Sunteți pe pagina 1din 24

UNCLASSIFIED

AD NUMBER
ADA800276

NEW LIMITATION CHANGE

TO
Approved for public release,
unlimited

distribution

FROM
Distribution authorized to U.S. Gov't.
agencies and their contractors;
Administrative/Operational Use; FEB 1951.
Other requests shall be referred to US Air
Force School of Aviation Medicine,
Randolph AFB, TX.

AUTHORITY
AFSAM ltr,

12 May 1970

THIS PAGE IS UNCLASSIFIED

v~r

Reproduced

fin IlP~A I'r


do

NVW

ail Is

by
acu

WIGHT.-PATTERSON

f
o

AIR FORCE IUASE-

DAYrOo't,,iC

Y~Government or other drawings,I anecifficatiol-s or


~ d or an~y purpose other than inVcnnnetmiojjn wit"
2~~ dGovernment pro)curemnent operatio-n,
the
q.
%ri~ncursnio responsibility, nor any obligatin"i
A-b: fact that the Governmenit may have formulated,
4 ".4ny way supplied, the said drawings., spe.,cifications'
t ~~b
be .%garded y implicatio ootherwise as
'cc~ndsirmy the holder or any -%ther pro
orQroa
<ma any rights or Dermis,zun to manufactaureI 1ise
or
Sin-ve-ntion that may in any- way be related ~thereto."

403

72

NC'w

TIDo

>O

fC~b.l

~~co

DISCRIMINATORY ANALYSIS
,,

. DISCNItNATION". CONStSTENGY PROPERTIES

Evelyn Fix, NI.D.


J.L. Hodges, Jr., Ph.D.
U i versity of Cap!,fornia, BerkeIey

PROJECT NUMBER 2!-49-0X.


REPORT NUMBER u,*
"(Freeired at the University of'Californi
under Contract No. AF4J (128)-31)

I ICAIC"- Qtr. W AA
fl

tC

RANDOLPH

AII ATI/nI

MEm D Of
01 N E

F!ELD, TEXAS

FEBRUARY

1951

NONARAMETRIC DISCRMNATION:
CONSISTENCY PROPERTI.E.S

1. Introduction
The discrimination problem (two populatin ease) may
be defined as follows: e. random variable Z, of observed
value z,

Is distributed over some space (say, p-dimensional)

eitner according tu
bution G.

la trb

The problem is

%-A
1r

sC-rM4_ng

to d&stri-

to decide, on the basis of s,

which of the two distributions Z has.


The problem may be classified in varl-.s ways intn
subproblems.

One pertinert method of classification is

according to the amount of Information assumed


available about F and G.

to be
three stages:

'te may distinguish

(1)F an4 G are completely known


(ii)
(Iii)

? and 0 are known except for the values of one


or more 'parrm.ters
F and G are coxple~ely unknown, except Deosei.ly
for as.umDtions about existence of densities, etc.

Subproblem (i)
The soluticn is

bas been,

.mplie-it in

in a sense, completely solifd.

the Neyman-Pearson lema

was mad6 explicit by Welch [2],

[11,

We may without loss of

gererality aisuMe the ,xltence of density funotli;s, say


f and g, eorresponding to F and G, since P znd ',, ar
solutely ccntinuous with rspect to F + G.

If

3b.

f and g

h9a

PHOJECT NtBOR 21-4"k0

ape

RESoP

NUMBER 4

known, the- discrimination should depend only on

f_(A)
g(s)

An approuriate (positive) constant c is chosen, and the


following rule
If

.t

Is observed
> at we decide in fwor of P
t

If
If

,.a<te
_flIt=a,

decide In favor of G

the decision may be made in ta

arbitrary

These procedures are known to have optimm properties with


regard to control of probability of mdsolasslficatlon (probability of wrng decision).

We shell rotor to this as the

"ilkelihood ratio _rocdurege,

and denote it

For simplicity, we shell a35o


that tre borderline ca&e t(s)

tw

by L(c).#

throughoot tha papar

= eg(s) can hm neglected-

Formally, we vostulate that

f
j

Pf f(Z) a rm(Z)) a 0
regardless of vhether Z coaes froa F or G.
classification Is arbitrary when

f(S) n *g(S).

