Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
IN A PROBABILISTIC SENSE
BY
THESIS
Urbana, Illinois
(;,2 \.~~t-t\o5
R ~G(l. a
UNIVERSITY OF ILLINOIS AT URBANA-CHAMPAIGN
August 1972
IN A PROBABILISTIC SENSE
1I 1 7-
);.
rk
J~ITICharg~ of Thesis
./ G..,-_t; .
------------- 'o../~
. . ead of Department
Committee
on
Final Examinationt
-----------------------
D517
F ILLINOIS
-=..-npaign Campus
- -- ate College
- ~ .istration Building
FORMAT APPROVAL
~,- ysis.
= r typing and publication, and to Mrs. Burns and Miss Anderson for
'ng. Special thanks are expressed to Mrs. Brown of the Graduate College
~he author.
...
iv
TABLE OF CONTENTS
Page
NOTATION 0 0 1
STATEMENT OF PROBLEM 3
MINIMIZATION OF E{e} 0 11
CONCLUSIONS 0 53
_ ENDIX A 0
55
-=-_ENDIX B 58
~= T OF REFERENCES 60
I
I"
.-
v
LIST OF TABLES
Page
- LE
3. COMPUTED EIGENVALUES OF G •• , 42
4. RESULTS OF OPTIMIZATION OF s , . 50
,.
~
[
. "t
..
..:
"
vi
LIST OF FIGURES
re Page
I. NOTATION
is a vector.
J (1)
.Y:.l 1
J(l)
.Y:.Z z
V> J >
l
process Q.
2.
1m Q is the.imaginary part of Q.
j denotes 1=1.
jwt is the time variation of all the fields and currents
e
axpressed as
r'r
-jk 0-
E(r, 8, ~) = -j(k Z /4nr) f(8, ~) e
- 0 0 -
ere (r, 8, ~) is the point of observation and i(8, ~), called the
jk r' or
i(8, ~) = -r x r x II ~(r', 8', ~') e 0 dr'd~'
- ative in nature, that is, as the mean currents grow, the probability
-- large errors also grows. For this type of errors (although if built
e desired pattern very closely) the large currents necessary for this
-_ecision will likely give rise to large error c~rrents. Since large
Ii
~ _or currents usually give rise to large errors' in the pattern function
thesized, it appears that asking for too much precision can do more
-~ than good. Large currents also create other problems such as high
4
c losses and extreme frequency sensitivity.
uz,
tions .!:!.l' . N ; i •e., l. = J! ul + . • • + J~ .!:!.
.!:!. N' The
= as [iI' i
z] = f ft iz w d~, where wee, ~) is positive for all directions
:). With the norm defined as 11.1 IZ = [., .] the performance index
11!u - i liZ can be expanded as
- e
G = f V><V w d~
I are given by
N
(3)
-1 -1 i\;-l~ utC>,
J > = U Diag [A I ' A Z ' • ••
, N (4)
o
~cating that J > may be strongly influenced by the last few small
o
~ values of G and that the level of currents, of which <J o J0> is a
5
~~ can be shown [5] that as more basis functions are added to P (which adds
re rows and columns to G) that the smallest eigenvalue can certainly never
sense.
~ rmer required the use of complicated functions and neither answered the
utationally for large sample sizes and subject to question for small
- ectly by assuming that the error currents are random variables and
6
_:> ative in nature. Under this assumption, E: becomes a random
:> iminated. By making use of an asymptotic series derived from the normal
..
" ,
7
III. MODELING OF ERRORS AND PHILOSOPHIES OF OPTIMIZATION
as random variables J>. For convenience, the random error currents, oJ>,
are defined as oJ~ = J> - J> where J> ~ E{J>} are the mean currents.
The first moment matrix of oj> is zero. The second moment matrix of
'=orming <Y A Y> = ijL (Y.J.)(Y.J.) * X .. = <ZXZ> > 0 where Z. = Y.J .. This
1. 1. 1.
