You are on page 1of 14

# 1828

## Distribution of the Phase Angle Between Two Vectors

Perturbed by Gaussian Noise
R. F. PAWULA, S. 0. RICE,

## Absfracr-The probability distribution of the phase angle between

twovectors perturbed by correlatedGaussiannoises
is studiedin
detail. Definite integral expressions are derived
for the distribution
function, and its asymptotic behavior for large signal-to-noise is found
for small, near lt/2, and large angles. Theresults are applied
to obtainnew formulasforthesymbol
error rate in MDPSK,to
calculate thedistribution of instantaneousfrequency,tostudythe
errorrate in digital FMwithpartial-bitintegration
in thepostdetection filter, and toobtain a simplified expresion for the error rate
in DPSK with a phase error in the reference signal. In the degenerate
case in which one of the vectors is noise free, the results lead to the
symbol error rate in MPSK.

AND

J . H. ROBERTS

## narrow-band representation of the noise at timeti with assumed

statistical properties Xi = yi = 0 and var xi = varyj = ui2 .x 1
and y1 are, in fact, theI and Q components at timet , , whereas
these components at time t, are the real and imaginary parts
of (x2 + i y 2 )exp (-ioct).
Here O 1 and O 2 are assumed to lie between -n and n.
Consequently, 19~ - d l can range from -2n to 277. However,
we shall be concerned with the angle \$ defined as the modulo
271 difference between O 2 and e l ,viz.
\$=(13~-e~)mod2n;

I. INTRODUCTION
HE distributionofthe phase angle between two vectors
arises
in avariety
of applicationsin
communications
systems [ 6 ] , [ l l ], andcommon cases are those inwhich
information is transmitted in the phase orfrequencyof
a
sinusoid.Typically,
thenoise-corrupted version of such a
sinusoid has the form

~<e1,e2<~.

(3)

## The modulo 2n appearing in this definition will be understood

to mean over some 2n interval where the appropriate choice
of the 2n interval may vary from application to application.
We shall consider theprobability P {\$, < \$ d \$ 2 } where
G 1 < \$ 2 and the interval (\$,, G 2 ) is assumed to be contained
entirelywithin the particular 2n interval of definition being
used.
z ( t ) = A ( t ) cos [act+ \$(t)] + n(t)
(1)
Previous workswhich have considered the distribution of
the
phase angle between two vectors perturbed by Gaussian
inwhich a, is the carrier radian frequency, A ( t ) and @(t)
noise have been concerned primarily with problems in angle
are the time-varying amplitudeand phase ofthesinusoid,
and
demodulation.
References  and [113
and n(t) is a narrow-band Gaussian noise. In dealing with modulation
contain
extensive
results
in
this
area
along with many refersignals ofthisform,the
vector (or phasor) representation
ences to related studies. Although the present paper may be
is convenient to employ. With subscripts to denote different
timeinstants,vectorformsof
z(t) for the two times t l = considered as extensions and generalizations of some of the
results found in these references, along with some new applit and t2 = t + T are
cations, we have attempted to make the paper self-contained.
Another areainwhich
the phase angle betweentwo noisy
vectors arises is that of phase comparison monopulse radar
~41.
Expressions for the probability P{G1 < \$ < \$z} are presented in Section I1 for three different cases of signal parameterconditionsand
noise correlation. Theseresultsare
derived by making use of an unconventional approach which
Graphical representations of these two vectorsare
shown relates the desired probability to a functionalofthejoint
of
the
narrow-band
waveform. A
in Fig. 1. The quantities xi and yj; j = 1, 2, are Gaussian vari- characteristic function
oneofthe simpler cases follows tradiables related to the in-phase and quadrature componentsin the secondderivationin
tional lines and isgiven in an Appendix. Using the results,
various asymptoticformulasfortheprobability
P{\$ 2 \$ o }
Paper approved by the Editor for Communication Theory
of the
that \$ lies above some value \$o are obtained in Section 111.
IEEE Communications Society for publication without oral presentation. Manuscript received September 29, 1981; revised March 25, 1982.
The results of Sections I1 and 111 are then applied in Section
The work of R. F. Pawula was performed under subcontract to RanIV to thestudy of the calculation oferror probabilities in
dom Applications, Inc., San Diego, CA.
MPSK, MDPSK, and digital FM communicationssystems,
R. F. Pawula is at 552 Savoy St., San Diego, CA 92106.
S. 0. Rice is withtheDepartment of ElectricalEngineering and
and to the distribution of instantaneous frequency. In particuComputerScience, University of California atSan Diego,LaJolla,
lar,
new formulas forthesymbolerror
ratein MDPSK are
CA 92093.
previously known results. The
J. H. Roberts is with Plessey Electronic Systems Ltd., Roke Manor, obtainedandcomparedwith
Romsey, Hants., England.
performanceofnarrow-band
digital FM is considered for
0090-6778/82/0800-1828\$00.75 0 1982 IEEE

1829

## spectrum. In this case, ifthe

- f c ) l / 2 , then

f c ) + S(-f

## r = r,(T) cos W,T - rs(T) sin w,r

A = r , ( ~ sin
) W,T

Fig. 1.

Anglebetweentwovectorsperturbedby

Gaussiannoise.

## limiter-discriminator detection which

in
the
postdetection
filter integrates for only a fraction of the bit time. In Section
V, the Fourier series approach is briefly compared, by way of
several examples, withthe preceding treatments.The final
section summarizes and discusses the results.
11. THE DISTRIBUTION P{ \$ 1

the analysis:

(44

## where pi = A?/2u? is the instantaneous signal-to-noise ratio.

A@ is the signal phase change in the absence of noise and will
be referred to simply as the signal phase change. Note that
A@ includes the carrier phase shift o c r , forgenerality,
although this term is frequently eliminated by receiver processing. The parameter U is the average signal-to-noise ratio of
the two vectors and W is their geometric mean. i V I is a measure of the spread in the SNR values.
When the noises perturbing the ends of the two
vectors
are correlated, the correlations between the x and y components will be assumed to take on the specialized forms

## As is known, when the noise is a wide-sense stationary random

process, r and A can be expressed in terms of the noise power
1 Again, the modulo 27r is to be interpreted so that the resulting
A@ lies within the 27r interval of interest. However, in practical cases,
A@ will usually be determined first, and the interval of definition then
chosen. Frequently, this interval will be (A@ - 7r, A@ + n). It is suf-

## ficient for our purposes here that

-7r
values of A@ require that the intervals
appropriately modified.

