
@ E ff(X )g = @ E E ff(X )g = E E ff(X )g W =
@x x t @x x Xs t s x Xs

t s s

= Ex E Xs ff(Xt s )g EfWs jXsg = Ex f(Xt ) E fWs jXsg ;
hence, for s < t  :
E fWt;x jXt;xg = E E fWs;x jXs;xg jXt;x
If we mark
W~ t;x = E fWt;x jXt;x g
then previous relation could be rewritten as
W~ t;x = EfW~ s;x jXs;xg; s < t:
Let us dene backwards ow of -algebras as
F t;x = fXu;x ; u  tg; t > 0; x 2 R;
23
and let for (x1; : : :; xn) be n-dimensional bounded Borel function. For s < t  t1 <


< tn we have

~ s;x(Xt1 ;x; : : :; Xtn ;x )g = Ex EfW~ s;x (Xt1 ;x; : : :; Xtn ;x )jFtg
Ex fW
=

= E x W~ s;x E f(Xt1 ;x; : : :; Xtn ;x )jFtg =

= E x W~ s;x E Xt;x f(Xt1 t;x; : : :; Xtn t;x )g =

= E x EfW~ s;x jXt;xg EXt;x f(Xt1 t;x; : : :; Xtn t;x)g =

= E x EfW~ s;x jXt;xg (Xt1 ;x; : : :; Xtn ;x ) ;
hence
~ s;x jF t;xg = E fW~ s;x jXt;xg = W~ t;x ;
E fW i.e.
process Z t
W~ t;x = t 1 Ef a(Xs;x ) 1 DXs;x dBs jXt;xg
0
is a backward martingale, relative to ow of -algebras
F t;x = fXu;x ; u  tg; t > 0; x 2 R:
Having in mind the denition of W~ t;x and 3 of Theorem 3, we deduce
Z 1 @
 pt(x; y)dy = Ef W~ t;xg < 1:
1 @x
Moreover, for s < t
Z 1 @
 pt(x; y)dy = EfW~ t;xg = E EfW~ s;x jF t;xg
1 @x
     
 E Ef W~ s;x jF t;xg = E W~ s;x =
Z 1 @
=  p (x; y)dy s
1 @x
We have proved the following Corollary:
Corollary 1. For each t > 0 and x 2 R process
Z t
W~ t;x = t 1 Ef a(Xs;x ) 1 DXs;x dBs jXt;xg
0
is a backward martingale, relative to ow of -algebras

F t;x = fXu;x ; u  tg; t > 0; x 2 R:

Consequently, function Z 1 @
 pt(x; y)dy
1 @x
is nite and decreasing with t. 

24
 POINT CUT AS SPATIAL HOMEOMORPHISM
It is well known [10, Theorem 2.20] that, for each t > 0, the mapping x ! Xt;x is a.s. homeomorphism
of R onto R, hence we can dene inverse random transformation
Xt 1 (
) : R ! R
such that
Xt 1 (Xt;x) = x; x 2 R:
In the next theorem we shall investigate the distribution of random variable Xt 1 (y), i.e. the random
function Xt 1 (
) applied to the point y.
Theorem 4. Let everything be as in Theorem 3. Then:
1 for each t > 0 and y 2 R function E fDXt;
jXt;
= ygpt (
; y) is integrable and
Z +1
E fDXt;x jXt;x = ygpt (x; y) dx = 1; (16)
1
2 for each t > 0 and x; y 2 R
Zx
PfXt 1(y)  xg = EfDXt;z jXt;z = ygpt (z; y) dz; (17)
1
3 probability density
t (y; x) = E fDXt;x jXt;x = yg pt (x; y) (18)
of the random variable Xt 1 (y) satisfy Chapman-Kolmogorov equations:
Z +1
t+s (y; x) = s (y; z) t (z; x) dz (19)
1
for t; s > 0 and x; y 2 R.
Proof. 1 As for each t > 0, Xt;
is a.s. homeomorphism of R onto R and DXt;x , partial derivative in
L2 (
) sense of random variable Xt;x w.r.t. x, is a.s. positive, we have
lim Xt;x = 1;
x!1
and from relation (15) we have, for x1 < x2
Z x2
PfXt;x2  yg PfXt;x1  yg = EfDXt;z jXt;z = ygpt (z; y) dz: (20)
x1
Now, letting x1 ! 1 and x2 ! +1 one can conclude that
Z +1
EfDXt;z jXt;z = ygpt (z; y) dz = 1;
1
what was to be proved in 1 .
2 >From the fact that a.s. Xt;x is homeomorphism and that Xt 1 (y) is its inverse we have the following
equivalence
Xt 1 (y)  x , Xt;x  y;
25
so, using relation (20), with x1 = 1 and x2 = x
Zx
PfXt 1(y)  xg = PfXt;x  yg = 1 PfXt;x < yg = E fDXt;z jXt;z = ygpt (z; y)dz:
1
3 Denote with Pxf
g the probability measure generated by the Markov process fXt;x; t  0g starting
from x. Then for t; s > 0 the Markov property yields
Z +1
PfXt+s;x  yg = PxfXt+s  yg = PzfXs  ygpt (x; z) dz = E fPXt;x fXs;z  ygg:
1
Using relation (15) we have
@ PfX @ 
t+s (y; x) = @x t+s;x  yg = @x E PXt;xfXs;z  yg =
 @ P fX  yg DX =
= E @x Xt;x s;z t;x

= E E Xt;x fDXs jXs = ygps (Xt;x ; y) E fDt;x jXt;xg =
Z +1
= E z fDXs jXs = ygps (z; y) E fDt;x jXt;x = z gpt(x; z) dz =
1
Z +1
= E fDXs;z jXs;z = ygps (z; y) E fDt;x jXt;x = z gpt (x; z) dz =
1
Z +1
= s (y; z) t (z; x) dz;
1
and the theorem is completely proved. 
References
1. N. Bouleau and F. Hirsch, Formes de Dirichlet generales et densite des variables aleatoires reelles sur l'espace de
Wiener., J. Func. Anal. 69 (1986), 229{259.
2. I.I. Gihman and A.V. Skorohod, Stochastic dierential equations, Naukova Dumka, Kiev, 1968. (Russian)
3. S. Kusuoka and D. Stroock, Applications of the Malliavin Calculus. I , Proc of the Taniguchi Inter. Symp. on Stoch.
Analysis. Katata, North-Holland, 1982, pp. 271{308.
4. P. Malliavin, Stochastic calculus of variations and hypoelliptic operators, Proc. Inter. Symp. on Stoch. Di. Equations,
Kyoto 1976, pp. 195{263.
5. D. Nualart, Noncausal stochastic integrals and calculus, Proc. of Workshop on Stoch. Anal. and Rel. Topics. Silivri
1986, 1988, pp. 80{129.
6. D. Nualart and M. Zakai, Generalized stochastic integrals and Malliavin Calculus, Probability Theory and Rel. Fields
73 (1986), 255{280.
7. D. Pantic, Stochastic Calculus on Distorted Brownian Motion, J.Math.Phys. 29 (1988), 207-210.
8. D. Pantic, Stochastic Calculus and Schroedinger Semigroups Admitting Ground State, J. Math. Phys. 31 (1990),
1221-1225.
9. D. Pantic, Toward the Existence of Density for the Distribution of a Brownian Functional, submitted for publication.
10. D. W. Stroock, Topics in stochastic dierential equations, Tata Institute of Fundamental Research, 1982.
11. S. Watanabe, Lectures on stochastic dierential equations and Malliavin Calculus, Tata Institute of Fundamental
Research, 1984.
12. M. Zakai, The Malliavin Calculus, Acta Appl. Math 3 (1985), 175{207.

