A Method of Estimating Plane Vulnerability Based On Damage of Survivors" by Abraham Wald (CRC)
A Method of Estimating Plane Vulnerability Based On Damage of Survivors" by Abraham Wald (CRC)
A Method of Estimating Plane Vulnerability Based On Damage of Survivors" by Abraham Wald (CRC)
smmhhhhmmml
EIIIIIEIIEIII
IIIIIIIIIIIImI
IIIIIIIIIII
ollllllllllll
Ot F
!9 0
A S T-
'"Afl 4 t Ali
:ON DAAE.O
ACAM
Abhp W
~~Ai-
1ww
ii itI , iW -R
See Block IS
7. AUTHOR(@) 2. CONTRACT OR GRANT NUMMER(#)
Abraham ald
17 PER61RMING ORGANIZATION NAME AND ADDRESS 10. PROGRAM ELEMENT. PROJECT. TASK
AREA II WORK UNIT NUMMERS
Statistical Research Group/National
Defense Research Commwittee,
Applied Mathematics Panel
I I. CONTROLLING OFFICE NAME AND ADDRESS 12. RFPORT DATE
Center for Naval Analyses ui'1980
2000 N. Beauregard Street 3 EAOF PAGES
Alexandria, Virginia 22311
14. MONITORING AGENCY NAME & AODRESS(il different from, Controlling Office) IS. SECURITY CLASS. (of this report)
17. DISTRIBUTION STATEMENT (of the abstract entered in &lack 20. it different froat Report)
20. ABSTRACT (Contineue on reverse side 11 necessary mid identify by block imober)
145-
LUj
0
z0
Z oMEMORANDUM FOR DISTRIBUTION LIST
0
X Subj: Center for Naval Analyses Research Contribution 432
CL
0
. Encl: (1) CRC 432, "A Reprint of 'A Method of Estimating
< Plane Vulnerability Based on Damage of Survivors'
< by Abraham Wald," July 1980
>N
> T I. Enclosure (1) is forwarded as a matter of possible
interest.
Z 4 2. This Research Contribution contains a series of
-memoranda written by Abraham Wald of the Statistical
Research Group at Columbia University during World War II.
Unfortunately, this work was never published externally,
although some copies of his original memoranda have been
available and his methodology has been employed in the
I I. analysis of data from both the Korean and Vietnam Wars.
-It is published by CNA not only as a matter of historical
interest but also because the methodology is still relevant.
(. PHIL E. DeP6Y
Director
Operations Evaluation Group
By--'' -
By ..... ....
. ..
Dis k A'
DISTRIBUTION LIST
Department of the Navy
SNDL
A2A CNR
B3 National Defense University (NDU)
B3 Armed Forces Staff College (AFSC)
E3A NRL
FF38 USNA, Nimitz Library
FF44 NAVWARCOL
FKA1A COMNAVAIRSYSCOMHQ
FKR3C NAVAIRTESTCEN
FT73 NAVPGSCOL
Other:
Defense Technical Information Center (DTIC) (12 Copies)
Department of the Air Force (SAMI)
Air University
Air Proving Ground Center, Eglin AFB
Institute for Defense Analysis (IDA)
Mathematics Research Center
The Mitre Corporation
The Rand Corporation
University of Rochester (10)
Navy Tactical' Support Activity
LSRC-43 July
1/ 1980
1/
Aa REPRINT OF
---- AMETHODOF STIMATING
PLANE VULNERABILITY §ASED
_ ONQAMAGE OF 5URVIVORS''/
BY ABRA-HAM/AL D'.
Abraham Wald /1
Ll;
Vr 104 A*r#10"
-- A:I t . . . 01 .f
$4%%~r-
l PAW-
*0*4..
tt b'St w
va4u Wtttj4U w~
7f7
4:li t
A 0r. *WA' o-,
A A
P ~ ~ ~ ~
*'A r" 4WfA7~AQ-
@4~qf
ir 0*ro -~r$Wufl ths m4,6taft
i~art Ito,
Aft ct
P.t4 njr~jp0y'' Th nasber.e
kpsaet gboe.
iu~ar
tfi wre aw.uir4 Aroflh tin fltina~l
Y'I
. :V;-
0A A
PART I
=
Qi qlq2"" "gi (1)
and
P1 = I - q 1 q 2 ... qi • (2)
A.
Denote the ratio R- by a i (i = 0,1,2,...) and let L be the
proportion of planes lost. Then we have
a 1(3)
i=0
.......
The purpose of this memorandum is to draw inferences concerning
the unknown probabilities pi and Pi on the basis of the known
quantities a0 , al, a 2 ,..., etc.
-2- el
is
CALCULATION OF x. IN TERMS OF a,0 a,,...,a,
np .I.,pn
x Pi(I
1 a). (7)
hle proportion of planes that received at least two hits and the
firs.t hit did not down the plane is obviously equal to
II- aa -x. Hence, x) i23..r 8
Putting
c I - a0 - a1 .. - a. 1 (1u)
11 + (xI+ .. + x i- c
C- (i =3,4,...,n). (13)
1 i- i-
-3-
Adding x 1 -1 I x i-I to both sides of
+ q~i-1 14
1i-i i- - 11(4
Hence,
and
= d, and = ~ 1 + d2 (20)
-4-
Prom equations 19 and 20, we obtain
x- 1
x =i
i (d.) + d
i L j -i p. i ' P i
Hence, in particular
n qn [n-i tn ) + (23)
Xn Z x = cn - n E (djtj+l'' d =]b= .
X
j=l
=1 [J=n
Pn
Since c L = an, and since t t ...V q .. qn- we
nnj+l n y q
-5-
an a+ a ~ - n-I1 d
an an- 1
q, ... q1 "'" q.n-
(25)
n- a_ 1
+ = 1q, ' qj-1
d1
Pl
n a |
= .. qj (I - a O ) =.0
j=l q.
or
1 -a . (26)
l q1 "' qj 0
n a.
