Linear Diversity Combining Techniques

D. G. BRENNAN
Classic Paper

This paper provides analyses of three types of diversity combining systems in practical use. These are: selection diversity, maximal-ratio diversity, and equal-gain diversity systems. Quantitative
measures of the relative performance (under realistic conditions)
of the three systems are provided. The effects of various departures
from ideal conditions, such as non-Rayleigh fading and partially
coherent signal or noise voltages, are considered. Some discussion
is also included of the relative merits of predetection and postdetection combining and of the problems in determining and using
long-term distributions. The principal results are given in graphs
and tables, useful in system design. It is seen that the simplest possible combiner, the equal-gain system, will generally yield performance essentially equivalent to the maximum obtainable from any
quasilinear system. The principal application of the results is to diversity communication systems and the discussion is set in that context, but many of the results are also applicable to certain radar and
navigation systems.

I. INTRODUCTION
When a steady-state, single-frequency radio wave is transmitted over a long path, the envelope amplitude of the received signal is observed to fluctuate in time. This phenomenon is known as fading, and its existence constitutes one
of the boundary conditions of radio system design. It is observed that if two or more radio channels are sufficiently separated in space, frequency, or time, and sometimes in polarization, then the fading on the various channels is more or
less independent; i.e., it is then relatively rare for all the channels to fade together. The standard techniques for reducing
the effect of fading—known as diversity techniques—make
use of this fact. The object of these techniques is to make
use of the several received signals to improve the realized
signal-to-noise ratio, or to improve some other performance
criterion.

Manuscript received April 21, 1958; revised January 14, 1959. The research reported in this paper was partly supported by the Army, Navy, and
Air Force under contract with the Massachusetts Institute of Technology.
This paper is reprinted from the PROCEEDINGS OF THE IRE, vol. 47, June
1959, pp. 1075–1102.
The author, deceased, was with the Lincoln Lab., Lexington, MA, and
Dept. of Mathematics, M.I.T., Cambridge, MA.
Digital Object Identifier 10.1109/JPROC.2002.808163

Several diversity combining and switching techniques are
known, and there have been numerous papers on this subject
in recent years. (A sample of these, with comments, is indicated in a Bibliography at the end; these papers will be referenced by numbers in square brackets, running footnotes by
superscript.) However, very few of these have provided quantitative comparative data on the relative performance of the
various techniques, especially the two significant techniques
(maximal-ratio and equal-gain) investigated since 1954. The
major exception to this is a paper by Altman and Sichak [8],
which is not widely known and even less understood.
Furthermore, there has been little attempt to explain the
fundamental concepts and principles involved. For such
reasons, therefore, it appeared desirable to provide an
expository treatment of a comparative analysis, within a
unified framework, of the three most promising diversity
techniques presently known. An earlier memorandum [17]
aimed at these objectives indicated that such a treatment
might be of fairly general interest.
Of course, in an undertaking of this kind, several previously published results are naturally included as individual
cases, though the available information will also be rounded
out in a number of ways. Specifically, this paper includes
the following material that the author has not seen published
elsewhere:
1) A careful statement of the idealized circumstances required for canonical performance of coherent combiners (Section II),
2) Simple expressions for the mean signal-to-noise
power ratios of various combiners [(18), (28), and
(44); Fig. 8; Table 1],
3) Probability distribution curves for equal-gain combiners for 3, 4, 6, and 8 channels (Figs. 10–13, Table 2)
4) Estimates of the relative performance of three standard
combiners for non-Rayleigh fading (Section VII),
5) A discussion of the relative performance of three standard combiners for correlated fading (Section VIII),
6) Estimates of the degradation of the average performance of equal-gain and maximal-ratio combiners
caused by correlated noise voltages (Section X),

0018-9219/03$17.00 © 2003 IEEE

PROCEEDINGS OF THE IEEE, VOL. 91, NO. 2, FEBRUARY 2003

331

7) A bound (due to Stein) on the degradation of coherent-type combiners with imperfectly coherent
signals (Section XII),
8) Certain aspects of the determination, meaning, and use
of long-term distributions (Section XIII).
In addition, some previously published material has been
simplified or otherwise clarified.
It should be mentioned that the criteria employed below
are expressed entirely in terms of SNR. This has sometimes
been taken to mean that the results were principally applicable to continuous signals, although they are also applicable
to certain binary or other discrete signals and can be translated into error rates once a suitable detection characteristic
is either theoretically or experimentally known. But in the
case of binary systems, it is possible to obtain more specific
and precise results on error rates for specific systems. Such
results have been extensively studied by Pierce [10], [15] and
others and are not considered below. Neither is there a discussion of the considerable benefits obtainable by coding or
other signal preprocessing techniques designed to counteract
fading, several of which are currently under investigation by
other workers.
On the other hand, it should be noted that radar and navigation systems in which a repetitive-addition signal-enhancement technique is employed are closely similar, in some respects, to certain diversity systems. Although radar and navigation systems are not discussed in detail below, many of the
results and discussions set forth there are directly applicable
to such systems.
II. BASIC ASSUMPTIONS AND OTHER PRELIMINARIES
The principal background required of the reader is a basic
acquaintance with certain elementary notions of probability
and statistics, essentially equivalent to those developed in
the first six pages of a highly readable tutorial paper given
by Bennett. [18] No advanced techniques are required here.
However, we shall make frequent use of a few ideas and techniques that were not particularly emphasized by Bennett, and
a brief exposition of these is given in Appendix I. All probability distributions used in this paper will be interpreted as
explained there.
We shall be concerned throughout with random variables
given as functions of time (waveforms) in various intervals.
In this setting, time and distribution averages are one and the
or or
for such averages will
same thing so or
be written interchangeably, but it is important to note at the
outset that our averages will refer to intervals of different
durations. Specifically, intervals of three different durations
will be considered: 1) Short intervals, whose duration will
be denoted by . The requirement for is that it must be
short in comparison to the time required for the fading amplitude to change appreciably, but long in comparison to the
period of the lowest frequency of interest in the signal. Specific representative values of would range from a few microseconds to a few milliseconds. 2) Intermediate intervals,
whose duration will be denoted by . The requirements
must satisfy are rather complicated and will be explained at
332

various points below. Specific suitable values of
would
range from a few minutes to a few hours. 3) Long intervals,
would
whose duration will be denoted by . Values of
range from one month to one year or more.
These intervals will be employed as follows. The short intervals of length will be used to form “local” statistics. For
is the instantaneous signal voltage
example, suppose
is the instantaneous noise voltage on some circuit.
and
Then
(1)
and
(2)
would be the local rms signal and local rms noise, re-specand would be the local mean-square signal
tively, and
and noise voltages. Letting denote the circuit resistance,
would be the local average signal power at time , obover the last seconds to find
.
tained by averaging
This averaging could be performed, for example, by feeding
into a suitable linear filter. Alternatively, one could de] and
termine the distribution of in the interval [
as the second moment of the distribution, though
obtain
distributions in intervals of length will not actually be of
concern here.
Local statistics such as (1) and (2) will generally fluctuate
in time because of fading and other effects. For example, the
and the local signal-to-noise power
local rms SNR
ratio
(3)
will usually vary over wide limits, though they will be much
better behaved than the (meaningless) instantaneous ratio
. The behavior of variables such as the local
, will
statistic (3) in intervals of length , where
be studied. In particular, various distributions and averages
will be considered. Such
relative to intervals of length
-distributions and -averages will also change with time,
in ways discussed in Sections VII and XIII. Performance
relative to -intervals under standard conditions is summarized in Section VI.
Finally, the variability of certain -averages will be con. This is
sidered in intervals of length , where
done in Section XIII. It is usually assumed in system design that, for suitable values of , all distributions for the
system in question will be essentially the same in every corresponding interval of length . (A suitable value might be
one year, for example.) This is in marked contrast to the situation for -distributions. However, it is found experimentally that this assumption is a reasonable first approximation;
moreover, if this assumption were not satisfied, there would
be no method available for predicting the performance of the
system, at least at the present time.
PROCEEDINGS OF THE IEEE, VOL. 91, NO. 2, FEBRUARY 2003