26
Electronic Journal: Southwest Journal of Pure and Applied Mathematics
Internet: http://rattler.cameron.edu/swjpam.html
ISSN 1083-0464
Vol. 01, 1995, pp. 27{29.
Submitted: July 19, 1995. Published: September 18, 1995.

A LOGICAL ALTERNATIVE TO THE

EXISTING POSITIONAL NUMBER SYSTEM

Robert R. Forslund
Abstract. This article introduces an alternative positional number system. The advantages of this alternative
system over the existing one are discussed, and an illustration of the use of the system to re-interpret apparent
errors in ancient archaeological documents is presented.

(1991) A.M.S. (MOS) Subject Classication Codes. 11, 01.

Key Words and Phrases. positional number system, archaeology, zero, placeholder.
Introduction
To simplify this article, the history of positional number systems will not be discussed, and interested
readers can nd this information in the references listed. In addition, the discussion is restricted to the
set of positive integers; however, the extension to decimal and negative values, if necessary, is easily
accomplished.
The existing positive integer number system in a selected base (b), is represented by the following
mathematical notation :
X
N= (aj  bj ) where 0  aj  (b 1); b > 1; and the sum is over j
with 0  j  (n 1), and aj are the n digits of the positive integer N in base b.
For example:
10203 = 1  104 + 0  103 + 2  102 + 0  101 + 3
Traditionally we have selected base : b = 10, as in the example, and commonly use other bases such as
binary (b = 2), octal (b = 8) and hexadecimal (b = 16). Ancient civilizations introduced other bases such
as 60 used by the Babylonians [1], and 18 and 20 used by the Mayans [1].
The alternative system
There is a problem with the existing number system. Base 1 is illogical since the digits are equal to
zero, and therefore we must select base, b > 1 when dening the system. In addition, the digit integers
starting at one are concrete concepts, while the concept of zero (nothingness) is abstract. Therefore, I
sought a system that did not require the use of zero as a placeholder. As you will see shortly, there is such
a system, and the concept of a zero placeholder adds unnecessary complexity.
As a result, I reasoned as follows. First of all, and most importantly, why not use digits from 1 to b
rather than 0 to (b 1)? In base 10, this simply means introducing a 10th symbol for the digit \ten". Any
Ministry of Natural Resources Centre for Northern Forest Ecosystem Research c/o Lakehead University, 955 Oliver
Road, Thunder Bay, Ontario, Canada P7B 5E1
E-mail Address: mjohnsto@sky.Lakehead.ca
c 1995 Cameron University
27
symbol will do, however the symbol \A" borrowed from the hexadecimal system provides some familiarity.
This is the only change to the notation used in the introduction for N : viz : 1  aj  b and b  1. This
alternative system has no need for a placeholder (zero) as I illustrate below. I will use bolded symbols for
digits in the alternative system to distinguish between the two systems.
In the alternative system (base A), the number 10 is now represented by the digit A. The number
20 now becomes : 1A = 1  A + A, and 90 is 8A. The number 100 now becomes : 9A = 9  A + A,
and 1000 = 99A = 9  A2 + 9  A + A = 900 + 90 + 10, and so on. Notice that, with the exception of
integers that have digits equal to zero in the existing system, all other numbers remain the same as in the
alternative system as long as the digits lie between 1 and (b 1) inclusive. Therefore the smaller the base,
the more the dierences will be between the two systems. Base 2 for example will remain the same as the
existing system only when all digits are 1. An example in base 2 is : 21 = 10101 in the existing system,
and 1221 in the alternative system. In comparison, the number 31 in both systems remains at 11111 in
base 2.
Advantages of the alternative system
Base 1 is now permissible and completely logical and is simply a tally system. For example, 7 =
1111111 (base 1) in the alternative system. Secondly, for number theorists, the re-arrangement of digits
in the existing system causes digits to be lost. For example the reversal of 12300 in the existing system is
321 and two zero digits are \lost". In the alternative system this does not occur. For example, 12300 =
1229A, which is A9221 upon reversal. As this simple example illustrates, this alternative system lends
itself to further research in number theory. In addition, in the existing system, powers of the base are
identied by the number of zeros in the number - for example 10000 = 104 . Powers are similar in the
alternative system and are identied as in this example by 999A = A4 . More generally, in base b, powers
= : : :(b 1)(b 1) : : :(b 1)b. Notice that fewer digits are often needed to represent the numbers -
this too is an advantage, and one of the main reasons positional notation was rst introduced. Also, in
base 10 for example, instead of having 9 numbers with one digit, 90 numbers with 2 digits, and so on,
the alternative system now has the rst 10 numbers having a single digit, the next 100 numbers having 2
digits, and so on. This too is simpler since the number of integers with n digits in base b, is bn rather
than bn b(n 1) in the existing system.
Application of the alternative system to archaeology
It appears that this alternative system would have been a slightly more logical rst step in the develop-
ment of a number system in ancient times. One possible use of this system might be in the re-interpretation
of ancient calculations found in archaeological documents. History tends to identify the introduction of
the zero placeholder into the number system, as an advancement in mathematics. However, the alterna-
tive system, with its inherent simplicity, illustrates that the introduction of zero may have unnecessarily
complicated the use of integers, and numbers in general.
Perhaps by re-interpreting ancient mathematical documents using this alternative system, mathematical
archaeologists will reveal that some of the ancient calculations, previously thought to be in error, really
were correct. I will illustrate using the commonly used digit symbols in base ten, realizing that ancient
documents used other symbols for digits. The ancients may have used 10 symbols (if base 10 was used);
however, the two symbols \10" may really have meant 1A - which is twenty (not ten). In other words, it
may be that the zero symbol represented the digit ten - not zero. To detect such discrepancies, ancient
documents must contain calculations such as addition. For example : 10 + 11 = 31 - this appears to be
incorrect. However, if the symbol \0" actually represents the digit ten (A) and not zero, then the equation
is correct and equivalent to 1A + 11 = 31 = 20 + 11 in the existing system. Another example : 101 +
9 = 100. Once again, this appears incorrect, however, if \0" symbol is re-interpreted as \A", then this
is: 1A1 + 9 = 1AA, which is 201 + 9 = 210 in the existing system. In base 10, these apparent errors
would only exist when the symbol for ten (and interpreted as zero) occurred in the digits of numbers.
Larger bases such as 60 used by the Mesopotamians [1], would be even more void of such \errors",
and therefore it would be harder to determine if the alternative number system was used. I leave it to
28
archaeologists specializing in the interpretation of these ancient documents to examine this usage of the
alternative positional number system.
In summary, the alternative system appears to be a slightly more logical system relative to the existing
one. As illustrated, it has immediate application to the eld of archaeology. Readers should have little
problem nding other elds for the application of this system especially in the theory of numbers. If this
article motivates others to explore the possibilities of this alternative system, then it has succeeded.
References
1. Boyer, C.B. and Merzbach, U.T. : A History of Mathematics (Second Edition), John Wiley & Sons
Inc., New York, 1991.
2. Hellemans, A. and Bunch, B. : The Timetables of Science - A Chronology of the Most Important
People and Events in the History of Science, Simon & Schuster Inc., New York, 1991.
3. Swetz, F.J. : From Five Fingers to Innity, Open Court, Chicago and La Salle Illinois, 1995.
Acknowledgements
I wish to thank Dr. Mark Johnston and Mr. Luke Dalla Bona at the Centre for Northern Forest
Ecosystem Research for providing me with access to the use of electronic communication via the Internet
and for their encouragement and comments regarding this paper.

