621.396.33: 621.391.812.3:621.376.
55
PHASE-SHIFT KEYING IN FADING CHANNELS
By H. B. VOELCKER.
(The paper was first received 4th April, and in revised form 30th June, 1959.)
SUMMARY
Phase-shift keying (p.s.k.) is discussed as a modulation technique
for transmitting digital data over radio circuits subject to fading. The
modest bandwidth requirements of p.s.k. modulation suggest that it
can not only alleviate spectrum crowding, but can also transmit traffic
with fewer errors. The theoretical results presented here indicate,
however, that the random phase perturbations inherent in fading radio
signals cause unavoidable degradation in the performance of p.s.k.
systems. Experimental data which partially support the theoretical
results are cited, and comments relevant to the improvement of p.s.k.
systems are included.
(1) INTRODUCTION
Long-range radio circuits have long been used for the trans-
mission of teleprinter traffic and other forms of digital data.
Marked improvements in the reliability of such circuits began to
accrue about two decades ago, as frequency-shift keying (f.s.k.)
gradually replaced amplitude modulation for digital data trans-
mission. However, the steadily increasing crowding of the radio
spectrum, together with demands for ever lower error rates, has
fostered a continuing search for better modulation techniques.
It was recognized long ago that phase-shift keying (p.s.k.) offers
attractive possibilities for bandwidth conservation and for
reducing the number of errors caused by white noise, but prac-
tical difficulties delayed its exploitation until quite recently.
In 1957, an interesting field test was conducted over a 5000-
mile radio path to evaluate the performances of a modern f.s.k.
system and a new p.s.k. system. While the operating procedures
and many of the results of the test have already been reported,1
a brief summary of this information is given in Section 9. Per-
haps the most striking result was the comparative performance
of the two systems. For relatively low values of signal/noise ratio,
the p.s.k. system demonstrated a fairly constant superiority;
at relatively high ratios, however, the f.s.k. system showed a
superiority which increased steadily with the signal/noise ratio.
This curious comparative performance occasioned the theoretical
analysis and the comments set forth in the paper.
(2) PRELIMINARY DISCUSSION OF PHASE-SHIFT KEYING
The use of p.s.k. modulation for constant-rate data trans-
mission requires that a stable carrier wave be available at the
transmitter. At equally spaced instants in time, the carrier may
be shifted in phase either 180°—the binary case—or ±90°,
180°—the quaternary case—to transmit either one or two bits
of information. If quaternary keying is used, it is easy to show
that the time-bandwidth product of the modulation is 1/2; this
represents maximum channel loading if inter-symbol interference
is to be avoided. (For simplicity, only a single phase-keyed
carrier is considered in the paper; the extension to frequency-
multiplexed systems is obvious. Furthermore, p.s.k. data pulses
need not have a 'square' shape in the time domain, since the
Written contributions on papers published without being read at meetings are
invited for consideration with a view to publication.
Mr. Voelcker is in the Department of Electrical Engineering, Imperial College of
Science and Technology, University of London, and was formerly at the U.S. Army
Signal Research and Development Laboratory, New Jersey, U.S.A.
The Institution of Electrical Engineers
Paper No. 3125 E
Jan. 1960
associated sin x/x spectrum is sometimes inconvenient. Costas,
for example, has proposed a p.s.k. system which radiates signals
with essentially 'square' spectral characteristics.2)
Demodulation of p.s.k. signals requires a phase reference with
which incoming pulses can be compared. Schemes for providing
phase references fall into two classes, namely stable sources and
transmitted references. If the transmission medium has con-
stant phase characteristics, very stable frequency sources, e.g.
caesium clocks, may be used for modulation and demodulation.
After initial synchronization, the inherent stability of the sources
should ensure reliable operation with only occasional adjustments
to compensate for drift. Unfortunately, phase perturbations in
most physical media require that transmitted phase references
be used.
Schemes for transmitting phase references are many and
varied, and the paper is concerned mainly with p.s.k. demodula-
tion via 'differentially coherent detection'. With this technique,
each signal pulse is detected by phase comparison with the
(stored) preceding signal pulse. Information is thus conveyed
by the phase differences between successive pulses, and redundant
information need not be transmitted for reference purposes.
Section 10 explains why differentially coherent detection is well
suited for p.s.k. transmission over fading radio circuits, where
phase perturbations are more or less continuous.
Before proceeding with an analysis of p.s.k. detection, it is
worth while to review some of the statistical characteristics of
fading radio waves.
(3) STATISTICAL CHARACTERISTICS OF FADING RADIO
WAVES
Although the statistics of fading signals have received wide
attention in the literature,3"5 this brief review is included because
the bivariate statistics which are important in p.s.k. studies are
unfamiliar to many engineers. The mathematics are identical with
those of noise theory, since it has proved convenient to view
a fading signal as the envelope of narrow-band normally-dis-
tributed noise which may or may not contain a steady (specular)
component.
Suppose that a continuous stable carrier signal is transmitted
through a fading channel. At some instant, t, the received
signal can be represented by the vector sum of two independent,
quadrature vectors, xt and yt, each of which exhibits stationary
normal statistics with zero mean value and average power a2.
The probability density function (p.d.f.) of the received signal
vector is therefore
1 / xj + y2^
By converting this expression to polar co-ordinates (Et, 6t) it is
easy to show that the vector modulus, Et, exhibits the well-known
Rayleigh p.d.f. and that all values between ±TT are equally
probable for the phase, dr6>7
Ostensibly it would appear that p.s.k. modulation cannot be
transmitted through a fading channel because of the random
phase characteristics of the received wave. This is not true,
[31 ]
32 VOELCKER: PHASE-SHIFT KEYING IN FADING CHANNELS
however, for the preceding result tells nothing about the rapidity
of phase changes; it merely shows that the phase of a single
random sample is unpredictable. Temporal variations can be
evaluated by describing x, +T and yl+t, the quadrature com-
ponents at time / + T, in terms of conditional p.d.f.'s involving
x, and y,; these take the following form :7
1
exp 2a2(l -
. . . . (2)
where <£T == Av [V(/)V(f + r)]/of and 0 < <£T < 1.
A p.d.f. may now be written for the received signal at the two
instants, t and t + T, in terms of x,, yt, xt+i: and yt+T. This
expression can in turn be converted to polar co-ordinates
(Et, Et+X, A0T), where A0T is the random phase change over
r seconds., Finally, by integrating over all values of Et and -180 -120 -60 0 60 120
El+T, the following rather formidable p.d.f. can be obtained RANDOM PHASE-SHIFT OVER T SECONDS, DEG
for A0 T :
Fig. 1.—Phase stability of the 5 Mc/s WWV carrier received over a
short E-layer path, as calculated from the experimental data of
2TT(1 -
1+ R. Price.
. . . . (3) to phase perturbations. While coherent or phase-sensitive f.s.k.
where fx = (j)T cos A0T. systems can indeed be constructed, it is easy to show that f.s.k.
This expression was derived by Bramley in his investigation of detectors should usually operate on an energy- or envelope-
space-diversity reception, although he was concerned with phase comparison basis in the presence of fading and noise. It follows
differences between aerials rather than with temporal phase from this that phase perturbations which degrade a p.s.k. system
shifts.8 While the investigation of p.s.k. detection in later seriously will have a smaller, usually negligible, effect on an f.s.k.
sections does not involve eqn. (3) directly, some discussion of system.
eqn. (3) is worth while to illustrate certain natural constraints
governing the performance of p.s.k. systems on fading circuits. Eqn. (3) also shows that phase perturbations induced by fading
cannot be minimized by radiating more power; they are a
The key parameter in eqn. (3) is <j>x, the correlation coefficient function only of (f> . If the radio path and transmitting frequency
T
of each quadrature component over the interval T. Price has are fixed, (f) is determined by Nature rather than by man. The
T
reported experimental results which provide a good deal of use of very-short-pulse transmissions ostensibly offers a way to
9
physical insight into <f>x. To determine <f>T for a short E-layer maximize (f>T and thus to 'peak' the p.d.f. of eqn. (3). This
path he heterodyned the 5Mc/s signal from WWV to 20c/s method is useful, however, only when the fading is essentially
with a very stable reference oscillator; this yielded a 20c/s sine flat or frequency-non-selective. Frequency-selective fading is a
wave modulated by the noise-like envelope of one of the quadra-
ture components of the received signal vector. He then auto- manifestation of 'coarse' multi-path propagation, and it is clear
correlated the 20c/s signal, and obtained <f>T directly from the that contiguous short-pulse transmissions will be seriously
envelope of the autocorrelogram. His results show that the degraded if the relative path delays are a significant fraction
correlation function was roughly Gaussian in shape with a of the pulse length.
standard deviation in the order of 1 sec. More generally, his Finally, readers should note that the foregoing equations and
results agree with predictions based on the physics of turbulent comments are strictly applicable only to fading carrier signals
media, in that 0T should be an even, monotonically decreasing which contain no specular components, and thus exhibit Rayleigh
function of T. fading. While most of the following Sections deal only with
Rayleigh fading, the effect of a specular component will be
A plot of eqn. (3) for Price's values of cf>T is shown in Fig. 1. partially evaluated in Section 8. In the analysis of p.s.k. detec-
It is clear that when T is very small and <f>T ~ ], phase changes tion which follows, </> is taken as a constant of the environment,
greater than a few degrees, say, are possible but improbable. and selective fading isTignored (i.e. the data pulses are presumed
Conversely, as T increases and </>T decreases, large random
excursions in phase become increasingly probable. Direct to be long compared with multi-path time delays). Furthermore,
experimental measurements of temporal phase fluctuations have the 'speed' of the fading is constrained to be markedly slower
been reported recently by Brennan and Phillips for medium- than the data rate; this means that the power spectrum associated
frequency E-layer propagation,10 and by Landmark and Hagfors with a fading, non-keyed carrier must be markedly narrower
for ionospheric scatter propagation.11 These authors' results than the spectrum of the same carrier when keying is imposed
generally lend credence to eqn. (3) and the theory underlying it, upon it. This constraint has the effect of restricting cf>T to values
and to the comments on the nature of </>T. in the range between, say, 0 • 9 and 1 • 0. The restriction to 'slow'
Certain conclusions can be drawn immediately from eqn. (3) fading not only simplifies the analysis, but also is a necessary
regarding the performance of p.s.k. transmission in fading condition if phase keying is to be feasible in practical situations.
channels. The transmission of information requires time; thus,
if T is the basic pulse length of a p.s.k. system, eqn. (3) shows that (4) THE BASIC BINARY P.S.K. DETECTOR
errors will occur in binary or quaternary phase-keyed signals The first problem to be attacked is the specification of the
even though no other disturbances, e.g. additive noise, are basic binary p.s.k. detector. By what processes should trans-
present. This follows because the probabilities of large random mitted phase shifts be inferred from the received signals?
phase excursions exceed zero for <f>- < 1. It should be noted Certain assumptions simplify the answer to this question.
that f.s.k. systems, for example, need not exhibit this sensitivity First, additive white noise and fading are assumed to be the
 VOELCKER: PHASE-SHIFT KEYING IN FADING CHANNELS 33