## where rc(7) = :1 So) cos ( 2 n f ~ ) d and

f
rs(T) = :1 S o
sin (2nfr)df and the spectral density has been assumed t o be
normalized t o contain unity power.
Certain special cases of the above parameters will be considered in the following. In the case ofequal signal conditions, the two vectors will be meant to have equal SNRs
withnoise, andto lie in the same directionwithout noise,
so that p1 = p2and \$1 = C#J2.Consequently A@ = 0 and
V = 0 (in this case, W = U also). In the case of uncorrelated

both = and A = O.

R
Distribution
of a Modulo2n Difference
.
.
Theprobabilitydensityfunction
p ( \$ ) ofthemodulo
2n difference \$ = (0, - 0,) mod 271;-n < 01, O2 < n, can
be found by integrating the joint probability densityp ( O 1 , 0,)
of the phase angles J1 and O2 over the regions of the 0 , , O 2
plane in which (0, - 0,) mod 2n lies between \$ and \$
d\$ and O1 and O2 lie between k IT.These regions are indicated
in Fig. 2 for a particular value of \$ and, in general, are determined by the intersections of two of the three lines O 2 = O1 t
\$ + 2nk; k = 0 , f 1, with the square of side 2n centered at the
origin. Since the joint density p(O1, 0,) is of 2n periodicity2
in each of its arguments, the small shaded region in the lower
right-hand corner of the squarecan be raised by an amount
2n so that the resulting shaded region will be one continuous
strip. Consequently, p ( \$ ) can be determined from p(O1, 0,)
by the integration

A@ = (O,T

## < A@ < 7r. Extensions to other

of definition of e l and e2 be

## In the more general case when 8 , and O2 are not restricted

to values in (-n, n), the density of \$ = (0, - 0 , ) mod 2n can
be expressedinterms of the densityof \$, = O2 - 01, the
actual phase difference without the modulo 2n condition.
The relation is

where pa(\$,) is the density of \$,. In either case, the probability P{ \$1 < \$ < \$ 2 } is then given by

## 2 Strictly speaking, it is not the joint density that is periodic, but

rather a new function defined as the periodic extension
of the joint
densityfunctionthat
is periodic.Adistinctionwill
bedrawnonly
when necessary.

1830

## where (Gmin, \$ m a x ) is the interval of definition. Here, the

signum fuction cancels thediscontinuity in F(\$), and the
-F(\$min)
term makes &(\$,in)
= 0, k(\$max)
= 1. At all
points of continuity o f q + ) , p ( \$ ) = a~(\$)/a\$ = a\$+)/a\$.
We shall find it more convenient to deal with F(\$) than with
\$(\$). In this section F(\$) is stated for three4 different combinations of signal condliions and noise correlation. A derivationofthe
results will be given in Section 11-E. Particular
forms ofF(\$) are as follows.
Case 1-Equal Signal Conditions (A@ = 0, V = 0), Uncorrelated Noises (r = X = 0):

92

8,

F(\$)=-

-sin \$
47-l

rt2

e- U(1- C O S \$

cost)

(10)

dt

-n2

I - cos \$ cos t

## Case 11-Unequal Signal Conditions, Uncorrelated Noises:

,-[U-~sint-Wcos(A~-\$)cosf]

U- VS~~-WCOS(A@- \$)cost

(1 1)

Fig. 2.

W sin (A@ - \$)

## Regions of integration for probability density of

JI = (e2 - e 1) mod 2n.

## As will later be seen, the probability P{ \$ < \$ < 1 4 ~can

) also
be calculated without recourse totheintermediatestepof
first obtaining p(\$); however, (6) will indeed be employed in
an alternativederivationwhichfollows
themoretraditional
and conventional approach.

## < \$ < \$2)

G 2 , with G1 < Gar be

C. Expressions for P( \$

Letand
angles lying within
theparticular2n
interval of interest.Thentheprobability
P{\$l < \$ < G 2 } canbe expressed in terms of an auxiliary
function F(\$):

## By its definition, F(\$) is periodic with period 2n, and is discontinuousatthe

singular point \$ = A@ with a jump
F(A@+) - F(A@-) = -1. When either G1 or \$2 equals
A@, the value of F(\$) is to be interpreted as aright-hand
or a left-hand limit; i.e., F(\$,) = F(A@+) if G1 = A@, and
F ( I ~=~F(A@-)
)
if \$2 = A@.
F(\$) and the classical distribution function of probability
theory, say \$\$), are related through3

U-Vsint-Wcos(A@-\$)cost

r sin \$ - X cos \$
1 - (r cos \$

+ X sin \$) cos t

where
E=

(1 2)

U-Vsint-Wcos(A@--~)c~~t
1 - (r cos \$

+ A sin \$) COS t

## Note that in Case I1 F(\$) is a function of \$ - A@ only;

however,this
simple relation no longer holdsin Case 111
because of the noise correlations. The behavior of F(\$) in
Case I is illustrated in Fig. 3.
Equations (10)-(12) are written ina
manner which illustrates how increasingly complex signal and noise conditions
affectthe results.Somesimplification
of the integrands, at
the expense of complicating the limits, is possible in the most
general case through a change of variables by a transformation
due to Euler and Legendre (see (3 1) in Section 11-E)which results in
Case 111-Simplified Integrand Representation:

4 The correspondingprobabilitydensityfunctionsaregiven
in
Appendix B [ c f .(B-1) and (B-6)].
5 A subscript i will be used to indicate explicitly \$ dependence
whenever otherquantities in the sameexpressionare
J, dependent.
Also, because angles will be restricted to their principal values, the signum function is used in (14c) to give the correct sign of the square root.

1831

\$ol=

Pi\$

if Go < A@

-F(Q0);

if \$, > A @ .

## In all three cases, F(\$) is of2n-periodicity in \$ so that

F(\$) - F(\$ k 2n) = 0. This guarantees that the probability
density function p ( \$ ) integrates t o unity over any 2ir interval
(containing A@).
In Case 111, the difference F(\$) - F(\$ - n) is expressible
in terms of tabulated functions. Upon forming the difference,
and using (13) and the definitions (14a)-(14d), the difference
terms can be combined into integrals with limits ranging over
2n intervals, and there follows

-.5
The function F ( \$ ) in Case I (U = 1).

Fig. 3.