Since the
It hardly

seems worth while to introduoe complications Into the methods


to allow for it.

Howver, it

is not di*-?i*uLt to extend

our methods to take care of the situation which aries vChen

P( f(Z)

> 0.

The cholce of a depends on coraslerationh relatu.g


to the relative Importance of tW.u two possible errores
saying Z Is distributod according to 0 vwn In fact it
Is distributed according to FP

aWd conversely.

of c have been widely advocatedt

Two choices

P~CT?4U~BE
F~

.-49O)

roRFP NUORWTEER 4

(a) Take c = 1
(b) Choose a so that the two zrobabilities of e:ror
Choice (a)

has been cailed "logical-"

this miptimax procedure.

choice (b) yields

In this D&per We sahall not concern

ourselves with the- choice of c, but shall assume that a


given positive c is a datum of the problem,
The usual apDroaoh to subproblea (ii)

is

as follows.

We assume there are available samples from the two distributions,


say

X1 ,,

P.

sI3ple fro. F

Y* S sazple from G.

Y21 "

.e as!-su!me further that F and 1

in

'-.

forvi

that is.

that we know them exempt for the values or some real parxwtors.
which may be denoted collectively by G.
dCstributions correspondiL

4-r

aet'.104-.

)%V,

Say,

0.

IbA

using tho distributions F,

VUO uy Fat

to a

Drocedure ;urrenLiy etployed is

We may denote the


,G

Lo use the X's and Y's


th.Vt

fn

VPnOOAB~ A*

C%aa though they

tIlniA?

(11.

sre knowu t6

be correct.
The most fam_ir oxanl.e of this wrocess is
discriminant function L3].

There,

it

is

the linear

(tacltly) seSemd

that F and 0 are p-variats nnrnal ;* stributicns having the &am

Cunkcn~wzn) cevarlance mat:.'&-..

. a

urnknewn expeetstion vetors,

The two expectation vectors and the covariance matrix are


estirated frow
cedure is

the samples,

then empluyed,

and the likelihood ratio aro-

using the estlmated values as

PROJ=C

NU 1 P.ER 21-40-004 W'O"RT NUMBER 4

thuauh they vere knoviD to

.orrts6cU

Not such is ]mown about the desfrability of the usual


method of-attack on (Ii). We give in Section 3 a theorem
concerning asy

+totic properties of the wthod* Undoubtedly,

this proeddure is reasonable


metric form is

correct.

the asmed

pWa-

Bt the validity of t4e use of

the linear discrizinant function with data obviously not


normal or, if

normal, with obviously unequal covariance

matrices has been oZ g

bad results my

Ora.-c--u.-ern.

Presumably, very

ensue if a proedure Is usned, based on

certain assumptions about paamtr1 .n tomz

wbn those assump-i

tions are not even approximtely oormeto


There seems to be a need for 41bowlrulMion pro.4iw*u
whose validity does not r:equir* tA iaut
iJelied by the normality ass

of Mavl~4&s

ption, the homosoedast*.eG

assunption, or any assu;mption of parametric forn.


present paver is,

as far as the authors

Me

are aware, the

first one to attack sugproblen (311)s sw reasonable discrizination procedures

be found which will vuk even if no

rarawetric form can be assumed?


It is not to be expected that a&W proeedm

or= ov

guaranteed to give good reaults wltbout any rYatruLa,'4o

whatsoever on the dittributions F and 0.

To

lalo

this

point, we need to state a preise ueaening for 4good resultsee


This is dose in ftction 29 with the Introduction of th. concept
of "consistency."

We then proceed In Section 4 to prove,

under weak restrictions cnt~e dentitleo f and so the consis-

i-'oET NUP1rAR 2 1-49-O4 REPORT NUMBat


P

Tenoy

of a

class

,!oLu.. W

01

J-.%-.
iVXVUWrAl

A modification of these proaedure. is

--,

then considered !a

Section 5.
It may be noted that all of the methods and results
of this paper can be extended without difficulty to the
situation in vhi eh tho"

m.

utr.A than tvn "aOlstIS

to

be discriminated.
1he authors are engaged In further york along the
lines here laid dovL

Specifically,

so@

experim

saBdanp]i

Rents are being conducted, Intended to throw some light


on the performance of the proeedurc
sizes;

fcr moderate sample

and asymptotic properties of a class or sequential

nonparametric discriminatory procedures is b*ia.