only forms of signal degradation. Section 9 contains comments The process implied by eqn. (5) cannot be implemented
on the sufficiency of this description of the radio channel. directly, since neither Eo nor xx is known. Eo must be inferred
Second, the communication system is assumed to be time- from £0 and xx from £j. To see what the inference processes
synchronous. In practice, the additive noise should be mini- entail, consider the log-likelihood function for EQ and <x0, where
mized in an optimum complex-conjugate filter, although this part a0 is the random phase-shift induced in the reference vector
of the detection process will not be discussed in the paper. The by the additive noise (see Fig. 2):
combination of time-synchronism and optimum filtering, when
coupled with the restriction to non-selective 'slow' fading, allows log L(£o, ao|£o) = log [p(xo|£"o)p(^o)]
each data pulse to be described quite simply, after filtering, in
terms of the magnitude and phase of a single vector sample. cc ^-2{El - 2£0£0 cos a 0 + £§) (6)
Third, the two binary states denoted by M (mark) and S (space)
are assumed to be equally probable; also, the losses and gains
Denoting maximum-likelihood estimators12 by circumflexes:
involved in choosing between M and S are assumed to be
symmetrical.6 sin a
Since the receiver is to employ differentially coherent detec- ^— log L(£o, a o | £0) = - ^T o
tion, the binary detection process consists in determining
whether the current signal pulse connotes a 0° or a 180° phase Therefore a 0 = 0 , EQ£,0 ^= 0
change with respect to the preceding (stored) signal pulse.