1 -F(\$o);

where

F(\$i) -F(\$i - n)
= 1 sgn [sin (A@ 2

[ 1- d-ze

(Oi/ei,ai)]

+ r sin \$i - X cos \$i

e- a 'Io(pJ
2 sin yi
where Zo(-) is a modified Bessel function. We note that this
last equation can also be expressed in terms of the Marcum
Q-function (171 and [ l l , p. 851).

## < \$ < \$z}

of P{ \$, < \$ < \$2}

E. Derivation ofP{

## and the principal values of the arccosine on ( 0 , n), and of the

arctangent and arcsine on (-n/2, n/2), are t o be taken. This
form is notonlyhelpfulfor
numerical work,but will be
foundtobe
useful inSection 111 in obtainingasymptotic
formulas,andinSection
IV in deriving thedistributionof
instantaneous frequency.

A derivation
for the most gederal case,
Case 111, of a signal with unequal parameters and correlated
noises will be given using the followingunconventional approach which makes use of the joint characteristic function
of the
narrow-band
waveform under consideration. This
approach has found wide application to problemsrelated
to noise zero crossings, instantaneous frequency, and FM and
PM demodulation [ 111 . An alternate derivation for Case I1
is given in Appendix B.
For a general, noise-corrupted, narrow-band waveform with
sarilples Z , = R , COS (OJ e,), z2 = R 2 COS (act+ e,),
the jointcharacteristic function @(u,v) is [ 11, ch.21

## D. Some Special Forms for F( \$)

~ ( u , = E[eiuzl+i"z21

In thissection,certain
special situations are identified
in which F(\$) andthe difference F(G2) - F(\$,) assume
simple forms.
First, in Case 1, F(\$) has the special values [see (lo)]

F(0k) = T

l
2'

F(+n/2)= ?;e-',

F(kn) = 0

## and if the interval of definition of \$ is (-n, n), from F(n) =

0 we get

+ HF terms
(19)
in which the expectation in the final form is over the random
variables R i , R,, e l , and O2 and, also, the high-frequency
terms will be assumed t o vanish by virture of the expectation.6
By subtracting some angle, say E, from the argument of the
cosine prior to doing the expectation, a closely related conditional characteristic function @(u,u I t ) can be defined as
@(u,u IC;)

## In particular,P{\$ >, 0 ) = 1/2 and P{\$ > n/2l = exp (-W/4.

In Case 11, from (11)

F(A@ - n/2) =

+ [1 - (W/v)Ze (V/U,v)]

I,

=G

u-

vcose

(20)
that @(u,u) = @(u,u Ig = 0). This conditional characteristic
function may be regarded as the joint characteristic function
of the two random
variables z1' = R1 cos (act+ 81 + C;)

SO

e-(~-vC~~e)
de

=E.[Jo(tlu2R12+~2R22+2~~R1R2~os(ez-el-C;)))~]

(1 6 )

## where the re(*) function is definedin Appendix C. Also in

Case 11, F(A@ f n) = 0, so that if the interval of definition

## 6 Without the loss of generality, the oscillator which produces wc

can be assumed to have a uniformly distributed random phase; and, as
aconsequence, the highfrequencyterms
willvanishwhenaveraged
over this phase.

1832

## and z2 = R 2 cos ( a c t 02). Then, using this definition,

Roberts' general result forthedistribution
of themodulo
2n reduced phase angle \$ = ( 0 , - 8,) mod 2n can be stated
as [ l l , sect. 2.41

## integrals 0ver.R gives

where

Aderivation of thisrelation is given in Appendix A. Physi- With the changes of variable 8' = n/2 - 8, z = tan (8'/2),
c.ally, G(\$) is related tothe classical distributionfunction
the integral in g(\$J can beevaluated straightforwardly and
yields'
F(\$) by F(JI) = G(\$) - G(\$min).
In order to apply this result, it is necessary first to calculate
the joint conditional characteristic function which,in the case
of independent signal and noise, can be written as the product
@ S + N F @s X @ N , where the subscripts "S'and "N" refer
to the signal and noise components, respectively. In the case
of the particular narrow-band waveform of interest [cf. (2)] ,
these signal and noise components are readily shown to be
(cf. [ll, eq. (2.10)])
Hence, for \$1 < \$ 2 ,

## where yi is defined by (14a), and where7 A@ is the signal

phase change asgiven by(4a). Now, it is a matter of using
these in (22) and performing the differentiation and integrations. These operations are simplified somewhat by first making a transformation to polar coordinates bymeans of
u=-

fiR
81

find

8
cos-;

u=-

fiR
a2

8
sin .2

(25)

## which completes the derivation of (9) and (12). Equation (13)

follows from (1 2) by use of achange of variables due to
Euler and Legendre [4,p. 3 161

## where the transformation is from the variable t to the variable

8, and where d = r cos Gi h s i n \$i= -cos 7i.

## 111. ASYMPTOTIC FORMULAS FOR P{ \$ > \$o}

The probability P{\$ > IL0} that \$ lies above some value
\$o is frequently of interest in certain communications problems since this probability can often be related to the probability of making an error in detecting an information symbol.
In many cases this probability is not expressible in terms of
tabulated functions and, in addition, values of this probability
for a large parameter, the signal-to-noise ratio, are of primary
concern.Hence,in
this section,asymptotic
formulas for
P{\$ > GoJ will beconsidered for large U. For the sake of
definiteness, the interval of definition of \$ will be taken to
be A@ - 71 < \$ < A@ n, since \$ = A@ in the absence of
noise. Then, P( \$ > J l 0 } = P{\$, < \$ < A@ n} so that (9)

## Now, performing the differentiation and

employing Hankel's
exponential integrals (Appendix C) todothe
remaining
'The symbol "@" is beingusedboth
todenote acharacteristic
function and inthedefinition
of the signal phase change. The arguments of the characteristic function identify it and there should be no
danger of confusion.

## 8 A similar integral is evaluated in Appendix A. Cf. (A-4). The value

of g(\$3 at the jump \$ i = A@ has been chosen as the limit from theleft.
The actual values of g ( \$ 2 ) - g(\$ ,)at the jumps A@ = \$ and A@ = \$ 2
are of noconsequenceincomputing
P(\$ 4 5 \$2f ifthediscontinuity in g ( \$ i ) is chosen 'to cancel thatin F(\$i), so thatthe sum
F(\$ J + g(\$ i) is continuous.

