• 검색 결과가 없습니다.

Optimal Weights of Linear Combinations of the Independent Poisson Signals for Discrimination

N/A
N/A
Protected

Academic year: 2021

Share "Optimal Weights of Linear Combinations of the Independent Poisson Signals for Discrimination"

Copied!
9
0
0

로드 중.... (전체 텍스트 보기)

전체 글

(1)

2 002 , V ol. 13, N o.2 p p . 307~315

Optim al W e ig h t s o f Lin e ar Com bin ation s o f t h e In de pe n de n t P oi s s on S ig n al s f or D i s c rim in ation

Jo o - H w an K im1 ) A b s tra c t

S u pp ose on e is g iv en a v ect or X of a finit e s et of qu ant it ies X i w h ich ar e in depen dent P ois s on sig n als . A nu ll hy pot h esis H0 ab out E ( X ) is t o b e t est ed a g ain st an alt ern ativ e hy p ot h esis H1. A qu ant it y

i ixi

is t o b e com put ed an d u s ed for th e t e st . T h e opt im al v alu e s of i ar e calculat ed for t hr ee ca ses : (1) sig n al t o n ois e r at io is u sed in t h e t est , (2) n orm al appr ox im at ion s w it h un equ al v arian ce s t o th e P ois s on dist ribu tion s ar e u s ed in t h e t est , an d (3 ) th e P ois son distribut ion it self is u s ed. A com p arison is m ade of th e opt im al v alu es of i in t h e t hr ee ca s es a s par am et er g oes t o infin it y .

K ey w or d s : P ois son sign al, H y pot h e sis t estin g , P ow er , Opt im al w eight s ,

1 . In tro du c ti on

A b eam of n eu t ral p articles can b e u sed t o est im at e th e den sit y or m a s s of an obj ect (F eller (1970)). A m et h od of dis crim in at ion h ere is t o u se a n eut r al part icle b eam (N P B ) aim ed at t h e obj ect , an d a sm all nu m b er of n eu t ron sign als ar e cou nt ed at t h e det ect or . Bey er an d Qu alls (1987 ) sh ow ed t h at t h e n um b er of r et u rn n eu tr on part icles fr om an obj ect in t err og at ion for a g iv en dw ell tim e follow P ois son distribut on . Kim (1995, 1996, 1997 an d 1998) st u died an applicat ion of sign al pr oces sin g of r et urn n eut r on sig n als fr om an obj ect irr adiat ed by a n eut r al part icle b eam in t h e ca se of on e pr ob e - on e obj ect - on e det ect or .

1. A s sociat e P r ofes sor , Dep artm en t of S t at istics an d In form at ion S cien ce , Don g gu k Un iv er sit y , Ky on gju , 780- 714. K orea .

E - m ail jhk @m ail.don g g uk .a c.kr

(2)

T o ex t en d th e pr ev iou s stu dies , w e con sider t h e ca se of on e pr ob e - on e obj ect - k det ect or s . In depen dent P ois son coun t s in k b in s t h at h av e differ en t m ean s in each b in are con sider ed . T h e obj ectiv e is als o t o dis crim in at e b et w een a r e - en t ry v ehicle (RV ) an d a decoy u sin g t h e r et urn sign al. F igu re 1 sh ow s t h e sit u ation w e con sider in t his st u dy . T h is dis crim in at ion pr oblem is form u lat ed a s a t est of h y poth esis :

H0 : obj ect is an R V v s . H1 : obj ect is a decoy

(T h e t h eory of t estin g hy p ot h es es is giv en in Leh m ann (1986).) T h e ob serv ation s are form ed int o a v ect or of n eut r on coun t s in sev er al en er gy bin s :

x = ( x1, x2, , xk) .

