.I
セ@
..
,セ@
,/.
,.:. ...
ᄋ セ@
/
...
/
',•
This tJtle · •• • ·
·-
:,..,,ゥ ウ QQ、ヲL セエZZ@
• • 'in Scopus ·: • ,. セᄋ@
• / ;. :.. • Zセ@ ,• ../ r :. '
--:- -
.
.
--
,...
:""SC:JPU• ,. , ... セM 1,.
,, . . ! · '
r・ーイゥセエ
O・セ セ ヲゥZッュ@
エィ ᄋ・ セᄋ M
.. :·.
·:
:.
J • •
,,
.,
<· ,
·•
($t·
•
.
, /
セ@ ..
.. , ,. :I,. ,
:
• r . . / r
Fal·East
Jpurrial
of·
.
Mathern.atical Sciences (FJMS) ··
/ Volu.me 95, Nu·mbe-r 1
/ 2014,
PP'
VYセXP@
/
, / .
..
, ,; ,.
,. / ' ,; Li Mセ@ , . /
,,,, . ..
'""·'
セhセrper@
,1'ANALYSIS '(j'F
.upセ ᄋer@
•.8'ouNo'
. .FORtTHE
.... ... /
/
ITERATION COMPLEXITY .OF AN INTERIOR-POIN..T.
. ... ,, , / ,.,
.
.. _,,,,,,.....
. .
.
.I' .. • • ,- '
.. セ@ ,
....
,,.
,,
" ,. -
METHOD USING PRIMAL-DUAL FULL-NE.WJ"ON .
/
r'·"'
.
· ...,
..
/
/'
,
,
....
·
STE,P AtGO_RITHM
,r
,·
'
...:
b •
Y.,,
.r. _,'
.
,.
//
.
Bib Paruhum
sゥャ。
ャ 。ィセ@
.I ...
·'
! '· , ,...
J ii""
' .
, , ,. I
,.
, , , , •.· "'
, I
..
.•
I セᄋM
·'
Pustlpa l>(.blishing
Hou$e · ;
··
VIJaya Ntwas, 198 MumfordganJ Allahabad 211002, INDIA
..
エGNャGBB G ュMQ[ゥ[ QイNQセ@•
.,-ff,.i;.,
セᄋ@ http://pphmJ.com!Joumals/flms.htmfjmsOpphmj.com & arun@pphm).com . :.
..
-,·
.•
.
'· セ@
... , .
,
.
•
. lnfonnat1gn for Authors
Alms and セッー・ Z@
The
Far East Journal of Mathematical Scie.J&1i/f JMS) is devoted to publishingoriglnll research papers and critical survey articles In the field of Pure and Applied Mathematics,
セaーーャゥ」。エゥッョウ@ and Sta6stics. The FJMS is a fottnightty journal published in twelve volumes
annuttly II/Id each volume comprises of two issues.
セQQY N@ Indexing and Reviews: Global Impact Factor : 0.692, Scopus, M athematical Reviews,
MalhSciNet. ProOuest, lndexCopemicus, EBSCOhost. Zentralblatt MATH, Ulridl's web, Indian
Sdence AbslractS. SCIRUS, OCLC, Google Scholar, Exoellence in Research for Australia (ERA), .k8demlcl<eys.
Submlalon of Manuscripts: Authors may submit their papers ror consideration in the Far East
Joumel of Mathematical Sciences (FJMS) by the following modes:
1. Online 1ubmluton: Please visltjoumal's homepage at http://www.pphmj.comfJOUmalslfjms.htm
2. Elec1ronalty: At the &-mail address: tjms@pphmj.oom or kkazad@pphmj.oom
S. turd copies: Papers in duplicate with a lette< of submission at the address of the publisher.
The セ@ must be typed only on one side in double spacing wi1h a generous margin all round. Ari
elb1 II made to pubfish a paper duty recommended by a refereG within a period of three months. One
Mt ol galley proofs of a paper will be sent to the author submitting the paper. unless requested
ollelwt9e. without the original manuscript, for cooectioos.
セ。ョ、@ Refennce1: Authors are requested to provide an abstract of not moni than 250 WOl'ds --' 18test Mathematics Subject Classification. Statements of Lemmas. Propositions and Th80fems セ@ be set in itafics and references should be arranged in alphabetical Of'der by the surname of the
lr'ltalthor.