1. 1. J. J 1.J
should be noted.that at this point,. except for the form of the first and
second moment matrices of'oJ>, no assumption has been made as to the precise
One observes from the form of (5), that as the amplitudes of the
orows. Large error currents give rise to large errors in the pattern
everything could be done ~ith precision, the better the match to the desired
_attern function, the higher the current level becomes. Thus it appears
If the currents are random, then the pattern function they produce
(6 )
.out the minimum of E. Rathert one must choose some attribute of the
find a set of currents J> which minimizes E(J». This scheme will be
_espect to J>. The three methods are illustrated in Figure 1 (a, b, and
: the three possible solutions given in the illustration, J1>, J2>, and
-3> are the optimum solutions to the firstt second, and third schemes,
=espectively.
scheme would coincide with those of the first two methods. Thus, it appears
9
E;J»
VERTICAL
OPTIMIZATION
(a )
I
I
I
I HORIZONTAL
: OPTIMIZATION
I
I
€MIN
( b)
"II"
i::::::1
I:;;;;C:
1:::::1
-
IllIn
::~
E
k
= MEAN OF E
ImH-
REFERRED TO F(E;J
k
»
OPTIMIZATION OF E{E}
(and many others that can be derived) depends, among other things, on
It has been tacitly assumed that the solutions exist to the three
the only way for solutions to the two problems not to exist is for the
since the errors one is prone to make grow proportionately to the size
it appears that the current levels are bounded and hence solutions exist.
11
and provides a simple and direct way of determining the proper amount of
presented below.
2 -
8 = 11!u - ill + 110i I I - 21<e. [!ct - I, 0I] (7)
since E{oJ>}= 0, E{oi} also vanishes and, therefore, the expectation of any-
E{llofI1 }
2
= E{<oJGoJ>} = L E{oJioJj} Gij
i,j
added.
(8) reduces to e = I I~ 2
- II 1 +a<JJ>, which is the regularized performance
12
index used by Cabayan ~rlO] in the discrete array case.
were also considered by Cabayan, 1e~ ou(x) be a random process with mean
zero with each sample a continuous function ,defined on -a < x < a, Let
a a
where for all u, 0 < J J X(x, y) u(x) u(y)* dxdy < 00, the mean of the
-a -a
norm,squared of of is
a a
= J J X(x, y) G(x - y) u(x) u(y)* dxdy
-a -a
where
k
G(x - y) = J 0 e(~)e*(~) ej(x-y) d~.
-k
o
2 a a .
e = I Ifd - fl I +J J K(x, y) u(x) u*(y) dxdy, (9)
-a -a
2
e ;;: IIfd - f 11 + aJ~ Iu(x) ] 2 dx
-a
13
=rom Equation (10) that eigenvalues smaller than the level of errors
The performance index may be calculated from (I), (8). and aO) to be
(13)
for the solutions. In this case the performance index given in Equation (1)
is modified to
(14)
(16)
an arbitrary choice of the phase and polarization of~. For the given
(18)
(G + K)J> G(J»>
Ll
1,,'::111
l1fl'lI
;1:,"'11
16
V. APPROXIMATE CALCULATION OF THE DISTRIBUTION FUNCTION OF E
where E is given by (13) and X> oj> + ~J> where ~J> = J> - J >
o o
[Re.<oj, - Im<oJ]
is given by
On the other hand, the second assumption considerably reduces the work
<XGX> = <XIX'>. The covariance matrix of X'> is A' = SASt which is,
(20)
T
then
s = <X"X"> + E (21)
o
i = 1, 2, •• N.
N -1 N 2
= [TI =l (1 - jv.t)] exp {jt [€ + I I~J~I /(1 jv.t)]}
i 1 0 i=l 1
1
(23)
N
(24)
'\ = E{E} = I
i=l
N k-1 .