in view of the fact that the will have fluctuating local statistics. such copies More generally. . The simplest of these is that in which a single transmitting antenna furnishes a signal to several well-separated remote receiving antennas. that is. Certain navigation systems in which a periodic signal is transmitted also use this storage-and-addition principle. Another method. as indicated in Fig. Thus. the cases to be considered will be tics of the those in which the are locally constant. it is possible to keep the relation between theory and experiment clearly visible. If is written for the output of the storage device and for the signal currently being received. certain radar systems operate by storing the signal received during one scan and adding this to the signal received during the next scan. which is proportional to the channel gain and may be allowed to vary in accordance with the fluctuating local statis. there are circumstances in which it is useful and has actually been used. with higher local SNR . Since the may be allowed to vary. This method is economical in terms of antennas and real estate. in some sense.) In one form or another. the fading on the various channels is approximately independent. it is not true of radar and navigation systems. so far as communication systems are concerned. in the form of a resultant message component ( ). By combining these two methods. bidirectional. 333 . although this terminology has not been much used in the radar field. its chief disadvantage is equipment complexity. it should be added that recent experimental evidence indicates that the fading on the crossed pair of paths is more highly correlated than on the other pairs of paths. The transmitters are on different frequencies. the composite and similarly signal may be written (4) ) i. the individual . In general. one may be negative part of the time the other is positive. in particular. may have a than either . Now. this method is called space diversity. but expensive in terms of transmitters and required bandwidth.. then the composite . (However. Let us first consider briefly methods of obtaining several suitable . each of the form and one may form the sum (5) which may outperform. may consist signal is simply and an undesired adof a desired message component . defined as in (1)–(3). 1. The sum (4) may then be a better signal or alone. The circles marked D denote diplexing filters. The adjective “linear” in the title of this paper BRENNAN: LINEAR DIVERSITY COMBINING TECHNIQUES Fig. 1 Four-channel bidirectional space diversity system suitable for UHF and SHF systems. A variant of this. it will be convenient to consider weighted sums of the . as the method discussed in the opening paragraph of this section is essentially time diversity. If these are sufficiently separated. full duplex space diversity system that requires only two reflectors at each end. may be simply to add them together. than either or alone. involves transmitting the same information on two or more carrier frequencies.e. and several known methods of combining these to obtain an improved signal. It has been discussed more often than used. (However. including. This procedure is therefore vital to a complete and realistic analysis of communication systems in general and diversity systems in particular. the term “diversity system” refers to a system in which one has available two or more closely similar copies of some desired signal. ditive noise component . Hence. and a single receiving reflector with two feed horns or dipoles to separate the vertical and horizontal received signals. space diversity has been the most commonly used form of diversity communication. In radio communication systems. the general linear combination will be considered: (6) in which each is weighted by a combining coefficient . However. the noise components quite dissimilar. If the message plus a resultant noise component ( components and are closely similar. For example. and may be On the other hand. in particular. so they may partially cancel for part of the time. called frequency diversity. stems from (6). however. one should perhaps speak of (6) as locally linear or “quasilinear. However. one may have . or at least approximately so. their sum will simply approximate an enlarged copy of either or . in which In diversity communication systems. depending on the . time diversity involves transmitting the same information two or more distinct times. one of which transmits vertically polarized radiation and the other of which transmits horizontally polarized radiation. and vice versa. Altman and Sichak [8] obtained a fourth-order. Signal paths are indicated for one direction only. uses two separated transmitting antennas. there are several known methods of obtaining two or more signals . the practicable experiments required to verify theoretical predictions. one and way of using two similar or suitably related copies.) This is also true of the method called time diversity. as in the case of space diversity. so that . When this is instrumented for automatic operation. all of the latter methods in current practical use are special cases of (6). .By concentrating on system behavior relative to such prescribed lengths of intervals.” Evidently (4) is simply the case of (6) . suitable for use in systems operating at UHF and above. However.

a very sophisticated communication system. Furthermore. However. since . in which we observe the an interval of duration as new random variables. 2. are locally coherent. or. See especially Fig. Note that this assumption automatically implies that at least two intervals are considered: first. constant. at any given time. [27] These conditions are as follows: assume that simultaneous functions. and then indicate the way in which the results are modified by other circumstances or. We shall find it most economical to consider first a particular class of circumstances.. but long in comparison to the . as used by many commercial CW stations. as will be seen. where the are positive real numbers that change with time because of fading. assume that the variations of do not change appreciably within any period of duration . p. (7) C) D) is simply the local rms value of so that . sents the corrupted signal in the th channel containing the . the fading of the two components is far from independent if only a single transmission path is involved. and may be regarded as a form of time diversity in which the diversity is provided by the transmission medium itself. An even more important difference arises in the case of FM or other bandwidth-exchange systems. but at a rate that is very slow in comparison to the instantaneous . is actually a primitive but useful form of time diversity. in tropospheric transmission at UHF and above. At the other extreme.e. p. Once the method of providing a multiplicity of signals is decided. currently under development. where the duration of the averages is the same as in (7). and additive: and are the signal and noise components. uncorrelated) and have zero means: if .e. in the th channel. indicate where such modifications are treated elsewhere in the literature. the period [of (7)] involved in the definition of the . Evidently . The local rms values of the signals. [20]. We shall also assume that the are slowly varying. The balance of this paper is principally devoted to this problem. where predetection combining can lead to substantial improvement over post-detection methods.. each . B) The signals . Then. [22. Instrumentation problems as such are 334 not considered here. For example. A method that will sometimes yield two approximately independent fading signals is called polarization diversity. such combining methods require special phase-control provisions when used in predetection applications. only one of the . it is clear that the simple addition scheme (4) can fail grievously if and are not in the same the message components phase. The diversity combining techniques employed subsequent to this stage may be classed in two groups: predetection combining methods and postdetection combining methods. 20. Whichever of these methods is used. 1331] and very little effect of this type takes place. In normal ionospheric transmission at frequencies of a few megacycles. 20. [19] which is designed to eliminate effects due to multiple transmission paths between fixed antennas. In those methods in which. local mean square noises sometimes. [23] A somewhat similar approach at UHF and above is currently under investigation by several workers. or some other steady test signal with a constant local . taken over the last seconds before the present time. See especially Fig. it is found that the received signal includes both vertically and horizontally polarized components. . the basic problem confronting the designer of a diversity system becomes one of choosing the most appropriate method of combining these signals on the basis of reasonably accurate quantitative estimates of the performance of the various techniques. The circumstances initially considered are those often applicable to postdetection combining in an AM system. and RF or IF diversity signals will not usually be in the same phase unless special measures are taken to insure this. 91. and second.however. and the fading of these components is approximately independent. where is the duration of the interval employed for the local averages. . [24]–[26] but the efficacy of this technique is not yet firmly established. i. More precisely. 1331] Another method that has been used (infrequently) in the high-frequency region involves the combination of signals arriving with different angles of arrival (the Musa system). Consequently.. a switch of some kind. in (6) is different from zero. or a single-sideband system in which provision is made for maintaining coherence of the postdetection signals. important differences arise when the combining method is one in which two or more of the may be different from zero at the same time. period of are locally incoherent The noise components (i. NO. papers which describe certain instrumentation techniques are indicated. the signals obtained will initially be at radio frequency. even if both horizontal and vertical components are transmitted and separately received. VOL. For convenience. the polarization of the transmitted signal is quite well preserved [22.e. respectively. however. are statistically independent. while this is not always the case in postdetection applications. the distinction is basically unimportant. [21] However. actually sorts out the various multipath contributions and recombines them with suitable delays. It is clear that must be short in comparison to the time required for the fading amplitude to change appreciably. PROCEEDINGS OF THE IEEE. That the following conditions are apmean square proximately satisfied is also assumed: A) The noise in each channel is independent of the where signal. FEBRUARY 2003 . suppose originally transmitted signal is a steady test tone at a representative midband frethat quency. the simple practice of sending each word twice. represent the signals received in different diversity channels as correprerupted by noise and fading. i. in some cases.