## PAWULA e t al.: VECTORS PERTURBED

1833

BG GAUSSIAN NOISE

applies with
= \$o and \$ = A@ + n. In the following, where
we take \$o > A@so that P{\$ 2 \$01 = F(A@+ T ) - f l \$ o ) .
po = 4 ~ cos2
2 Go
V2 sin2\$,
Furthermore, in Cases I and 11, F(A@+ T ) = 0; and in Case
111, F(A@ n) will be negligible in comparison to -F(\$o) and, in (35c), yi, ai,and pi are defined by (14a)-(I 4c) with
unless Go is near A@ + n. Thus, P{\$
\$0) = -F(\$o) in
\$ i = \$0.
Cases I and 11, and in Case 111, when \$o is not near A@ n,
B. \$0- A@ = n/2
PI\$ > \$01 -F(\$o).

A. Small J/o - A@
The term ai - pi cos 8 appearing in the exponents of the
integrands in (13) will have a minimum at 8 = 0 provided
that pi > 0. This minimum will lie within the range of integration whenever 0 < yi - 6i < 2yi, and the condition pi > 0
will be satisfied when [cf. (14c)l

## The point \$o = A@ -I-n/2 is of special importance, for this

point is a turning point in the vicinity of which the character of the asymptoticbehavior ofP{\$ > \$o} changes. In Cases
I and 11, closed-form expressions can be obtained and are
i) Case I:

P{ \$ > n/2} = 1
e4

(36)

## P{ \$ > A@ + n/2} = [ 1 - (W/Vye( V/U, v)] .

The left-hand side of this inequalityis largest when A@ - \$ j =
0. Furthermore, at \$i = A@, the minimum of ai - pi is zero
as can be seen from (14c), viz.

C. Go

A@Near n/2

## Simple asymptotic formulasin

(33)
Assuming the conditions for a minimum within the range of
integration to be fulfilled, cos 8 in (13) can be approximated
by 1 - 812, the range ofintegration extended to fm,and use
made of theidentity [l , eq. (7.4.1 l)]

(37)

## \$o = A@ + n/2 evidently cannot be obtained,except in

Case I, withoutfurther assumptions aboutthe relative sizes
of the system parameters. For Cases I and 11, the asymptotic
formulas are
i) Case I:

P{ \$ > \$o}

+e-

+ L,(UCOS

\$0)]

(38)

## ii) Case 11:

(34)
which then leads to the asymptoticformulas
i) Case I:

## ii) Case 11:

(39)

sin \$o, 60 =
where Po = sgn (cos G0)@U2 cos \$o + V
sin-l (V/o0),and V/U Q 1. Lo(-) is a modified Struve function [ I , ch.121 . These last twoequations followdirectly
from (IO) and (1 1) by approximating the denominators by
their values at \$ = A@ + n/2.Equation (39) is notmuch
simpler than the exact expression (13) and, hence, for computational purposes (13) is t o be preferred.
D,Large Go

- A@

## Since the left-hand side of (32) is negative for \$i - A@ near

T , pi will have a negative sign in Cases I and I1 and also in
Case 111 for small correlations r and X. Asymptotic formulas
in these cases can be obtained directly from (13) by noting
that the major contributions to the
integrals come from the
endpoints; i.e., the minima of the exponent with a negative
pi. Expanding about these points, we are led to the asymptotic
formulas (cf. [8, Appendix]):
i) Case I:

iii)Case 111:

## erfc z = ( 2 / 6 ) 1 , exp ( - t 2 ) dt.

10 Equations (35a), (35b), and (35c) can be further simplified with,

## however, loss of some accuracy, by use of the asymptotic behavior of

the complementary error function; i.e., erfc z (l/z&) exp ( - z 2 ) for
large z .

## ii) Case 11:

1834

IEEE TRANSACTIONS
COMMUNICATIONS,
ON

## iii)Case 111: Althoughtheasymptotic

formulacanbe
worked out in thiscase, the end result is longand cumbersome and will not be given here in detail. It can be shown that

P@

## where p is the symbol signal-to-noise ratio, related to the signal

energy per bit-to-noise spectral density ratio by
P
---.
No - log2 M

(42)

## where the asymptotic form of F(A@ n) is the sum of two

nonzero t e p s and that of F(\$,) is composed of the sum of
WheriM= 2,(44) gives,upon inspection, pE(2) = exp (-p)/2.
four.
For large p , the asymptotic formula12 for PE(M) follows
IV. APPLICATIONS
from the results qf Section 111-A for Case I and small \$,
Thissectionpresentssomeillustrativeapplications
of the and is [cf. (35a)l
preceding results. In the first of these, new formulas for the
symbol error. probability in MDPSK are given, and are compared with previously known formulas. The secondexample
shows howthedistributionoftheinstantaneousfrequency
of a narrow-band wave canbe obtainedfromthe
Case 111
M > 3.
(46)
simplified representation. In the third application, use of the
Case 111 distribution is illustratedinthe
calculation of the
Several otherapproximations
have been developed for
biterrorrateperformanceofnarrow-band
digital FM, with PE(M) and are listed herefor comparison withthe above
limiter-discriminator detection
and
partial-bit integration result. For large p , Fleck and Trabka  derived the approxipostdetection filtering.Theseresultsare
also compared with mation
those for an ideal frequency detector by making use of the
results forthedistributionofinstantaneous
frequency ob(47)
tainedinthe
second application. In a fourthapplication, a
simplified expression is obtained for the BER in DPSK with a
phase error in the reference signal. The final example considers
where X = d m ( n / n / r > lArthurs
.
and Dym  later
the error probability in MPSK.
gave the estimate
A . Symbol Error Rate i n MDPSK
In M-ary differential phase shift keying (MDPSK), information is 'transmitted as the phase difference between two consecutive sinusoidal pulses, and thus the. receiver decision variable can be viewed as being the phase difference between two and,finally, Bussgang and Leiter  obtainedtheupper
vectors perturbed by Gaussian noise. For a maximum p poster- bound
iori probability receiver employingdifferentially
coherent
detection of equiprobable, equal
energy MDPSK signals, the
(49)
probability of a symbol error is [6, ch. 51
'

'E(M) =

(43)

P ( \$ ) db!'