1=1 qJ 0
It is easy to see that there exists exactly one root between zero
and one. We can certainly assume that q1 - q2 I-... > qn" We
shall investigate the implications of these inequalities and
equation 26 later.
-6-
.-. v
Yi - Pi b i Pi(ai + yl) (i = l,2, ...,n). (27)
Hence,
P. 1 - q, "'" qi ai
yi = O- a a. = -a.. (28)
n
Since Yi = L, we obtain from equation 28
y
1=1
n ai n
.. =q + a i =I - a . (29)
NUMERICAL EXAMPLES
a. = A.
t-(i = O,1,2,...,n)
14N
-7-
loom
" The bombing mission is representative so that there is no
sampling error.
" The probability that a plane will be shot down does not
depend on the number of previous non-destructive hits.
n a.
j=l q1 0
reduces to
n a.
2
E - -= 1 -a 0
j=l q])
Substituting values of a.
1
-8-
or
.200q 5 - .0}80 4
- .050q 3 - .010 2 - .005c - .005 = 0.
X i l - ao -CAI - . . - al
= Vl -X x 2
2 - . - xi- )
1 1 0
(i = 2,3,...,n)
for n = 5, we obtain
x I = P)(I - a0 ) = .030
x = P2(I - a - - x = .013
xo 1 - 2 x 2 ) = .004
3 = P 4 (1 -
a
a - I 2 - a 3I1 -
-1 x - x2 - x3 ) = .002
x5 p 5 ( -a aO - a I - aa 2 -a -- a 4 x - x I -
-x x 2- - x - x )
5
4ap55(I .001
-
4 - 3 2 3 3 x).0
4
-9-
TALE 1
I. Assume q I -Y
A
- 2 - - - 1 - .0990 - .9010
2. Assume q - .9010 - 2
ab
3. Assume q - .858887 - 73
C
4 73 .859807 - .007632 - .851255
4. Assume q - .851255 - 74
p
Y5 - Y4 " -" .051255 - .000234 - .051021
., .. . .~ .- ...
PART II
Aj > 0 and
Aj+l = -Pj+l Aj
Hence,
Hence,
In general
)
Aj + Aj+ 1 + ... + Aj+k = (1 - Pj+l) ... (I - Pj+k Aj > 0
(k = ,...,i-j)
-12-
°
-*
planes is not changed, Now let pr =P
i.e., it is equal to L.
r I
for r > i. Then the proportion L' of lost planes corresponding
to P ,..,pA is less than L. Hence, there exists a positive
A so that the proportion L" of lost planes corresponding to the
probabilities pr = p. (I + A) is equal to L. But, since p" > p
i i i
(r = l,...,i)we must havejL x! > L x = Lx Hence, we
j=1 I j=l I j=l j
j= - I - a . (30)
Z q . qj 0
qI ... q j 0
-13-
Hence, there exists a positive factor A < I so that
n a.
=1 a 0
j=l 1
Now we show that there exists at most one value j such that
1> qo > qo Suppose there are two integers j and k such that
) 1
>
I> qo> q. qO Let j' be the smallest integer for which
= qJ0 and let k' be the largest integer for which q = q°
q
Let qj, = (1 + £) qj,,
0 1 q 0 , (E > 0), and q =
j k' IT+ k r =r
n a
q . 0. "" qi and -< I - a0
r=l q1 ... q
q[
Hence, there exists a positive factor A < I such that
n ar a
r=±1 r -
<
where q= b ut q* ... qt < ql "'" i= ... qo, which
-14-
II
"" ....... ,.............................
-"'" .......
'" ....
......
°"' .... " "-'"-' -i.................. '....... ' . - : ... ,"
It follows from our results that the minimum of q is the root of
the equation
r=.~r=i a0 (32)
_/ a a a
q~d 1
+1 +
2
~n . qn-l
1 1- a 0 . (33)
2 2 q2 )/
__( a2 a an
q2 j (a + 2 3 +2 + n 0 (34)
(Lagrange multiplier = X)
(1 -a)=0; A l 2
-1.5-
. . .. . . .. .'
or
q( 1- a2 2a3
2 + + (n-n- 1)a n (37)
0 q2 q 2q 2 -
'
q1 =
o
a2
1 + aq2 +
q2
a___n__
a, + ... +
q2
)- (38)
a3 2a 4 3a 5 (n- 2)an
2 + 4 n-i - a1 = 0. (39)
q2 q2 q2 q2
a1
value q, > T- equation 38 has exactly one positive root in
0
q2 " Denote this root by O(q1 ). Hence, O(ql) is defined for
a1
all values q, > I - a It is easy to see that
-16-
t.
'.. ,......................................................................................."......
. ..... ..
, .
•li ra (q i + o"
Hence (assuming
a1 a I > 0)
lir (q) =+
°
q,
where )
*(q = ql ONi )
must be reached for this root. This proves our statement that
of equations 38 and 39 satisfy the inequalities
if
q 1 roots
1 the > q2' then for these roots qlq 2 becomes a minimum con-
-17-
. .-
.....
.
.
.~
n a
r o
S+ 2 3 + +
n I - a (40a)
+ + 2 n-I 0
ql qlq 2 qlq 2 qlq 2
q __ a
_ + ** + n
aa/0o ~ i-1l(-
-qq qlq n-i 2 q (41)
2
and
(i - 1)qlq i-2
2 -
x
. (9
a2
2 +
2a3
+ +
(n - 1)an
+'" 0
1 q2 (41a)
qlq2-
Substituting 1 - a for X (the value of X obtained from
0
equation 41), we obtain
(i - 1 1
- aO a2+23+(
+ 2- - + + n-I )an 0
-ql
0 q2 q2 q / (42)
-18-
°,
L-------------. . .. "'-. . '5 '5"" '.:, - •,
From equations 42 and 43, we obtain -
3- =1-a (45)
j(qS)J
and q" is the root of the equation
a 2 +a3 2+ ,+ ann1a0(6
(q11 2 (q")n = - .(6
qf
n n- r
n +I
=o 0 +a I + .+ ar-l
a. = a.jrl1,,,
qj qj+r-1 (
i ii- r +
3=1 (v')l 0
a2 a3 a **
a1+- + -~ + 0.. - I- a .(46*)
1 v" (v") 2(v") n-
(c~ ~ = nO,
(c 0) q 1 and q for all r ,'n.