Therefore EQ = £ 0 , where Eo = E0(oi0) . . . . (7)

T h u s , £ 0 is t h e best (in t h e maximum-likelihood sense) estimate

of Eo. Similarly, it c a n be shown that t h e best estimate of xx
is p r o p o r t i o n a l t o £ p Substituting these estimates in eqn. (5)
gives
PMls<xUZi (8)
Eqn. (8) shows that the decision between M and S should be
ib) (C) governed by the polarity of the scalar product of the reference
(a) and signal vectors. If £0£i > 0, M is chosen; if ^0^t < 0, S is
Fig. 2.—Binary phase-keying in a fading, noisy channel. chosen. While this is certainly not a surprising result, it does
(a) The reference vector, i.e. the preceding signal pulse. justify the use of product detection in p.s.k. demodulation. The
(b) The signal vector on the hypothesis of an M (0° shift) transmission. transistorized p.s.k. equipment used in the 1957 field test
(c) The signal vector on the hypothesis of an S (180° shift) transmission.
employed synchronous differentially-coherent demodulation via
Fig. 2 illustrates the situation for a given pair of vectors, £0 product detection; it is described elsewhere.13
£j. The various vector quantities in the figure are defined as The form of the basic binary p.s.k. detector having been
follows: specified, the next step is the derivation of the p.d.f. of the
product detector output. As shown in Fig. 3, the output, r, is
£0 = Reference vector, which represents the preceding
corrupted signal pulse.
£j = Signal vector, which represents the present cor-
rupted signal pulse.
Eo = 'Pure' signal component of the reference vector;
this is used to orient the axes in Fig. 2.
Elx = 'Pure' signal component of ^ which is in-phase or
anti-phase with Eo, depending upon whether M
or S was transmitted.
Elyly = 'Pure' signal component of £j which is in quadra-
ture with Eo due to propagation turbulence.
n
0y Independent, additive, zero-mean, normally-dis- Fig. 3.—Binary phase-keying in a fading, noisy channel.
n
\y tributed noise vectors of mean power cr2. (a) The reference vector.
ib) The signal vector.
(c) Symbolic representation of the product detector.
If M and S are viewed as two hypotheses, the detection process
can be specified via the ratio of their likelihoods.6'12 Denoting the scalar sum of the products of the respective quadrature
likelihood functions by L( ): components. The derivation of the p.d.f. will only be outlined
L(M| £»,£,) _ p(Eo)p(xo\Eo)p(yo)p(xl\Eo,M)p(yi) here, for it is tedious in execution although simple in principle.
A transform pair is defined as follows:
US\ £0,
r 00
(4)
p(r\E0) = J
Pr(f\E0) exp {2irjfr)df . (9)
— OO
p(Xl\E0,S)
If PMjS (preference for M as opposed to S) is a logarithmic {-2irjfr)dr (10)
measure of the likelihood ratio, then substitution of eqn. (2),
after modification to include additive noise effects, yields
where p(r|is0) is the p.d.f. of r, given Eo, and Pr(/|£"0) >s tne