< F ig . 1> On e pr ob e - on e obj ect - k det ect or s ca se T h e sign al X i are in dep en den t P ois son r an dom v ariables . Let

E ( X ) = ( t1, t2, , tk) un der H0 E ( X ) = ( d1, d2, , dk) un der H1 T h en

H0 : E ( X ) = t

H1 : E ( X ) = d di< ti for all i

(3)

T o dis crim in at e an obj ect u sin g k sig n als fr om k det ect or s , w e con sider a sum m ary st at ist ic th at is a lin ear com bin at ion of X i

Y =

k

i = 0 iX i

w h er e t h e w eigh t i ar e t o b e ch osen lat er . S in ce t h e i ar e t o b e ch osen p osit iv e (see (2.2), (3.3), an d (4.1) b elow ), w e r ej ect H0 if Y som e crit ical v alu e c . T hr ee m et h ods of ch oosin g th e i ar e g iv en an d com par ed .

T h e fir st m et h od is b a sed on a sig n al - t o - n oise r at io S/ N . M ax im izin g S/ N u su ally is int en ded t o m ax im ize t h e p ow er of th e st at istical t est , defin ed by th e crit erion for r ej ect in g H0. In t h e pr esen t sit u at ion t h e v arian ce of Y (t h e t est st atist ic ) is n ot th e sam e un der H0 an d H1; con s equ en tly , m ax im izin g S/ N an d m ax im izin g pow er ar e n ot equ iv alen t . T h e secon d m eth od of ch oosin g w eig ht is b a sed on m ax im izin g pow er for n orm al dist rib ut ion appr ox im at ion s w it h un equ al v arian ces un der H0 an d H1. T h e t hird m et h od m ax im izes pow er for th e ex act P ois son distribut ion s . N ot e t h at w e on ly con sider th e optim al w eigh tin g of th e lin ear com bin ation s of P ois son coun t s , bu t it h a s n ot b een v erified h er e a s b ein g able t o dis crim in at e t h e hy p ot h esis at r ea son able risk s an d .

2 . S ig n al - t o - n o i s e R atio M e th o d

S in ce t h e X i ar e in dep en den t P ois son sign als , t h e m ean is E ( Y ) =

k

i = 1 iti u n der H0, b ut

E ( Y ) =

k

i = 1 idi

u n der H1. T h e v arian ce V ( Y ) is V ( Y ) =

k

i = 1 2

i ti, u n der H0

V ( Y ) =

k i = 1

2

i di, un der H1

In th e t h eory of t est in g hy poth eses con cern in g m ean s 0 an d 1 w it h com m on v arian ce 2, th e pow er (probability of rej ectin g H0: = 0 w h en H1: = 1) is an in creasin g fun ct ion of t he sign al- t o- n oise r atio (i.e. m ean differ en ce b et w een H0 an d H1 ov er comm on st an dard deviation )

(4)

| 0- 1|

.

T his su g g est s t h at in th e pr esen t ca s e of t h e h y pot h e sis t estin g , on e m ig ht ch oose i t o m ax im ize

S

N =

k

i = 1 i( ti- di)

k

i = 1 2 i di

(2.1)

w h er e t h e den om in at or of (2.1) w a s ch osen t o t h e st an dar d dev iat ion of Y un der H1. S in ce w e a s su m e th e st an dar d dev iat ion u n der H0 is gr eat er t h an or equ al t o t h e on e u n der H1, th e m ax im u m of sign al- t o- n oise is obt ain ed w h en w e u se th e st an dar d dev iat ion of Y un der H1 a s den om in at or .

By t h e Cau chy - S ch w ar z in equ alit y (Roh at g i 1976 p .165 ), w e h av e

k

i = 1( i di) ti- di

di (i = 1k 2idi) (i = 1k ( ti- ddi i)2 )

w it h equ alit y h oldin g if an d only if

i di = K ti- di

di , i = 1 , 2 , , k ,

for som e con st an t K . In ot h er w or d s , t h e sign al- t o- n oise r at io is m ax im ized by t h e ch oice

i= ti- di

di (2.2)

S in ce an y con st an t m u lt iple of t h e i also m ax im izes S/ N , t h e i of (2.2) can b e r es caled s o t h at

k

i = 1 i= 1.