= (k - 1)! I v
i'
. (v . + k
.. i
\,~J'.'I>2),
1
.
k 2, 3, • ••
i=l
(25)
the frequency function of F(€; J») and the cumu1ants of ~(t) characterize
Since one is dealing with chance anyway, if the results of the simulation
where
Z
HZ (z) z - 1
3
H (z) z - 3z
3
5 3
HS(z) = z - 10z + lSz
of Hz),
z = (e - E)/cr
but (26) may also be used in other optimization schemes. Since the
be applied, but not without some difficulty. Every term and factor
although one could work through the various definitions and obtain the
here, and in spite of: the apparent numerical complexity of their implemen-
After all, if one is only going to build one antenna, minimizing the mean
E does not seem to make as much sense as maximizing the probability that
is even more desirable than vertical optimization, since it does not require
have the distributions of probable errors, and they are included in some
of each ring of which there are N and each ray of which there are 2M uniformly
2M jpicos(<P-<Pk)sin 8
V.
1.
sin 8 L e
k=l
M
2 sin 8 I cos (picos (<P-<Pk) sin 8), (27)
k=l
--here (8, <P) is the direction of observation in the usual spherical 'ill
'illl'
:IU
Since V. is periodic in <P with period ~<P, this type of array is I ~1I1
""I'
1. I!lli
~11I
1 ~III
l.ulliO:
~he type considered here. Specifically, let the desired pattern function
{ sec II'
o
(Because of the symmetry of the problem about the 8 = 90 plane, angles
22
RAY 4 RAY 3
RAY 5 RAY 2
RAY 6 z '"I,
I'll
::::\,
1Il!1I
Hlilti.
)"11
)1"1
,
""
....
:: > 900 need not be considered.) The motivation for this type of problem
TI/2 2TI
2 2
II • 11 = (1/8TIM) f fl. 1 sin e d~de.
o 0
For all the examples considered, the errors are assumed to obey the
n the ith ring are identical and are given by oj., where the real and
1
ariables both having mean zero and standard deviations of 01J.1 I. Errors
atrix is K
0
= 202 Diag[Gll, G22, ••• , GNN].
In order to see the effect of ° on regularized optimum solutions,
-1
.tth J> given by J> = (G + K ) C> • Thus Z[0l, 02] is the mean of €
°2
24
for an actual error level of cr of an antenna optimized for an error
1
level of crZ'
' d for N = 10 , Pi = 2
Patterns were synth eSlze 1 TI -
( 1 )10-i - 1,
2
700, and a number of different values of M and cr, The evolution of
all polar plots of the mean pattern functions in various planes super-
imposed on plots of the desired pattern. The plots in the first and
which are E-p1ane patterns. The p10ffiin the lower two quadrants of these
By using the law of large numbers, it can be shown that a random variable
IIt:m:
:::llUI
1I\~:1:
variables is approximately normally distributed. In this case, twelve :::~
III11UJ
1I111111
method [14]) were used:. With the samples of oj> computed in this way,
the real and imaginary parts of each component were independent and
and q= E{q}, the corresponding sample means of E and q denoted by <E> and
<q>, respectively, s[O,l percent, 0.1 percent], and the mean currents, J>,
25
ep= 0
90 PLANE ep =0° PLANE
1.0 8 = 0° PLANE
1.5
2.0
M= I
ep =0° PLANE
1.0 e=0 0
PLANE
1.5
2.0
2.5
M=2
Hl'lI
lIilll
\.0 e = 0° PLANE
1.5
2.0
2.5
M=3
o o
ep = 22.5 PLANE ep = 0 PLANE
1.5
2.0
2.5
M=4
Iii! II
Itilli
1111 I!
1111
illl
iii!
"II
1111
1111
1.0 e = 0° PLANE
1.5
,
,.I"
"
2.0 "
2.5.