Pege Charges and セーイゥョエウ Z@ Authors are requested to arrange page charges of their papers @ USO
40.00 per page for USA and Canada. and EUR 30.00 per page for rest of the worid from their
lns1ilutions/researcti grants, if any. However. for authors in India this charge Is Rs. 800.00 per page.
No eldra charges for colour figures. Twenty-five reprints in print version and a copy in soft version are
provided to the COfT'espooding authof ex-gratis. Additional sets of reprints may be ordered at the time
°'proof comtdion.
Copynght It is assumed that the submitted manuscript has not been published and will not be
セウエケ@ submitted or published elsewhere. By submitting a manuscript, the authors agree that
h copyright for their artides is transferred to the Pushpa Publishing House. Allahabad, India, if and
when. the paper Is accepted for publication. The pubfisher cannot take the responsibility of any loss of
nwnusa1pt. Therefore. authors are requested to maintain a copy at their end.
Subscription Information for 2014
lnstllutloNI Price for all countries except India
Eledronlc Subscription € 720.00
Print Subscription includes Online Access € 1195.00
USS995.00 USS 1695.00
,., - . . . HI^セ。@ lcel'M b セウI@ ol lie F• &st_,,,,, ol ....,__ s... cfNimsセ@ ... lacillly ioセ@
... h セ@ loll be ...,.,, fvough ... lnsVlulioNI 9 digits II' eddr9u 10 be prcMcled by Iha セ@ -...ly The
. . . "' _..., _ . ..,,.,.,. bl ... lfld o1 ... neJCI セ@ )4W ll'Om 111t last ia-ol Ille \l()lume ""'9cribed. F0< セ@
NNNNNNLセ ioBMエャQ・セッャ@ ... ⦅LLNセMNNNNNQッイMゥMMMセit|YIG「・ィ。、ッョ。セ@
...
Price In Indian Rs. (For Indian Institutions 1n India only)
( Print Subscription Only
I
Rs. 15000.00I
The IUbecription year runs from January 1. 2014 through December 31, 2014.
lnfonnatlon: The journals published by the "Pushpa Publishing House· are solely distributed by the セ@ Boolts and Jourmil• Dl1tributors·.
Cone.ct Person: Subscription Manager. Vijaya Books and Journals Distributors. Vijaya Niwas, 198 mオュヲッイ、ァ。セ N@Allahabad 211002. India; sub@pphmj.com : arun@pphmj.com
Far East Journ1I of M1them1tle1f Sciences (FJMS) C> 2014 Pushpa Publishing House, Allahabad, India Available online at http://pphmj.com/joumals/fjms.htm Volume 95, Number I, 2014, Pages 69-80
SHARPER ANALYSIS OF UPPER BOUND FOR THE
ITERATION COMPLEXITY OF AN INTERIOR-POINT
METHOD USING PRIMAL-DUAL FULL-NEWTON
STEP ALGORITHM
Bib Paruhum Silalabi Department of Mathematics
Faculty of Mathematics and Natural Sciences Bogor Agricultural University
JI. Meranti, Kampus IPB Dannaga Bogor, 16680, Indonesia
e-mail: bibparuhum!@yahoo.com
Abstract
The use of interior-point methods to solve linear optimizarion problems has become a great attention to the researchers. The most important thing is that the interior-point methods have the best complexiry compared to other methods and also efficient in practice. The worst upper bound for the iteration complexiry of this method is polynomial. Roos, Terlaky and Vial presented an interior-point method using primal-dual full-Newton step algorithm that requires the best known upper bound for the iteration complexiry of an interior. point method. In Lhis paper, we present their method with a slightly better iteration bound .
1. Introduction
Recently, the use of interior-point methods (IPMs) for solving linear
r・」・Qカ」、セjオャケ@
8, 2014; Revised: August 20, 2014; Accepted: September J. 2014 20 I 0 Mathemaiics Subject C lassificaiion: 90COS. 90C5 I.70 • Bib Paruhum Silalahi
· optimization (I:.0) problems has become a major concern of the optimization researchers. This happens mainly due to the fact that interior-point methods have polynomjaJ complexity, which is the best compared to other methods and these methods are also efficient in practice.