3 . N orm al A pprox im ation w ith U n e qu al V ari an c e s

A s sum e t h at th e in dep en den t P ois s on dist ribu tion s P ( i) of t h e X i can b e appr ox im at ed by n orm al distribut ion s . T h en , appr ox im at ely ,

X i N ( i, 2i)

w h ere i= 2i = ti or i= 2i = di accor din g t o w h eth er H0 or H1 is tru e. Also, appr ox im at ely ,

Y N (

k

i = 1 i i,

k

i = 1 2 i i).

T h e det ection rat e is 1 - , w h er e

(5)

= P ( r ej e ct H0 whe n H0 is tr u e)

= P ( Y c E ( X ) = t)

P Z c -

k

i = 1 iti k

i = 1 2 iti

an d Z N ( 0 , 1) . T h is is equ iv alen t t o c -

k

i = 1 iti k

i = 1 2 iti

= - 1( ) = - - 1( 1 - ) , (3.1)

w h er e is th e cu m u lativ e distribut ion fu n ction of t h e st an dar d n orm al r an dom v ariable Z . T h e false alarm r at e is

= P ( a ccep t H0 wh e n H1 is tr u e)

= P ( Y > c E ( X ) = d)

W e a s su m e di>0 . T h erefor e, t h e di w ill r em ain in th e an aly sis . S o

1 - P Z

c -

k i = 1 idi

k

i = 1 2 idi

or

c -

k

i = 1 idi

k i = 1

2 i di

= - 1( 1 - ) (3.2)

T h e com bin at ion of (3.1) an d (3.2), for fix ed , sh ow s t h at satisfies

- 1( 1 - ) =

k

i = 1 i( ti- di) - - 1( 1 - )

k

i = 1 2 iti k

i = 1 2 i di

(3.3)

T h e qu ant it y

k

i = 1 i( ti- di) is t h e "ex ces s " sign al u n der hy p ot h esis H0 ov er t h at

u n der H1. N ot e t h at t h e fir st t erm in (3.3 ):

k

i = 1 i( ti- di)

k

i = 1 2 idi is t h e sign al - t o - n oise r at io an d is ju st (2.1).

(6)

It is r equir ed t o fin d { i} t h at m ax im ize t h e fun ct ion h ({ i}) defin ed by t h e righ t - h an d side of (3.3 ). On e can a s su m e t h at n ot all i= 0 , sin ce ot h erw ise

= 1. On e fir st a s su m e s t h at 1= 0 an d inv est ig at es all m ax im a of t h e r esultin g fu n ct ion of k - 1 v ariables . T h en a s su m e 1 0 an d div ide nu m er at or an d den om in at or of (3.3) by 1 an d put i= i/ 1, i = 2 , 3 , , k . T h en h ({ i}) b ecom es a fu n ct ion k - 1 of v ariables : h ( { i}) . A su fficient con dit ion (s ee Lu enb er g er (1973)) t h at th e p oint h ( { i

*}) b e a st rict local m ax im u m for h is t h at h ( { i

*}) = 0 an d t h at t h e m at rix 2h ( { i

*}) b e n eg at iv e defin it e. T his con dition can b e v erified b y ch eckin g t h at t h e eig en v alu es of t h e m atrix

2h ( { i

*}) ar e n eg ativ e. F in ally , h av in g ch eck ed all st rict local m ax im a , it is n eces sary t o in su r e t h at t h e fu n ction does n ot b ecom e elsew h er e g r eat er t h an it s v alu e at on e of t h e st rict local m ax im a . T h er e are ot h er pos sibilit ies t h at m u st b e ch eck ed su ch a s n on st rict m ax im a .

T h is pr ocedur e w ill b e illu st r at ed in only on e ca se : k = 2 . A s sum e 1 0 . P u t x = 2 = 2/ 1. T h en

h ( x ) = x ( t2- d2) + ( t1- d1) - - 1( 1 - ) x2t2+ t1

x2d2+ d1 (3.4)

A ft er calcu latin g h ' ( x ) = 0 , t ran sp osin g a squ ar e r oot an d s qu arin g b oth sides , a qu adratic equ at ion in x is obt ain ed :