The choice of scale. simply It will frequently be assumed that the variables follow a Rayleigh distribution with density and distribution functions (9) respectively. over periods of length . 2 while the and are shown separately.. This saves considerable cluttering of will often be taken the landscape below. which is approximately constant over such periods. the local amplitude ratio. 2 for “exploded. and could not be successfully illustrated there because the period BRENNAN: LINEAR DIVERSITY COMBINING TECHNIQUES required for approximate independence of the is much were greater than the total scale length of Fig. 2. might be a few milliseconds. 2. the requirements on cannot be too long. it will simplify the appearance of various expressions if the context is relied upon to make clear whether a short-term or intermediate-term average or distribution is meant. until Section XIII.) There are four principal types of diversity combining systems in practical use. Many of the combiners in actual use are not pure examples of one of these types. A discussion of and is provided in Section XIII. Writing the Rayleigh distribution in the form (9) implies a particular . By are illustrated in exploded fashion in Fig. for example. in which case the local amplitude ratio is numerically and . However. 2. 2 Signals and noises in two diversity channels. in practical cases. approximately true of the waveforms Assumption (D) is not particularly illustrated in Fig.Fig. . . however. e. 15. and expressions such as (9) are to be understood as referring to positive values only. they involve approximations to.e. say (10) . and from here on the word “ratio” is to mean “SNR. If the or IF signals. where the this “locally” refers to intervals of length . are not locally coherent in the sense of (B). i. Note that values implies that the have the assumption are RF the same zero crossings. it will often be assumed that over every interval of length . 2. it will often be required that the noises be essentially different. Such averfrom. one of these types. The meaning of the locally coherent assumption (B) is that. independent of and . Most of the work below is concerned with signal-noiseratios. In suggested by the waveforms (if ) particular.g. In particular.” This will be qualified as an amplitude ratio or a power ratio will be written for as the context requires. In contrast to the . and (11) (Distribution functions will always be written with upper case letters. except that the would be nonnegative and would not usually be symmetric about their mean values.” we mean that the actual signals given would be . A plot of the Rayleigh density function is given in Fig. it will have constant sometimes also be assumed that the noises local average power.e. 2. . at least roughly. is. and similarly. or modifications of. when . and are in phase. but ages could be distinguished by suitable notation. . However. e. this is the meaning of (C). We shall often take . AssumpIt is important that tion (D) is that. the signals and are essentially identical except in amplitude. The local rms are indicated by the dashed curves. in which case no variation of the would be perceptible within the scale of Fig. refer to intervals of length . If the represent base-band signals. there is no generally accepted 335 . which has the distrihas the simple distribution bution (9). it implies that Rayleigh distribution is often written with an arbitrary scale factor. in particular. density functions with lower case letters. the resulting and curves would resemble the waveforms illustrated for . the variables statistically independent. The circumstances characterized by assumptions (A)-(D) . (The terminology used here is not entirely standard—indeed. the local power ratio in the th channel. the data below are given in which case in a form that is completely independent of such scale factors. as and of Fig. then instead of is exactly the local amplitude ratio. that (8) is a constant. For. . i.g.. In addition. . 2. Similarly. when observed in intervals with a are duration on the order of . the period might be several microseconds or more. might be a few milliseconds approximately 30 minutes. This would be at least and of Fig. All distributions considered in this paper are zero for negative values. If the and the graph plotted throughout an interval of length were then compressed to the length of Fig. the effect of such modifications can often be estimated. Note that disor quantities derived theretributions or averages of the .

in terms of (6). VOL. Furthermore. and combiner diversity. the authors made it clear that they were thinking in terms of selection diversity. (This result has several times been quoted in the literature as requiring the weighting to be proportional to the amplitude ratio. Very often. NO. 91. at any time. The general arrangement of a two-channel maximal-ratio system suitable for postdetection combining is shown in Fig.) It is clear that (13) is a definite improvement over either scanning diversity or selection diversity. C. Maximal-Ratio Diversity This system is defined by the property that.. uses that signal only until it drops below threshold. the others do not then . i. Then a maximal-ratio system realizes standard terminology— but is the result of careful consideration and discussion with several colleagues.e. (B). the maximum output ratio (13) is realized if and only if the gain of each channel is proportional to the rms signal and inversely proportional to the mean square noise in that channel. however. 3. However. in (6) is different from zero.Fig. let denote the local power ratio of . the result (13) is equivalent to the requirement that the coefficients in (6) be proportional to (14) i. and then scans the other channels in the same fixed sequence until it again finds a signal above threshold. then this type a channel for which of system is characterized by the design criterion for for . [2] was actually a maximal-ratio system. which is why it is sometimes called antenna selection diversity. (15) PROCEEDINGS OF THE IEEE. A selector device scans the channels in a fixed sequence until finding a signal above a preset threshold. [2]. This will be proven below. FEBRUARY 2003 . More precisely.g. Thus. (It is quite possible that the diversity system discussed by Peterson. Radar systems of the type discussed in connection with (4) and which employ square-law detection are essentially maximal-ratio systems. et al. We shall not consider this type in the present paper. in which case it would be less misleading to speak of weighting proportional to the rms signal. . the maximum power ratio realizable from any linear combination (6) is equal to the sum of the individual power ratios. D. and uses that one alone. it is characterized by the property that all channels have exactly the same gain. A three-channel selection diversity system is depicted in Fig. optimum diversity. Selection Diversity This is also a switched technique. (12) in terms of (6). and closely similar results have been used in radar systems for some time.) Maximal-ratio diversity has sometimes been called ratio squarer diversity. but at least one of the following techniques will always outperform it. It is often applied to the case of two antennas supplying a single receiver through the switch.) The four “pure” techniques are as follows: A. This ) by Hausman [4] technique has been analyzed (for and most recently and most extensively by Henze [13]. (12) is often a good approximation to such cases. It should be noted that this is correct only in the case where the local noise powers are all equal. 4. among all systems of the type (6). and that only one of the one is equal to 1. it yields the maximum SNR of the output . 2. 3 Selection diversity. This is essentially the classical form of diversity communication [1]. at any given time. provided assumptions (A). . Equal-Gain Diversity This is probably the simplest possible linear diversity technique. a predetection combiner would require the addition of phase-control circuitry to satisfy assumption (B). confining our attention to the next three. The design criterion here is that. isfied. although the form stated here can be traced to [6]. with the same proportionality constant for all channels.. one of the terms in the sum This observation is essentially due to Kahn [5].e. and (C) are satsignal 336 (13) i. by using a common detector in such a way that the strongest signal cuts the others off) and is not quite as sharp as (12) would indicate. the selection in such systems is by electronic means (e. which can yield only as the output power ratio. Scanning Diversity This technique is of the switched type.. It does not require a separate receiver for each channel. whatever their actual instrumentation may have realized. B. 4 Maximal-ratio diversity.e. but of a more sophisticated sort.. Fig. let denote the index of contribute to . More precisely. the system simply picks out the best of the noisy signals . The f may be predetection or postdetection signals.

By (12). etc. it is possible to instrument unconventional FM detectors for postdetection equal-gain combining. and the local mean square noises BRENNAN: LINEAR DIVERSITY COMBINING TECHNIQUES The distribution function for an -channel selection diversity system is particularly simple to obtain. and an equal-gain system is decidedly different both in instrumentation and performance. (17) This integral is evaluated in Appendix VI. As in the case of Fig. equal-gain systems will outperform selection diversity. (16) is the distribution function of . However. Thus. A basic two-channel equal-gain system is illustrated in Fig. this is irrelevant to the performance of the system. conversely.e. for an -order selection diversity system. 6. and the way in which the results are altered by various modifications of the conditions as they occur in practice will be indicated. The important feature is that the channel gains are all equal. these values will be used below. III. adding the th channel increases by only . a common-detector combiner is essentially a selection diversity system. 5 and 4). which may include conversion and detection gain. the noisy signals are simply added together. 6 Equal-gain diversity. but not a common signal detector of the usual type. under the conditions discussed. independent AGC systems. In the following three sections. under conditions often occurring in practice. and will perform almost as well as maximal-ratio systems. In this case. [28] An arrangement suitable for use with AGC is shown in Fig. where it is shown to reduce to the remarkably simple form (18) . 4. Now. provided the are constant. . Hence. the channel gains can be made to vary in such a way that the resultant signal level is approximately constant. by (D). the application of an equal-gain combiner before detection would require the addition of phase control provisions to Figs. and. Since the are independent. must be approximately constant.. and assume that the are Rayleigh-distributed. it should be noted here that neither the functioning of a selection diversity system nor the statistics developed in this 337 . SELECTION DIVERSITY Fig. an equal-gain system may well use a common AGC detector. Thus. in consequence of (B) and (C). Then. Note: it is important to observe that this is not the case with conventional common-detector type diversity systems.Fig. the probability that all channels the is simply the product of the separate have power ratio probabilities that each channel individually has a power ratio . 5 and 6. see Figs. Let . in addition. Also. which. the output power ratio of tribution the combiner is simply the largest of the individual . it follows that equal-gain diversity should be more widely known than is presently the case. The average value of will be required for this system. In view of the simplicity of the instrumentation required for (15) as compared to (14) (equivalently. as will be seen. it may not perform as well as selection diversity. the channel is is precisely probability of having the power ratio of be the probability of having the individual channel ratios all simultaneously. 5 Basic equal-gain diversity. It has been pointed out by Sichak [8]. must be fixed. then the power ratio of if the largest power ratio is . the Rayleigh-distributed. It will be seen that the next two systems to be considered can perform much better in this respect. this fact is of great practical importance. as the detection gain of an FM receiver depends on the signal level. . they could not include separate. Since. Thus. however. these conditions are often (approximately) satisfied. the principal features of diversity combiners of types 2)-4) above will be developed. and mean values power ratio of the composite signal of . distribution functions will be obtained for the local . as is also the case with maximal-ratio systems [5]. The conditions required are that assumpmust be tions (A)-(D) must be satisfied. then so is the power ratio of . so are . it is most easily obtained from the distribution (16). however. local noise powers . In particular. Note that the blank boxes representing receivers must have the same gains. if the power ratio of every every channel is . they could not be conventional FM (or similar) receivers. The boxes “variable gain” must have the same gain. therefore. the realized local power ratio. including conversion and detection gains. [29] that. However. It is clear at once from (18) that increasing the number of channels in a selection diversity system yields rapidly diminishing returns. in Section VII – XII the results will be compared and evaluated. In applications of this technique. i. have the disThen the individual channel power ratios of (11). However. Under other conditions. 5. hence.