It

Is 'Intended to prepaw

Ia6&stigat~ed.

further reporta t~tting :forth

results.
t2, qn.a notion of' eonsistency.
In setting out to define an optiium property In stvtisscal

inference,

it

Is useful to have in mind Va

of excellence beyond which it

is not PosuIble to go*

procedures L(c) described in Seetion 1


"On

the case of

TM

provide suob a limdt

nonqpamtric discriminatioiw wo cannot, with

any nonparamueic classification procedure,


better than the best which is

tuitively obvlo'as,

exct

to do

pcssible when tCe densities

thesselves are assuned to be known.

given.

!=it

This fact is

io-

but If desired an exact proof Is easily

Wren f and g aie known, Z Is suffloiant foe the

o.asslfLcation,

with respect to (Zi X1,X2,ooo,]m~lley;1s...,

!D),

Ph0J E"'C

*_r

N N4!Ei H 212-44-004 R.'TORT NUIMRHF 4

wo may (by using randomization) exactly duplicate (wiv.h

. procedure based on Z) the per-rormance characteristic o@


any procedure based on (Z;X

1X
2 ,...,XM

Y1 ,Y 2,*..&,syn).

Thus, no nonparametric procedure can have probabl"iiies


of error less than thoss of a likelihood ratio nz,-dure.
the other hand, we shall procose in

On

Sections 4 and 5 classes

of (sequences of) nonparametric procedures which,

in

the limit

as m and n tend to infinity, have the same p.obabilitie- of


error as the procedures L(c),

We may therefore reasonably

say that cur procedures are consistent w

the

16kllt.ao,

ratio procedures.
There are two differenm
sequences of .tatistic3l

notions of consistency for

-ecsion ftunction.i,

worth while to distinguish thea.


space is

fignite (as is

when there n,

may bR

Surrosc t..,at the as-"--..-

the case In discriminatory analysis

finitely many povulatitis).

decisions be..r-,,-.t,

ane! It

,y JI1

C'''"

(r"

Let the possible

Now suppose we

are c~nsidering two sequences or decision functions# say

{6%.'

and {L. *}. How should we d,

.n. the notion that

these two sequences terd to ;t4rez "th


conist,.ht

witri each other; as n -"

we might reou-,re t~'et In

each other, or be
co?

On the one hard..

t;ne limit there should be c.l.ose

aeyeement between the probebilities of decision; on the other


hand we might reqrire that in

the limit there be MO.. proba-

bility of aireement of decision.

The former requIremeat re-

lates to the porformannce characteristics of the decision


functlcns;

t~e latter reqktrement relates to the decision

I N4 1 1110 1, ;91-43j4MO

hf
W-

AUMW.Z41V&~dV3

Deo

I:!4'Whr NiUPAJ4
S. --

Im

--

1...

ition 1. We shall say that the soquenaes


are ecssetin

"(.L

-,'be *ones of popfoltAUCf ohaMtgrstits

Ifv whatever be the true distributioznst and whatvewe


there exists a mamber

for every decision


Delitin,2.
fA

wid

V such that vionaver

be

g a, Ot

aui N ^M nd 3o Ne

i
We shall say that tho eaq~aafi

and

a"e 'ou~satent In the sense or decision tfuctions It,

wh-atever- be the true distributionis, and whatever be

exists a number

X seuohthat~h~'.

a z-

aWnd

Ot0 the"

31.N-

We observe that consistency In the second sense IMplis that

in the first#, 46-e P(A


qu~antities

?(A'

A)
andS

Is n~ot loss than each or thej

Aip

an

P(

~?.!A'

The definit ions are not equivalent 0owevero as the following


trivial examnple *tiows

ir

a~

and

any men) the process of ohooslMa


aiby

each denotes (for

between two altermtives

to&sla~g * oin tben P tLA^

while
&V( A

fo

U Is 2

PIFXCJLX ITNUMBER~ 21-49-"1 PR!2"N L NUMBEIC 4

ImAuc1 P-L 't Is customary to evaluate deoision functions


enloly in terms of their perrormnae characteristiost Dtfl-

However all ""'f, of Conulstency

Iis the more nataraA.

iti.')i

given In this paper pr-Tide comsistency in the stronger sense


and consequently we staak!

of the isecond dt.nitton,

"6

-a-c

Since our prCoedures are based on two stmales, ve must


consider a double lmit

process as both

and

to

tnd

To avoid difficultlea which wouia Wcatherwise arix.