M=6
ep= 0
0 PLANE
1.0 e=a 0
PLANE
1.5
2.0
2.5
M=8
:1::::1
1111111
:::::~
1.0 e = 0° PLANE
1.5
2.0
2.5
M = 10
M 1 2 3 4 6 8 10*
J
1 .5387 -.1557 -.1651 -.3509 -.8022 -.3383 -.3845
J2
1.473 1.353 1.012 .9788 1.184 .5907 .5878
J
3 -.5773 -.7060 -.1293 .5093 1.465 .7157 .7846
J
4 -.8400 -.7632 -.9258 -1. 866 -3.427 -1. 541 -1. 693
J
5 1.195 1.135 1.187 2.070 2.376 .9383 1.020
J
, 6 -.1516 -.1264, .06008 -.3540 1.615 1.149 1. 249
J
7 1.105 1.093 1.488 -1. 245 ' -4.726 -2.530 1-2.912
J8
1.450 1. 421 1.81,2 1.593 4.937 2.431 2.994
J
9 -.9182 - .8805 1.103 -.9550 -2.813 -1. 228 1-1.719
J10
.2572 .2401 .3019 .2464 .7428 .2946 .4984
t[5%, .1%1 .04066 .01952 .01651 .02149 .1028 .02533 .04215
<E> .04168 .01921 .01649 .0-2161 .1041 .02476 .04323
q .01464 .01230 .01451 .02077 . .1026 .02519 .04205
<q> :01566 .01199 .01450 .02090 .1039 .02463 .4314
l
£ [,1%, .1%] .02603 .007220 .002001 7.211 • 10-~ 2.323 • 10. i 1. 423 • 10.~ 1.09 • 10 ~
S. D. .007343 .006610 .008473 .01181 .06200 .01658 .02771
Y1 1.226 1.333 1.460 1.:j35 1.467 1.583 1.594
Y2 2.546 2.951, 3.504 2.876 3.531 4.009 4.075
I<JJ> 3.030 2.879 3.185 3.781 8.909 4.411 5.217
* 2500 samples used W
N
33
probability means. One also notices that as the number of rays increases,
performed for each case. S.D., Yl' Y2' E[5 percent, a], q,<f::>,<q>,da, a],
mean currents J. 's and <JJ>1/2 are listed in Table 2 for the value~ of
~
a given above. The pattern funTtions corresponding to the currents in
E[5 percent, a], with its corresponding <s>, against a. Notice that.
note that even for the simple error model used in this example, the
The method used to solve the system of equations, (G + K)J> = C> was
C'>, (28)
3% 5% 7% 10%
0 0% 1%
J1 .04127 .04502 .04585 .04557
2.790 .02200
J2 .1657 .1626 _.1574
-4.989 .1808 .1693
J3 -.003275 -.01517 -002019 -.02464
-.9968 .04477
J4 -.02591 -.01559 -.01535 -.02045
-3.809 -.09016
J .1112 .1143 01123 .1061
5 4.613 .1006
J6 .1232 . .1001 .09032 .08179
-.5083 •2124
J7,- -.08305 -.07037 -.06700 -.06599
4.136 -.1607
J ' --. -.02768 -.01155 .•..•
008039 -.009804
8 5.853 -.06149
J .1751 .1386 .1241 .1126
9 -4.064 . .2824
JlO -.06183 -.03736 -.02643. -.01538
1.346 -.1377
5.003 • 10-~ 3.915 • 10-
4 3.852 • 10-
11
3.891 • 10-
4 4.147 • 10-4
d5%; 6] .4718
5.007 • 10-' 3.903 • 10-' 3.887 • 10-
l
3.898 • 10-4 4.082 • 10..:q
<E> .4662.
2.401 • 10-' 2.244 • 10-
l
2.156 • 10-'" 2.024 • 10-4
q .4718 3.631 • 10-'
3.634 • 10-
11
2.388 • 10-4 2.278 • 10-
11
2.161 • 10-l 1.975 • 10":'''
<q> .... .4662
6.791 • 10-~ 1.518 • 10-
11 .