P oc x E (5) characteristic function of p(r| EQ).

VOL. 107, PARTB. 2
34 VOELCKER: PHASE-SHIFT KEYING IN FADING CHANNELS
Because r is the s u m of two independent, r a n d o m variables, with the results of Pierce, who investigated ideal coherent detec-
tion of fading f.s.k. signals.14 Pierce found that when the
P r ( / | £ 0 ) = P*o.*i(/|£o)P,o,i(/) . . . (ID phases and amplitudes of f.s.k. signals can be predicted before
where, for'example, detection, the samples from each receiver should be weighted
by the expected amplitude, and the weighted samples should be
CO 00
added. The weighting operation prescribed by Pierce is implicit
?xOxi(f\Eo) = J J
p(xQ, x^o) exp (-iTTjfxoxJdXidxQ in differentially-coherent product detection.
— 00 — 00

. . . . (12)
P E ancl (6) ERROR PROBABILITIES IN DIVERSITY RECEPTION OF
If xOxi(f\ o) Vyoyi(f\E0) are evaluated by substituting the BINARY P.S.K. SIGNALS
appropriate normal p.d.f.'s, the following expression for eqn. (10)
is obtained: Error probabilities can now be derived for mth-order diversity
reception of binary p.s.k. signals. Because of symmetry in the
?r(f\E0, M) = decision rule and signal statistics, only the p.d.f.'s for M (or S,
but not both) transmissions are required to derive the various
expressions in this and following Sections.
Eqn. (13) gives the characteristic function of the p.d.f. of the
. . . . (13) output of one product detector. Since m product detectors
2 2
where c = cr + cr^l — </>) 2 operate on m independently fading paths, one may write by
a2 = Mean signal power of either quadrature signal inspection
component. PS(/|M, WJ =
(f> = Correlation coefficient of either quadrature signal
component over the pulse length. (This is [l+(277/can)2]"wexp|-
</>T of Section 3.)
. (18)
Inverse transformation of eqn. (13) and averaging over all values
of Eo, which is Rayleigh-distributed, yield where P5(/(M, Wm) = Characteristic function of p(
s = Combined output of the product
detectors.

Expressions for p(^|M, m), where m is the order of diversity,

r > 0
can be found by inverse transformation of eqn. (18) and by
where p(r|M) is the p.d.f. of r, given an M transmission. averaging over all values of Wm. Because the various Eo terms
fade
The expressions for p(r|5)
| are identical except for the algebraic independently,
p the p.d.f. of Wm (which is required for the
sign preceding <f> in the exponents. averaging) is14

(5) DIVERSITY COMBINING

Since the paper is concerned with phase-shift keying for long- 0 < Wnnm = 1 , 2 , 3 , . . . (19)
range radiocommunication, it is appropriate at this point to Thus, for example,
consider diversity reception. In particular, how should the
signals in an rath-order diversity receiver be processed? This
question can be answered by again examining a likelihood V(S [-J^-J] \M, 2) = exp
ratio. Suppose that m basic binary p.s.k. detectors operate
upon signals which fade independently and have independent To 4
additive noise. The kth product detector output is denoted (20)
by rk+) if it is positive, and by rk~} if it is negative. Suppose
further that some number, /, of the outputs are positive at a given Average error probabilities are found by integrating p(^|M, m)
time, and the remaining m — I are negative. The likelihood from —oo to 0. For the non-diversity and dual diversity cases,
ratio for this situation is the following results obtain:
_ p{\ ~<f>)
L(M\rx . . . rm) n P., = (21)
= k=x
(16)
• • r m )
np(/i
I
+>
|s) k=i+\
nki+ Pe T = (22)
20,
If a logarithmic measure of the likelihood ratio is denoted
by Pj$ s and the appropriate p.d.f.'s are substituted, it follows that where P€m = Probability of an error in mth-order diversity
reception of binary p.s.k. signals.
p = cr];l<J%, a mean signal/noise ratio.
k=l - s k=\
• (17)
The structure of the error probabilities shows that errors, are
Thus, the combination of the product detector outputs should induced by two sources—noise and propagation. When
be achieved by linear addition, and decisions should be deter- <>
/ ~ 1 and p is small, the error probabilities decrease with
mined by positive and negative excursions of the sum. increasing signal/noise ratio, as one might expect. Unlimited
It is interesting to note that this combining method agrees increases in p are useless, however, because at some point errors
 VOELCKER: PHASE-SHIFT KEYING IN FADING CHANNELS 35
induced by the transmission medium begin to exceed those both 1-bit and 2-bit errors. These expressions follow imme-
caused by additive noise. The irreducible error probabilities diately from the foregoing decision rules:
characteristic of binary p.s.k. transmission in Rayleigh-fading
media, which were predicted in Section 3, may be found by r00 rw , r* ,
taking the limits of eqns. (21) and (22), as p approaches infinity. Pe0,m = 2\ pOVm)\ p(s,\WJj o p(sq\Wm)dsqdSidlVm • (25)