29
Electronic Journal: Southwest Journal of Pure and Applied Mathematics
Internet: http://rattler.cameron.edu/swjpam.html
ISSN 1083-0464
Vol. 01, 1995, pp. 30-36.
Submitted: May 12, 1995. Published: October 16, 1995.

RESULTS ON CONTROLLING THE RESIDUALS OF

PERTURBED NEWTON-LIKE METHODS ON BANACH
SPACES WITH A CONVERGENCE STRUCTURE

Ioannis K. Argyros
Abstract. In this study, we provide sucient conditions for controlling the residuals of perturbed Newton-like
methods, in such a way that the convergence of such methods is ensured. In order to achieve this, we apply a
theorem due to Kantorovich on a setting of a regular space with a convergence structure.

(1991) A.M.S. (MOS) Subject Classication Codes. 65J15, 65B05, 47D15, 47H17, 49D15.
Key Words and Phrases. Newton-like methods, residuals.
I. INTRODUCTION
In this study, we are concerned with approximating a solution x of the nonlinear operator equation
F(x) + Q(x) = 0; (1)
where F is a Frechet-dierentiable operator dened on a convex subset D of a Banach space X with values
in X, and Q is a non-dierentiable nonlinear operator with the same domain and values in X.
We generate a sequence fxng (n  0) using the perturbed Newton-like method scheme given by
xn+1 = xn + n (n  0) (2)
where the correction n satises
A(xn)n = (F(xn) + Q(xn)) + rn (n  0) (3)
for a suitable rn 2 X (n  0) called residual. The importance of studying perturbed Newton-like methods
comes from the fact that variants of Newton's method can be considered as procedures of this type. In
fact, the approximations (2) and (3) characterize any iterative process in which the corrections are taken
as approximate solutions of the Newton equations. This happens when approximation (2) is solved by any
iterative method or when therein the derivative is replaced by a suitable approximation. In [4] we provided
sucient conditions for the convergence of iteration (2) to a solution x of equation (1), by assuming that
X is a Banach space with a convergence structure (to be precised later).
Here we derive sucient conditions for controlling the residuals rn in such a way that the convergence
of the sequence fxng (n  0) to a solution of equation (1) is ensured.
We also refer the reader to [5], [6], [8] and the references there for relevant work, which however is valid
on a Banach space X without a convergence structure. The advantages of working on a Banach space
with a convergence structure have been explained in some detail in [4], [6], [7].
Department of Mathematics, Cameron University, Lawton, OK 73505, U.S.A.
E-mail Address: ioannisa@cua.cameron.edu
c 1995 Cameron University
30
 II. PRELIMINARIES
We will need the denitions:
Denition 1. The triple (X; V; E) is a Banach space with a convergence structure if
(C1 ) X; k
k) is a real Banach space;
(C2 ) (V; C; k kv) is a real Banach space which is partially ordered by the closed convex cone C; the norm
k kv is assumed to be monotone on C;
(C3 ) E is a closed convex cone in X  V satisfying f0g  C  E  X  C;
(C4 ) the operator j j : D ! C is well dened:
jxj = inf fq 2 C j(x; q) 2 E g
for
x 2 D = fx 2 X j9q 2 C : (x; q) 2 E g;
and
(C5 ) for all x 2 D kxk  kjxjkv.
The set
U(a) = fx 2 X j(x; a) 2 E g
denes a sort of generalized neighborhood of zero.
Denition 2. An operator L 2 C 1(V1 ! V ) dened on an open subset V1 of an ordered Banach space
V is order convex on [a; b]  V1 if
c; d 2 [a; b]; c  d ) L0 (d) L0 (c) 2 L+ (V );
where for m  0
L+ (V m ) = fL 2 L(V m )j0  xi ) 0  L(x1 ; x2; : : :; xm )g
and L(V m ) denotes the space of m-linear, symmetric, bounded operators on V .
Denition 3. The set of bounds for an operator H 2 L(X m ) is dened to be
B(H) = fL 2 L+ (V m )j(xi ; qi) 2 E ) (H(x1; : : :; xm ); L(q1; : : :; qm)) 2 E g:
Denition 4. A partially ordered topological space V is called regular if every order bounded increasing
sequence has a limit.
We will need the following proposition due to Kantorovich [3], [5].
Proposition. Let V be a regular partially ordered topological space (POTL-space) and let x, y be two
points of V such that x  y. If H : [x; y] ! V is a continuous isotone operator having the property that
x  Hx and y  Hy, then there exists a point z 2 [x; y] such that z = Hz.
III. CONVERGENCE ANALYSIS
We will need the following basic result:
Lemma. Let V be a regular partially ordered topological space, an operator L 2 C 1(V1 ! V ) with
[0; a]  V1  V for some a 2 V , an operator M 2 C(V1 ! V ), operators R < I, B > I, T, K 2 L+ (V ), a
point c 2 V with c > 0 and a point p 2 [0; a].
Assume:
(a) The equation

g(q) = BT L(p + q) L(p) L0(p)q + M(p + q) M(p) + K(p + q) K(p)

(I R)q + c = 0 (4)
31
 has solutions in the interval [0; a] and denote by q the least of them.
(b) Let G 2 L+ (V ) be given and R+ 2 L+ (V ), c+ ; p+ 2 V be such that the following conditions are
satised:

R+ < min (G 2I)BTL0 (p) + BTGL0 (p+ ) + G(R I) + I; I = 
(5)
0 < c+  BT G(L(p+ ) + M(p+ ) + K(p+ )) + GBT(L(p) + M(p) + K(p))
2BT(L(p) + M(p) + K(p)) + (R+ + G I BTGL0 (p+ ))c =  (6)
and
0  p+  p + c; (7)
where  and  are functions of the operators and points involved.
(c) The following estimate is true
M(p)  M(p + q) (8)
for all p; q 2 [0; a].
Then the equation

g+ (q) = BGT L(p+ + q) L(p+ ) L0 (p+ )q + M(p+ + q) M(p+ ) + K(p+ + q) K(p+ )

(I R+ )q + c+ = 0 (9)
 
has nonnegative solutions and the least of them, denoted by q+ lies in the interval [c+ ; q c].
Proof. Using the hypothses g(q ) = 0 and R < I we deduce from (4) that c  q . We will show that
g+ (q c)  0: (10)
From equation (9) and using (6), we obtain in turn

g+ (q c)  BGT L(p + c + q c) L(p+ ) L0 (p+ )(q c)

+ M(p + c + q c) M(p+ ) + K(p + c + q c) K(p+ ) (I R+ )(q c) + c+

= g(q ) BT L(p + q ) L(p) L0 (p)q + M(p + q ) M(p) + K(p + q ) K(p)


+ BT G L(p + q ) L(p+ ) L0 (p+ )(q c) + M(p + q ) M(p+ ) + K(p + q ) K(p+ )

+ (I R)q c (I R+ )(q c) + c+

= BT L0 (p) BTGL0 (p+ ) + (I R) (I R+ ) q

 
+ (G I) (I R)q c BT L(p) + M(p) + K(p) + L0 (p)q


+ BT L(p) + M(p) + K(p) GL(p+ ) GM(p+ )

+ BT GL0 (p+ )c + (I R+ )c + c+ c

= (2I G)BT L0 (p) BTGL0 (p+ ) + R+ + G(I R) I q

+ 2BT(L(p) + M(p) + K(p)) BTG(L(p+ ) + M(p+ ) + K(p+ ))
GBT(L(p) + M(p) + K(p)) + (BTGL0 (p+ ) G + I R+ )c + c+  0
since (5) and (6) are satised.
Moreover from (9) for q = c+ , we obtain
g+ (c+ )  0: (11)

By inequalities (10), (11), the fact that g is continuous and isotone on [c+ ; q c], and since V is a regular
partially ordered topological space, from the proposition, we deduce that there exists a point q+ with
g+ (q+ ) = 0 (12)
and
c+  q+  q c: (13)

We can assume that q+ denotes the least of the solutions of equation (9).
That completes the proof of the lemma.
The following result is a consequence of the above lemma.
32
 Theorem 1. Let fcng 2 V , fTng, fRng, fGng 2 L+ (V ) (n  0) be sequences and V as in the above
lemma. Assume:
(a) There exists a sequence fpng 2 [0; a]  V1  V for some a 2 V with p0 = 0 and
X
pn+1  cj for n  0: (14)
j =0;1;:::;n
(b) R0 < I and the function

g0 (q) = BT0 L(p0 + q) L(p0 ) L0 (p0)q + M(p0 + q) M(p0 ) + K(p0 + q) K(p0 )

(I R0)q + c0 = 0 (15)
has roots on [0; a], where B, L, M, K are as in the above lemma. Denote by q0 the least of them.
(c) The following conditions are satised for all n  0
Rn+1 < n+1; (16)
0 < cn+1  n+1 (17)
and
0  pn+1  pn + cn : (18)
(d) The linear operators Tn are boundedly invertible for all n  0, and set Gn = Tn+1 Tn (n  0).
1
(e) Condition (8) is satised.
Then, the equation

gn(q) = BGn Tn L(pn + q) L(pn) L0 (pn )q + M(pn + q) M(pn) + K(pn + q) K(pn)

(I Rn)q + cn = 0 (19)
has solutions in [0; a] for every n  0 and denoting by qn the least of them, we have
X
cj  qn (n  0): (20)
j =n;:::;1
Proof. Let us assume that for some nonnegative integer n, I Rn > 0, gn(q) has roots on [0; a] and denote
by qn the least of them. We use induction on n. We also observe that this is true by hypothesis (b) for
n = 0. Using (14), (16), (17), (18), the lemma, and setting c = cn, c+ = cn+1, R = Rn, R+ = Rn+1,
G = Gn we deduce that qn +1 exists, and
cn+1  qn +1  qn cn : (21)
The induction is now complete and (20) follows immediately from (21).
That completes the proof of the theorem. From now on we assume that X is a Banach space with a
convergence structure in the sense of [4].
The following result is an immediate consequence of Theorem 1.
Theorem 2. Assume:
(a) the hypotheses of Theorem 1 are satised;
(b) there exists a sequence fxng (n  0) in a Banach space X with a convergence structure such that
jxn+1 xn j  cn. Then
(i) the sequence fxng (n  0) converges to some point x;
(ii) moreover the following error bounds are true
jx xn j  qn (22)
and
jx xn+1j  qn cn for all n  0: (23)
We can introduce the main result:
33
 Theorem 3. Let X be a Banach space with convergence structure (X; V; E) with V = (V; C; k kv), an
operator F 2 C 1(D ! X) with D  X, an operator Q 2 C(D ! X), an operator A(x) 2 C(X ! D), an
operator L 2 C 1 (V1 ! V ) with V1  V , an operator M 2 C(V1 ! V ), an operator K 2 L+ (V ), and a
point a 2 C such that the following conditions are satised:
(a) the inclusions U(a)  D and [0; a]  V1 are true;
(b) L is order-convex on [0; a] and satises
K + L0 (jxj + jyj) L0 (jxj) 2 B(A(x) F 0(x + y)) (24)
for all x; y 2 U(a) with jxj + jyj  a;
(c) M satises the condition
0  (Q(x) Q(x + y); M(jxj + jyj) M(jxj) 2 E; M(0) = 0 (25)
for all x; y 2 U(a) with jxj + jyj  a;
(d) for the sequences fcng, fTn g, fRng, fGng, fpn g (n  0) the hypotheses (14), (b), (16), (18) and (d)
of Theorem 1 are satised;
(e) the following conditions are also satised
jnj  cn  Tn j (F(xn) + Q(xn))j  n  n (n  1); (26)
j rn j  Tn Rncn
1 (27)
where

n = Tn L(pn +cn ) L(pn ) L0 (pn)cn +M(pn +cn ) M(pn )+K(pn +cn ) K(pn ) +Rncn (n  1):