Infinity.

proach

nfinity at the same speed.

Precisely

and

are both bounded away frow

"nn--*

ever we write

co'

na

as

ap-

w and

in Section 5, we shall assume t-u'oul.cut that

va
When-

-- + co.

this r'estriction should be under-

Our restorIction has the effect of rLiucing tne limitLt-k

stood,

prc,.ss from a do.tble to a s!4le

one.

In the :equel we -.hae.1 be ucaparing certain discritr:-.nry


with proccdur:

projetu-ere

DrXio. 3.
cedures, based on
itad v,#
L(c)

k2&'"

Yn

A aeqz~o
Z and on samjlAs
from

L.cl, .........

of tChe ty~pe

Xl, X2 ;'' '1 X

G, and whatever be

P(A

and

L(c)

and

Z is die -ribu.ted accordig t*o

Ing to

f r-v

0, Is said to be consistent wLth

Ifp whatever be the ds1trioutionn

less off whetuier

or dllsrilmnatory pro-

G. regardF ora caord-

e > 0, we can assure

yield the sane classificatlOn Cf

Z)

> 1

provided z'y W"U


1

UL" .A U

1(%!, AL

We may also deti


istutecy.

a corresponjing notion of unifom con-

.i

Utp In Deftnttie- 3, the bound on probability of

agreement can be assured for all


sLze specification on

3.

and

and

G with a single

no we say that

IA

I is

Consistency for the parametric case.


We shall now damonstrate that; the analogy of the notion

of consistency just intz-xhd,-ed with the like-named notion i.z


po~nt eotimation, In more than fozmal.
of paramotric discimination
We 2lhall f

'11m

Consli'r the problem

(ubpra-jea

(ii))

to time have Ocoasion LO consider

prababilites 4omputed under tha

i-ia
1"--

outed Accordin

G.

to l1t

t;

and

P,

of section 1.

Ps Or according to

that
i

Z. i. ziicz-

Ala convteZmisn

denote! nrobabilttles cwutsd under the&*

respective assumptions.
Let

and ;

be classes of densities pareuetrize4 hy

parameters denoted collectively by


or convergezc

given a soeqence

beingMMtn
a function of X-,3 P,, .-':hsorm1.

Let there be a Uotioni

intro4ied "- the space 0

Suppose there is

.n

O.

af0az'mter

ialues.

oi e-stimte,

for

a1

0o

If

rumenions of 0 for every

exoop t pEhao f or

a e Z

2, re
'5

F'RO:KCT NLMBZR :1-0-M0

0,1-

i~mlln

~iCed~W*8

RF~-FRT NIdi.MBR A

ht theLApqtnce of d1aur~imition pcj'

2,

Obtal&n"e

l)j

by IWlPI2

C> 0

&0tof

vwtn critleal TCU

prnil

lhike~ho4d ratio

(Z)

69.

and

Z)

W~.3~
i~th IN(C 1 Ak

j~if

-AA

Frci

The idea of t-he proof in Yro'y simple:

w,~
ill probably be near

CODAi~tbnt,

-~iS

avxo large.

prnbably be neart

z~ I1

cgo.

i teandic

-cgowZ!

faszz
III:
Kaim

-1

-tz=

]LAA *

contimuoas ti,.nction of
'.

IIfA(Z)

like f %metiAcr.

when

boPthU

f,(Z)

#0 a

.-

z; ve can associate wit'"

whotevar)

f I(z)i <

L7

such that

(z)~ > 0

uantity
-

for a.1l

Avlmn

1'

takes on the valme

.lading an evont of zero proloarUity

every

CWill

4s the cumulat-Ive -Ounct Ic..-of

.j

;)

varl.aois

We. no

and

this masms

-lherefores Vt, ia not likely to make mich

~rih~,&.~dby assamptiom

f@,

az4

ame coif~ntimouis

ge

%.

ca()Cans~ih1p

rj-f{r;,(Z)

jsinee

and

will probably be near

tht a

r.

f~

But sLLCC

Sitmee

arles

if

Is replac*d by cg.