(23) ,m = 2 PCJ,| Wm) \ pCj | Wm)dsqdsidWm • (26)

«,2|p-c
_ /* ~ 0 1 - 0) (24) ^2,m) (27)
] • • where Peo>m = Probability of a 'zero error', i.e. a correct
Diversity reception is obviously useful in reducing errors decision, in wth-order diversity reception of
caused by both noise and propagation. Although the analysis quaternary p.s.k. signals.
has been restricted to paths which fade independently, Pierce's
P€2,m = Probability of a 2-bit error.
work on correlated fading implies that the diversity gains will
Pe\,m — Probability of a 1-bit error.
still be substantial when moderate correlation exists between
the paths. 14 The various p.d.f.'s in eqns. (25) and (26) are either given
earlier [see eqn. (19)] or can be found readily by the charac-
(7) ASYMPTOTIC ERROR PROBABILITIES IN DIVERSITY teristic-function method outlined in Section 4. Eqns. (25) and
RECEPTION OF QUATERNARY P.S.K. SIGNALS (26) can be evaluated for any reasonable order of diversity by
The investigation has dealt thus far with noisy binary p.s.k. successive integrations by parts and by use of the identity
signals. The analysis of quaternary p.s.k. transmission in this
Section is limited to noiseless signals; the error probabilities exp ( —) exp ( — —) dtdr = - arc tan ( —
which are derived are thus attributable solely to the fading n \ 2 / Jn \ 2 / c \c
medium. While solutions including noise effects can probably
be found, the asymptotic values given here are easy to obtain For the non-diversity and dual-diversity cases, the following
and are of considerable practical interest. expressions obtain for the noiseless but fading channel:
The basic receiver is again assumed to be a time-synchronous
differentially-coherent device. The dual product detectors A i i = I ~ 20(2 - < £ 2 r 1 / 2 " arc tan [<f>(2 - <f>2)~ll2] (28)
required for quaternary p.s.k. detection are shown symbolically / 77
in Fig. 4, together with vector diagrams of the signals. Diversity
^ - l~ a ^ tan (29)

Pe\ 9—

arc tan (3o)

V x oyi

2
2-<f>

Fig. 4.—Quaternary phase-keying in a fading, noiseless channel. (2 - _ i . j (31)