: .2.379 • 10
-~ 3.852 • 10-
04
5.961. 10
..:.~
1.022 • 10-';
;[0, 0]
11
1.639 • 10-'1 1.609 • 10":-4
S. D. .3647 2.168 • 10-
11
1.673 • 10-l 1.660 • 10-
1. 728 1.804 1.826 1.841
Y1 1. 865 1.386'
Y2 4.768 5.116 5.216 5.278
5.394 3.087 w
.2657 .2514 ~
/<JJ> 11. 81 .4763 .3174 .2806
35
o
4>= 9 0
PLANE ep = 0 PLANE
15
2.0
2.5
(j' = 00/0
0
ep= 9
0
PLANE ep = 0 PLANE
/.5
2.0
2.5
(J = I 0/0
ep== 9 PLANE
0 ep== 00 PLANE
1.0 e ==
0
0 PLANE
1.5
2.0
2.5
(J == 30/0
ep = 9 0
PLANE
2.0
2.5
o
ep= 9 PLANE
2.0
2.5
(J = 70/0
ep = 9° PLANE
1.0 e= 0° PLANE
1.5
2.0
2.5
0- = 10 %
4
•
o
o
Figure 16. Mean E'S for Solutions Regularized to Various Noise Levels
but in the Presence of an Actual cr = 5% Level
42
of linear equations, it does avoid some of the numerical difficulties
The last two entries of Table 3 can be considered as nothing less than
its actual value is not much greater than the.levelof machine. precision;
Since the currents obtained from Equation (4) are strongly dependent on
0
2 which can be much greater than the level of numerical errors.
(22). This raises the question: How can the moments and approximate
how can the method of Chapter V even be applied in view of the form
Although small positive eigenvalues may also be erroneous,. one must take care
1.1/2
1
o. . .0
0
S 1.1/2
2
o. . •0 • U t.
0
'.~.1.1/2
r
o• . •0
Al = SASt and its unitary matrix Ware now r X r matrices and the
-1
Al
-1
0 .. 0
1..
2 0
TJ > =W
t
Q Q .u t . U C'>
o
. -1
.
0
.
0
A
r
..
0
X...,l
N
t [,-1/2 -1/2
= W Diag h
1 '.~2 '
the second and higher moments of E and hence the shape of the
1111
III
nit
illl
lii~
1111:
III!
iii
of the distribution function depends on s which, in turn, depends
o 111"1
III
III
computed from (11) which does not involve J >. Thus all the moments
\ a
of E can be computed accurately if the assumption that G is very nearly
Since one has no way of knowing what the smaller inaccurate eigen-
values of G actually are, the best and perhaps the only way to determine
how small r should be and how good the approximations made above are
This was done for the case where N 10, M = 10, and cr = .1 percent.
45
The results displayed in Figure 17 show excellent agreement between
this example, two independent simulations with sample sizes of 2500 were
examples considered.
has the minimum mean but generally has the most desirable properties of
all the distributions plotted. The figure also indicates that although,
It is now shown that: for the antenna structure used in this chapter
all stationary points J> of (17) are real. Moreover, it is shown that
all solutions obtained in this chapter are stationary points of (17). Due
to the symmetry of the antenna, v> and consequently G are real. Equation (15)
46
F(E)
1.0
o
X
x
.9
.8
.7
.6
.5
.4
.3 EDGEWORTH SERIES
.2 X G SIMULATION
.1
o
o .01 .02 .03 .04 .05 .06 .07 .08 .09 .10
- -
F (E)
1.0 • 0* .8. 0
0.9
0.8
0.7
0.6
SIMULA TIONS
X 0" = .01
0.4
.• (J = .03
0.3 o (J = .05
A (J = .07
0.2 .0"=.10
0.1
x 10-4
0
0 2 3 4 5 6 7 8
E
.l>-
Figure 18. Distribution Functions for Solutions Regularized to Various -...J
Noise Levels but in the Presence of an Actual a = 5% Level
48
(G + K) J> HJ>.
then J> is real. The plots of the patterns obtained in this section
show that f(8, ~) = <JV> > O. Evidently, for these patterns, G>
If any such points actually exist~it seems very doubtful that they
would produce an E from (17) which is less than that produced by the
type of behavior will occur for structures with certain symmetry. For
in general, be complex.