Interior-point method first appeared in 1984, when Karmarkar [4] proposed a polynorrual-time method for LO problems. In the worst-<:ase, for a problem with
n
inequalities and L bits of input data, Karmarkar's algorithm requires O(n3·5L)
arithmetic operations on numbers with O(L) bits.In [8], Renegar improved the number of iterations to
0(
,/;
L) iterations. Other variants of IPMs, called potential reduction methods, require also onlyo(.J;iL)
iterations. This was shown by Ye [13], Freund [I], Todd and Ye [12] and Kojima et al. [5].Sonnevend [ 11] and Meggido [ 6] introduced a class of IP Ms which uses the so-<:alled central path as a guide line to the set of optimal solutions. These methods are called path-following methods. A variant of path-following methods was presented by Gonzaga (3], Monteiro and Adler [7] and Roos and Vial [I 0). Their methods require 0(
..In
l) iterations, which is the best known upper bound for the iteration complexity of an IPM. Roos et al. in their book (9) obtained the same upper bound by using an algorithm which is a so-called primal-dual full-Newton step algorithm. Their upper bound for the number of iterations isイeョQョョセ
PゥN@
(I.I)where t is the absolute accuracy of the objective function and µ0 > 0
denotes the initial value of the so-called barrier parameter.
In this paper, by careful analysis, we reduce the upper bound by a factor
セ N@ The iteration upper bound that we obtained is
r
..r;;
logョセッ@
l
ᄋ セ@
Sharper Analysis of Upper Bound for the Iteration Complexity
2. Primal-dual IPM with Full-Newton Steps
The standard fonn of an LO problem is as follows: min{cT x : Ax == b, x
セ@
O},
71
(P)
where c, x e '.Rn, b
e
'.Rm and A e '.J?.mxn. Any LO problem can be transfonned into standard form, by introducing additional variables (2). The problem (P) is often called the primal problem. Associated with any LO problem is another LO problem called the dual problem, which consists of the same data {A, b,c)
arranged in a different way. The dual of (P) ismax{bT y : AT y
+
s=
c, sセ@
0},
(D)
where s e J<n and y e '.Rm; (D) is called the dual problem.
The feasible regions of (P) and (D) are denoted by P and V,
respectively. The (relative) interiors of P and V are denoted by P0 and V0 •
Finding an optimal solution of (P) and (D) is equivalent to solving the following system (9):
Ax== b, x セ@ 0,
T
A y +
s
=
c, s
セ@ 0,xs=O, (2. 1)
where xs is the component-wise (or Hadamard) product of the vectors x ands
and 0 denotes the zero vector. The first line in (2. 1) is s imply the feasibility constraint for the primal problem (P) and the second line represents the feasibility constraint for the dual problem (D) . The last equation is the so-called complementarity condition.
72· Bib Paruhum Silalahi respect to µ . The resulting system is
Ax= b, x セ@ 0,
T
A y +
s
=
c, s
セ@ 0,XS= µe. (2.2)
If system (2.2) has a solution for some µ > 0, then a solution exists for every µ > 0 [9). This happens if and only if interior point condition (IPC) is satisfied. The solutions are denoted as x(µ), y(µ) and s{µ). We call x(µ)
the µ-center of(P) and (y(µ), s(µ)) the µ-center of(D).
When µ runs through (0,
oo),
then x{µ) runs through a curve inP
0which is called the central path of (P). Similarly, the set {(y(µ), s(µ)) : µ e
(O, oo)}
is called the central path of(D).Ifµ-+
0, then x(µ), y(µ) and s(µ) converge to a solution of (2.1 ), which means that the central path converges to the set of optimal solutions of (P) and (D). On the other hand, ifµ --. oo, then x(µ) and (y(µ), s(µ)) converge to the so-called analytic center
of (P) and (0), respectively.
Next, it will describe how Newton 's method can be used to obtain an approximate solution of system (2.2), for fixed µ . Given a primal-dual feasible pair (x, (y, s )}, we want to define search direction
tu,
fly and t:.ssuch that (x
+
flx , y +fly, s +t:.s)
satisfy (2.2).Since Ax = b and Ary+
s
=c,
system (2.2) is equivalent to the following system:aセク@
=
0,AT fly
+
t:.s=
0,stu
+ xt:.s +tu!:Js
=
µe - xs.Sharper Analysis of Upper Bound for the Iteration Complexity 73 The third equation is nonlinear, due to the quadratic tenn
t:.xt:.s.