Let

~t
~'

BF~21-0.0.04 RU'OT XT.-AtJER 4

z mn t 1?iL's'

and find )f~0 such that

2())

11P2.
eI ()
Tsing finally the GanSisi0MCy Of the estimates,

and

Large e6oi0gh so that whenever

P1-

. 21

between

than

L(c)

and

n> W1,

wll

w'ith -- oablit4e.

aAls

'1s3

E.
(1)

Remarks.

on
that

and

Combining the. above, a disagrezmn1t

)~1

a > I

Choose

!s important.

fylz)

and

a fixed set,

eqs

The dependence of the disoontinuitT sets

g0 (z)

Were we to d4'm*n
be continucua

Pi(Z) = 0. 1

in

i, 2,

the stronger
0

for

pro-.ert_

z Z,

all

we should exclude many

which are inoluded under the theorem as giZ.


,-;

TWO notiOn* Of c'On-rg.eae la OR

are involved!

vi

with ressect to which the e-tlratos are conistent, and that with
ti

61-

daaAt&

4.

LLW
C~fl~O
aLW~~
j.;@

the samet, urovided the former Lpi1es the latter.

(3)
I,

it

If

and

to thhe hpothess

the

ntly of"

iMrsnda
using th

is,,

cont1ixous f"=ctolon

0 Uroof
he

-~mt

If

1mfom17l

of
0?

OW

' .orf 01 %beuro 1 ma

value o r

f.

e cft -,a h n

onaonc

if

the e-nIties

be .L..xQ

Q. then that proof goes through fory all

cUsarencly ofi

mNi

of theore

Speclficall,

,_ a"
Se unnifrorl1, consistent,

a.-1e
'unlformly
andI" L

it.

z;

adde

may also be added to the conclusions.

the estimates
f

Is

u---ozIt-'

e tu
too
e

0
.ui,.jr

CLmse"

QA tifaidi.nz!
.-

anid itscnitcy

Let uz next consiaer the discrimination problea of tbe


Il'

PROXFC

NUMBEli -i-4-:ri4 it

:RMgi

third kt.-d delineated in Sention I. we


that Lhe densities
~

for

av

and

for

ty

Admlt the niblh

may be any in

i Lc

a4

large'

ara *-

to be cherateorized by a, finite number of


V anA

Thus,
Wiauthr.

ra7y consist of ell uniformly continuous densitie,

or of all continuous densities,

or of all

densities continuous
I

criminatlon procedures which art. reasonable to ..... .. ,n


little

is
r el

assumed about the ronulation-s bpldI-

r ',a

A.C
.. )

served to have the value


Carry out t19e procedure

Fnd

g(z).

for

1. J~
.
L.

..

s-ini-tsoSd ,,
...

L(c)

In the procedure

cre the two real niners

ectimatea

no

Once estimate: have been cbtained,


-

. .

f(s)

(S)

*,-:I

to be estimated, but

we may Inatead proceed to esuolmate the numbers

-_-

. ..

?%.(C)j we employed the estiwarl

as a reann M- obt.inin,

directly.

-I&i,,

s, the onZ', ".'2ormation needed to

tha wotiparametric case there is

i,,

'.- 4

" ,

so

-a.ll designate such proce-vies by

f(s)

g(z)

e*d

we may appl7 the


TjAtej

-f# wh-re

.ia g

ax'; tb' estimnates for

a..-i

g.
g

Befor, considerin6 the problem cf esatimatrag the densitler,


let 4c note the properties whi.cn such estimates stould have
we are ro be able to prove the consistency of

Theorem 2.

mates for
t2

fz)

I?

and

1Pz)

g(t) for all

SMon ()

L* e,

)
g,

f
with

a.re CoBIsteCit esti-

excet
.poSIbly_

& E Zf,

4
S-,...-I
21"4.4-004 E.IPGRT NUr4BERI

..
.. ,.,,"-1u-.:..

i -a Ir

)mO,

PS

f , g))

tben

2#

L(e).

s.',_-twont with

The pruof follows lines similar to that of theorem i

and

will be omitted.
Our problem is now to find consistent estimates for
emd

g(z)-

gous remarks apply to

the same for eaoh value,

We fIx

only, an analo-

f(z)

We shall for brevity o,)flder

a(z).

f(z)

since the argwsnt ii

Z,

Our basiC d'a is

the propil-aln

tt

z X's which tall in a stated (small) negi~borhood of

of the

may be used to eatimatE the

q ina "-t.a

X-proIai

neighDornood.