which is quite short. We now consider the statistical properties of the local power ratio .section depend on assumptions (B) or (C). side the average and since (23) Furthermore. In order to do this. is given in Ap. each with the distribution (11). it will be convenient to use a mathematical device known as the Schwarz inequality. [18] as indicated in Appendix III. On the other hand. that the are independent Rayleigh variables . can be written (27) (22) or the possible without regard to the distribution of the . which increases much less rapidly with than (28) does. This is not specifically related to statistics. dependence of these variables. without reference to characteristic functions. as discussed in Appendix I. then if . The first point to be noticed is that implies so that (21) Now in (6). but is a general result of great importance in many fields of pure and applied mathematics. 91. FEBRUARY 2003 . No use was made of distribution functions or any similar apparatus. NO. which proves that cannot exceed . However. (18) is approximately log N . . VOL. then is the local power (28) ) and by assumption (B) and since is locally constant and can be taken out. we shall employ the same assumptions used for the selection diversity case. If. Alternatively. namely. (25) 338 if and similarly if for any . only a more formal way of including assumption (B). and so that ratio of . let us write . pendix II.) The average value of the local power ratio of the output of db an -order maximal-ratio system is simply above a single channel. This is. Note that if then (19) takes the form (20) which. Thus we are interested in the of independent random distribution of the sum variables. it can easily be solved by using the geometric approach mentioned in Appendix I. IV. MAXIMAL-RATIO DIVERSITY The first order of business is to establish (13) and (14).] It is interesting to note that the only purely statistical fact used in this development is that averaging is a linear operation. In order to obtain an explicit distribution of . PROCEEDINGS OF THE IEEE. But (all sums from 1 to (26) This behavior is in marked contrast to the corresponding relationship (18) for selection diversity. however. so that the have the and that exponential distribution (11). it is important to observe that each of the assumptions (A)-(C) entered in a very vital way. . (24) by (61) and assumption (C). This problem can be treated by a simple application of characteristic functions. that the notation is a different function of in (18) than it is here. since . then numbers and (19) The proof of this. (One integrates the joint density function 1For large N . . The significance of this fact will be discussed. thus proving (13) and (14). Thus.1 (It should be clear is used in a somewhat flexible way. or even complex . using (21). [Readers acquainted with the Schwarz inequality for complex numbers will recognize that it may be used to include the case of positive or negative . 2. in particular. which are not required here. are any real One form of this states that if are any real numbers.

. (31) which can also be written from for an equal-gain system. is accordingly of interest. p) = 0 1 p! for an equal-gain system. From (34) and (35). provided assumptions (A)-(C) are satisfied. Some communication engineers express (34) by saying that coherent signals “add linearly”. p so that his p is here N 1. EQUAL-GAIN DIVERSITY Recalling that the relations of (23) and of (24) did not depend on a choice of the [i. each with the distribution (9). [8] (It also appeared independently in an unpublished memorandum [17] at about the same time. Pearson tabulates I (u.e. in general.) were subsequently Curves of (31) for several values of published by Staras [9]. Put Rayleigh-distributed. 2In his notation ([30]). See. while (35) similarly expresses the fact that the average power of a sum of uncorrelated signals is equal to the sum of the individual average powers. 7). In order to develop compara. we have (34) and (35) and. Since (37) It is clear that the distribution function of will follow immediately from that for . by (31) for arbitrary Altman and Sichak. this problem is not nearly as tractable as in the maximal-ratio case. Table 7. In the context of maximal-ratio diversity combiners. 3Short tables of the  distribution are given in many other statistical tables and in most textbooks on statistics.) It is easy to see that this is t e dt. The distribution of a sum of Rayleigh variables. The characteristic function of a Rayleigh variable is not expressible in an immediately useful form. it is again assumed that the .. the result. they hold for any combiner of the type (6). and hence hold in particular . Let denote the distribution function of (37). writing for the associated density function. however. p. That (31) is the distribution function of sums of squares of Rayleigh variables has been known in radar circles for quite some time. we have (32) (36) The utility of the form (32) is that it indicates the approximation (33) are accurate for sufficiently small . The distribution (31). Unfortunately. the result was first published in March. and his u is here p= N .] 3 The origin of the maximal-ratio distribution (31) has sometimes been incorrectly attributed (in Pierce [15] and Packard [14] among others). tive statistics for this. ]. for example [32] BRENNAN: LINEAR DIVERSITY COMBINING TECHNIQUES (38) 339 . We are here essentially forced to rely on the geometric approach mentioned in Appendix I. [31. (We can stay in the first quadrant because the density function is zero in all other quadrants. known as the gamma distribution. 1956. easily verified by an integra- By using tion by parts. we have for (30) and the recursion relation . K. The relation (34) is simply the well-known fact that the rms value of a sum of coherent signals is equal to the sum of the individual rms values. and that the are independent and . and the of the – plane bounded by the line coordinate axes (Fig.over the -dimensional volume bounded by the hyper-plane and the coordinate hyper-planes. it can be seen that is given by the integral over the region of the joint density function.) In for the desired distribueither case. is easily computed for the integral values of of interest here and has also been tabulated. tion function and is (29) V. this language is both formally meaningless and conducive of an imperfect understanding of the situation and is better replaced by “add coherently” if some such expression is necessary.2 It can also be identified with the chi-square distribution with 2 degrees of freedom. say . 122 ff. For .

FEBRUARY 2003 . The maximal-ratio points are values of (28). [32]. numerous workers—going back to Lord Rayleigh in terms of tabulated himself—have tried to express . 4. By completing the square in the exponent and making a few other routine manipulations. and the selection diversity values are from (18). these give the increase in decibels in the average local power ratio over a single channel.785 are relatively infrequent in observed fading distributions. 3. it is easy to see that the distribution function of the sum of independent Rayleigh variables is [see equation (41) at the bottom of the page]. it is nevertheless easy. The constant on the distribution of the . if . 6. . as in Fig. present paper (in Section VI) for An outline of the method of computation is sketched in Appendix IV. However. the equal-gain points are from (44). i. e. 3. [35] and curves of have been constructed from these tables for the . and so where (44) is the error function and is tabulated. See [34]. and 8. The . CANONICAL ONE-HOUR PERFORMANCE The three systems will first be compared simply on the basis of the average values of the local power ratio of the chanoutput. following . 2. (36) Appendix I. For the Rayleigh distribution. as was also the case for maximal-ratio systems. for from nels. For any distribution. is . this becomes [8] (39) . . 6. Although these averages depend on the 4The error function or. bution function of (40) and is readily plotted. 91. but with no success if has recently been tabulated. 4. depends the desired average value. is a dimensionless constant between 0 and 1. 7 for . [33] 4 Thus the distri.. which is simply the integral of the joint density function over the -dimensional volume bounded by the hyper-plane and the coordinate hyper-planes. 8. it is seen that there are terms. But the absolute maxslope imum by which (28) can exceed (44) is db.. but values of much less than 0. for that matter. for the We shall next obtain the average values equal-gain case. each equal to . (39) itself. Unfortunately. the integral (41) is quite as frightful as it appears.g. and this only in the limit of an infinite number of channels. Brief tables of the normal distribution function appear in most statistical texts. By considering the terms of the sum as the entries from an by matrix with the main diagonal such deleted. Corresponding to (38). so (43) Fig.distribution of and the distribution of is not given in a particularly explicit form. VOL. can also be expressed in terms of the much tabulated normal (Gaussian) probability distribution function. and 8 are presented from data of Fig. Since . functions. The only difference is that (28) increases with slope 1 while (44) increases with for Rayleigh fading. 8 for (41) 340 PROCEEDINGS OF THE IEEE. NO. This is done first in Fig. to obtain the becomes (42) Since the are independent. Since . It is thus seen that increases linearly with . VI.e. using the fact that . Let . 7 Region of integration for (38).