By neglecting this quadratic tenn, according to Newton's method for solving nonlinear equations, we obtain the linear systemAtu
=
0,AT fly + t:.s
=
0,stu
+ xt:.s=
µ e - xs. (2.3)The resulting directions of (2.3) are known as the primal-dual Newton directions. By taking a step along these directions, one finds new iterates
(x +, (y +, s + )) such that x + and s + are positive. The new iterates are given
by
+ + +
x
=x +tu,
y=
y+
fly ,s
=
s
+
t:.s.
In the process of following the central path to the optimal solution, by using Newton steps, we generate a sequence of points within the neighborhood of central path. We need a quantity to measure the proximity
of (x, (y, s)) to the µ-center.
Before defining this proximity measure, we refonnulate the linear system
(2.3), by scaling
tu,
fly and t:.s to d.x, d .1• and ds as follows:d x -- - -
vt:.x
x '
where
fly
dy
=
-iµ·
v=J!f.
d s -- - -vt:.s
s
'
lfwe define D
=
diag(J;/S), then system (2.3) is equivalent toADd.x
=
0,(AD)T d y
+
ds=
0,74 Bib Paruhum Silalahi
The first two equations of system (2.4) show that the vectors dx and ds
belong to the null space and the row space of the matrix AD, respectively. These two spaces are orthogonal, therefore, d x and ds are orthogonal. The orthogonality of dx and d5 implies
II
dxセ
R@+II
dsセ
R@=II
dx + ds11
2=llv-
1-vf
Note that dx , ds (and also dy) are zero if and only if v-1 - v
=
0, which happens only if v = e, and then x, y and s coincide with the respective µ-centers. Therefore, to measure the 'distance' of (x, (y, s )) to the µ-center, we use the quantityO{x,
s; µ) defined byo(x, s;
µ)
:=o{v)
:=エセ@
v -1 - vf
(2.5)For any t セ@ 0, the t-neighborhood of the µ-center is given by the set
{{x,
y,
s): x E 'P,(y.
s) EV,o(x.
s; µ) :5<}.
After a full-Newton step, the duality gap at the new iterates always assumes the same value as at the µ-centers, i.e., (x +)Ts+
=
nµ (cf. [9, Theorem 11.47]). An IPM with full-Newton steps can be described as in Figure 1.Primal-dual IPM with full-Newton steps Input:
an accuracy parameter E > O; a proximity parameter t, 0 :5 t
<
I;strictly feasible
(x
0•yo,
s0) with(x
0 )rs
0=
nµ0 and o(x0,s
0; µ 0)::;; r : a barrier update parameter 0, 0<
0<
I .Sharper Analysis of Upper Bound for the Iteration Complexity
begin
x
:=xo; s :=so; Y
=yo; µ:= µo ;while nµ セ@ E do µ := (1 - 0)µ;
x := x
+
6x;y := y
+
6y;s
:=s
+
65;end while
end
Figure 1. Primal-dual IPM with full-Newton steps.
3.
Analysis of the Primal-dualIPM
with Full-Newton StepsWe first deal with the effect of a full-Newton step on the proximity measure. The next lemma implies that when
o(x,
s;µ)
is small enough then the primal-dual Newton step is quadratically convergent, as stated in Corollary 3.1.Lemma 3.1 (cf. (9, Theorem 11.50]).
If
o
:= O(x, s;µ)
:5 I, then theprimal-dual Newton step is feasible, i.e., x+ and s• are nonnegative.
Moreover,
if
o
<
I, then x + and s + are positive and52
S(x+ • s+; µ) :5
セRHQ@
-o
2 )Corollary 3.1.
If
o
:=O{x,
s;µ)
s*,
then o(x+, s•;µ)
so
2 .
76 Bib Paruhum Silalahi
Lemma 3.2 (cf. (9, Lemma 11.17]). After at most
r
tnµo1
9log7
iterations, the algorithm stops and we have nµ s; &.
We have the following lemma that will be used in proving the next two theorems.