-20-
n a
r=l q, ... qr
where
qr =
r
* *
an
-a o -a- an-i
-21-
PART III
a
It is clear that for any value u >1 r .. aoar_
lim ([,u]i-r' =
U. r~u G (48)
-22-
Vj ,'. . : , _ . _ . . L . , _ _ ... - - , .i +
and
u
It follows easily from equation 47 that if u-* -a -r
o .. "- ar-l
then #r(u) - +w. Since i > r, we see that equation 48 must
hold. It follows easily from equation 47 that lim *r(u) = 0.
U=+ OD
r(u n- r
Vie also see from equation 47 that if u-*w, the product u
must have a positive lower bound. Equation 49 follows from this
and the fact that lim *r(u) = 0.
U-0
We have seen in part II that equations 43* and 44* have exactly
one positive root in the unknowns, q and q 2. Let the root in
* 0 * o
q be u0. Then the root in q 2 is equal to # (u0).
I ir r ir
From equations 48 and 49 it follows that u [*r (u) Iris
a
strictly decreasing in the interval 1-a - aI u
U< U
- - a. ~ i
0
and is strictly increasing in the interval u. < u < +o.
Denote by u'
r the positive root of the equation
ar dr+l an
+ +
+ = ... n-r+l I - ao - ... - a rI. (50)
u u u
O
u ri#~r
[l)]iur, [l)i-r,
[r and u?r rir r)ir "
[r(u
-23-
Amu
-&
L7.
In this case,
ir
= u'
r
IrlU)i-r if U 0(u) i > 0 for u = u'
-- r
and
In this case,
and
In this case,
M
= 0
M.
r ur . jjYlu
o~i-r
0
ir I
d~ru) dv
and - u- =- can be obtained from equation 47 as follows.
-24-
Iar
Denote -+
ar. 1
- + .. + -jj
an
by G(u,v). Then
-
rr
(52)
1(ar + lan
+- + **+ n-r
u u uV /
r+1 r+2 (n
- + -3 + . . n-r+l
vv
+ V3 n-r+l/
# (1)
value of d
diu [,(u)]i-r
rin for u =u,
r and
and u = 1 if u, r
an ~ a
4.-i_ .+ + -r 1 a -al - . ar-
v vn-o 1
r(1)inU41 (53)
-25-
Thus, for carrying out the investigations of cases A, B, and C
for r = l,...,i-1, we merely have to calculate ui ...,u'.
Since u(
r Yr and we can say that Mir is
the smallest of rr)
the three values u'~,
where
ti = Min
1.2.,
(uP)i, (u )i-.. (ui_
(U! 1 2 ,'(
V U! 5
(55)
2o
0
To obtain a lower bound z i of Qi' denote by y" (j = Ol,.,i-l)
the proportion of planes (number of planes divided by the total
number of planes participating in combat) that would be downed
out of the returning planes with j hits if they were subject to
i - j additional hits. Then
+
Pi *Yo + y + Yi- + x. + x 2 + "'" + xi " (56)
-26-
It is clear that ajPi> y. (j = O,1,...,i-l) and consequently
Hence,
Y + yI + * ++ < Pi (57)
a0 + aI + ai-l
i+
Equation 56 can be written
(58)
( - a0 - - a. _ x a+
1- 0. + x.
* . o - -9i_
i.
0i a0 a-
y. + .". + Yi-l
Hence, P. is a weighted average of YO + + ai-l and
. 0 a + + ai-
Px + . . + i5
i 1- - a I - ... - a (59
x + ... + x(
1 + 9"0 + xi < pi < I-a 1 i (60)
Hence,
1l - + .. + - ai_ Q <
o 1 - (x + + x1 ) (61)
-27-
In part II we have calculated the maximum value of x I + ... + x.-
NUMERICAL EXAMPLE
Qi = qlq 2 - qi
q, = q 2 = ... = qi = q0 (say),
O
Qi < o i
If
0
Pi - I -0 0 ,
-28-
bad""
0
PCo is the least upper bound of P.; that is, the probability of
" 0
The calculation of ti > Qi'
" The calculation of z Q
0
" The exact value of QO.
For r - 1, we obtain
a I + a22+ a33+ a1
4
-T+ a 55 aO0
u U u U u
-29-
......
..
which reduces to
u' = .851
For r = 2,
a2 a3 a4 a5
U U U U
which reduces to
4 3 2
.12u - 05u - Olu -005u =0
u2 = .722.
For r =3,
a3 a a5
-+- + -- la a a
U U2 U302
which reduces to
.07u3 - .0h
l 005u -. 005 = 0
u;= .531
For r = 4,
a4 a5
-+ 2~=1 a0 - a1 -a 2 - a3 ,
U U
which reduces to
u 4 =.333
-30-
t Mmi 1
(up), (up'1 1 ... 2,
(u!) (up]
We have
Hence,
t1=Min [(U!] =U
= .851
4
t
4
=Min
I (u (u2, (u3. 2 , (u4.)]
= Min [.524, .376, .282, .333]-
= .282
2. Calculation of z. (z. Q0
L1 i i i
a1 a2 a3 a4 a5
q + 77+ -+ -+ -9. I -a 0
q q q q q
-31-
.-~*
..
. ' i..
x, .030, x 2 = .013, x = 04, x .002, x5 = .001.