[ d2( t2- d1)x - d1( t2- d2) ]2( t2x2+ t1) - ( d1t2- t1d2)2 - 1( 1 - )x2= 0 . (3.5) In t h e sp ecial ca se t h at d1= 1, d2= 2 , t1= 3 , t2= 4 , - 1( 1 - ) = 4 , equ at ion (3.5 ) r edu ces t o 16x4- 16x3- 12x + 3 = 0 w hich h a s t w o r eal r oot s , on ly on e of w h ich giv es a v alu e zer o t o h ' : x = 1 . 3998 : an d h ' ' (1 . 3998 ) <0 . T h e con dit ion

1+ 2= 1 fin ally giv es 1= .4274 , 2= . 5726 . S in ce h ( 0) <0 an d h ( ) = - 2 , t h er e is n o ot h er m ax im um v alu e of h ( x ) t o t h e sp ecial ca s e.

F or t h e ca se k = 3 , h in (3.4 ) b ecom es biv ariat e fun ction of 2* an d 3* an d w e m ay fin d t h e m ax im u m poin t for s om e specified v alu es of t h e p ar am et er s . F or g en er al k , h b ecom es k - 1 v ariat e fun ct ion of 2*, *3, , *k.

(7)

4 . Optim al W e ig h t s f or P oi s s on D i s trib uti on s

F or P ois s on coun t s in on e en er gy bin , Bey er an d Qu alls (1987) g iv e a r at h er com plet e an aly t ical an aly sis . F or t w o en er gy bin s , t h e discrim in at ion surface, an alog ou s in t h e dis crim in at ion cu rv e in Kim (1997 ) is a m appin g of R2 t o R2. F or th is an aly sis , on e n eed s t o dev elop on e or m or e t est st at istics . In t his sect ion w e giv e an opt im al t est st at istic b a s ed on t h e ob s erv at ion t h at t o m in im ize is t o m ax im ize t h e p ow er 1 - . W e seek th e m ost p ow erful (M P ) t est of t h e h y poth esis H0. T h e N ey m an - P ear s on lem m a [Lehm an n , p .74] com put es th e M P t est in t erm s of a

r ej e ction r eg ion = {x p ( x ; d)p ( x ; t) >K },

w h er e

p ( x ; d) p ( x ; t) =

k

i = 1 dixie- di/ xi!

k

i = 1 tixie- ti/ xi!

= ex p{- i = 1k xi log dtii }i = 1k e ( ti- di)

is t h e lik elih ood r at io. T h e M P t est h a s t h e form R ej e ct H0 if y =

k

i = 1 ixi c,

w h er e

i= log ti

di (4.1)

A g ain , w e can re scale th e i s o t h at

k

i = 1 i= 1.

5 . Com p ari s on an d R em ark

W e n ow h av e t hr ee m et h ods of calcu lat in g w eigh t s : (1) S/ N , (2) n orm al appr ox im ation s , an d (3 ) P ois s on M P t est . In t his sect ion w e com p ar e t h e optim al w eig ht s for th ree m eth od s .

Com p arison is m ade in t h e lim it for lar g e P ois son coun t s . Let ti= pidi w it h pi> 1 for all i. T h en con sider t h e lim it of t h e i for t h e t hr ee m et h od s of t his p aper a s t h e decoy cou nt s b ecom e infin it e, i.e . a s m in (di) .

F or th e S/ N m et h od, w e ob t ain

i= ti- di

di

= ti di

- 1= pi- 1 pi- 1 as m in ( di) (5.1)

(8)

F or th e P ois son M P t est m et h od , w e obt ain

i= log ( ti

di ) = log ( pi) log ( pi) as m in ( di) . (5.2) F or t h e n orm al approx im at ion m et h od , w e furt h er sp ecialize t h e lim itin g pr oces s s o t h at t h e decoy cou nt s di= d0i for d0i 0 for all i w it h . Ex pr es sion (3.3 ) b ecom es

ipid0i

2 id0i

- - 1( 1 - )

2 ipid0i

2 id0i

(5.3) T h e s econ d t erm of (5.3 ) for v ariou s ch oices of th e i is b oun ded ab ov e (an d b elow ), sin ce t h e m in (d0i) >0 . Con sequ ent ly , a s , th e fir st t erm dom in at es an d t h e i th at m ax im ize (5.3 ) conv er g e t o t h e S/ N w eig ht s ; i.e.

i ( pi- 1) as .