Lemma 3.3 (cf. [9, Lemma 11.54]). Let (x,
s)
be a positive primal-dual pair and µ>
0 such that xr s = nµ. Moreover, leto
:= S{x, s;µ)
and letµ+ =(I - 9)µ. Then
+ )2 ( 2 92n
O(x, s;
µ=
I - 9)o+
4(1 _a)"
The next theorems present iteration bound of the primal-dual IPM with full-Newton steps.
Theorem 3.1.
If
t =1/.Ji
and 9] iOセN@
then the algorithmrequires at most
r
Fn+i
logョセッQ@
terations. The output is a primal-dual pair (x, s) such that x Ts s; &.
Proof. Let us take t =
If.Ji.
By using Corollary 3.1, since we have5(x, s;
µ)<;If.Ji,
after the primal-dual Newton step we have o(x+, s+;µ)
S 1/2. Then, after the update of the barrier parameter to µ + = (I - 9)µ with
} =
1/Fn+I, by using Lemma 3.3, we get the following upper bound of>(x+, s+; µ + )2:
+ + 2 I - 9 92 n
=
_!. .
O(x•, s
;
µ ) s -4- + 4(1-9) 2Sharper Analysis of Upper Bound for the Iteration Complexity 77
The last equality follows by substituting 0 =
t/..r;;+T.
Hence, we obtaino(x+, s+; µ+) s;
1/./2
=
t . This means that the property o(x, s;µ)
s; tis maintained after each iteration. Therefore, combining this with Lemma 3.2, we obtain the theorem. 0 Note that Theorem 3.1 holds for every n セ@ 1. In practice, n is much larger. For such cases it is worth mentioning that slightly bcner iteration bounds can be obtained, as the following theorem.
T heorem 3.2.
If
te
[0.6687, 0.6773] and 9=
1/Fn, then for n セ@ 47,the algorithm requires at most
r
Fntogョセ
P
Q@
iterations.
Pr oof. Let us take 0 =
1/Fn.
By using Lemma 3.1 and Lemma 3.3, we can verify that if(I - l/Fn)'t4 + 1 <t 2
2(1 -
t2 ) 4(1 - 1/Fn) - •(3.1)
then o(x, s;
µ)
s; t is maintained. The region in the ( t , n )-space defined by (3 .1) is depicted in Figure 2, where the smallest value of n is around 46.6274 at t=
0.6731. Therefore, n = 47 is the smallest integer value of n which satisfies (3.1 ). We can find out that for n セ@ 47, the inequality (3. 1) holds forBib Paruhum Silalahi
to
j
.
(1-1/v'ii}iA 1
I
1
n
I
2(1 - T2)
+
4(1 - l/ Jn} $TJ
10•.l
i
sof.
"
/1
I
..,
___._- - __j
oc: OM OM O• 01
o.n
[image:7.1024.129.788.11.597.2]t
Figure 2. The region defined by (3. 1 ).
The value of n in Theorem 3.2 can be improved to n
セ@
6, as stated in the following theorem.Theorem 3.3.
If
't E [0.7433, 0.8289) ande
=
1/../n,
then for nセ@
6,the algorithm requires at most
rJniog
nn
iterations.
Proof. We use (9, Theorem 11.52], a sharper quadratic convergence result of a primal-dual Newton step. This theorem states that if
o
:= O(x, s;µ)
< I,then
52
o(x•,s•;µ)S
J1(t-04)·
Then,
by
using Lemma 3.3, we obtain that fore
=
t/../n
the propenyO(x, s;
µ )
セ@ 't is maintained ifSharper Analysis of Upper Bound for the Iteration Complexity 79
(1 - l /
fn)-r
4 I 22(1 - 't4 ) + 4(1 - 1/Tn) $ 't .
(3.2)
Figure 3 depicts the region defined by (3.2). The smallest value
of
n is around 5.1971 at 't=
0.7968, and we can verify that for the smallest integervalue n = 6, the inequality (3 .2) holds for 't
e
[0.7433, 0.8289]. Thus, weget the theorem. 0
..,
35 ·
:io
25
-n ( l - l /y'ii)r
4
l <
TJ
2(1 - r4)+ 4(1 1/Jii)
20-IS •
10 •
51..
lセ@
o• OM ,...,___ 07 on 0 1
D
ッセ@t
Figure 3. The region defined by (3.2).