A, = x, = .030
x
A = +x = 043
2 1 2
A3 = x + x 2 + x3 =.047
A 4 = X1 + x2 + x 3 + x 4 .049
A = x I + x2 + x3 + x + x5 .050.
5
A. i-:
zi = 11 - a - - ai 1 Qi
Then
A1 _ .030
z = 1 1-a 1 .0= .850
o '
a a.0
0
A3
0
z33 1 -1 - a=
aO - a -a a 2 1 =4.329
I .07
=1 A4 .049 = . 183
1 - '04
4 1 - ao - a1 - a 2 -a 3 .06
-32-
1- 0<
i1-<t t 1 i-12..5
0
The exact value of Q. is obtained as follows:
M. - Min ( ,)i-r+l i i-r F o i-rl
}
ir Ur'(r+l) 'ir [fr(Uir
0 0
where u. and 0r(u.r) will be defined below.
ir r ir
Q = Min [Mil,...,M i , i -l ]
O = Min It .851
22 = Min it22'
, u20
211 [ (u21 2)]h o
3 = Mn ft 3 , u3 1 [ *l(U3 1 )] 2 u32 [ * 2 (u32 )]}
oo o 2 o 0
4= M tV
4 , u4 1 * 1 (U41) , u42 [ 0 2 (u4 2)J , u 3 0 3 (u43)]}
*
31 A(U2.
032 2U320
u41 .l(U41)
42 2 42
u42 (u42
u43 3u43)
-33-
The following equations have exactly one positive root in q* q*.
1 2.
O in qo2hiss
root inq
The root in q* is u ir;; the roo (ir)0
(u
a*2 a*
3 n____
,a* + + + * + -In (1 a *)q*
1 q2(q) 2(q*)nl
1 * ~ 2 o2
where q* satisfies
where
n* =n- r+ 1
a* =a + a + .. + a
o o r-i
a j+r-I
S=i - r + 1
TABLE 2
u i r n* 1* a* a* a* a*
_ir o 1 2 3 4 5
u31•
o 3 1 5 3 .80 .08 ...05 .01 .005 .005
U41•
o 4 1 5 4 .80 .08 •
.05 .01 .005 .005
u42
o 4 2 4 3 .88 .05 .01 .005 .005
where
-34-
3v
A
I-
a 0 0
on
o
*.4
400
10c
91
a. 0
8 a
s. +4 am4
aI I 9
I ~ a oftI OS
o on0
,_I.o
-o ,V *,. o.o
a- + -
* 4
u ~ 9 9- 0.2 '
* - -* 4
-35-
O N
r I
ft
-"
, 4
44 U
++
|6 OOm
Off"
S0
00
-- 36
S-"a
*" ml
a 1' . -
Substituting the values from table 3 in equation A and neglecting
several terms as explained in table 3, we have
0.
Ql = .851
0 = Min 1.722, .7211 = .721
0 = .091
TABLE 4
i ZQ t. q
>
q, I q 2 ... > q5 '
Note that
Z Q' = tl = qo °
-37-
This is always true.
i
Qi = q (i = 1,2,...,5).
o o o 3
In table 4, 01 = qo and Q2 is very close to q 02 Q3 and q0
differ by approximately .1 and the agreement between Qio and
i
qo gets progressively worse. It will usually be true that
0
and Q are approximately equal for small values of ii but will
differ widely as i increases.
-38-
-e
PART IV
q0 ql= xq0
A2 qj (j
j- i,...,n-l) (64)
(4
must hold.
1
This part of "A Method of Estimating Plane Vulnerability
Based on Damage of Survivors" was published as SRG memo 89 and
AMP memo 76.4.
-39-
LetLtqr
r r - l,...,i and q+
qofrfor r'n - A q!3 ffor j = i,...,n-1.
Then we have
n a.
q.. q!= qL..and.
o < I.- a .(65)[
n a.
z= 1 =j
0 0
in contradiction to our assumption that q.' q? is a minimum.
Hence, Lemma 1 is proved.
Proof: Assume that Lemma 2 does not hold and we shall derive a
contradiction. Let u be the smallest integer greater than one
0 o
such that qu > Xlqu_ 1 it follows from the definition of the
integer u that if u > 2, then q-= Alqu
0 2 From assumption 63
it follows that q, < 1. Hence, if we replace qu-I
- A2 q I n-1) is ful-
for r P(u, / v, then lq _jqk+l (k
filled. Furthermore, we have
=qoj..q
. 0 and aa <
S" ii j=l qj "'" 0
n a.
j -- ql ... q" o
-41-
r+l a n-r-1 a
+ r+l+j =1-a
j=l jti- 1 ) r(r+l)+rj j(j+l)
Al 2 q] X1 2 2 qr+l+j (66)
(r6
Then
r(r+l ) (i-r)(i-r-)
2 i
E*2 r 1 2 gr
(67)
MINIMUM OF Q
0 0
Let ql,...,q n be values of ql,...,q n for which Qn becomes a
minimum. We shall prove that qj = (j =
0lqj ,...,n-).
oXq o
> Xq0
Assume that there exists a value j < n such that qj0
ql ... qn n a.
ql ... qnjand j
= 1'"
q ... qo
-42-
IT-
But then q" ... q < q~
0. q 0in contradiction to the assumption
1 n 1 n
that q° ... qn is a minimum. Hence, our statement is proved.
S2 q]
n(n-l)
X1 2 qn
then the minimum of Qn is equal to
MAXIMUM OF Qi (i < n)
q = Xq* i,...,n-.) (8
must hold.
n a.
q "'" q! = q .. q! and F q ..--- ] > I- a.
j=1
n a.
jEl q1 ... q
-43-
I'
Lemma 4: If for some j < i we have qj > then
q*+l= Xq for k = l,...,j-l.