(a) The reference and signal vectors.
(b) Symbolic representation of the quaternary p.s.k. detector.
Thus, propagation-induced errors occur in quaternary p.s.k.
combining is effected by linear addition of the product detector signals whenever <f> < 1, even though the signal/noise ratio is
outputs. Because two bits of information are conveyed in infinite. As in the binary case, diversity is useful in minimizing
quaternary p.s.k. signals by transmitted phase shifts of 0°, 180° these errors.
or ±90°, the following rules govern the decision process when
the four phase states are equally probable: (8) THE EFFECT OF A SPECULAR COMPONENT
(\s,\ > \sg\, s, > 0 ) ^ 0 ° All the preceding results have been predicted upon Ray-
(\si\ > \sq\, Si < 0) - * 180° leigh fading. This implies that the received signal contains no
steady or specular component whose phase is constant (except
(\Sq\ > \Si\, Sq > 0) ->90° for transmitted shifts and additive noise effects) compared to a
(\sq\ > \Sj\, sq<0)-+ -90° stable local reference oscillator. It is known from experience,
where the symbol -» means 'implies'. however, that fading signals frequently do not exhibit ideal
Sj = Combined output of the / (in-phase) product detectors. Rayleigh statistics, and sometimes follow a Rice distribu-
sq = Combined output of the q (quadrature) product tion. 5 ' 7 - 10 This situation will be investigated for non-diversity
detectors. binary p.s.k. signals in a noiseless channel.
Fig. 5 shows a specular component, ii^, as well as the fluctua-
In contrast to the binary case, errors need not result in a total ting components Ex and Ey used previously; note that the refer-
loss of information; therefore, expressions must be found for ence axes are now fixed by Es. By slightly modifying the
36 VOELCKER: PHASE-SHIFT KEYING IN FADING CHANNELS
than about 30 dB, measured in a 3kc/s bandwidth containing
12 channels), while the f.s.k. system showed a superiority at
high ratios which increased steadily with the ratio. The
1 (x,. y, )
results of the preceding Sections, coupled with existing
theoretical analyses of f.s.k. systems, e.g. Reference 14, seem
to explain this phenomenon as follows. At low signal/noise
ratios, noise was apparently the main cause of errors and
the smaller noise bandwidth of the p.s.k. system gave it
a relative advantage. At high signal/noise ratios, however, the
(C) p.s.k. system apparently had reached the error-rate 'floor' set
by the phase stability of the medium; thus, only the f.s.k. system
Fig. 5.—Binary phase-keying in a noiseless channel with fading profited from further increases in signal power. Theoretical
superposed on a specular component. curves were fitted to the comparative-performance data to obtain
(a) The reference vector. crude estimates of the <>/ parameter.
(b) The signal vector.
(c) The product detector. The comparative results discussed above were presented in the
characteristic-function method outlined earlier, the following form of variations of the error-rate ratio of the two systems with
p.d.f. can be derived for the output of the product detector: signal/noise ratio. (It should be noted that this type of presenta-
tion requires very careful data processing if the results are to lie
- 0) + 2r , r < 0 (32) within a reasonable confidence interval.) Other results were the
individual performances of the two systems. These were less
p(r \M) = —, 2a](l - reassuring from a theoretical viewpoint, for the experimental
The average error probability is found by integrating eqn. (32) data did not fit theoretical curves of error-rate versus signal/noise
from — co to 0, giving ratio.
Two major discrepancies were noted. First, the slopes of the
-3) • • • «> experimental curves were considerably smaller than expected;
about 20-25 dB of signal power was required, on the average,
The exponent of eqn. (33) might be termed a 'propagation' to reduce the error rates by a factor of ten. Second, both
signal/noise ratio, for it involves the power in the specular systems seemed to approach an irreducible error rate at very
component and the mean fluctuation power. When Es = 0, high signal/noise ratios, though this asymptotic behaviour was
eqn. (33) reduces to eqn. (23). When the fluctuation power is more marked in the p.s.k. system. This author believes that
zero, the error probability vanishes in the noiseless case. these discrepancies between theory and experiment in the indi-
Eqn. (33) indicates that a strong specular component can effect vidual performances are directly attributable to the assumptions
a marked reduction in the errors due to propagation turbulence. underlying the theoretical analyses. In particular, the descrip-
Unfortunately, values of EJas significantly greater than unity tion of additive disturbances in the long-range radio environ-
occur but infrequently on most long-range radio circuits. ment in terms of stationary, normally distributed white noise
It is worth noting that the model shown in Fig. 5 applies to seems unrealistic. The assumption of white noise has, however,
the case of a non-fading signal in additive noise if cf> = 0 and two advantages: it is mathematically tractable, and one can be
of is interpreted as cr%, the mean noise power. With these modi- assured that white noise is always present, although it may not
fications, eqn. (33) reduces to be the dominant form of additive disturbance.
Why did theoretical analyses based on a seemingly unrealistic
I non-fading (34) assumption succeed in predicting the comparative performance
of the two systems, and yet fail to predict the individual per-
This result was derived by Lawton in his study of binary p.s.k. formance of either system? The author suggests that the answer
and other digital modulation techniques under non-fading and lies in the nature of the systems. The characteristic of phase-
additive noise conditions.15 shift keying which makes it attractive is bandwidth conservation;
thus, one might expect a p.s.k. system to combat most forms of
wide-band additive disturbance more efficiently than an f.s.k.
(9) FIELD-TEST RESULTS system. If one views impulse noise and many other types of
The 1957 field test mentioned in the Introduction was con- transient phenomena simply as wide-band additive disturbances,