References
(I] R. M. Freund, Polynomial-time algorithms for linear programming based only on primal scaling and projected gradients of a potential function, Math. Program. 51 ( 1991 ), 203-222.
[2] D. Goldfarb and M. J. Todd, Linear programming. Optimization. G. L.
[image:7.1024.74.415.44.287.2]80
.
Bib Paruhum Silalahi
f3) C. C. Gonzaga, An algorithm for solving linear programming problems in
O(n3L) operations, Progress in Mathematical Programming: Interior Point and Related Methods, N. Meggido, ed., pp. 1-28, Springer-Verlag, New York, 1989.
[4) N. Karmar'kar, A new polynomial-time algorithm for linear programming, Combinatorica 4(4) (1984). 373-395.
[5) M. Kojima, S. Mizuno and A. Yoshise, An O(../nL) iteration potential reduction a.lgorithrn for linear complementarity problems, Math. Pro&f31!1., Series A 50
( 1991), 331-342.
[6) N. Meggido, Pathways 10 the optimal set in linear programming, Progress in
Mathematical Programming: lnterior Point and Related Methods, N. Meggido, ed., pp. 13 1-158, Springer Verlag, New York, 1989. Identical version in: Proceedings
of the 6th Mathematical Programming Symposium of Japan, Nagoya, Japan, 1986, pp. 1-35.
[7] R. D. C. Monteiro and I. Adler, Interior-path following primal-dual algorithms. Part I: Linear programming, Math. Program. 44 ( 1989). 27-41.
(8] J. Renegar, A polynomial-time algorithm, based on Newton's method, for linear programming, Math. Program. 40 ( 1988), 59-93.
f9] C. Roos, T. Terla.ky and J.-Ph. Vial, Interior Point Methods for Li near
Optimization, Springer, New York, 2006. Second Edition of Theory and Algorithms for Linear Optimization, Wiley, Chichester, 1997.
f IO) C. Roos and J.-Ph. Vial, A polynomial method of approximate centers for the linear programming problem, Math. Program. 54 (1992), 295-306.
r
11) G. Son.nevend, An "analytic center" for polyhedrons and new classes of global algorithms for linear (smooth, convex} programming. A. Prckopa, J. Szelezsan and B. Strazicky, eds., System Modelling and Optimization· Proceedings of the 12th IFIP-Confercncc held in Budapest, Hungary, September 1985, Vol. 84 of LccrureNotes in Control and Information Sciences, pp. 866-876, Springer-Verlag , Berlin, West Germany, 1986.
{ 12) M. J. Todd and Y. Ye, A centered projective algorithm for ltnear programming, Math. Oper. Res. 15 (1990), 508-529.
(13) Y. Ye, A class of potential functions for linear programming. Management Science Working Paper 88-13, Dept. of Management Science, University of Iowa, Iowa City, IA 52242, USA, 1988.
FAR EAST JOURNAL OF MATHEMATICAL SCIENCES (FJMS)
Editorial Board
Gunjan Agrawal, lndia
Natig M. Atakishiyev, Mexico
Antonio Carbone, Italy
Hasan Coskun, USA
Zhenlu Cui, USA
Manav Das, USA
Sbusheng Fu, China
Wei Dong Gao, China
Toshio Horiuchi, Japan
Lisa M. James, USA
Koji Kikuchi, Japan
Victor N. Krivtsov, Russian Federation
Alison Marr, USA
Jong Seo Park, South Korea
Alexandre J. Santana, Brazil
A.
L.Smimov, Russian Federation
Chun-Lei Tang, China
Carl A. Toews, USA
Vladimir Tulovsky, USA
Qing-Wen Wang, China
Peter Wong, USA
Pu Zhang, China
George S. Androulakis, Greece
Carlo Bardaro, Italy
Yong Gao Chen, China
Claudio Cuevas, Brazil
Maslina Daros, Malaysia
Massimiliano Ferrara, Italy
Salvatore Ganci, Italy
Demetris P. K. Ghikas, Greece
Jay M. Jahangiri, USA
Young Bae Jun, South Korea
Hideo Kojima, Japan
Yangming Li, China
Manouchehr Misaghian, USA
Cheon Seoung Ryoo, South Korea
K. P. Shum, China
Ashish K. Srivastava, USA
E. Thandapani, India
B. C. Tripathy, India
Mitsuru Uchiyama, Japan
G . Brock Williams, USA
Chaohui Zhang, USA
Kewen Zhao, China
!:
"·
...
r!'