(28)
Then
(i) the sequence fxng (n  0) generated by
xn+1 = xn + n with x0 = 0
remains in U(x0; t0 ) and converges to a solution x of equation F(x) = 0;

(ii) moreover, the error estimates (22) and (23) are true where qn is the least root in [0; a] of the function
gn(q) dened in (19), with pn = kxn x0k (n  0).
Proof. Let us assume that xn ; xn+1 2 U(x0; q0 ), where the existence of q0 is guaranteed from hypotheses
(d). We note that j0j  c0 . Using the approximation
(F(xn+1) + Q(xn+1)) = (F(xn) F(xn+1) + A(xn)(xn+1 xn)) + (Q(xn) Q(xn+1 )) rn; (29)
(24), (25), (27) and setting pn = jxn x0j, we obtain in turn
j (F(xn+1) + Q(xn+1 ))j  jF(xn) F(xn+1) + A(xn )(xn+1 xn)j + jQ(xn) Q(xn+1 )j + j rn j
Z 1 
 L0 (pn + tjxn+1 xnj) L0 (pn ) + K jxn+1 xnj dt + M(pn + jxn+1 xnj) M(pn) + j rnj
0
 L(pn + cn ) L(pn ) L0 (pn )cn + Kcn + M(pn + cn) M(pn ) + Tn 1Rncn :
Hence, by (26) we get
cn+1  Tn+1 j (F(xn+1) + Q(xn+1 ))j  n  n (n  1);
which shows (17).
It can easily be seen that by using induction on n, the hypotheses of Theorem 2 are satised. Hence,
by (24) and (27) the iteration fxng (n  0) remains in U(x0 ; q0) and converges to x so that (22) and
(23) are satised. Moreover, from the estimate
j (F(xn) + Q(xn ))j  jA(xn) F 0(xn)jcn + jF 0(xn)jcn + j rnj;
(24), (27), the continuity of F, F 0, A, Tn , Rn and cn ! 0 as n ! 1, we deduce that
F(x) + Q(x ) = 0:
That completes the proof of the theorem.
34
 IV. APPLICATIONS
We will complete this study with an example that shows how to choose L, M, in practical applications.
For simplicity, we assume that A(x) = F 0(x) for all x 2 D.
Example. We discuss the case of a real Banach space with norm k k. Assume that F 0(0) = I and there
exists a monotone operator
f : [0; a] ! IR
such that
kF 00 (x)k  f(kxk) for all x 2 U(a) (30)
and a continuous nondecreasing function ` on [0; r], r  a such that
kQ(x) Q(y)k  `(r)kx yk (31)
r
for all x; y 2 U( 2 ).
We showed in [3], (see also [9]) that (18) implies that
kQ(x + e) Q(x)k  h(r + kek) h(r); x 2 U(a); kek  a r; (32)
where Zr
h(r) = `(t) dt: (33)
0
Conversely, it is not hard to see that we may assume, without loss of generality, that the function h
and all functions h(r + t) h(r) are monotone in r. Hence, we may assume that h(r) is convex and hence
dierentiable from the right. Then, as in [9], we show that (32) implies (31) and `(r) = h0(r + 0). Hence,
we can set Zq Zs
L1 (q) = L(q) = kF(0) + Q(0)k + ds f(t) dt (34)
0 0
and Zq
M(q) = `(t) dt: (35)
0
Dene the functions f1 , f2 , f3 on [0; a] by
Zq
f1 (q) = L(q) q; f2 (q) = f1 (q) + h(q) and f3 (q) = f(t) dt:
0
Choose B = T0 = 1, R0 = K = 0 and p0 = 0. Then by (15) we get
g0(q) = f2 (q):
It can easily be seen that with the above choices of L and M conditions (24) and (25) are satised.
Suppose that the function g0 has a unique zero q0 in [0; a] and g0 (a)  0. It is then known [3], [9] that
equation (1) admits a solution x in U(q0 ), this solution is unique in U(a), and the iteration fxn g (n  0)
given by (2) is well dened, remains in U(q0 ) for all n  0 and converges to x. By applying the Banach
lemma on invertible operators we can show that A(xn ) is invertible and kA(xn ) 1k  f10 (kxnk) 1 =
Tn (n  0).
Moreover, suppose that (30) is satised. Furthermore, assume that instead of conditions (16) and (17),
the weaker condition (10) is satised. Using (30) and the approximation
 
rn = (F 0(xn ) F 0 (x0)) + F 0(x0 ) n + F(xn) + Q(xn)
we obtain

krn k  f3 (kxnk) + 1 cn + Tn 1 n (n  1):
The above estimate provides us with a possible (but not the only) choice for the Rn's. Indeed, (27) will
be true if the Rn's (n  0) can be chosen so that
(f3 (kxnk) + 1)cn + Tn 1 n  Tn 1 Rncn (n  1):
Using the above choices and Theorem 3, it nally also follows that estimates (22) and (23) are satised
for all n  0.
35
 REFERENCES
[1] Argyros, I.K. The secant method in generalized Banach spaces, Appl. Math. Comp. 39, (1990),
111{121.
[2] . Newton-like methods in generalized Banach spaces, Functiones et Approximatio
Commentarii Mathematici, XXII, (1993), 13{20.
[3] Argyros, I.K. and Szidarovszky, F. The Theory and Applications of iteration methods, C.R.C. Press,
Inc. Boca Raton Florida, U.S.A. 1993.
[4] Argyros, I.K. Perturbed Newton-like methods and nondierentiable operator equations on Banach
spaces with a convergence structure, Southwest Journal of Pure and Applied Mathematics (SWJPAM)
Vol. 1, No. 1 (1995), 1{12.
[5] Kantorovich, L.V. and Akilov, G.P. Functional Analysis in normed spaces, Academic Press, New
York, 1978.
[6] Meyer, P.W. Newton's method in generalized Banach spaces, Numer. Funct. Anal. Optimiz. 9 (3
and 4), (1987), 249{259.
[7] . A unifying theorem for Newton's method, Numer. Funct. Anal. Optimiz. 13 (5
and 6), (1992), 463{473.
[8] Moret, I. A Kantorovich-type theorem for inexact Newton methods, Numer. Funct. Anal. and
Optimiz. 10 (3 and 4), (1989), 351{365.
[9] Zabrejko, P.P. and Nguen, D.F. The majorant method in the theory of Newton-Kantorovich approxi-
mations and the Ptak error estimates, Numer. Funct. Anal. Optimiz. 9 (5 and 6), (1987), 671{684.

36
Electronic Journal: Southwest Journal of Pure and Applied Mathematics
Internet: http://rattler.cameron.edu/swjpam.html
ISSN 1083-0464
Vol. No. 01, 1995, pp. 37-45.
Submitted: June 8, 1995. Published: October 26, 1995.