of of (40) (equal-gain) and (31) (maximal-ratio). 8. i. Table 1 Comparative Average SNR (Same Conditions as in Fig. the vs . such as coding.. respectively.Fig. The reason for using the median value of the Rayleigh distribution as a reference here is that this is commonly presented as an experimental datum. although they would get farther and farther away from the selection diversity curve and the base axis. The last entry in the equal-gain column is essentially the assertion that no matter how far the curves of Fig. since 341 . from which . A brief discussion of the significance of these data is in order. Let us next compare the probability distributions of realized by the three systems for different orders of diversity. in estimating relative average system capacities. the top two would never differ by more than 1.g. This approach is appropriate whenever high reliability is a primary requisite.e. 8 were continued. given criterion. for independently fading Rayleigh-distributed locally coherent signals in locally incoherent noises with constant local rms values. The Rayleigh fading curve is of (11). for example. in important military communication systems. a different point of view in Table 1. together case with the distribution of for a single channel with Rayleigh fading for comparison.05 db. i. 9 is expressed in decibels relative to this . (For the distributions considered here. a value for which variable for which for 50 per cent of the time and for 50 per cent of the time. This requires information about probability distributions. That curves of Fig. 9 Fig. . 8. the data of Table 1 may be more meaningful than results based on the distributions to be presented. Most recent diversity systems have been designed for a specified percentage of reliability. counting the zero axis as a curve. These results are useful. is the per cent of time ordinate exceeded. conditions of Fig. a specified percentage of time during which the system performance will exceed some BRENNAN: LINEAR DIVERSITY COMBINING TECHNIQUES Dual diversity distributions. 8. which gives the differences between the maximal-ratio values and the lower curves of Fig. 9.e. 8 Diversity improvement (in decibels) in average SNR. or in relay systems carrying commercial television programs. Thus.6 db. e. 9 refers to a value of a random . it should be pointed out here that some systems do not require very high local reliability or they may effectively achieve it by other means. The term “median” in the designation of the ordinate scale of Fig. where is. the median of the one-channel distribuand solving for tion (11) is obtained by setting . or in other circumstances where the average value alone of is of interest. In such circumstances. The ordinate scale of Fig.. 9 are plots of is. However. The (dual diversity) is illustrated in Fig. the median values of do not differ from the corresponding mean values by more than about 1. of (16) (selection).

342 Fig. 9 only if one were fighting for the last decibel. the equal-gain diversity distributions can be approximated quite well by translating the maximal-ratio distributions downward by the values in the second column of Table 1.” The term was apparently introduced by Jelonek. That is. etc. VOL. which must be satisfied for the equal-gain and maximal-ratio systems to work properly. 10 –13.99 per cent point. 9 –13 are applicable if the relevant conditions are satisfied. Suppose a high-reliability communication system is to be designed for a fixed information rate. For example. However. it is desired to maintain the local power ratio above a certain value for. as can be seen in Figs. the differences in the performance of the various combining techniques become more important as of channels is increased. 4.4 db above the selection curve at the 99.median values can be read directly from the distribution function determined by a totalizer. et al. 91. Similarly. NO. This offer may be withdrawn at any time. Referring to the 99 per cent exceeded values of Fig. while the maximal-ratio curve is only 1. it can be seen that the difference between the Rayleigh fading curve and the dual selection diversity curve is about 10 db at the 99 per cent point. and 8. Requests on postal cards or form letters will not be honored. 9 –13 are available from the author to those having serious need of such graphs. FEBRUARY 2003 .) However. 10 N = 3. part of the reduction could be applied to the antenna gains. Letters requesting the same should describe the nature of said need. indeed. Evidently one would choose among the three types of two-channel systems on the basis of Fig. which cannot be maintained whenever the received local power ratio drops below a certain value. 9 –13 are useful in the design of radio communication systems and radar and navigation systems of the type discussed in the Introduction. the difference between . the 99. Of course. say. 9 are quite small. 2. 99 per cent of the time during an interval of length for which the curves of Figs.5 give the distributions for (Note that the ordinate scales of Figs. 11 N = 4. Here PROCEEDINGS OF THE IEEE. The maximal-ratio the number and equal-gain systems improve much more rapidly than selection diversity does. as will be seen. that is. the selection diversity curve does not depend on the important assumptions (B) and (C). The data of Figs. 5High-resolution graphs of the curves of Figs. 9. [3].99 per cent exceeded level of the selection diversity curve is almost 20 db above the Rayleigh curve. One such application is as follows. a transmitter of 10 db less power would be adequate if dual selection diversity were employed at the receiving terminal. It can be seen that the maximal-ratio and equal-gain curves differ approximately by the constants in the equal-gain column of Table 1. This reduction in transmitter power required for a given grade of local reliability has been called “diversity gain. But this implies that whatever transmitter power was required for a single-channel system. 11 shows that the use of fourth-order maximal-ratio diversity would enable a reduction in transmitter power of 19 db relative to that required for a single-channel system. 9 –13 cover different ranges. which . the maximal-ratio and equal-gain curves remain quite close together. especially in comparison to the difference between any one of them and the Rayleigh fading curve. Even then. one would wish to make very sure that that last decibel could actually be realized.) It is clear that the differences between the various dual diversity curves of Fig. This is one of them is hardly significant even for the facts that makes equal-gain diversity quite attractive and suggests that there are many applications where it should be exploited. reference to the 99 per cent values of Fig. 6. respectively. Fig.

Table 2 Local Reliability Gains (in db). however. for 99 Per Cent and 99. 9 –13.e. The maximal-ratio curve of Fig. such values would not be meaningful. In order to discuss the effect on the distributions of Figs. The dependence of the local reliability gains on the percentage selected can be seen in Table 2.9 Per Cent Local Reliability .” is used to emphasize the fact that it is not a gain in the usual sense and that it depends very heavily on the local reliability percentage chosen. NON-RAYLEIGH FADING DISTRIBUTIONS Fig. Among other things. Similarly. other distributions are often observed. Now. of the quarter circle is obtained by inthe equal-gain distribution function tegrating the same density function over the triangle bounded . The probability that a . i. the fact that the maximal-ratio system outperforms the other two is intimately connected with the fact 343 . when observed in intervals of length . Various modifications and extensions of these considerations as they occur in practice will be considered next. 3.99 per cent or higher percentages. It is therefore of interest to discuss the way in which these results are modified by other distributions. 14. the term “local reliability diversity gain” or simply “local reliability gain. VII. and for which the results can be used without significant modifications. and 8 corresponding to local of systems for reliability percentages of 99 per cent and 99.9 per cent range.1 per cent to 99. considering the present or immediately foreseeable state of the art. For short-range UHF circuits and normal or scatter ionospheric transmission at VHF and below.. the maximal-ratio system has a local power ratio .9 per cent. 6. at frequencies of a few megacycles and below. Let and be the local case amplitude ratios of the two channels. The corresponding region for the by the line is the square bounded selection diversity distribution . 4. They can be made to appear even larger by computing the local reliability gains corresponding to 99. is obtained by inmaximal-ratio distribution function tegrating the joint density function of and over the interior in the – plane. It can be seen that the values of Table 2 are much larger than those BRENNAN: LINEAR DIVERSITY COMBINING TECHNIQUES Only in the case of long-range UHF and SHF tropospheric transmission does it appear that observed fading distributions are most often Rayleigh. Conditions of Fig. Fig. 8. an accurate fit to the Rayleigh distribution is more nearly the exception than the rule. However. These three regions are shown together by in Fig. the Rayleigh distribution does not provide an accurate model for actual fading distributions outside of the 0. Certain of these results are easily discussed. . it should be noted that there are many practical situations in which the conditions assumed above are realistic approximations. 8. Indeed. assuming that the other conditions still hold. 12 N = 6. of Fig. 8 and the last column of Table 1 do not in any way depend on the fading distribution and are not modified at all. 13 N = 8. which gives the values realized by the three types . it will be convenient to return to the geometric approach mentioned in Appendix I and consider the channels.