I J' • ; ' .. • 7 BGGᄋGセ@ , •-
r
· ""·
/'
·.
· ·
M}·
r"
i"
ャ セZ@ セ、セ」・ウ@
セaュjゥセッセセセ」セGHセZ@
E_Wセ
MMセG@
. _:,_.I._
'1-;'lJ
:,,
2:
Qセセセ@
aーーセセッッウ@
..
inaキ、
ᄋ セセセサgifZ@
ッNセセ[@
sooセオ@
..
s) _ ,.セM
3.LL aゥヲカセ@
aod(Applicatioosftt-Staff.stics
HセNウA・ヲ@
. " ,
p;-
·J,
セ@
r:
•J/ 4., '"
セLェョcッューオエVェNMs」ゥ」ョ」・BBGセァ
セ@ - ..,-:,7 •ᄋ Hッif
Z P N Wャkャ
/--v
N@
,
/ · ·-
..%?.:
;z.•/
.I _,) '
Sr
AdvaiaS
iii
Diffcn!iiQal Equations 。ョ、セャ@ セウ」ウ@ (GIP:..O.<i07) · ,_;__ -'..
;j
-セ@ セ@ .I i" • ·< ;,r • - •
,.
セNセvゥオクZ・ウ@ in Fuzzy Sefs-and Sntems ( : u.703) .. ·.,... .t/· , · • -._
£
t-. ,..rn /, . ,,.·-:{: WN ᄋ セセセidセHセQイZセXWエ L@ f-/·,..,-·· LL⦅_LZセMNLL@
8. Far
e。ウエ⦅jセ。ャ@
NッVキャゥセセguZ
LZ@
0.7Z7) .)セM M
_#.
-
.
セェ@ , . . l·:.·"!i!, 9. f。イeNNLuッオョェセオ^ヲャ^ケ。。ュゥ」。エ@ ウセ@ HgゥZ・NN[j^NVXUス セ@ ( ..• セNNL@ セ@
...
'··-:... 10.', Far
e。ウエjセ@
ッヲセセセセ@
セセgw@
Z PNQPAL
A セウセ
N@
11.f。イe。ヲエGjッオュ。ャqヲセッイエHエjキZP
N VYTI
セ
N@
r.セ MN@
0. i;.{_:/lr2.
fセ@ セ@ セ@
_セュ。エゥセ
M セ・ャjiC@
Hfセセ@
Hgセ Z@ セNセGサ@
セ
ᄋ@
..セᄋ@
.•セ@
./ 1Y.
f。イe。ウエjッオュ。ャ
G ッヲGセᄉ」。ャ@
Statistics (GIEdU67) _, • . ,__ ·.·
セ| N QT [ GNZGNZ Mョ[@ セセッヲ@
aャセ
セ セュセセヲッ@
BエyNpャゥ」セセGgif@
セN VXI Gス Ms」ーォイセ
ヲG@
·
セLNN@
"ts.,
JP ff"l'"'•l ""?:·- ofBiostiti.stics Hgイf セ Z@ O.S8l) 7 .. . I, • ' · {/ ./_.cql ; セ@
16.
j セ ヲッオセセセセ
セpッゥ、
ヲ セ
エN セ@ aーーャゥセッゥAェBHgifZ@
PNVセ@
;f:;..
セセ@
,, , · .-17 . •jpjセセヲセセLootッセHoif
ゥL⦅ セNAIs[セpusI@
;r...
,.
ZN セ@
18.jp Mjoオ}ゥイ。ャッヲh・。エ。ョ、セGャGセウヲ」イHgセセセ[セpusI@
·J··;.,_ .·
19.
ィu・ュ。エゥoョヲャIセ@ ofPunctionit
AW
_··;s.
Ope:ritQtlfheory
1¢ApplicafiOns
'•·1 .A "lh •
r.
rf'> Iz./'!P,'7' .
r ..
;i;
{GIP: 0.597) セᆱNGᄋ@ • ' ()1· , ., • • セ@20 . .
iョセセ@
セ。ヲ⦅アセイエ。ィ@
eョセ
N@
g
。ョウャMj・」セャッァェ@
(GIF.i
oNセI@
- 2J.