ducted over a 5000-mile high-frequency radio path from the it follows that analyses based on white noise plus fading should
Hawaiian Islands to Fort Monmouth, New Jersey, U.S.A. predict f.s.k.-p.s.k. comparative performance quite accurately,
The equipments involved were a 16-channel f.s.k. terminal and a although the same analyses may be useless for predicting the
12-channel quaternary p.s.k. terminal. The f.s.k. system was individual performance of either system.
of the asynchronous narrow-band limiter-discriminator type
with 'ratio squared' diversity combining, while the p.s.k. system
employed synchronous differentially-coherent detection with (10) AVERAGED-PHASE REFERENCES
additive diversity combining. Single-sideband radio equipment It was noted in Section 2 that many methods have been pro-
and dual space-diversity reception were used throughout the test. posed for transmitting phase references; therefore the restriction
About 1-4 X 107 words of teleprinter copy were received, and to differentially-coherent detection (phase comparison with the
the error statistics were subsequently analysed on a digital com- preceding pulse) which obtains throughout the theoretical
puter. Plots of many of the results and a comprehensive analysis may seem arbitrary. One can envisage, for example,
description of the test are published elsewhere.l p.s.k. systems which ostensibly derive 'clean' phase references by
A striking result which emerged from the field test was the averaging several of the preceding signal pulses. The statistical
comparative performance of the two systems as a function of procedures used previously can be extended to analyse phase-
signal/noise ratio. As noted earlier, the p.s.k. system demon- averaging schemes, and, in fact, an optimum 'memory' function
strated a fairly constant superiority at relatively low ratios (less can be found as a function of cf>(r) and the signal/noise ratio
 VOELCKER: PHASE-SHIFT KEYING IN FADING CHANNELS 37
which minimizes the error probability. As one might expect, are applicable to the 1957 field test, wherein character (non-
the mathematics shows that a large number of data pulses should redundant multi-digit group) error statistics were collected, only
be used to establish a phase reference when additive noise is the on the rather questionable assumption that the correlation
preponderant source of errors, but that only the preceding pulse coefficient over the character length was small.
should be used when errors arise mainly from propagation
fluctuations.
Although a perfect phase reference can never be established (12) COMMENTS
for the detection of signals transmitted through a fading The preceding analysis shows theoretically that p.s.k. trans-
medium, it is worth while to determine the maximum improve- mission through fading channels is subject to unavoidable
ment which a perfect reference would afford. It is well known degradation, and that this degradation cannot be overcome by
that the error probability for ideal coherent detection of a non- radiating more power. Many questions remain unanswered,
fading binary-phase-keyed signal in white noise is however. The analysis tells nothing about frequency-selective
propagation effects or sequential error statistics. One might
Efcn
even question whether the fluctuations in received phase-keyed
1
P
i )ideal = i — V27T exp(--)dt signals can be described by stationary, ergodic statistical models;
perhaps the underlying physical processes are non-stationary
even over short time intervals. Nevertheless, from a practical
Since fading will affect only the signal amplitude in the ideal- viewpoint, the limited experimental data available imply that
reference case, the error probability for ideal non-diversity p.s.k. systems are indeed limited by the phase stability of radio
reception of a Rayleigh-fading signal is channels.
What can be done to optimize long-range radio systems using
p , p2 N f i , f2x phase-shift keying? From the comments in Section 3 on short-
(PfJldeal = i - | "^2"exP (^- 2^2^ —p^ e X P ( ~" 2) pulse transmissions, it is clear that p.s.k. systems should use the
shortest pulse length possible, subject to the limitations imposed
by 'coarse' multi-path time delays. Thus, the signalling speed
(35) should be a compromise which minimizes the sum of the errors
caused by phase fluctuations and multi-path delays. This means
Comparison of eqn. (35) with eqn. (21) under the necessary that the signalling speed should be adjustable to cope with a
condition that p(l — 0) < 1 shows that a differentially-coherent variety of propagation conditions, and this in turn requires
p.s.k. system requires about 3 dB more signal power than a p.s.k. terminal equipment which is not only flexible but which can also
system using an ideal phase reference in order to equal the per- monitor continuously the characteristics of the radio path. The
formance of the latter. Thus any physically realizable scheme design of channel-matching radio systems (i.e. systems which
for deriving a 'clean' phase reference can approach a 3dB adjust themselves to propagation conditions) is in its infancy;
improvement only under the most favourable conditions. one such system for highly redundant signals is described in
Furthermore, it is not difficult to show that a phase-averaging Reference 16.
scheme which has too long a 'memory' (compared to the phase However, there is reason to believe that redundant coding
stability of the medium) can degrade the performance of a p.s.k. may be more effective than channel matching for improving
system severely. p.s.k. performance, although the development of simple codes
These comments on phase-averaging strictly apply only to for phase-shift keying poses some unusual problems. For
Rayleigh-fading signals. If a strong specular component is example, the very nature of the differentially-coherent detection
present, phase-averaging is obviously more feasible. Lawton process promotes error doublets. This characteristic means
has shown, however, that in the limiting case of no fading the that simple codes, such as single parity checks and constant-ratio
advantage offered by ideal coherent detection is only in the groups, must be interlaced if they are to be effective. Moreover,
order of ldB. 1 5 it was noted in Section 11 that p.s.k. errors are bound to occur