T.e .atio of this estimated probability to the measure of tba

This is

z.

near

in

t'in

an estimate cf

MRak6 zome assumption about t1e


we ma. let the -ael

w1l1

appro&ch

will- 'v

pjt-o4 off

v2

t havo the ne1lhborhoo4


to
.;s

tre,4-a

MTis will assure that the Prc-

--s earnnot hope to astiit..t.


rths

f(z)

c&-inE1v the distribution of


i::g the distribuUion1

CI

t(z

&zrr

1-1-eIo'"
=AAA.
Fort

t i'aUL" ty

e"or

otherwise we could alter

X20I

To obtain
dowr to >; ab

In the neighborhod is e con.stent estimate of

*,e probab~lity.
It ia obvious
X1W

.
Otfn
f

over the nelhborhoud

but we will take ca

expectation.

a -nftW1

Xis

itself

. so t-hat ."b propo-1tion of the

..

:.,4awl
,

sl,-Ir

f(z),

f(x)

we

if

f(z'

bo-hood sn.K

so that the average of

m--+ so,

average value cf. f(x)

zr
W,
f
a
then an est.a.

neigNborhood is

arbitr&rily without in any way


X 1 , X2,'V

viid thus wLhouL charn-

any sequence of saziuJtes based ot

PRtCJecr NUMIISR 11-49-OO4 REMRT NUMBF R 4

denote Lebosgue meazore In our (p-dimonslonai)

Now let

between points

donute the

iid.a)

d.Mo

7, of tals space.

ard

f(xi

It

yI

x-

and let

sample space,

a- anttnuou

a#

cOf sets such tsat


Is a-sogeno.

and

lin a /4,(40 ) w a,

wb,h

.o in

=Ad if

cnd

-a.

to~ bt shown.

thu

--. k

,,

and we conclude

l .ma

Xt99 xf

Is a constent eotimatoor

the.n

as was

N is -the nwber ofX

Cmbining

--

tM

0LQ2

If f(S)u or 1
completes th. pr-rwf,

ire have in 'Loma 3 a clas3 of a-stiates, ay of which, by


virt~ue of theorem 2# wiil ----

ccnsistent d.i.-R&mination
e..A.

any (nonparaetric) eltaseso

and I

oP

(o-^s'

yw..ibly

for a set of values

While the poe*dures

44

whose meubers are continuofzOlO measi;.e).

of

thb

last

1:1
Fin-Mr NUMN73;' 21J-g-004 REP4HT NUMBER~ 4

of their

~~mu
~*quosIiont

applicability is,ia
--

"ap

and

a"e not large remaai s

be

o;ire

n,.

(Lik.

any a..a.. totio terem,)

We shall ia tMa present section su&3..st sogia alternative P timates for

f(:)

arnd

g(z), whicli seem cn intfltive growids moLoe

likely to glvo good results than the estimates proposed ;*fore.


6'

....... e t

th

--

simplicity of consistency proofs,

"

but need not b@ d;3frable in

Mhe main precticai difficulty in using %ne former estlmates


.ies ir

the choice of the regions

reeions for

g,

saty

these regionm

M and

eL sample

tOO small.. on thaut

;z-cpcrttone

3rali, the numbert

i!

If

)).