A PROBABILISTIC APPROACH TO THE ESTIMATION OF THE DISTRIBUTION

OF THE TARGET LOCATION POINT

M.A. Tabatabai
Abstract. A probabilistic approach is used to estimate the probability distribution of the target location
point. The results can be used in missile-testing and target acquisition re support model simulation. The
target may be a mobile multiple launcher rocket, an antiaircraft gun or a ship in mid-ocean. In each case, there
is a time lag between the detection of the target and the arrival of the projectile.

(1991) A.M.S. (MOS) Subject Classication Codes. 60, 62.

Key Words and Phrases. Probability density function, Gamma function.
I. Introduction

The target location problem frequently appears in missile-testing and artillery eld programs. In this
article, interest focuses on the probability distribution of the target location point for a multiple launcher
rocket and a Howitzer. Our main concern is to nd the probability density function of a mobile MLR
location which is deliberately moving so as to confound prediction of its position. Let us now describe
the situation by calling the two parties Red and Blue. The mission of the Red multiple launcher rocket is
to strike by ring and move as quickly as possible in a specic direction with an average speed R. The
reaction of the Blue commander is to predict the target location and launch a round of artillery at an
anticipated target location.
Let the initial target location point (the point at which the Red initially res) be represented by
a bivariate normal random vector (X; Y ), where X denotes the horizontal coordinate, Y denotes the
vertical coordinate and X and Y are assumed to be independent random variables with a mean vector
(X ; Y ) = (0; 0)
and a standard deviation vector
(X ; Y ) = (100; 100)
where the mean and standard deviation for both X and Y are in meters.
Let T denote the Blue artillery reaction time. This random variable is the total response time from
the Red multiple launch rocket ring to the rst rounds of the Blue artillery landing on target. The
distribution of the random variable T will be derived in Section 2. The Red displacement time using
daylight values ranges from 1.5 minutes to 2 minutes.
Department of Mathematical Sciences, Cameron University, Lawton, OK 73505
E-mail Address: mohammad@tiger.cameron.edu
c 1995 Cameron University
37
 II. Distribution of Reaction Time

A sample of 103 response times from the Red ring to the rst rounds of the Blue artillery landing
on target is included in the analysis. The data was provided by the Fire Support Center at Cameron
University. The goal of this section is to nd the distribution of Blue reaction time.
Before making a specicaiton of the distribution of response time, it is a good practice to look at the
data in a graphical form in order to see if they could have come from one of a known family of distributions.
One of the simplest ways of looking at the data is to make a stem-and-leaf plot. Figure 1 displays the
stem-and-leaf plot for Blue reaction time (in minutes). In this display we immediately see the asymmetry
of the main part of the data. The distribution is positively skewed and the gure indicates the presence
of some extreme outliers. In Figure 2, we have the Boxplot. It provides a visual impression of several
important aspects of the underlying distribution of the Blue reaction time. From a Boxplot we can pick
out the following features of a data set:
Location
Skewness
Spread
Tail length
Outlying data points
Stem-and-leaf of Blue Re N=103
Leaf Unit=0:10
2  1246777789999
3  00001112222233333444555566666666666777777888999
4  001111222334444446668899
5  00014
6  2
7  24578
8 
9  79
10  49
11  44
12  1
13 
14 
15 
16 7

Figure 1. Stem-and-leaf Plot of Blue Reaction Time

The mean reaction time for this data set is 4.554 minutes with a standard deviation of 2.396 minutes.
The data ranges from a low of 2.160 minutes to a high of 16.770 minutes. The rst, second and third
quartiles are 3.300, 3.740, and 4.630 respectively and the trimmed mean is 4.238 minutes.

38
 I+ I  0 00 0 0 0
+ + + + +
3: 0 6: 0 9: 0 12: 0 15: 0

Figure 2. Boxplot of Blue Reaction Time

Let T be the Blue reaction time and M be the Red movement time. More specically, M is the dierence
between Blue reaction time and Red displacement time. If we use 2 minutes for the Red displacement
time, then we can write the random variable M as M = T 2. Figure 3 displays the stem-and-leaf graph
for Red movement time.

0  1246777789999
1  00001112222233333444555566666666666777777888999
2  001111222334444446668899
3  00014
4  2
5  24578
6 
7  79
8  49
9  44
10  1
11 
12 
13 
14 7

Figure 3. Stem-and-leaf Plot for Red Movement Time

To have a more reliable estimate for the parameters of underlying distrubtion of random variable M,
we examine the Boxplot for the Blue reaction time. Figure 4 shows the Boxplot for M.

0
I + I  000 00 0 0
+ + + + + +
0: 0 3: 0 6: 0 9: 0 12: 0 15: 0

Figure 4. Boxplot of Red Movement Time

The small hollow circles in the Boxplot indicate the presence of eight extreme outliers. There is a
widespread awareness of the dangers posed by the occurrence of extreme outliers, which may be a result
39
of keypunch errors, recording or transmission errors, misplaced decimal points, exceptional phenomena
such as a lack of adequate training in ring rockets, or members of a dierent population slipping into
the sample. Outliers occur frequently in real data. When the data contain outliers, one approach is
to remove the outliers and then calculate your sample statistics. A second approach is to use robust
and resistant estimators to estimate the population parameters. Robust estimators limit the in uence of
extreme observations and are insensitive to outliers. For more details about robust statistical techniques,
see Hoaglin, Mosteller, and Tukey (1983); Roussaauw and Leroy (1987); Tabatabai and Argyros (1993).
In this article we rst remove the extreme outliers and then calculate some statistics.
Table 1 shows a summary of some relevant estimates of the parameters of Red movement time.
N MEAN MEADIAN TRMEAN STDEV SEMEAN
95 1.960 1.690 1.838 1.124 0.115
MIN MAX Q1 Q3
0.160 5.850 1.270 2.410

Table 1. Summary of Red Movement Time

Let us now turn our attention to the theory underlying estimation of the density function of M. Let
the range be partitioned into k class intervals Ij ; j = 1; 2; : : :; k and Ij = (aj 1 ; aj ); j = 1; 2; : : :; k where
a0 < a1 <


< ak . Here we assume that aj aj 1 = 1 minute for j = 1; 2; : : :; k. Dene
(  
f = n(aj aj 1) for aj 1 < m  aj ; j = 1; 2; : : :; k
f^M (m) = j
0; elsewhere
Pk
where fj is the frequency of the class interval Ij and n = fj is the total number of observations. The
j =1
function f^M dened on ( 1; 1) is the relative frequency histogram for M. We note that
f^M (m)  0 for all m 2 ( 1; 1); (1)
Z1 Z ak X k Z aj fj X
k f
j
f^M (m) dm = f^M (m) dm = n(aj aj ) dm = n = 1: (2)
1 a0 j =1 aj 1 1 j =1
Hence f^M denes a density function and can be used as an estimate of the underlying density function
fM . In particular, an estimate of the probability of an event A is given by
Z
^ =
P(A) f^M (m)dm:
A
^ is an unbiased and consistent estimate of P(A).
Moreover, P(A)
For our data on Red movement time, the estimate of probability distribution of M is given by