## in which the noise has been taken to be uncorrelated at the

sampling instants. Using the results of Section 11, for the Case I
distribution with \$2 = n and \$ 1 = n/M, we find the symbol
errorprobability' to be
(./M)
n/2
(?--PI
I--os(nlM)costl
PE(4 =

2n

L,,,

dt

1 -cos(n/M)
cos

(44)

## which is also a good approximation forlarge p . Although all of

these three approximations are good for large signal-to-noise
ratios, none is asymptotic in the sense that the error in using
it becomes vanishingly small as p becomes infinitely large.
TheFleckandTrabkaapproximation
is closest inform
totheasymptoticformula,(46).
Neglecting the1/8inthe
denominator'of (47), employing theasymptoticexpansion
for the erfc function in both (46) and (47), and then forming the quotient of these two gives (for M > 3)

## 11 We are indebted to M. K. Simon for pointing out the following

relation between the bit and symbol error probabilities in the
case of
quaternary DPSK, M = 4. In DQPSK, a symbol error results in a single
bit error whenever n/4 C I @ 1 Q 3n/4, and resultsin two bit errors when
3n/4 Q I @ I < n. Using these, we can show that the bif errof- probability
Pg(4) is related to thesymbol error probabilityPE(4)by

## Pg(4) = ( 1 / 2 ) P ~ ( 4+) P { 3 n / 4 < \$ Q n}.

The probability P{3n/4 4 @ < n} can be evaluated exactly from ( 9 )
and (10);however, it is sufficient for our purposes to use the asymptotic
form given by (40), namely,
~ { 3 n / 4< \$

cT}

e-~/(2np).

## Consequeptly, for the region of interest ( p greater than 10 dB),PB(4) =

( 1 / 2 ) P ~ ( 4 with
)
little error. This approximation is then better than the
commonly used Pg(4) = (2/3)P,~(4),which corresponds to orthogonal
signals; i.e., PB(M) = MPE(M)/~(M
- 1) for M = 4.

## 12 Another asymptotic formulacan be obtained from (44)by noting

+n with an error that
that the limits of integration can be extended to
becomes.vanishinglysmall
as p getsinfinitelylarge.This
gives the
asymptotic formula

## which has been obtained previously whenM = 4 [ 11, p. 2471.

PAWUL.4 PERTURBED
et al.: VECTORS

1835

NOISE
BY GAUSSIAN

## which is independent o f p . For M = 3 , 4, 8, and 16, values

of this ratio are

0.9186
0.9768
0.9987
16

## Hence, although Fleck and Trabka's result is not asymptotic

in the strict sense of the term, it
differsin its leading term
from the asymptotic formulasby less than 9 percent in the
worst case when M = 3, and by less than 3 percent in the
more ordinary situation when M is a power of 2 [M> 2 is
tacitly assumed, since an exact expression is known forPE(2)].

## as 7 + 0 is to note first from (14a) that yi -+ n as T -+ 0. Then

the limits of integration in the integrals of (13) are over 2n
ranges and, consequently,
1) the values of lii in (13) are unimportant,
2) the algebraic sign of pi is unimportant, and
3) the integrals in (13) are
expressible in terms of Io and
l e functions as in (1 8).
Thus, employing(13)
to evaluate the limit of F ( 0 0 7 )
as 7 + 0, and using the result in (52), we find (cf.[ 11, eq.(6.6)])

(53)

## B. Distribution of Instantaneous Frequency

Thedistributionoftheinstantaneousfrequency
of a in which a and are the limits as 7 + 0 of (14b) and ( 1 4 ~ ) ,
narrow-band signal may be regarded as a limiting case of the and are
distribution of the phase angle between two vectors perturbed
(wo - i ) 2
by Gaussian noise, if the vectors are appropriately regarded
a = qI+
(54)
as samples of a narrow-band waveform with vanishingly small
2
wo2 - r - 2wox
timeseparation.The
Case 111 simplified representationof
Section 11-C for P{ G1 Q \$ < \$2} provides a convenient mechU2(Oo - &)2
1'2
p=
(5 5)
anism for readily determining the probabilityP{w< w o } ,that
wo2 - i: - 2woX
theinstantaneous
radian frequency lies belowsome value
w o . In the remainder of this section, wewill consider only the
case that the noise power is not time-dependent so that u1 = and A = dA(t)/dt and U = U(t) [cf. ( l ) ] . Equation (53) can
also be written as
02.

+ (A/#

[?

of a narrow-band

Theinstantaneousradianfrequency

signal is thetime

derivative ofitsinstantaneousphase;

i.e.

## in which 8, = 8(t), O2 = O(t T) and, in the final limit,

O2 - O 1 has been replaced by its modulo 2n difference \$ =
(8, - 8,) mod 277 (cf. [ l l , ch. 61).Then,formally,the
probability P { o < w o } can be determinedusing (a) as the limit

P{ 0 < wo}
=

1- P { o

= 1 - lim

x,

## where v = w o - w, and r have been expressed in terms of

os,u s 2 , a n d w , . H e r e h = w , + G s a n d r 2 =-h2-u s 2 . The
quantities Gs and us2 dependonthe normalizedbaseband
power spectrum S(f) through YS; = f ? w u S ( f ) d f and us2 =
J:,(w
- Gs)2,S(f)d& The expressions for a and 0 corresponding to (53a) can be obtained by replacing w o - & by
v - in (54) and w o - Y- 2w0h by us2 + (v - Os)2in (54)
and (55). Equivalent results, expressed in terms of theMarcum
Q-function, were obtained by Salz and Stein [ 121 . l

> wo},
P{\$ > w ,

7},

.
0
'
7

0'7

+ lim F(w07)-

7-f 0

if wo >a,

## The bit error probability for narrow-band digital FM with

limiter-discriminator detection and integrate and dump (I&D)
output filtering is consideredindetailin
 ; however,this
reference exclusively treats
the
uncorrelated noises case
of an output filter (see Fig. 4) which integrates over the entire
bit time. The more involved problem ofpartial-bit integration,
i.e., that in which the output filterintegrates over less than
the full bit time, requires additional analytical tools that are

## where, by (4a), & = w, + and = lim,,o 7-'(\$2 - \$1).

The limitsin(52)
can be evaluated using (13) byfirst
regarding all ofthe parameters U, V , W , A@,r, and X as
functions of T, and then letting T -+ 0. For small T , (7) x
1
iLr2/2, where i: = [d2r(T)/dT2]7=o;
X(7) 2 kr, where
= [ d X ( . r ) / d ~ ] , =and
~ ; (13) shows immediately that F(A@+
13 The probability density function of the instantaneous frequency,
n) + 0 as T + 0. Hence, only F(w07) needs to be further
which can beobtained by differentiating (53a),
has been given by Gatkin
considered. The key to the evaluation of the limit of F(w07) et al. [ 2 1 ] .