(and tho aorresponding

}
It

are made too

oon-& failing ir o t!er.

nat

anolUj

bc %;curate estieatas for the aorr*esponding probabilities


t he other hand, 1L. the -eoions aoo made

,'' (-.An'
2

;oo itrge, the e probabilitase will not ba good appr-oxinations


I
R-

IL

0r,d

JLf

mu t ateer a mli(dle ,ourae.

the swallest velues

.oint_.

of

and

li,%,1,ebi

We might, fonr
N

zXapleo

wa could '.oer.te,

decie
"

choose

to include the choen ntm'oer of


just big enough
AI

-/IL

e nd

But to do so alters the probabilistic p-cpartles; now


and

ara fixed and Z

and A.

are randam.

Are th6 result&


V

of ienmw'9 A still
Even if
happen tht

vmllid?

they are wR may still


near

be in difficulties.

there are naueroias


t

X's, but few

It nay

Y's; but
'C

"-!MRIP. 1

21-49-004 RF7ORT

MrW!-%k1 WNa".B

by going a little further we fint the situation rovera*.4

cea-17Vo

1=4cotin

A.the

ostiatod

tf

but if we take si&p


may bo clos.,

and g

containing a total of

To avoid thise

P*or k.

i a sizwle region# A~n

k points of sitbaz oempi.

th~in procedure seems

'.i an

CThooe a a

lifficulty the following Laea is suggested:


and take, in the nolghbov-hood of

Tb.

bu.t. s.ince

Intui!tivelly

= kwo

haointro-

d~cod do;; 'dence of our ostizatos and fur'ther' altered the p!'70)s
bilistic properties. The question which now arises is whether
or not estimates for' f(s) and S(s) based an N and go
when so determinede ar'e still consistent.

As a first stop In answering th-vso questions* obserwe that


we :nay by mains of a prelan
p dimensions to onve

troui

Let

real valued4 function Of pairs

space.

Suppose

IV

tvA

es~z)and

(xey)

is

wa.vAsom.

n~ow

the probl1em of

each

not bot-h C at 0.

e(Yl92,' .%t a"y'


859*p
It%

etfs

.1le, by

deci~hil

MTese

PzD).'z.)

whet~ar

fts)

L Ps12~'~
'FZ
10

(1.z)t respfctlvely.

0r

eg(s)

f (0) or

cgzu'in larger; and further v~m1moe the saup1v


'l

so *veopt

JR

a.t"e 'Lars er b7 the problem of declidng wbether

lo .2;

"

a,(x)t continuous and

roplace

4*bcz'F

Pj(Zf~g) a 00

*Ie

Pz'cpertles are satisfied. for ex


I.

of points in the sample

so eccostrusted that when z-.

x t Z

" I

asksTow

~I...denoto

0. and suppose fuwta,*r that

9 x)---"*
r'htp~j froi

yiiar~
transformation reduce our space

as

lj,'es.
and

We may nows with-

21A'1-U4 R EPORT 4 UM5:-R 4

IROJEC'T YAJBi

""0M.
I-"#..'---

that f

gnrltasm

are denxi--

an

Zies of non.-nogativo univawlate random vtriables, andd that


ba nail. nowtivee

Xi.~
-a

_,,

cT

d function
,
such
-n .3----+
that

n1 e-1

a~r)=--p 0.#

".nd

restricted so
i

it 0.

a# a

and

k(mina)

Let

k("

n--4oo.
from

U n kth sna oat valu* q't CSubimd BaM


IK= _nvw+
Of Xsa
,
S a= number of Ys
U.
Then
Is a consistent estimatetor

515teflt estimate Lot


;g:10

~~Pi

k . ((mtn )

by

etl.n v(Vn

be .aitL

(This tandenc1 telta


0

and

com )6

a 0OIts

Def

And

Y's,

is a con..

f (0) mid

g(O).
Qand~~l
n-1

k. (a eon ) + k-

-
-

In bounded !Mt

.tha

Lot

(0
M.

= (u .n )

"g 7

k 1 (anu)

at(O)
- - -- -- -- = - - 0

a nd

()+

-;,

(a) a-:
_

'

(z-1)

n)

Observ

~i

tt

Define

number of X's % wual)p


Sa
son

hnumber of

Y'sa

v(Sie),

j~

21-W~-004 R WOR.T NUMBf~

PROJ3t'CT "Ua~

= r~wber of

Yfa cwiiazi).