40
 m f^M (m) m f^M (m)
.5 13/103 8.5 2/103
1.5 47/103 9.5 2/103
2.5 24/103 10.5 1/103
3.5 5/103 11.5 0/103
4.5 1/103 12.5 0/103
5.5 5/103 13.5 0/103
6.5 0/103 14.5 1/103
7.5 2/103
where mj = (aj 1 + aj )=2 is the class mark of class interval Ij . The mean of f^M (m) is given by
Z ak X
k Z aj mfj X 
k f a +a  X
k
^M = mf^M (m) dm = n(a a ) dm = n
j j
2
j 1 = fj mj :n
a0 j =1 aj 1 j j 1 j =1 j =1

So far we have answered th following questions: What does the underlying probability density function
fM look like? How do we estimate fM ? The next question that one needs to answer is: Does fM look
like a known or specied distribution? The probability density function of M is Gamma distribution with
parameters  and . We have
8 m 1 exp( m=)
< ; m>0
fM (m) = : () 
0; m0
R1
where () = x 1e x dx.
0
The mean M =  and standard deviation M =  p. To estimate the parameters  and , we use
the method of moments technique. Hence, the moment estimators ^ and ^ are
(sample mean)2 and ^ = sample mean :
^ = sample variance ^
If we use Table 1, then the moment estimates are
2
^ = (1:960)
(1:124)2 = 3:0419 and ^ = 1:960 = 0:6443:
 3:0419

Now, we use the chi-square goodness-of-t test to test the following hypothesis?
3:0419 1 exp( m=0:6443)
H0 : fM (m) = m (3:0419)(0:6443) 3:0419
vs 3:0419 1 exp( m=0:6443)
H1 : fM (m) 6= m (3:0419)(0:6443)3:0419 :

The chi-square test statistic is 4.9153. The associated degrees of freedom is 3 and the P value is
0.1781. Therefore, we have to conclude that the Gamma distribution with parameters  = 3:0419 and
 = 0:6443 does provide an excellent t for the data. Similar conclusions were reached when we used the
Kolmogorov-Smirnov test.
41
 The maximum likelihood estimates of  and  are solutions of the following equations
d  1X n m
ln d () = () = lnm n j =1 ln mj and  = 
where m is the sample mean.
These type of equations are not easily solvable and it is necessary to resort to numerical methods.
Let FT (t) denote the distribution function of Blue reaction time, then we have
FT (t) = P(T  t) = P(M + 2  t)
Zt 2
= P(M  t 2) = fM (m) dm
1
Z t 2 m 1 exp( m=)
= ()  dm:
1

Then, the probability density function fT (t) of Blue reaction time is given by
8 (t 2) 1 exp( (t 2)=)
< ; t>2
fT (t) = dtd FT (t) = : () 
0; elsewhere
where the values of  and  can be estimated by 3.0419 and 0.6443 respectively. Then mean for Blue
reaction time is estimated by ^T = 3:960 and the standard deviation is estimated by ^T = 1:124.

III. Distribution of Target Location Point

Let the random vector (D1 ; D2 ) be the target location point and be dened by
D1 = X + RM cos 
D2 = Y + RM sin
where R is the average speed of Red, M is the Red movement time,  is the Red angle of movement and
X; Y are independent normal random variables with X = Y = 0 meters and X = Y = 100 meters.
Our goal is to nd the joint probability density function of (D1 ; D2). We begin this section by rst
nding the probability distribution of RM.
Let W = RM and FW (w) denote the distribution function of the random variable W. Then
FW (w) = P(W  w) = P(RM  w) = P(M  R1 w)
Z R1 w Z R1 w m 1 exp( m=)
= fM (m) dm = ()  dm:
1 1
Then the probability density function fW (w) is given by
8  1 exp(
d <W w=(R)) ; W >0
fW (w) = dW FW (w) = : ()(R)
0; elsewhere:
42
Thus, we can write
D1 = X + W cos 
D2 = Y + W sin 
where the random variable  is uniformly distributed on the interval [0; 2) with the probability density
function g() dened by 81
< 0   < 2
g() = : 2
0; elsewhere
and the random variables X, Y , , and W are independent reandom variables with joint probability
density function given by
f(x; y; ; w) = fX (x) fY (y) g() fW (w)
8
>
>   1 <y < 1
>
< w 1 exp w=(R)2 12 (x2=100 2 + y2 =1002) ; 1 <x < 1
=> 40000 ()(R) 0   < 2
>
> 0 <w < 1
: 0; elsewhere
Let
D1 = g1(X; Y; ; W ) = X + W cos ;
D2 = g2(X; Y; ; W ) = Y + W sin ;
D3 = g3(X; Y; ; W ) = ;
D4 = g4(X; Y; ; W ) = W;

A = f(x; y; ; w) : 1 < x < 1; 1 < y < 1; 0   < 2; 0 < w < 1g;
B = f(D1 ; D2 ; D3; D4) : 1 < D1 < 1; 1 < D2 < 1; 0  D3 < 2; 0 < D4 < 1g:

1 1
Dene the transformation g : A !B
onto
as
g = (g1(x; y; ; w); g2(x; y; ; w); g3 (x; y; ; w); g4(x; y; ; w))
= (D1 ; D2; D3 ; D4)
1 1
then the inverse transformation is g 1 : B !A
onto
which can be written as

g 1 = (g1 1 (D1 ; D2 ; D3; D4 ); g2 1 (D1 ; D2; D3 ; D4); g3 1 (D1 ; D2; D3 ; D4); g4 1 (D1 ; D2 ; D3; D4 ))
= (x; y; ; w)
where
X = D1 D4 cos D3
Y = D2 D4 sinD3
 = D3
W = D4 :
43
Let  @x @x 
 x; y; ; w   @D@y1


@D4 
@y 
J D ; D ; D ; D =  @D 1


@D4 
@D4 
@


@
1 2 3 4  @w1
@D



@w 
@D1 @D4
be the Jacobian of inverse transformation g 1 . Then
 cos D3 
 x; y; ; w   01 0 D4 sin D3
1 D4 cos D3 sinD3  = 1:
J D ; D ; D ; D =  0 0 1 0 
1 2 3 4 0 0 0 1 

The joint probability density function of random variables D1 ; D2; D3 ; D4 is denoted by h(D1 ; D2; D3; D4 )
and is given by


h(D1 ; D2; D3; D4 ) = jJ j
f g1 1 ; g2 1; g3 1 ; g4 1
8 1   
>
< D4 exp D4 =(R) D12 + D22 + D42 2D4(D1 cos D3 + D2 sin D3) =20000 ; (D1; D2; D3; D4) 2 B
=> 400002 ()(R)
: 0; elsewhere.