For purposes here. which are associated with shallow fading. VOL. Consequently. for ). i. (a) Rayleigh density function. 6 and even for this case. 91. which are associated with frequent or persistent deep fades. 14. so that their joint density function tributed. with a joint density function on the square . say . therefore.e. [18] say on and . one of a family of distributions given by Rice. its performance will more nearly approximate a maximal-ratio system. out” along the diagonal where it will contribute more to the integral over the selection region than to the integral over the equal-gain region. . 9 in the fact that the maximal-ratio curve is strictly above the others. or special applications. Returning now to Fig. 14. (b) Representative Rice distribution. that output ratio is and . illustrates the less disperse or shallow fading often encountered at frequencies below UHF. together with the Rayleigh distribution. the application of equal-gain diversity should be viewed with a modicum of caution in cases where very disperse fading distributions might be encountered. 9 would be above the others for any fading distribution. It is thus seen that the relative performance of selection diversity and equal-gain diversity depends to some extent on the fading distribution involved. two cases may be distinguished fading distributions more disperse (broader or more smeared-out) than the Rayleigh distribution. Of course. the exponential distribution used above is probably extreme in this respect. However. Hence. 14 Regions of integration for three types of dual diversity systems. the maximal-ratio curve of Fig. the equal-gain and selection systems yield identical performance. and distributions less disperse (narrower or shallower) than the Rayleigh distribution. as the relevant distribution functions are easily evaluated.. indeed. for any fixed . or selection regions fit inside this square ( i. NO. This can also be seen directly by considering the basic operation of a two-channel equal-gain system. 15 Representative fading distributions. for postdetection distributions in FM systems. Curve (b). It is then easy to see that for values of for which both the equal-gain and . 14. the probability that the maximal-ratio is smaller than it is for the others.. strictly less than it is for dual selection diversity. the equal-gain system is not signifi6So far as conventional applications are concerned. the local reliability gain of dual equal-gain diversity is. no matter what fading distribution is involved. Hence. It should be noted that it is not extreme. the joint density function will still be nonnegative and. this result could also be seen from the fact that the maximal-ratio system yields an output power ratio that is indeed maximal. Noting that the contours of constant height of this is constant. after Altman and Sichak [8]. and visualizing the case where the two signals are approximately constant. this is not true for the more disperse distributions.Fig. 2. Next. will still be smaller than for the others. This would lead one to suspect that the relative performance of selection diversity and equal-gain diversity depends on the form of the fading distribution. FEBRUARY 2003 . However.e.and medium-frequency systems. Thus. in Fig. the region of integration for the maximal-ratio system is smaller than it is for the others and interior to both of the others. suppose the independent amplitude ratios are exponentially disand . and the bulk of the density function will be “pushed 344 Fig. it is easy to see that the integral of this density function over the equal-gain region is strictly larger than it is over the selection region. It can be seen there that the areas of the selection and equal-gain regions are identical and that neither region includes the other. an equal-gain combiner will continue to outperform selection diversity in the presence of shallow fading. parallel to the density function are the lines boundary of the equal-gain region. This is because the height of the joint density function will be small in the region near the origin common to both the equal-gain and selection regions. The reason for this can be seen at once in Fig. 15. In order to discuss this. 14. that. for exponential amplitude fading. This is reflected is. consider the nature of the possible departures from the Rayleigh distribution. This can also be verified by direct computation. Thus. their distribution functions are identical. These cases are illustrated in Fig. for any percentage. it is not difficult to see that independent shallow fading will tend to improve the performance of an equal-gain system. or even sufficient. That is. But there is no similar fact to use as a guide in comparing the other two. with respect to the smaller values of . such as that of Price and Green. [19]. [36] Curve (c) illustrates the more disperse case sometimes found in short-range UHF circuits and in high. Consider first the case where the individual amplitude ratios and are rectangularly distributed. Fig. (c) Typical distribution of the unpleasant sort often observed at frequencies below UHF. its integral over the maximal-ratio region of . PROCEEDINGS OF THE IEEE. for which we must rely on Fig.

cantly poorer than selection diversity. zero as the In space-diversity communication systems. it is not difficult to visualize the way in which the maximal-ratio and equal-gain curves of Fig. Given these facts. indicating that the average local power ratio of an equal-gain system is not substantially degraded by even very disperse fading distributions. The first effect is simply an expression of the fact that as the correlation increases. or other standard sources. It is known 7 that and if and only if . 9. 14 that appreciable correlation will. but only to the extent of the parameter of (44). selection curve for With respect to the average values of Fig. RELATIVE EFFECTS OF CORRELATED FADING Two smoothly varying random variables such as the cannot. respectively. It is also known that the vanishing of  does not necessarily imply that x and y are independent. as approaches 1. in general. when the density function . It can 7[32. (Variables with the same distribution are being considered here. 8. 8 were independent of the fading distribution.) Indeed.) Packard [14] and Bolgiano. For two random variables and with positive variances [18] and . It is convenient to estimate this in terms of a parameter called the correlation coefficient. as is easily seen. un-correlated if . distributions considered above. 16. or by noting that the terms of (42) are replaced by . Now. if anything. an antenna separation of 30 to 50 wavelengths is typically required to obtain correlation coefficients consistently less than 0. (See Appendix V for certain questions related to this subject. BRENNAN: LINEAR DIVERSITY COMBINING TECHNIQUES be seen that approximately half of the uncorrelated local re. [11]. the themselves may well follow the Rayleigh distribution. 16. Van Wambeck and Ross [39] measured the performance of certain HF selection diversity systems directly. in a form suitable for direct comparison with Fig. (Of course. there are two major effects of correlation: first. in terms of the joint density function. It was noted above that the maximal-ratio data of Fig. et al.6. In contrast. and are said to be correlated if . [21] have obtained numerous data bearing on this question in the 6. the local reliability gain of the dual maximal-ratio system over either the selection system or the equal-gain db system is exactly (approximately) for rectangular (exponential) fading. refer once again to Fig. tend to improve the performance of equal-gain diversity. while denotes the dual selection diversity curve of that figure. At equal-gain curves coincide and are uniformly 3 db above the . Some of Staras’ results will simply be reproduced here in Fig. Related material has also 345 . the probability that can differ appreciably from necessarily decreases. VIII. without measuring correlation coefficients. Of course. the mass of the density function tends to con. the maximal-ratio data are unaltered by correlated fading. 8 approach approach 1. and apparently found that even shorter spacings led to useful results. From these considerations. 265].3. Grisdale. et al. 16. the mass centrate around the diagonal line tends to be pulled back nearer the origin.to 18-mc region. not merely the first two moments. 9. the selection diversity values of Fig. If the noises are principally due to interference from remote sources. The problem of correlated fading in selection diversity systems has been studied by Staras [7] and others [3]. if If and are independent. The equal-gain values actually increase toward the maximal-ratio values with increasing correlation. For the rectangular and exponential and . it is not difficult to see from Fig. This can be seen either from physical considerations. and their results will be published in the near future. For high reliability percentages. [13]. This will not be usually strictly true and in many cases will not even be approximately true. then and are independent. be strictly independent. is negligible for To consider the relative effect of correlated fading on the other systems. But it is certainly clear that moderate changes in the form of the fading distribution could not lead to substantial changes in the selection diversity values of Fig. Pierce [37] and Stein [38] independently studied correlated fading in -channel maximal-ratio systems. It is therefore of interest to estimate the effect of dependent fading.) The second fact can be inferred from the behavior of Fig. p. Quite recently. 14. then . relative to selection diversity. and partially correlated if . no such simple and clear dependence of the selection diversity mean values on the form of the distribution exists. This can also be seen by considering the basic operation of the two systems. VARIABLE LOCAL NOISE POWERS Many of the data above were obtained on the assumption constant. However. they may fail to be even approximately independent. this is defined by (45) which reduces readily to (46) [18]. all three systems degrade in an absolute sense with increasing correlation. However. the dual maximal-ratio and sity curves of Fig. 10 to 15 wavelengths will often yield coefficients less than 0. The result of (18) is intimately wrapped up with the Rayleigh distribution. the mean power ratios of equal-gain systems do depend on the distribution. More recently. The curve is the Rayleigh fading curve of Fig. second. a case that has recently been studied by Bond and Meyer [12] for dual selection diversity. it is clear that the approaches zero except on the line equal-gain system approaches the maximal-ratio system in performance. 8. IX. and that the effect liability gain is realized even for . [24] have studied this problem for two-channel maximal-ratio systems. 9 follow the selection diver. Unfortunately. Hence. for and in fact approaches 1 as approaches 1.

FEBRUARY 2003 . the average local power ratio of that a two-channel equal-gain system would be less than for a single channel. Similar considerations show that the average local noise power of a maximal-ratio system —by which is meant one for which the coefficients are given by (14). may fail to There are essentially two ways in which be identically zero. and the performance gets worse as the number of channels is increased. 91.and the local average will fluctuate about zero if the noises are basically unrelated. Equations (45) and (46) become (48) . though this is no longer “maximal” —is increased by . In this frequency of the noise is large in comparison to terms will be negative as often as positive and case. the actual amount of fluctuation to be expected is a function of the noise bandwidth and the duration of the local averages. The troublesome case arises when the noises have a definite positive correlation. 2. the local noise power is increased instead of . which may not be trivial. then the imately constant. Then the output local noise power of an equal-gain system becomes (47) Fig. Rayleigh fading. Hence.. The amount of fluctuation will decrease as is increased and will be small if the lowest . which has no provision for cutting off a very noisy channel. In terms of the analysis above.) It is not difficult to obtain quantitative estimates of the degradation in particular cases. It is probably gratuitous to point out explicitly that. To see how untrivial it can be. then is the correlation of and . which is to say the output power by the factor ratio of (42) is decreased by this factor. 16 Dual selection diversity distributions. for which (18) would still hold. in a postdetection combiner when the noises stem largely from sources of external interference. let and . in of the discussed contrast to the correlation over length if in Section VIII above. (Note that the local correlation over intervals of length . However. it would be much better to use a selection diversity system.) Let . consider for an equal-gain system. NO. FAILURE OF THE NOISES TO BE LOCALLY INCOHERENT The failure of assumption (C) would have no effect on selection diversity systems. for example.e. is considered here. as can happen. This fluctuation has been studied by Rice. X. (The distribecomes more disperse as the noise power bution of the fluctuations increase. for which the assumption if has no relevance whatever. [40] If the are principally may be approxdue to receiver front end noise. been given by Clarke and Cohn. the first of which is simply due to the is over a short interval of duration fact that the average 346 . this assumption is of vital importance for the maximal-ratio and equal-gain systems. i. PROCEEDINGS OF THE IEEE. It should be pointed out that extreme fluctuations in noise power can lead to very poor performance of an equal-gain system. the principal effect of variis to modify the distribution of the able with results as discussed in Section VII above. in contrast to the maximal-ratio and selection systems. Hence . the will simply contribute a small perturbation to the output noise power of a maximal-ratio or equal-gain system. To consider this. This case is not troublesome. for various degrees of correlation. [41]–[43] whose results are quite useful in determining a suitable value of . as will be seen. VOL. in such a case. the factor It follows that the use of maximal-ratio or equal-gain diversity in circumstances where the noise voltages may be highly correlated is hazardous. if which reduces to (46) when — a situation by no means impossible — it would follow .