Proof: Assume that qj > X qj for some j < i and that there
j+1 1 j
exists an integer k < j-1 such that q* < A2q " We shall
k+l 2k
derive a contradiction from this assumption. Let u be the
smallest integer such that q*
U+l < 2q* Furthermore, let v be
2u .
the smallest integer greater than or equal to u + 1 such that
u
qv > Xlq*v It is clear that v < j. Let q 1 C > 0),
q (1 + c) q, and q' q* for r yi u, / v. Then for suffi-
v v r r
ciently small c we have
Furthermore, we have
n a.
=
qj... q! q ... ql and qq' > 1- a
(r a l,...,i-l)
Let Dir irq be the maximum of Q.
1 under the restric-
tion that q j+l = 1q.j for j = r+l,...,n-I and qj+l = X2q. for
j = l,...,r-l. From Lemma 3 and 4 it follows that the maximum of
Q. is equal to the maximum of the i - 1 values Dil,...,D _
ii
The computation of the exact value of D.i can be carried out in a
way similar to the computation of Mir in part II. Since these
computations are involved if n is large, we shall discuss here
only an approximation method.
-44-
Let D* (r = l,...,i-l) be the value of Q. if q = Alq for
* 1 j+1 1)i
j = r+l,...,n-l and qj+ X2 qj for j - l,...,r. Furthermore,
let D* be the value of Q. if qj+l = Xlqj (j = l,...,n-l). Then,
10and
if A is not much below one, the maximum of D and D#
I
(r = l,...,i-l) will be nearly equal to Dir. ir i,r-l
Hence, we obtain an
approximation to the maximum value of i by taking the largest of
the i values D ...,D*
10 1,i-I"
Then
r(r+l) (i-r-l) (i-r)
D =2 2 + r 2 1
ir 2 gr
MAXIMUM OF On
-45-
Furthermore, we have
n q.
S ' = q* and > 1- a
j= q1 ... q 0
n a
F= q . a0
j=l 1j**
n a.
j=l J (3- ) 0
A2 qJ
X2
NUMERICAL EXAMPLE
-46-
lower bound to the Q i could be obtained. The assumption here is
that the decrease from qi to qi+l lies between definite limits.
Therefore, both an upper and lower bound for the can be
obtained.
We assume that
Alqi -qi+l 2 qi
n a.
j~l<9jl) <1 a (A)
i 2
is satisfied.
ao = .780 a 3 = .010
a 1 = .070 a 4 = .005
a 2 = .040 a 5 = .005
1 = .80 2 -. 90
.07 • + .04
-- t .83
.01 .0056 10
.005 "209
.20529
.8 (.8) (.8) (8)
which is less than
1 -a = .22
-47-
-~ - -~ - -~-~ -. ~
a1 a2 3 a4 a 5
+
~+ l--* 3 + +6 1I05'-' = 1 - a o 2 B
q X2 q X2q X2 q
= 844.
a1 a2 a3 a4 a5
+ 2 3 .33 4 + 4 6 5 ===1aI
2 3 a0 C
(C)
q
q xlq X1 X2 q 1 2
xA 1Xx 2q
g = .904.
a. a2 a3 a4 a5
q X+q 2 X3 q 3 X5 xq4+ 7 3 = -a
Xq 2 3q3 2 X2
2= .941.
-48-
a, a2 a3 a4 + a5a(E
-+ + 3 -~ -3- 5 =1-a 0 (E
q AIq Ai~q Aq Al A2 q
9 3 = .964
where
b~ir)= r)(i -r - 1)
= .844
= .904
92= .941
93= .964
-49-
TABLE 5
Q. i r a(i,r) b(i~r) g ri
E
I1
Q4
3 4 3 0 320 6 .844 .507 .270
-50-
The lower limit of can be obtained directly. The lower limit
of
a a a a
2
1 + a 4. a
3
+ ~ 4
+
0 5
q = .974
2 33 050
g* .974
-51
al a2 + a3 + a4 + a5 =-
2 2 3 3 34 4 65 0
q q xxq x Aq A l
5 2 3 2
=* .905
al a2 a3 a4 a5
q ~ \2qq 2 l \xq 0o
g*= .869
al a2 a3 a4 a5
2+ 33
-4 - = 1- a
q 2qXq 0
-* .851
-52-
Next, calculate the i numbers defined by
b(i,r)
a(i,r) 19(r *i i-l),
Dtr )2 -0,1...
where
g .974
9*- .905
9*- .869
9*= .851
X10 *5
5 ~2 q
where q* is the positive root of
a1 a 5
q ,2q A2 q X2q A2 q
.07+
.4 .0 + .005 + .005 - .22
q .9q2 729q3 531441q4 348678q 5
-* .844.
-53-
TABLE 6 F
COMPUTATION OF UPPER LIMIT OF 0.
01 1 0 0 0.7 94 .7
Q2 2 o0 0 1 .974 .949 1 75
2 1j 1 0 .905 .819__ .737
Q4 14 4
0
1
0
3
6
3
.974
.905
.890
.671
.236
.250
-54-
The upper limit of
Q-(.9)1 (.844)~
-149
-56-
I.
assumption that dummy bullets are used. It can easily be shown
that it is impossible to estimate both Yi(il,...,ik) and
Q(ilb...,ik) from the damage to returning planes only. To see
this, assume that k is equal to 2 and all hits on the returning
planes were located in the area A.. This fact could be explained
in two different ways. One explanation could be that
Y(i,i2) =0 for i2 > 0. The other possible explanation would be that
Q(ili 2 ) = 0 for i2 > 0. Hence, it is impossible to estimate
both yi(il,i 2 ) and Q(i1 1 i 2 ). Fortunately, Yi(il,...,ik) can be
assumed to be known a priori (on the basis of the dispersion of
the guns), or can be established experimentally by firing with
dummy bullets and recording the hits scored. Thus, in what
follows we shall assume that Yi(il,...,ik) is known for any set
of integers il.,k
Clearly, the probability that i hits will not down the plane is
given by
-57-
O
-N NE~ N
6
The probability i(il ... ik) can be determined from the distri-
bution of hits on returning planes. In fact, let a(il,...8ik)
be the proportion of planes (out of the total number of planes
participating in combat) that returned with i hits on area Al,
i 2 hits on area A 2 ,.... and ik hits on area Ak. Then we
obviously have
Qia(il , ...,ik)
Si l k 1 (72)
n a.