1836

## IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. COM-30, NO. 8, AUGUST 1982

LimiterDiscriminator

; ; ;FI

t)
(+ clicks)
+(

## discriminator output noise . Part of this calculation,the

onlypartto
beconsidered here, requires the evaluation of
P{\$ < 0) for A@ > 0. The desired result for this probability
follows from P{ \$ < 0) = F(0) - F(A@- n) and (12), and is

PI\$ < 0)
provided by our results in Case 111 of the present paper. In this
example, we briefly indicate how the Case 111 results can be
used in the BER calculation in the partial-bit case.
Intheworst
case situation of an alternating one-zero
data stream, the operation of narrow-band digital FM can be
partlymodeledintermsofthe
phase angle betweentwo
vectors perturbedby Gaussian noise,when the underlying
sinusoid[cf. (l)] has an amplitudeandphase of theforms
(cf. [8, eqs. (12) and (13)] for a linear phase IF)

A ( t ) = d a o + a2 cos 2n

f_

7(
+

a,. sin n

(56a)

a, sinn-

1 - COS A@ COS t

dt

rsinA@
\$ 7 -

/,,,

exp i-u

n/2

dt

1 +cost
1 + r cos A@ cos t

1 + r cos A@ cos t

## In the limit 7 -+ 0, the output filter may be regarded as a

sampler which samples the instantaneous radian frequency, say
o,of the discriminator output. The above probabilitythen
becomes P ( o < 0) , which can be found in the limit of the
above equation, but is more directly given by (53) with oo=

O,X=O,k=O,viz.

a.

cy2

1-rcost

(60)

and

\$(t) = tan-

1 - COS A@ COS t

+ a2 cos 2n T

where aO, a,, and a2 are constants (ao > 0, a. -+ a 2 ) and Tis
the bit time. Letting 7, 0 < T < T , denote the partial-bit integration time, we can
express the phase angle of interest between thenoisy vectors as

and
i: = [d2r(~)/dr2],=o.

## where u2 is the noise power after IF filtering. The noise

correlations of (5) are determinedbythe
system narrowband IF filter; and for ( ~ ~mod
7 ) 2n = 0 and for a filter with
a symmetric low-pass equivalent H ( f ) , h = 0 and r = (7) is
given by

(59)
J-

## The overall bit error probability is composed of averages

over various bitpatterns,andcontainscontributionsfrom
boththecontinuousandthe
click componentsofthe

## D.The Effect of Phase Error on DPSK BER

In a recentpaper, Blachman [ 131 considers theeffect
of an error in the delay of the preceding signal, in calculating
the error probability in decoding the present signal, in differential phase-shiftkeying (DPSK). By recastingthis situation
in terms of the phase angle between two vectors perturbed by
Gaussian noise,asimplificationresults
fortheerrorprobability.
If we let 0 denote a constant phase error in the reference
signal, the error probability PB can be written as an infinite
series, withterms
involving modified Bessel functions as
[13, eq. (7)l

PAWULA PERTURBED
e t al.: VECTORS

1837

BY GAUSSIAN NOISE

## where p 1 and p 2 are theinstantaneous signal-to-noise ratios

of the two vectors. Using (18) with \$ i = -n/2,r = h = 0, and

where

## and Uand W are given by (4).

The equivalence between (62) and (63)14 can be shown directly by the following steps. From (4),(C-l), and Neumanns
additiontheorem(cf.
[ l , eq. (9.1.79)]), U-le(Y,
has
the representations

v>

U - Ze(Y,

-+

v>

rl

(-l)mem cos 2 m O

in which p = p 1 and -n < \$ < n (A@ = 0 since MPSK is implicitly coherent signaling). The limit as p 2 -+ 03 (or u2 0) of
the function F(\$)will first be shown to be consistent with this
p ( \$ ) , and will thenbe applied tothe MPSK symbolerror
probability calculation.
Taking the limitp 2 + m of (1 1) (with A@ = 0) leads to

m=O

dt

## in which x = p 1 / 2 , y = p 2 / 2 , eo = 1, and e , = 2 for m > O .

When (65) is used in (63), after some rearrangement (63) becomes a Fourier series in cos (2n + l)O, as is (62). Equality
follows by equating coefficients and use of the identity (which
can be verified by differentiation)

## which is essentially the same as the result of Weinstein [ 171 .

When M = 2, (71), with the aid of (69) and (34), reduces t o
the well-known form p ~ ( 2 )= (1/2) erfc
Also, from
(71), it is straightforward toshowthat P~(ln) erfc [&
sin (n/M)]f o r M > 2 (cf. 16,p. 2311).

(6).

## E. Symbol Error Rate in MPSK

Adegenerate case of the distribution of the
phase angle
betweentwo vectors perturbed by Gaussian noise is that in
which one of the vectors is noise free. This situation is representative of MPSK; hence, appropriate limits of our previous
results can be used to obtain the symbol error probability in
MPSK .
When the second vector, say, has zero noise, the probability
density function of the phase angle between the two vectors
14 We aregratefulto
N. M. Blachman fortheinterestingbit
history
that
a
special
case
of
( 6 3 ) was obtained,butapparently
published,
by
L. Lewandowski
in
1961.

## In taking the limit of (1 1) to get (69), it is t o be noted that

both U and V + 00 as p 2 -+ 00 and that the integrand of (1 1)
becomespeaked in the vicinity of t = t o , to = tan- { V /
(W cos \$)}. Expanding the integrand in powers of ( t - to) and
passing tothelimit
gives (69). Equation (70) then follows
by the change of variable x = tan 8 . That (68) and (69) are
consistent is readily verified byshowing that thesesatisfy
P(\$) = w \$ ) / a \$.
Turningnow t o MPSK: forequiprobable, equalenergy
MPSK signals, the symbol error probability for the optimum
receiver is PE(M)= 2P{\$ > n/M}. Consequently, employing
(70) to evaluate F(n) - F(n/M) gives

of
not

## V. RELATION TO THE FOURIER SERIESAPPROACH

In most of the preceding, definite integral expressions have
been presented for the probability densities, distributions, and
related quantities of interest.However, some of these densities
and distributions can be obtained in a quite general way as
Fourier series .It is the goal of this section to briefly explore the series approach and show its relation to our previous
work. Primarily the Fourier seriesmethod appears to be limited
to Cases I and I in which there is no
noises perturbing
endpoints
thevectors;
the
ofhowever,
the

between the

1838

IEEE TRANSACTIONS
COMMUNICATIONS,
ON

## method will be found to have someapplicationin Case 111.

Additionally, the Fourier series method quite naturally allows
for extensions to situations which fall outside the scope of the
main part of this paper, situations in which the phase angle of
interest is the result of the modulo 277 combination of more
than two components.
ThestartingpointfortheFourier
series approach is the
Fourier series of the periodic extension of the Bennett probability density, of Section IV-E, for the phase noise of the
angle of a single vector; i.e.,

-,MIG~.