Finally, the distribution of the target location point is denoted by h(D1 ; D2) and is given by
h(D1 ; D2)
8  2 2
2 1
>
> exp D1 + D2 =20000 Z Z exp  D2 2D (D cos D + D sinD ) + 20000 D  =20000 dD dD ;
>
> 4 4 1 3 2 3
R 4 4 3
< 400002(R) () 0 0
= 1 <D1 < 1
>
>
> 1 <D1 < 1
: 0; elsewhere.

IV. Some Special Cases

Case 1. Suppose the initial point (X; Y ) is predetermined and is equal to X = a, and Y = b. Then
D1 = a + W cos 
D2 = b + W sin

1 1
Dene the transformation S : A1 ! B1
onto
as

S = (S1 (w; ); S2(w; )) = (D1 ; D2 )

where
A1 = f(w; ) : 0 < w < 1; 0   < 2g
44
and
B1 = f(D1 ; D2 ) : 1 < D1 < 1; 1 < D2 < 1g:

Then the inverse transformation S 1 : B1 ! A1 is a one-to-one and onto function dened by

S 1 = (S1 1 (D1 ; D2); S2 1 (D1 ; D2 )) = (W; )
where
h i 21
W = (D1 a)2 + (D2 b)2
and

D2 b  :
 = arctan D
1 a

The Jacobian of inverse transformation is

 W;    @W @W 
J D ; D =  @D @D2  = 1
@1 @  h i1 :
1 2 @D1 @D2 (D1 a)2 + (D2 b)2 2

The joint probability density function of the target location point conditioned upon knowledge of
(X; Y ) = (a; b) is denoted by h1 (D1 ; D2) and is given by
  
h1(D1 ; D2 ) = J DW;; D 
fW (S1 1 )
g(S2 1 )
8 1 2

< (D1 a) + (D2 b)2 2 1
exp n (D1 a)2 + (D2 b)2 12 =(R)o ; if (D1; D2) 2 B1
> 2
=> 2(R) ()
: 0; elsewhere.

Case 2. If  = 0, then the joint probability density function of D1 and D2 is denoted by h0 (D1 ; D2) and
is equal to
h0(D1 ; D2 )
8 exp  D2 + D2 Z 1    
>
< 1 2  1
D exp D42 2D1D4 + 20000
1 <D1 < 1
= > 20000(R) () 0 4 R D4 =20000 dD4 ; 1 <D2 < 1
: 0; elsewhere.

In general, if  = C, then the joint probability density function of D1 and D2 is denoted by hC (D1 ; D2 )
and is given by
hC (D1 ; D2 )
8 exp  D2 + D2 Z 1    
>
> 1 2 
D exp1 2 20000
D4 2D4(D1 cos C + D2 sinC) + R D4 =20000 dD4;
>
< 20000(R) () 0 4
=> 1 <D1 < 1
>
> 1 <D2 < 1
: 0; elsewhere.

45
 References

[1] Hoaglin, D.C., Mosteller, F., Tukey, J., Understanding Robust and Exploratory Data Analysis, John Wiley & Sons, Inc.
(1983).
[2] Rousseeuw, P.J., and Leroy, A.M., Robust Regression and Outlier Detection, John Wiley & Sons, Inc. (1987).
[3] Tabatabai, M.A., and Argyros, I.K., Robust Estimation and Testing for General Nonlinear Regression Models, Applied
Math and Comp 58/1 (1993), 85{101.

46
 Information for Authors cont.
Copyright. Authors will be asked, upon acceptance of an article, to transfer copyright of the article to
the publisher. This will ensure the widest possible dissemination of information under copyright laws.
Preprints are available via anonymous ftp at ftp://rattler.cameron.edu/SWJPAM-preprints. Authoriza-
tion to copy is granted provided that the SWJPAM copyright symbol is not removed.

Tables. Tables must be numbered consecutively with roman numerals and have titles. They should be
prepared in such a manner that no break is necessary.

Figures. All graphs and diagrams should be referred to as `gure' (not abbreviated) in the text. A
detailed caption should be provided for each gure, and gures should be numbered consecutively.

Footnotes. The use of footnotes should be minimized.

Notication. The author will be notied of the posting of his/her work and the time of release for
SWJPAM.

References. References should be listed at the end of each paper. They should be numbered consecutively
using arabic numbers between square brackets in the text. In the list of references, the numbers should
appear without square brackets. A journal reference should comprise of name(s) and initial(s) of author(s),
title of article, abbreviated name of journal (following the style of `Abbreviations of Names of Serials
reviewed in Mathematical Reviews'), volume number, year of publication (within parentheses), and full
page numbers, e.g.
1. Argyros, I.K. and Szidarovszky, F.: On the Monotone Convergence of Algorithmic Models, Appl.
Math. Com. 48, (1992), 167{176.
A book reference should comprise of name(s) and initial(s) of author(s), title of book, publisher, town,
year of publication, e.g.
11. Ortega, J.M. and Rheinboldt, W.C.: Iterative Solution of Nonlinear Equations in Several Variables,
Academic Press, New York, 1970
An edited volume reference should comprise of name(s) and initial(s) of author(s), title of paper, name(s)
of editor(s), title of book, publisher, town, year, inclusive page numbers, e.g.
12. Prenter, P.: On Polynomial Operators and Equations, L.B. Rall (ed), Computational Solution of
nonlinear operator equations, Johy Wiley Publisher, New York, 1971, pp. 361{398.

Indexing and Abstracting services. Articles published in SWJPAM are selectively indexed or ab-
stracted in: Mathematical Reviews of A.M.S., Zentralblatt fur Mathematik, European Mathematical So-
ciety (EMIS).
 SOUTHWEST JOURNAL
OF
PURE AND APPLIED MATHEMATICS
VOLUME 1, JULY{DECEMBER 1995
CONTENTS
I. Argyros Perturbed Newton-Like Methods and Nondierentiable Operator Equations
on Banach Spaces with a Convergence Structure . . . . . . . . . . 1{12
G. Anastassiou Dierentiated Shift-Invariant Multivariate Integral Operators . . . . . . 13{16
D. Pantic Stochastic Calculus and Initial Value Analysis on the One Dimensional
Diusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 17{26
R. Forslund A Logical Alternative to the Existing Positional Number System . . . . 27{29
I. Argyros Results on Controlling the Residuals of Perturbed Newton-Like Methods
on Banach Spaces with a Convergence Structure . . . . . . . . . . 30{36
M. Tabatabai A Probabilistic Approach to the Estimation of the Distribution of the
Target Location Point . . . . . . . . . . . . . . . . . . . . . . 37{46

ISSN 1083-0464