E -2 = 1 - a0 (73)
j=1 Qj
-
We have assumed that q q2 "'" >- q n' This is equivalent to
stating that
-58-
........ M,
Q(i..,rl, r ...,
0") (75)
then
Q•'* " r
Q(Ierir,Ir iLr+" , li,...-,li~
k) < 01 * oj -'r+lJr+l'*~ k
I , -(l.-J~lJ,.,k
(76)
for r = l,2,...,k.
k 1 k I1
-59-
for all values il,...,ik for which iI + ... + ik = i, or
a(la •i ( l...iF
2 1
Q(il,...,ik) = [q(l)J [q(2)] ... q(k) . 78)
i i
[Y(1)] i .. [y(k)) (79)
Yi i Yi(l' k
,.,k) 1 1 . iki
-60-
- II
4 aR
I' k
6i'l,...,k ) i1 ! ... k! [6(1)1 ...* [6(k)] .(G0)
Furthermore, we have
6(i) =- k
y(i)q(i)
~ Y(i)q(i)
i=I
Since the probability q that a single hit does not down the plane
k
is equal to . Y(i)q(i), we obtain from equation 81
Y(i)
q(i) =q (82)
Jk 31
,3
6(i) = ~ l.(3
+ +
"'" Jl "" k)a(jl''''. k
S(j 1 +
*** +j
3k 31
j=l a- (84)
-61-
NUMERICAL EXAMPLE
* The probability that a hit will down the plane does not
depend on
= the number of= previous non-destructive hits. That
is, ql q2 "' qn qo
o Given that a shot has hit the plane, the probability that
it hit a particular part is assumed to be known. In this
example it is put equal to the ratio of the area of this
1
part to the total surface area of the plane.
N = 400,
whence
A = 240 a = .600
0 0
A = 68 a = .170
1 1.
A = 29 a = 072
22
A3 = 12 a3 = 030
A4 = 10 a 4 =. 025
-62-
As before, the probability that a single hit will not down the
plane is given by the root of
a2 3 4
q qo qo qo
which reduces to
°
.4q - - 0 72q .030q _ .025 0
0
and
qo= .850.
Ratio of
area of part
to total
Part number Description Area of part area (Y(i))
The ratio of the area of the i-th part to the total area is
designated Y(i). Given that the plane is hit, by the third
assumption, Y(i) is the probability that this hit occurred on
part i. Thus
-63-
Y(l) = .269
Y(2) = . 346
Y(3) = .154
Y(4) = .231
A i + 2A 2 + 3A3 + ... + nA
1 39 .193
2 78 .386
3 31 .154
4 54 .267
q(i) = qo
-64-
- .'
t.
" .
. -
whence
6(1) .193
q(l) = f-(1 q0 .269 .850) 61
6(2) .386 95
q(2) Y C2 So .346 .850)
YT3) =
q(3) 6(3) ' o .154"
.154 (.850) = 85
- 5
-65-
a"
.w - .. ..
.... .. . . . . .. ..... . ..
........ ........ .___
PART VI
SAMPLING ERRORS 1
(i = 2,3,...),
n a.
I - a0 , (07)
j=l q1 ... q
-66-
provided that x. = 0 for i > n. In part I we have shown that
x i = 0 for i > n if there are no sampling errors. This is not
necessarily true if sampling errors are taken into account. However,
in the case of independence, i.e., when qi = q (i = 1,2,...), x i
*2 i( i
a (a.) = Nbiq 1 - q ) . (e9)
-67-
From equations 88 and 89 we obtain
(91)
a2 b (1-q'
a1 a 2 an
Since the variates -, 2""' are independently distributed,
q q q
and since ai is nearly normally distributed if N is not small, we
can assume with very good approximation that the sum
n a.
(92)
i=1 q
E ) = E bi =1 - a (93)
) i=l b
i-a --
i b(-q n
a
~b1
+k
f
q -- -tq' F (95)
-68-
ft
is fulfilled forms a confidence set for the unknown probability q
with confidence coefficient a . However, formula 95 cannot be
used, since it involves the unknown quantities b,...,bn . Since
aa .
-r converges stochastically to b. as N--, we change the stan-
q a. I i-
dard deviation of E, only by a quantity of order less than
qi
the inequalities
-a - I - n --- i - a ° +( - q (96)
1z~
-- - s i - o a 2Tq1-
r') a.
q]0
*j q) o
1I - ao -'( j - n a) n al - 'q
-69-
O )
=- I - a0 -A a Y
i=l q Nq()
and the lower end point of the confidence interval is the root ot
the equation
n a n a (1-i
=1 a o 2i_
i=l q 0 +) i=l Nq - (o9
NUMERICAL EXAMPLL
-70-
mI
It
n a.
1 (A)
1 q
Ai
N =500 ai = Fr
A =400 a = .80
o o
A1 40 a I = .08
A = 25 a 2 = .05
A = 5 a 3 = .01
A = 3 a 4 = .006
A = 2 a 5 = .004
475
a a2 a3 a4 a5
-I + - - + = I - ao
ci q2 q3 q4 q5
We obtain
.20q 5 2
-. 08q 4 - .05q 3 - .0q - .O06q - .004 = 0
qo =850.
-71-
.~.. ~
Wit~'%S s~r .,J.fl. .a~t 4g.. t., ~ _
The next step is to calculate the standard deviation of
expression A. This can be shown to be approximately equal to
l 2 "q
S2 3 4
1 a(1-q0 ) a 2 (l-qo) a 3 (l-q O) a 4 (1-q 0 a 5 (-q 0 )
2
-- + + + 8 + 0-
°
Nq Nq Nq Nq Nqo
0 0 0 0 0
= .01226
n a.