セNjoッュ。ャ@
QtNumeri.cat
mセ@
add
A ·" ucadons;(GJF:
ッNセーj@
/ ./ .:-::n_
uイオvセャoオュ。ャヲZGヲセーエエセ。ョ、
L@
M"athematr:
s」ゥセセZ@
oセI@
G GNセ@
セ M sjャヲvセゥヲゥmjエ「」Nゥゥ」ウ@
セセ@
s」ゥ・P
」イ・DHセ
jA[UAX
エ[⦅@
_
¥,
·- ·
1Ai1
セセ@
Joomal of Nutrition aOctf)jetetics1?-;
' ·
7--
1f ., , セ M . ,_.r -
;?
Nイセ@. / ,.
セᄋ N@ r?;,L I , . J:- / , セ@ • • セ Gᄋセ@ セ@ ./r>
セZM n・キMTアオイョ。jウャGイッAゥQRPQN⦅M
セセ
Q
ᄋ@ セ@
i
' ···.
y.' .. ,
• 1. • i I .l·/' -"j , / ,-,• "-' 7 . · ' ( • セ@ - ' ? •. _, ?-' r ( I' •
NMセN@
r'.2SY
セエョ、ゥ」ウ@
inA.rtificia'i
QNョセZmョ」・@
,.,.,, ,, • .. . /·.
セ@ セ HN@ ェNセセセ@
ki:i
aーーャゥセセゥ
エ ュ@ ᄃ\^yBセfOsゥョ・・ョ
⦅@
·•
·rig
.:/,-".I·
' ..
r;:
•
,....-J'21.i 'J-.2.J; ...
, uuu•r v ., fャゥNセ\Iカ@
Mセ@·
セ@ ・オ、ゥ」。セ
,"$ • セ@M ・エセオョゥ」ャャャッッャN
セ@ • _ ) ' 1'!'NGLNNM
'
v ·
1
,,, .
, ,,,, I- ᄋ セセ ZM ャj。ゥカ」イウ。ャ@
lG ᄋ セ@ ,, / " J ·• , ( セ@ッヲセ
G@ G ャゥ・、 セ@
sics。ョ、 Nセ・、ゥョッャッァケ@
r
,I
セMM セ NZ Z セセjッオュXQY⦅A
NN@
,.
ュゥヲZGNN。ョ、bオセcsB
QN LmセァKエ@
. , 30
.-;/"u ...
QaセM]MQ@ P( :it:2 , .En .
. ,
and イMセ@. • ,
' ' ·,,. • 1 セ@
ruv ..
セ M セ@... o
Electfi
gmpenng
.
キッョョ。エゥ 」。NZLN セMHᄋN@,..
Nセ@セ@
ゥG セ j@ NᄋセN ャャヲ@
. I versalャセ@
,, 9f M . 1iCal Stieocesand
tセjュッャーァケ@
.&·
セ@
_,,-セ@ ... ,.._"7f . • /· .....
セ@
. . . .セセセ
N セセQAQQャ
G セセRエエゥ@
. . '·', •• セ@ ,..,. .. LN セ@ ;,:r',. セ@ ,; t" 1.
r._.. __ " ,... _ , ... セN@ セ@ ' セ@ J
'· • - · •l ..• _,. -.._,J - セ MG M セ Oᄋ@ • ,-;"'
/
.
-lv
'.<.
Gセ@'}·
.. •-
I '. ,.,
セ@ .-
J ' -,f. ; • #..
.
,.. __ ,. '- "1 i--!..C . 4 f . .)
..
, ,._., /r.- .. r ; " セZ@ MMセ@ •. r """, -,,,, イセZ@
/_,,,,.
• -"....:! • ' ,..., , セ@ . , • , ,
, /_ .. . . ,,. .. - _fi_°"" M セ@ セ@ . . ... -:.. •
· - LセM N O[NN@ :;y,· .. - Zセ@ セ@ • セ@
., c;
/'1 ,. ..
, .
,.._, --:-<.
Jr . ,
··-:'f""'.'."-'
''" . f , ,.
··'ff / - -;,
:.,--,... ,,_">_ ,__ ....