VaIzig the contin.uity and positiveness or


q O

4 4

Omx 10qo

soai=all thatwhon
rim

:,a
n

wzr.

end

1f

I ( 04

end

ou.dch'that When

< qta, a m.a

hes

gat Go find

a20a

roe tocI~~

VV
f(

1V(M

8)

VA.-

)l

(a)(L

mno
Similarly obeerve

ag(O) v(Mon)(l.. )

Z(iUcd

L4

Sk*pn
~4

wV ,j~~

< ng(O) W(0n0n) (le+

ki~sva)
TS

VIC'

fl-cindon variablos involve,' are b~inomW&Os


t~end to

.heae expectatious

ot but more 810w'7 th~an thO na-berw oft trials, as


rn

: ridelursaps

win tsr

cana assurt
Vi

I;

1ie~ti,.

kNUBE
4
2 (zen))

-49-06

P(N~

an soon as

At
ush~~il.
lugV

VV

msUz

-,W a !A

46Pul)

and
tb

n > l.-,t which r a at--i Ce


that

fnact

UIf

C(,w(upul))i

U > v(msu)

1.1

2e,

I~l.a the event th*t all

Ic Xvis

~~ the

imd

'a~ a44

\m,

-NO

t)&

t-o).(e(i
aUS-)
.(

)LO

S)

X's <
avio~nt

Restrictlibg

j.i' ,

H.I.

~~)i

xad

e. kiu)
(m #n)

P(U>
vut))

I'Ce event

IVA we ~wy make.

--

Y3

ni

*I

PROJECT NUiM3Eh 21-'490-iX

Uwe

flPORT NUILMBER 4

aancC'
v( 1 )

Ike'

the- Crnclusion~

,.(0
s at band.

II A.F

A siailar arg~ment shown

gIo

A situation in which one of the densitlee~ Is 0 at 0 can


be; dealt wilth by a ce-r-xespandiig but, simpler argimnt which we
omnit.

4.

rhe .efect of theorem

large DeMPIG1e

03U.LtS irt

1. to assures us 00. saa-isfactiory

we employ proced~-%za

k. a positive integer which is large but small. com-.

,3hoose8

p~ared to the saemle sizes.

Specify a metric in the saarnr e pame.

10.-exam-ple or(U-arby Euclidean distance.

an4 fin~d* of' the


-%
to
A-A

kc

a-=er

~ipth

"%

--

of'

P0ol the two

MWIT-61

valuies in the po~led samples wb-iah are nea'-

number m'-4ch ars Vs.


a-

oWff he follovtng

X which &~re

1'. .

'irocead with. t-1ie

Let

N m k-4

ihmnd ratio 21&-

Ukik

a.wa

*fl

ndolyI
01o11:

That Lao asig

g(z).

CuiaZ~

(COF ILS OSTA I NASE FMNO

110 63 3

ATI1

CWIVC(4 ITV OF CALWIO&NIA,

CACK))

OELETL

~:i~?i~i ATRYAMLYSI5 - NOWAPARAMTPIC DISCRINI FaATtON:

CONSISTENiCY

PWOECRT113 -

EVELYN

J:;~.L. HDOWGS, JP.

t'SAF

tf1W

8"KE. TEX..

"tAVATION

PFJ6CT RPOrTIf

21~P

NWDICik4L. PA#OVLP1

iwAr ccna. No.

PTAT IST;C4.AL ANALYSIS

FE8151

AP-41(1M)-31
%CI DeV,

%XMCtA0' I IF#L

10-

A 117 fOWE

(NEVORT

Wt.. 41

EWUAAL (33)

PR~jJECT

e~ 3.

VMWEEA 21-49-004 RWOIRT NUiMBEU

WAI!

3. ?EMAISONI,

em4 S3

of tost,,Irag

Contribixttiow to the theory

ev6ttiastical k'.ypothees*4 Stoat,

Res.

Kosoirap Vol* 1 (19.36)s pp. 1-37-

2s

Bo Le UWMHs *Note on disecimimnant- fu.toa'

Bicmetrlkas

VoL. X' (1939)t pp. 218 :;204


3o

Ho A. PIM

*Theus of mulwatiple- ma.aswraemnts ira taei-

nomic Problemse' Analsof

PP- 179-188,

~gnc-s

Vol.7(93i

S-ar putea să vă placă și