Knowing that I is approximately normally distributed with
i1l q
mean value 1 - a0 and the standard deviation 0, we can determine
f 2.95
-X .99
From the table or the areas of a normal curve, it is found that
A .95 = 1.959964
.9 9 = 2.575829
-72-
i -L- (1- a <
i 0a) < a O
n a.
i --L -a +x
- 3Z .0122678X 1 - a - a I - ° +X y
-a-_ a__ 1_0 a 0 aI
.95 1.959964 .024044 .175956 .224044
.99 2.575829 .031600 .168400 .231600
22-+ ~q3
- +. 4 4 +175956,
5
which reduces to
5 4 3 2
.175956q - .08I - .05q - .1q 2 - .OU6q- .004 = U
= .912,
and equation
aI a., a3 ri a
S- + + - + - + -= .224044,
c, 3 '14 S5
t which reduces to
-73-
3 -_.Olq 2
.168400q 5 - *08q4 - .05q _ .006q - .004 = 0
q = .935
-74-
PART VII
MISCELLANEOUS REMARKS 1
1. Factors that may vary from combat to combat but influence the
probability of surviving a hit. The factors that influence the
probability of surviving a hit may be classified into two groups.
The first group contains those factors that do not vary from
combat to combat. This does not necessarily mean that the factor
in question has a fixed value of all combats; the factor may be a
random variable whose probability distribution does not vary from
combat to combat. The second group comprises those factors whose
probability distribution cannot be assumed to be the same for all
combats. To make predictions as to the proportions of planes
that will be downed in future combats, it is necessary to study
the dependence of the probability q of surviving a hit on the
factors in the second group. In part V we have already taken
into account such a factor. In part V we have considered a
subdivision of the plane into several equi-vulnerability areas
A 1 ,...,Ak and we expressed the probability of survival as a func-
tion of the part of the plane that received the hit. Since the
probability of hitting a certain part of the plane depends on the
angle of attack, this probability may vary from combat to combat.
Thus, it is desirable to study the dependence of the probability
of survival on the part of the plane that received the hit. In
addition to the factors represented by the different parts of the
plane, there may also be other factors, such as the type of gun
used by the enemy, etc., which belong to the second group. There
are no theoretical difficulties whatsoever in extending the
theory in part V to any number and type of factors. To
illustrate this, let us assume that the factors to be taken into
account are the different pirts Al,...,Ak of the plane and the
different guns g 1 ,...,g m used by the enemy. Let q(i,j) be the
probability of surviving a hit on part A 1 knowing that the bullet
has been fired by gun gj. We may order the km pairs (i,j) in a
sequence. We shall denote q(i,j) by q(u) if the pair (i,j) is
the u-th element in the ordered sequence of pairs. The problem
of determining the unknown probabilities q(u) (u = l,...,km) can
be treated in exactly the same way as the problem discussed in
-75-
part V assuming that the plane consists of km parts. Any hit on
part A. by a bullet from gun g] can be considered as a hit on
part Au in the problem discussed in part V where (i,j) is the
u-th element in the ordered sequence of pairs.
2. Non-probabilistic interpretation of the results. It is
interesting to note that a purely arithmetic interpretation of
the results of parts I through V can be given. Instead of
defininq qi as the probability of surviving the i-th hit knowinq
that the previous i - I hits did not down the plane, we define qi
as follows: Let M i be the number of planes that received at least
i hits and the i-th hit did not down the plane, and let I].
1 be the
total number of planes that received at least i hits. Then
M1
= -. "Thus, qi is defined in terms of what actually hap-
a.
-a 0
> >
Assuming that ql q 2 > "" n ' the minimum value Oi .i
-76-
- - - -
3. The case when Y(i) is unknown. In part V we have assumed
that the probabilities y(l),...,y(k) are known. Since the
exposed areas of the different parts AlI...,A k depend on the
angle of attack, and since this angle may vary during the combat,
it may sometimes be difficult to estimate the probabilities
Y(1),...,Y(k). Thus, it may be of interest to investigate the
question whether any inference as to the probabilities
q(l),...,q(k) can be drawn when Y(l),...,Y(k) are entirely unknown.
We shall see that frequently a useful lower bound for q(i) can
still be obtained. In fact, the value q*(i) of q(i), calculated
under the assumption that the parts A.(j
KJ pi i) are not vulnerable
(q(j) = 1), is certainly a lower bound of the true value q(i).
Considerinq only the hits on part Air a lower bound of q*(i), and
therefore aiso of q(i), is given by the root of the equation
n a*
r =I-a*(I0
r 0 (100)
r7 q o
The lower limit obtained from equation 100 will be a useful one
if it is not near zero. The root of equation 100 will be
n
considerably above zero if L a* is not very small as compared
r= r
with 1 - a*. This can be expected to happen whenever both Y(i)
0
and q(i) are considerably above zero.
-77
-7"-
PART VIII
Let q(j) be the probability that the plane will survive a hit by
gun j (not knowing the part hit). Then obviously
-78-
6(ilj) = (iJj)q(i,j)
]y(i Ij )q(i,j) y(ilj)q(i,j)
q(j) (103)
q(i,j) = y((i7j)
6(i1 q(J) "(104)
Hence,
Y(ll
q(j) = (106)
-79-
vulnerable to gun j. If q(i.,j) has the same value for any gun
j, then q(j) is proportional to P(j). Thus, the error is perhaps
not serious if we assume that q(j) is proportional to p(j), i.e.,
p
q(j) = XP(j). (109)
The proportionality factor X can be determined as follows. From
equations 101 and 104 we obtain [
8(ij _______
y~~)q = AP(j) (i) (110)
Hence,
XY(i,j) = q 8k(i21P(i~ 11
8(ilj) = 8j~
( . (112)
Since
SY(ilj) = 1,
Aj Y(ij) q (114)
But
~ Y(ij) =1
Hence,
q 6((115)
F, - = 1 - a0 ,
j=l qJ0
NUMERICAL EXAMPLE
e The probability that a hit will not down the plane does
not depend on the number of previous non-destructive
hits. That is, q1 = q 2 = "'" = qo (say), where qi is the
conditional probability that the i-th hit will not down
the plane, knowing that the plane is hit.
o The division of the plane into several parts is
representative of all planes of the mission.
o Given that a shot has hit the plane, the probability that
it hit a particular part, and was fired from a particular
type of gun, is known.