In any event. that a predetection equal-gain combiner yields substantially less multipath distortion than is obtained with a postdetection maximal-ratio combiner. or in other bandwidth-exchange systems with a pronounced threshold effect. In such systems. namely XII. a predetection maximal-ratio or equal-gain combiner will require the addition of phase-control circuitry in order to satisfy the local-coherence assumption (B). with large a maximal-ratio system will not yield much improvement over selection diversity in such circumstances. it is easier to satisfy the conditions (B) and (C) required for maximal-ratio and equal-gain combiners in the case of a predetection system. It is clear that the operation is identical in either case. whether selection or maximal-ratio diversity is actually used. theoutputratiochangesfrom “completely useful” to “completely useless” with a few db change of input ratio. as before. An alternative point of view would be that the distribution of the amplitude ratios at the input of the combiner would be such as to eliminate the equal-gain combiner from consideration. an SNR at the detector input that is more than a few db above threshold yields a large output ratio. and by Mack [45] for FM systems. The problem of adequate phase control for predetection maximal-ratio or equal-gain combining leads naturally to the next topic. the actual local reliability gains of such maximal-ratio systems—which could be computed from specific detector characteristics. 8–13 and Tables 1 and 2 are not realistic for post-detection combining in FM systems. The phase control and adder circuits require only two semiconductor diodes and 16 passive linear elements. On the other hand. Of course. This may be regarded as due to an effective lowering of the detector threshold resulting from these techniques. and note the unfortunate effect if any one of the becomes large. both theoretically and experimentally. When dealing with postdetection combination in . then the selection diversity values of Table 2 are applicable whether the selection is predetection or postdetection. such as those given by Middleton [44] or obtained by measurement—would certainly be less than the maximal-ratio values in Table 2. If the are not all in phase. Phase control is established via a phase discriminator. Instrumentation for postdetection maximal-ratio combining has been discussed by Kahn [6] and by Morrow. The following treatment is due to Stein. the distribution-sensitive results of Figs. [27] for what amount to AM systems. Then 347 . there are always practical differences between predetection and postdetection combining. the differences between the maximal-ratio and selection values of Table 2 then illustrate the added advantage of predetection maximal-ratio or equal-gain combining. the full advantages of maximal-ratio and equal-gain combiners can be realized in FM systems when and only when they are employed before detection. that the are. This fact has important consequences. [47] where Recall that (B) was the assumption was the slowly varying local rms value of . et al. However.. equal-gain combiners are not even suitable for post-detection combining in conventional FM systems. a Rayleigh distribution of input signal strength for an FM system will emphatically not lead to a Rayleigh distribution of the postdetection amplitude ratio. An ingenious predetection maximal-ratio combiner has been devised by Granlund. indicated in Fig. the maximal-ratio and equal-gain data of Table 2 are completely applicable to predetection combining. Taking the selection values of Table 2 as being the gains obtained by a postdetection maximal-ratio combiner. Accordingly. e. 17. [16] BRENNAN: LINEAR DIVERSITY COMBINING TECHNIQUES If the local reliability gain is defined in terms of the transmitter power required to maintain the input level of the detector above a certain value for more than a specified percentage of time. Furthermore. [46] A particularly elegant predetection equal-gain combiner has been developed by the Federal Telecommunication Laboratories (now the ITT Laboratories). it would sharp-threshold FM systems. substantial changes are required in the case of FM systems with a large deviation ratio. That is. Furthermore. FAILURE OF THE LOCAL-COHERENCE CONDITION (B) It is obviously of interest to estimate the possible degradation in performance of maximal-ratio and equal-gain combiners when the local-coherence condition (B) is not satisfied. Consider the case the where are locally constant in the sense and . but otherwise will not usually make a significant difference in the operation of the system. a maximal-ratio system can be used for postdetection combining in an FM system. It will eliminate switching transients. To begin with. Of course. we must write . However. as is easily seen. An additional advantage of predetection combining in FM systems is that FM multipath distortion can be reduced by this method. where have different phases. this may be regarded as a consequence of the fact that the detection gain of such systems is not constant. while an input ratio that is more than a few db below threshold yields a very small outputratio. predetection selection diversity will sometimes produce smaller switching transients than postdetection selection. Once phase control is established. This corrects the phase of the local oscillator via Miller-effect changes in the oscillator tube capacity. when both are operated under the same circumstances.XI. A specific distribution computed on the basis of a highly simplified detector characteristic has been given. (36). It has been shown by Adams and Mindes [16]. This combiner. PREDETECTION VS POSTDETECTION COMBINING In systems where the power ratio at the output of the detector is essentially the same as at the input. is the same one used in the experiments reported by Adams and Mindes [16]. there is no fundamental change required in the conclusions developed above. cf. The requirement for the coefficients insures that any channel contributes very little to the output.g. at least for probably be best to use only the selection diversity values of Table 2. called simply a phase combiner in FTL literature. Hence. the output of which is applied as a bias voltage to one of the local oscillators. This can be seen on the basis of various qualitative considerations.

91. Then . It is an experimental fact that. in any combiner whatever of . FEBRUARY 2003 . it is only necessary to maintain the phases within 37. This illustrates the fact that the condition (B) cannot be ignored. Then . averaging over a few cycles (or more) of locally linear combiner of the type (6). . in UHF and SHF tropospheric systems. while 51 is sufficient to guarantee a maximum loss of 2 db. is the “phase degradation ratio” . for a suitable choice of . It is manifestly necessary to consider the behavior of diversity systems over longer intervals than those of length . (53) . 8–13. Then (assuming ) (C) still holds. the approach used above and all of the results developed above depend entirely on the use of finite intervals of observation that are neither too long nor too short. maximum of . neither assumption is satisfied. Let denote the output power ratio when of the general combiner (6) when (23′) holds. do that the magnitudes of the phase differences. it is necessary to understand that the results of Figs. the local power ratio is not reduced by more than . refer only to intervals of length . It is easy to see that may . not exceed 90 . both of these assumptions are often satisfied. Then. Specific values are roughly as follows for long-range transmission. relative to which the approximately Rayleigh-distributed and approximately independent. depend on the circumSpecific suitable values of stances. LONG-TERM VARIABILITY Recall that the distributions of the and . whatever this may be. This can be used with any type of modulation. 17 FTL predetection equal-gain combiner. . NO. (50) since . 2. PROCEEDINGS OF THE IEEE. and let . The previous results were obtained using a Rayleigh distribution (9) of unit mean square. actually vanish. VOL. etc. At frequencies below VHF. were to be determined in intervals were assumed to be of length . it is is actually conservative. not much can be Apart from the fact that said about in the general case (52) in the absence of addi348 That is. On the other hand. It is also an experimental fact that if is made too long or too short. and this reduces to (23 ) . or 10 db.g. Hence. It is important to understand the nature of this situation. e.5 of each other. primarily the carrier frequency and transmission distance. for any or (54) (49) where the last step used and the fact that Equation (49) may also be written . so Fig. and mean values of these quantities. In particular. Furthermore.. intervals of 30 minutes to an hour are usually suitable. it is not necessary that it should be satisfied with great precision.. Of course. so that (51) where (52) . this the general type (6). to restrict the reduction in due to imperfect phase control to 1 db or less. and denotes the same for the phase-degraded case (50). . This may be done as follows. Suppose . In VHF ionospheric scatter systems. Thus.tional information about the . clear that the estimate XIII. provided conclusion holds for equal-gain and maximal-ratio combiners. which is entirely to be expected when adding two signals of equal magnitudes and opposite phases. values of one or two minutes usually seem to be appropriate. intervals of five to 30 minutes are often used.