To illustrate how the pattern one obtains is improved by allowing
symmetric)pattern
o
sec e
where
'IT
n> f V><VJ> (1 - fd!l-<JV>I) sin ede + KJ>
o
Figures 20(a) and 20(b) are the respective phase functions. Figure 19(b)
I
J2 2.421038.6 1.525 ~9.74 I.
,,
J5 .5800U1.98 .7238047.9
e: .5457 .1994
0
1.5 1.0
( b) (a)
PHASE (DEGREES)
180
150
120
90
60
30
o 160 180
-30
-60 e (DEGREES)
-90
-120
(0)
-\50
-180
180
150
120
90
60
30
o 140 160 \80
-30
-60
-90 e (DEGREES)
-120
-150 (b)
- 180
Figure 20. (a) Phase Function of Pattern of Figure 19 (a)
(b) Phase Function of Pattern of Figure 19 (b)
53
VIII. CONCLUSIONS
APPENDIX A.
-2
Consider the random variable X formed by squaring the normal random
-2
n(t) E{eijtX}= 1/;;:;;. foo exp [jtx2 - (x - y)2/v] dx = [11 ~ jvt]-l
_00
r 00
[x _ ]1
(1 - jvt) J
2
and
imaginary parts of
1
2 vI' , 12 vN' respectively. In view of
By Shl'ftl'ng
X-2 as def'
lned' In th'
e prevlous paragraph by th e constant
k
,-k d
v = J --- ~n ~(t) k 1, 2, •••
k dtk
t=O
v 2 2.
j
-1 d.Q,n n 2+V
---+ j-----
V yt
dt 1 - jvt 2
(1 - jvt)
v2 2 2 2
2 + 2V v + j2 • 1 • V v t
(1 - jvt)2 (1 - jvt)3
which is proven by
k
-1 d .-k d ~n n
j dt J k
dt
57
,
+ j (k+l) • k! tvk ,}vl (1 - jvt)k+2
Setting t 0,
are the cumulants of net). It follows from this that the cumulants of
~ lj:I(t)
are
N
kk-l I ]J.] 12
v
k = (k - l)! L [v.
1.
+ kv.
1. 1.
i=l
APPENDIX B.
1.0 Calculation of G
2M
V.
1.
= I
k=l
exp [jPi cos (~ - ~k) sin e] sin e
2'IT'IT/2 *
G., = f f V,V, sin e de d~/8'ITM
1.J 0 0 1.J
G ..
1.J
where
and
1 'IT/2. '"
-- f eJzcos ~ d~ = J (z)
2 'IT 0 0
2M 2M 'IT /2
3
G,. = I L f J (p, 'k.Q, sin e) sin e de/4M.
1.J k=l,Q,=l 0 0 1.J
2M 2M
G. . L L 8 (p i .k£ ) 14M
~J k=l £=1 J
where
M
G .. = L 8(P"k) + -2l[8(P.+ p.)+8(p. - P.)]
~J k=2 ~J ~ J ~ J
where
2TI TI/2
Ci = J J fee) V. sin e de d~/8TIM
~
a a
1 2M 2TI TI/2
2
=- L JJ exp [jP. cos (~-<Ilk) sin e] sin e fd(e) de de»
8TIM k=l a a ~
60
LIST OF REFERENCES
[6] D. R. Rhodes, "The optimum line source for the best mean-square
approximation to a given radiation pattern," IEEE Trans 0 Antennas
and Propagation, vol. AP-ll, noo 4, ppo 440-446, July 1963.