It is int er est in g t h at t h e n orm al appr ox im ation s t o t h e P ois son dist ribu tion s im plicitly in clu ded in (3.3) an d (5.3 ) b ecom e b et t er a s ti an d di b ecom e lar g e ( ) but t h e i th at m ax im ize S/ N do n ot con v er g e t o th e P ois s on M P t est w eig ht s .

R e f e re n c e s

1. Bey er , W . A . an d Qu alls , C. R . (1987 ), Discrim in at ion w it h N eu tr al P art icle B eam s an d N eut r on s , L os A lam os N a tional L ab ora tory , LA - 8 7- 3140.

2. F eller , W . (1970). A n I n trod uction to P robability T he ory and I ts A pp lica t ions : V olum e I I , T h ir d E d., J oh n W iley & S on s .

3. Grav es, R. E . (1986), Sign al- t o- back groun d noise ratio optimization for an N P B n eut r on det ect or w ith en er g y m ea sur em en t capabilit y , L os A lam os N a t ional L ab., LA - UR - 86 - 3101.

4.“Kim , J oo - H w an (1995), P r op erties of t h e P ois son - pow er F un ct ion

Dist ribu t ion , T he K or ean Com m un ica t ions in S ta t is t ics , V ol. 2, N o. 2, pp . 166 - 175

5. Kim , J oo - H w an (1996), Err or Rat e for th e Lim it in g P ois son - pow er F u n ct ion Dist ribu t ion , T he K or ean Com m un ica t ions in S ta t is t ics, V ol. 3, N o. 1, pp . 243 - 255, 1996.

6. Kim , J oo - H w an (1997), T h e M inim um Dw ell T im e A lg orith m for t h e P ois son Dist ribu tion an d th e P ois s on - p ow er F u n ction Distribut ion , T he K orean Com m un ica tions in S ta tis tics , V ol. 4, N o. 1. pp . 229 - 241.

7. Kim , Joo- H w an (1998), M on ot on e Likelih ood Rat io P r operty of th e P ois son S ign al Dist ribution w it h T h ree S ou rces of Err or s in t h e P ar am et er , T he

(9)

K orean Com m un ica tions in S ta tis tics , V ol. 5, N o. 2, pp . 503 - 515.

8. Lehm an n , E . L . (1986 ), T es ting S ta tis t ical H yp othes es , 2n d E dit ion , J ohn W iley & S on s .

9. Lu en b er g er , D . G. (1973 ), I n trod uct ion to L in ear an d N onlin ear P rog ram m ing , A ddison - W esley P u blishin g Com pan y , 1973.

10. Lohat gi, V. K. (1976), An Introduction to Probability T heory and M at h em at ical S t at istics , J oh n W iley & S on s .

[ 2002년 9월 접수, 2002년 10월 채택 ]

참조

관련 문서

• We calculated effective capacitance of series or parallel combinations of capacitors.. • Batteries (Voltage sources,

The proposal of the cell theory as the birth of contemporary cell biology Microscopic studies of plant tissues by Schleiden and of animal tissues by Microscopic studies of

14 For individuals with a triglyceride concentration of 200–499 mg/dL, pharmacological therapy should be considered to lower triglyceride concentration after

z 차별에 대한 경제적 분석은 노벨상 수상자인 Becker(1957)의 연구 “The Economics of Discrimination” 에 기원하고 있음 Æ taste discrimination... z Becker의

_____ culture appears to be attractive (도시의) to the

It considers the energy use of the different components that are involved in the distribution and viewing of video content: data centres and content delivery networks

After first field tests, we expect electric passenger drones or eVTOL aircraft (short for electric vertical take-off and landing) to start providing commercial mobility

1 John Owen, Justification by Faith Alone, in The Works of John Owen, ed. John Bolt, trans. Scott Clark, &#34;Do This and Live: Christ's Active Obedience as the