in clusters, because propagation fluctuations must occur slowly,
compared with the signalling speed, for phase-shift keying to be
(11) SEQUENTIAL ERRORS feasible; this further complicates the design of simple codes,
It is important to understand that the preceding analysis deals although it is advantageous from the viewpoint of pure
only with independent errors. The error probabilities derived communication theory.
earlier can be applied in two contexts. In the first, consider a It is quite possible that phase-averaging techniques may be
multiplicity of p.s.k. receiving systems (where a system may be unexpectedly useful in both channel-matching and redundantly
an mth-order diversity receiver) processing information from coded p.s.k. systems. This follows because phase-averaging
statistically identical but independently variable paths. The over several data pulses can provide not only a measure of the
error probabilities are then a property of all sets composed of channel stability, but also a means of controlling error-clustering
one decision selected at random from each system. In the to some extent. Thus it might actually be advantageous to use
second context, consider a single receiving system emitting an averaged-phase reference which (incidentally) increases the
sequential decisions spaced T seconds apart. The error proba- probability of independent errors in order to enhance the
bilities are then a property of all sets composed of decisions efficiency of a particular redundant code.
spaced Kr seconds apart, where <\>{KT) ~ 0. These restrictions In summary, the author believes that p.s.k. systems offer at
are necessary because error-clustering has been neglected. That present a means of conserving bandwidth, but that they do not
errors will indeed tend to occur in groups or clusters is intuitively necessarily offer greater reliability for long-range radio service
evident from the high degree of coherence required for accept- than more conventional systems. Because the requirements for
able phase-shift keying. The transition probabilities which both accuracy and capacity in long-range data systems are likely
describe error clusters are very important in coding studies, to increase markedly in the foreseeable future, efforts should be
but their mathematics are less tractable in the p.s.k. case than directed toward the improvement of p.s.k. systems. In the
in the f.s.k. case. The error probabilities derived in the paper author's opinion, worth-while improvements can be effected by
38 VOELCKER: PHASE-SHIFT KEYING IN FADING CHANNELS
adjusting signalling speeds and detection mechanisms so as to Medium and High Frequencies', Proceedings I.E.E.t
match the channel, and by efficient redundant coding. The two Paper No. 891 R, November, 1949 (96, Part III, p. 517).
types of technique are related, and both require more knowledge (6) WOODWARD, P. M.: 'Probability and Information Theory,
about the statistics of actual radio channels than is now available. with Applications to Radar' (Pergamon Press, 1953).
However, some of the results reported here—eqns. (23) and (33), (7) RICE, S. O.: 'Mathematical Analysis of Random Noise',
for example—suggest that certain relevant propagation charac- Bell System Technical Journal, 1945, 24, p. 45.
teristics can be measured quite easily with a simple p.s.k. receiver (8) BRAMLEY, E. N.: 'Diversity Effects in Spaced-Aerial Recep-
and an unmodulated, stable, high-power transmitter. tion of Ionospheric Waves', Proceedings I.E.E., Paper
No. 1062 R, January, 1951 (98, Part III, p. 19).
(13) ACKNOWLEDGMENTS (9) PRICE, R.: 'The Autocorrelogram of a Complete Carrier
Wave Received over the Ionosphere at Oblique Inci-
Much of the work reported here was done while the author dence', Proceedings of the Institute of Radio Engineers,
was a Project Officer at the U.S. Army Signal Research and 1957, 45, p. 879.
Development Laboratory, Fort Monmouth, New Jersey, (10) BRENNAN, D. G., and PHILLIPS, M. L.: 'Phase and Ampli-
U.S.A. The author particularly acknowledges the encourage- tude Variability in Medium-Frequency Ionospheric
ment afforded by Messrs. R. A. Kulinyi, B. Goldberg, and H. E. Transmissions', Technical Report No. 93, The Lincoln
White of that Laboratory, and thanks the U.S. Department of Laboratory, Lexington, Massachusetts, U.S.A., 1957.
Defense for permission to publish the work. Professors E. C. (11) HAGFORS, T., and LANDMARK, B.: 'Simultaneous Variation
Cherry and G. A. Barnard of the Imperial College of Science of Amplitude and Phase of Gaussian Noise, with Appli-
and Technology provided valuable insight into some of the cations to Ionospheric Forward Scatter Signals', Pro-
statistical concepts used in the presentation. ceedings I.E.E., Paper No. 2743 R, November, 1958
(105 B, p. 555).
(14) REFERENCES (12) FISHER, R. A.: 'Statistical Methods and Scientific Infer-
(1) BRENNAN, A. T., GOLDBERG, B., and ECKSTEIN, A.: 'Com- ence' (Oliver and Boyd, 1956).
parison of Multi-Channel Radioteletype Systems Over a (13) DOELZ, M. L., HEALD, E. T., and MARTIN, D. L.: 'Binary
5000-Mile Ionospheric Path', National Convention Data Transmission Techniques for Linear Systems',
Record of the Institute of Radio Engineers, 1958, 8, Proceedings of the Institute of Radio Engineers, 1957, 45,
p. 254. p. 656.
(2) COSTAS, J. P.: 'Phase-Shift Radio Teletype', Proceedings of (14)
PIERCE, J. N.: 'Theoretical Diversity Improvement in
the Institute of Radio Engineers, 1957, 45, p. 16. Frequency-Shift Keying', ibid., 1958, 46, p. 903.
(3) RATCLIFFE, J. A.: 'Diffraction from the Ionosphere and the (15)LAWTON, J. G.: 'Comparison of Binary Data Transmission
Fading of Radio Waves', Nature, 1948,162, p. 4105. Systems', Proceedings of The Second National Conference
(4) BOOKER, H. G., RATCLIFFE, J. A., and SHINN, D. H.: on Military Electronics, 1958, p. 54; see also Proceedings
'Diffraction from an Irregular Screen, with Applications of the Institute of Radio Engineers, 1959, 47, p. 333.
to Ionospheric Problems', Philosophical Transactions of (16) PRICE, R., and GREEN, P. E., JR.: 'A Communication Tech-
the Royal Society {A), 1950, 262, p. 579. nique for Multipath Channels', Proceedings of the
(5) MCNICOL, R. W. E.: 'The Fading of Radio Waves of Institute of Radio Engineers, 1958, 46, p. 555.