-81-
These probabilities depend upon the proportions in which differ-
ent guns are used by the enemy. To overcome this difficulty a
method that does not depend on these proportions is developed in
part VIII. The assumptions necessary for the method of part VIII
differ from those of part V only in that the fourth assumption is
replaced by:
* Given that a shot has hit the plane, and given that it
was fired by a particular type of gun, the probability
that it hit a particular part is known.
The Data
The numerical example will be an analysis of a set of hypotheti-
cal data, which is based on an assumed record of dhmage of sur-
viving planes of a mission of 1,000 planes dispatched to attack
an enemy objective. Of the 1,000 planes dispatched, 634 (N)
actually attacked the objective. Thirty-two planes were lost
(L=32) in combat and the number of hits on returning planes was:
A. - number of planes returning with i hits
A = 386
A = 120 (A)
1
A2 = 47
A3 = 22
A4 = 16
A = 11
A1 + 2A 2 + 3A 3 + 4A 4 + 5A =
5
(B)
120 + 2x47 + 3x22 + 4x16 + 5xll = 399
-82-
B1 Flak
B 20-mm aircraft cannon
B 7.9-mm aircraft machine gun
C1 Forward fuselage
C2 Engine
A C3 Full system
C4 Remainder
The necessary information from the record of damage is given in
table 7.
TABLE 7
20-mm 8 7 17 18 50
cannon, B 2
7.9-mm 7 13 17 18 55
machine
gun, B 3
-443-
damage. Although a simplified method is used in this example,
more accurate estimates can be made if more technical data is at
hand. The first step is to make definite boundaries for the
areas C1 , C2 , C3 , C4 . Next, assume that each type of enemy fire
B1 , B2 , B 3 has an average angle of fire 618 02, 03* Finally,
assume that the probability of hitting a part of the plane from a
given angle is equal to the ratio of the exposed area of that
part from the given angle to the total area exposed from that
angle.
(C)
Y(C 1 81 ) = .058 Y(C1 B 2 ) = .143 Y(C IB 3 ) = .143
Y(C 2 B I ) = .092 Y(C 2 1B2 ) = .248 7(C 2 B 3 ) = .248
Y(C 3 B ) = .174 Y(C3 1B2 ) = .303 Y(C3 B 3 ) = .303
Y(C4 1BI ) = .676 7(C 4 IB 2 ) = .306 Y(C4 B3 ) = .306
-84-
Computations for Method of Part VIII
Let q(Ci,B j) be the probability of surviving a hit on part C. by [
gun Bj. By equation 104, we have v
6(C. lB.
q(Ci,B j ) = (i-Bj) q(Bj) (D)
(E)
-05-
- . ' . U .- - .. . .. . .. . . . . . . . . . . . . . . . . . . . . g
P(Bj) = ain (ClB) Y(C 2 IB j ) Y(C 3 1Bj) Y(C4 B
j~i~-l~~3
=~~~ mm C2 1B) D6C 3 1B) 'MT1TB
= .984
(G)
p(B2 )
S.143
( .160
.248 .303 .3061
' .140 ' .340 ' .360J
= .850
143 .248 .3U3 .3061
P(B3 = min
p(3=m .127
.14 ' .236
.2 '
8 .3091.0
,:.3 6.3271
= min > , >1 , .981 .936
= .936
6(Bj) I:'
A N
-86-
. *.-
The root is .930 (= qo, say).
The values (Bj) are obtained directly from table 7 by taking the
* ratio of hits of type B. on returning planes to the total number
of hits on returning planes.
6(B I ) S= 294
29 = .737
6(B.)
qo ZO(
- .9703
Substituting in equation F
6(C.ilB.)
q(CiB) B q(B
i.
-87-
4
. S,*. w. - -~ .. ,. ... -
q(CI,B I ) = (.058) (.955)/.058 = .955
q(C 2 ,Bl) = (.085) (.955)/.092 = .882
Comments on Results
The vulnerability of a plane to a hit of type Bj on parL C i is
-88-
j
ADOA091 073 CENTER FOR NAVAL ANALYSES ALEXANDRIA VA OPERATIONS EV--ETC FIG 1/3
A REPRINT OF 'A METHOD OF ESTIMATING PLANE VULNERABILITY BASED -- ETC(U)
JUL A0 A BALD
UNCLASSIFIED
ENDlK
CRC-432
"Tl I
on the engine
* 7.9-rn machino
conclusions thi
derived by the
9 guides for loci
prediction of I
tUNRB
Flak, B~
cannon, B2
7.9-Rm
machine
Vulnerability
specified area
a
specified
bThis is the p
hit, When neit:
specified.
MWICOO TESTCHART
RESOLUTO#
NAMPUMMSWEPJ
Of SIANSDARDS.1963-A
on the engine area. The next most vulnerable event is a hit by a
7.9-mm machine gun bullet on the cockpit. These, and other
conclusions that can be made from the table of vulnerabilities
derived by the method of analysis of part VIII, can be used as
guides for locating protective armor and can be used to make a
prediction of the estimated loss of a future mission.
TABLE 8
Vulnerability
to hit on
specified area
when type of
hit is un-
.074 .038 .070 b
specified a .114 .179
-89-