=% k=O

k 5 k ( P l ) z i Z k ( P 2 ) 6 k ( P 3 ) COS

k\$;

\$ = ( 2- e 32 )- m
e 1o d 2 n .

(77)

## Other combinations can be handled similarly.

Probabilities
stemming
from
the
series representations
of such densities can be determinedby termwiseintegration.Forexample,
Blachman's errorprobability PB given
by (62) is just the termwise integral of (76) over the intervals (-n, - n/2 - 0 ) and (n/2 - 0 , n), viz.

(72)

## The Fourier series corresponding to this& (8) can be shown to

be (cf. Prabhu , Middleton [18, p. 4171 and Lindsey [19,
P. 1901)
.

p(\$)

where'

## Further, distributions like P{\$ > ll,o} in Cases I and 11, of

concern in previous parts of this paper, can be obtained as
definiteintegralexpressionsbydirectlysumming
theircorresponding Fourier series. However, the ii-coefficient of the
combination density
becomes
increasingly complicated as
themodulo
2n combination gets more involved, so that
it becomes more difficult to sumtheFourier
series. For
example, we have not as yet succeededinsumming
either
the series for p ( \$ ) or its counterpart for P{\$ 2= ll,o} in the
case of 11, = (202 - 0' - 0 , ) mod 2n.
VI. SUMMARY AND CONCLUSIONS

(74)
From this series, the series for the density

## 2n follows by noting that p ( \$ ) = [p1(11,/2)

which leads to

\$ = (20) mod2n.

of \$ = (28) mod

+ p1(\$/2 + n)]/2,

(75)

## The probability distribution of the modulo 2n phase angle

between two vectors perturbed by correlated Gaussian noise
has been studied in detail. For various types of signal conditions and noise correlation, the distribution and its asymptotic
behavior have been presented. Some of the resultswere applied toproblemsinanglemodulationincommunication

systems,andinfinding
thedistributionofinstantaneous
frequency.
Even in the most complicated case, the expressions for the
distribution requireonly the numericalevaluation of single
integrals,andtheseare
easily done on a digital computer or
hand-held programmable calculator to ten digit accuracy
using the methods described in [ 101 . Such methods were used
to get numerical comparisons of the asymptotic formulas with
exact results and, ingeneral, the asymptotic results were found
to be quite good for U > 10 dB and over the entire range of

The secondingredientin
theFourier series approach is
the observation, made from (6), that the Fourier series for the
combination of twoindependent e's, with densities of the
\$0.
form of (73) or ( 7 9 , has Gcoefficients which are just products
By following thetechniques used, the resultscanbeexof the Z-coefficients of the component pdfs. Then, series for tendedtoother
special situationsofinterestin
phaseand
the'densities of combinations such as \$ = (0, - 0 , ) mod 2n frequency modulation (cf. [6, ch. 51).
and \$ = (202 - e l - 0,) mod 2n" [all the 0's are independent with densities given by (73)] can be written by inspection,
APPENDIX A
viz .
CHARACTERISTIC FUNCTION METHOD

11, = ( e 2 - 0 , ) mod 2n

## is the well-knownsignal suppression factor [ 6 , p. 501.

This combination arises in Manchester encoded digital FM .

1 5 ;El(p)

16

(76)

## Here we sketch the derivation of (21) and (22). The original

developmentin [ I 11 hasbeen changed somewhatand rearranged into a step-by-step procedure.
1)Defineaperiodic
sawtooth function s ( t ) of period 277
such that

PAWULA PERTURBED
et al.: VECTORS

1839

BY GAUSSIAN NOISE

- \$i n)] that can be used in (A-2) t o get an expression for P{ \$ 1 < \$ G G2}. The desired equations (21)and
(22) now follow when (20) is used to express EIJo(*)] as

0,

and useconstruct
toit
the
function

1
+n)1.

- [(\$2-\$1)+s(\$-\$2+~)--s(\$-\$l

2n

@(-u, conditional
u I \$J, the
tion 11-E.

characteristic
of function

## This function of \$ is discontinuous at \$ = \$ 1 and \$ = 1 4 ~ .

It is equal to 1 in \$, < \$ < G2 and to 0 elsewhere in - n <
\$ < n. Then (cf. [l 1, eq. (2.26)])
ALTERNATE
DERIVATION

P{\$lG\$G\$2}

\$ 2 -IL1

1
+E E S ( ~-ol
,
2n

=
7
-s(02

-81

- \$1

+ n)I

\$2

+n)
(A-2)

where theexpectationextends
over therandom variables
0 , and 0 2 . By the periodicity of s(-), the mod 2n in
\$ =
(0, - e,) mod 2n is accounted for.
2) Note the basic relation [ 11, eq. (2.20)]

Sec-

APPENDIX B
OF P{ \$1

## This Appendix presents an alternate derivation of P{\$l <

\$ < tJ2} which follows more along classical lines thanthe
derivation given
Section
in
11-E.the
For
sake of simplicity,
only CaseI1is considered in some detail. The derivation can
be extendedto
Case 111, althoughthe
algebra involved becomes tedious.
CaseII: Initially,the derivationproceeds along the lines
followed by Fleck and Trabka [SI, who start with the fourdimensional joint Gaussian probability density function
p ( x l , y l , x2,y2) of the four noise variables in (2). A change
is made to polar coordinates, and the resulting radius vectors
integrated over,leading to the joint density p(8,, 0,). Employing (6), the result for p ( \$ ) , \$ =
- e,) mod 2n, can
be written as (cf. [8, Appendix])

(19,

## which can be verified by changing to polar coordinates x =

cos @,
y = R sin @.Then the integration with
respect to
R can be performed immediately (since it is equivalent to
integrating Jo(cR), c being a constant).Theintegration
with respect to 4 can be done by setting 2@= n/2 - t and
using

where E = U - V sin t - W cos (A@ - \$) cos t. The observation is made that the negative exponent E satisfies the
identity

3) Replacein
(A-3) by O 2 - O 1 - E and change the
variables of integration to u , u where x = uRl , y = vR2 and
R , , R 2> 0. This carries (A-3) into

## so that multiplying and dividing the integrand of(B-I)by

E, using (B-2), and integrating over (\$1, G2) gives

+-4n

dt
-n12

## Takjng the expected value of both sides of (A-5) and setting

t = \$i - n (SO a/a\$ = a/aJli) gives a form17 for E[s(O2 17 Interchanging the order of expectation and integration
canbe
). If thedouble
justifiedbystandardmethodsofanalysis(see
integral is first expressed in the polar coordinates (R,@)then the most
difficult step to justify is the interchange of the expectation with the
R-integration. The validity of this step can be shown by first writing
the R-integral as one from 0 to M plus a second from M to m. Then
the interchange can be made in the first part with the finite limits and,
as M
-, the contribution from the second part can be shown to be
nil by use of the Riemann-Lebesguelemma (or theasymptotic behavior of the Bessel function).
--f

(___-)(T
a\$
-1

\$I

aE

d\$)

~cost

03-31
Uponintegrating thesecond integral by partswithrespect
to t , and the third integral by parts with respect to \$, it will
be found formally that all of the remaining double integrals
cancel, and that there results the single integral

1840

IEEE TRANSACTIONS
COMMUNICATIONS,
ON

## However,care must be taken if the singularity E = 0 lies

within the range of integration, for then the steps performed
in going between the last two equations would not be valid.
This singularity occurs when

## where I o ( - )is a modified Bessel function [see [9, eq. (A3-5)]

for (C-2)].
2) Hankel Exponential Integrals
( [ I , eq. (Z1.4.28)]:

rm

\$ =A@,and

t=tan-

(B-5)

and will be avoided if A@ lies outside the range of integration. If A@ lies withinthe range of integration,thenthe
abovederivation
mustberepeatedforthecomplementary
probability P{\$ <
or \$ > \$2). Evaluating the integrand
of (B-4) then leads to (9) and(1 l), which completesthe
proof.
CaseZIZ: Inextendingthis
derivation to Case 111, the
probability density functionp(\$) becomes

## VOL. COM-30, NO. 8 , AUGUST 1982

tJo(at)e-bf2 dt = - e2b
I
Jl(at)e-bf2 dt =-(1
a

a2/4b

-e-a2/4b).

## 3) The following two integrals can be obtairi\$d by use of

the Euler-Legendre transformation, (3 1):

exp[a+bsint+ccost

4-

1 + d cos t

2ir

I,

1 + d cos t

dt

[+
a

4-2

277

L,,

dt

= e-aIo(13)

+ c cos t
1 + d cost

b sin t

a+bsint+ccost

where
where

E=

U--~sint-W~os(A@-\$)co\$t
1 - (r cos J/

+ x sin \$) cos t

## and the analog of (B-2) is

REFERENCES
M.Abramowitzand
I . A. Stegun, Handbook of Mathematical
Functions. New York:Dover,1972.
E.
Arthurs and H. Dym,On the optimum detection of digital
signals in thepresence of whiteGaussian noise-A geometric
of threebasicdata
transmission sysinterpretationandastudy
tems, IRE Trans. Commun. Syst.,vol. CS-IO, pp. 336-372, Dec.
1962.
 J. J . Bussgangand M. Leiter, Error rate approximations for
differential phase-shift keying, IEEE Trans. Commun. Syst., vol.
CS-12,pp. 18-27, Mar. 1964.
 J. Edwards, A Treatise on the Integral Calculus, Vol. 11. London,
England:Macmillan.1922.
 J. T. Fleck and E. A. Trabka, Error probabilities of multiple-state
differentially coherent phase-shift keyed systems in the presence of
white, Gaussian noise, in Investigation of Digitul DataCommunication
Systems,
Rep. UA-1420-S-1, J.G.
Lawton,Ed.,
Cornell Aeronaut. Lab.,Inc., Buffalo, N Y , Jan.1961, Detect
Memo 2A; available as NTIS Doc. AD256584.
 W . C. Lindsey and M. K . Simon, TelecommunicationSystems
Engineering. EnglewoodCliffs. NJ: Prentice-Hall,1973.
 J. I . Marcum, A statisticaltheory of target detection by pulsed
vol.IT-6,Apr.
1960, pp.
radar, IRE Trans.Inform.Theory,
59-267.
[SI R. F. Pawula, On the theory of error rates for narrow-band digital
FM, IEEE Trans. Commun.. vol. COM-29, pp. 1634-1643, Nov.
1981.
 S. 0. Rice, Statisticalproperties of asine wave plus random
noise, Bell Syst. Tech. J . , vol. 27, pp. 109-157, Jan. 1948.
Efficient evaluation of integrals of analytic functions by the
[IO] -,
trapezoidal rule, Bell Syst. Tech. J . , vol. 52, pp. 707-122, MayJune 1973.
[I]

E[2U-2W(rcosA@++sinA@)-E(1-r2-A2)]
(B-7)
APPENDIX C

## SOME INTEGRAL RESULTS

This Appendix sumniarizes certain integral results needed in
the main part of the paper.Also listed are two interestingdefinite integrals which are related to the analysis, and which do
not appear tobe given in most tables.
1) TheIe(k, x) Function: This function isdefined and
tabulated in , and [ l l ] containsadditionaltabulations.
For Ik I< 1
(C-1)

__

77

e-x(~--kco~e)

de

1--cos0

(C-2)

1841

## design. intermodulation distortion, multipath, jamming, modulation

and
demodulation, and error correction codingand
decoding. Heis also
interested in long-range avionics development.
Dr. Pawula is a member of Sigma Xi, Tau Beta Pi. Eta Kappa N u , Phi
Eta Sigma. Rho Epsilon, and Alpha Eta Rho.He was an Alfred Sloan
Fellow at M.I.T. and aHowardHughes
Fel.low at Caltech. He also
belongs to the Aircraft Owners and Pilots Association. the Experimental
Aircraft Association, and the Baja Bush Pilots.