Then this expresses the actual local power ratio delivered by any combiner of the type (6) as the median does not depend on sum of a variable whose time and a variable that is approximately constant over every interval of length . on which figures a normal distribution would be a straight line) and since the distribution of is approximately normal. be written as relative to a period of length . and that the median value of is . (The factor only reason for writing this scale factor in terms of the medians is that these are easily determined experimental quantities. to apply the results of Figs. 9–13 must be modified in accordance with the discussion of Sections VII–XII if the circumstances so dictate. such distributions are often called distributions of hourly medians. it is necessary to have information about the long-term variability of the ( -) median . tions. the relevant data are largely scattered in (generally unobtainable) Signal Corps reports. Hence. with median values . Then this methe interval from dian function is a continuous function of time and the experis usually approximately constant imental fact is that over intervals of length . instead of taking the local rms signals to be the with a fixed median . Distributions of are usually studied in intervals of length one month to one year. though they should properly be distributions of -medians. The second method is applicable in circumstances where are approxall distributions of in all intervals of length imately the same. the combined output signal is sum and the local power ratio besimply multiplied by wherever was before. If is any random variable and is any nonzero constant. Rayleigh of or other. If the median function . we may take To apply this to our previous results. But this approximation is not very accurate for It should be added. Such distributions have been computed by Shimony [49] and Sichak [50] among others. another exare usually approxperimental fact enters. the -distribution of would therefore [18] approximate a normal distribution with a median and variance respectively equal to the sums of the medians and variances of the and distribu. The medians imately the same for different channels. One way to represent this fact is to let denote the median of the distribution obtained in to the present time . . where ( ).. However. is to estimate the lowest -median likely to be encountered on the circuit. is approximately constant over intervals of length . for the larger values of . Now. Staras [9] has observed that. (The great virtue of these . the experimental fact is that the fading distributions observed over different intervals of length will not usually have the same median values. that this second method has not been universally accepted by designers of high-reliability 8It should be noted that such circumstances. though a few such data have been published. Here. furthermore. Then the -distribution of the local power ratio would be the distribution of the sum of two independent variables with known distributions. Several such distributions for frequencies at VHF and above have been given. note that the local linearity of (6) implies that any common scale factor multiplying the signal components may be taken outside the . with a nonzero median has a distribution then it is easy to see that of the same form as that of . differing only in the scale . BRENNAN: LINEAR DIVERSITY COMBINING TECHNIQUES In order to describe the long-term variability of the actual local power ratio . however. for different combiners and orders of diversity. However. Observed distributions of are sometimes approximately Gaussian (normal) in form. the distributions plotted in Figs. however. and URSI and CCIR documents. it is only necessary to where translate their ordinate scales by is the median of the single-channel fading distribution for the interval concerned. there are two ways it may be used. Hence. are not too common in practice. and we may write . The first method. no comparable single source of information for MF and HF systems presently exists. which is why the distributions of said to be log-normal.with a definite median value . [48] Unfortunately. FCC hearing transcripts. above. . It would usually be found that the “exact” determination of the -distribution of would require numerical methods of integration in (67). which is applicable to high-reliability systems at VHF and above. etc. the variables and of (55) would then be independent. Of course.) Then. to a very high degree of accuracy. the and curves of Figs. especially at the higher are often frequencies. This may be used as follows. may be made on the basis of economic and other factors. the actual local rms signals may . all be of the same form. the relevant distributions of are approximately normal (cf. after which a rational choice among the various possibilities of transmitter power. the data of Figs. 12 and 13. and could be computed. 9–13 are precisely the distributions of the variable for the conditions of Fig. 349 . 9–13 are then translated to this value. relative to . the medians obtained in two adjawill not usually cent or overlapping intervals of length differ by very much. order and type of diversity system.) The orsystems is that this minimum value of is not dinate scales of Figs. however. for which the values are the medians. It should be clear that this does not depend on having the distributions.8 It is then easy to see that the distribution one year would also be of in an interval of length the same. Now. and on the basis of the local reliability percentage it is desired to maintain during such worst hours. 8. This becomes comes even simpler when expressed in decibels. Let (55) and be the local power ratio in db. 9–13 to any particular interval of length . Once a distribution of pertinent to the proposed circuit is available.

g. et al. FEBRUARY 2003 . but only 5 db at 600 miles. Computing the long-term distribution of serves to obscure the question of whether the periods of very low signal are a few long periods or many short ones. 2. the interdecile range is about 12. [24]–[26] in which it is not a reasonable approximation.. [52] Van Wambeck and Ross. It is quite simple to plot selection diversity distributions for unequal medians or even for dissimilar distributions. 350 Most of the material above presupposed that the -medians for the several channels were all the same. there are cases. It is hoped that some comparative experimental results will be available within the next year.98 per cent exceeded level. Unfortunately. and the validity of such a computation would seem to be open to question. computing long-term distributions of tends to obscure the gains (Table 2) that actually are realized by diversity techniques. and the one to be used will depend on the circumstances. Perhaps the best single datum presently available was obtained in unpublished experiments conducted by the Signal Corps a few years ago.[39] Glaser and Faber. The second point to be noted is that any long-term distribution of is highly specific to the circuit for which it was computed.25 db entered in . 18 gives only the distributions of the median “scatter” loss. VOL. e. because distributions are highly specific.) It can be seen that the dispersion of decreases with distance. clear-cut and unambiguous experimental data bearing on the comparative performance of the three combining techniques studied above are so rare to be as essentially nonexistent. In addition to the data set forth above. This could probably be traced to the failure of one or both of the conditions (B) and (C) during such periods.XIV. Two additional points should also be noted in connection with long-term distributions of . Maximal-ratio distributions for unequal-median Rayleigh signals can also be expressed explicitly. This is indicated in Fig. it would be possible to obtain the low-signal ends of such distributions by taking the first few terms of a power series. all things considered. XVI. In other words. Experimentally. However. the “selection” column of Table 1 for there were many periods during which the performance of the maximal-ratio system was inferior to the other. it should be especially noted that the PROCEEDINGS OF THE IEEE. It appeared that the maximal-ratio system yielded an average power ratio of 1. [20] and Grisdale. it is precisely the variability of that can be reduced by diversity techniques. the variability of will tend to be obscured by that of . Since two hours in a year corresponds to the 99. the free-space loss has been subtracted out. [21] It is unfortunate for our present purposes that most of these data relate to selection diversity only. XV. but this is simply a reflection of the dispersion contribution by . including Glaser and Van Wambeck. This question can be important. provided that the interval length short. (After Morrow [51]. The first is that the variability or dispersion of will increase as the dispersion of the -median increases. this is not true of complete equal-gain distributions (cf. one of which used dual maximal-ratio diversity and the other used something approximating selection diversity. 18. However. the problem mentioned in footnote 8 would often infect such a computation. [53] This compares very favorably with the value 1.5 db above the selection system when averaged over periods of about 30 minutes. However. e.999 per cent exceeded level in order to insure approximately the reliability obtained by the first method. Two high-frequency systems were compared. but which could be disastrous if concentrated in a single interval of two hours. Appendix IV). However. Thus. experimental data from several Lincoln laboratory circuits. CASE OF UNEQUAL MEDIAN SIGNALS Fig. most especially angle diversity. this is found to be a reasonable approximation in most is not made too cases. This indicates that a long-term distribution of would only be valid for the distance on which the distribution of was based. (Note that Fig.” All three have their merits and defects. CONCLUSIONS Perhaps the most important conclusion to be drawn is that. wintertime propagation at 400 mc. the empirical distributions on which this computation must ultimately rest are not known to anything approaching this degree of accuracy. no one of the diversity combining techniques studied deserves to be called “the optimum system. the of (16) are simply replaced by the identical factors proper distribution functions. However. there are systems in operation in which the loss of two hours in a year would not be troublesome if split into a number of separated intervals of a minute or two each. while that of cannot.0 to 1.) systems. EXPERIMENTAL RESULTS Experimental data relating to diversity systems have been given by several workers. it would be necessary to compute the -distribution of down to something like the 99. The first treatment of this problem apparently was given in [3]. [51] which shows percentage points of distributions as a function of distance at 400 mc. NO. the simplicity and efficacy of the equal-gain system suggest that this may well become the principal standard of the art.. but a detailed analysis of this problem would appear to be premature at the present time. It has been noted [50] that distributions for different orders of diversity show less difference than would be indicated by Table 2. which would not affect a selection system. for the following reasons. 18 Distributions of hourly medians as a function of distance.g. However.5 db at 200 miles. In addition. 91.

This is an experimental fact. all of the multidimensional probability theory given by Bennett [18] can be directly applied to our present circumstances. however long or short the interval of observation. it is not difficult to see that this will lead to the same as defined above. there is probably little to be gained by this formalism in the present context. but with a clear understanding that any realized distribution function can only approximate the Rayleigh distribution. but certainly not -distributions. (See Section XIII. Instruments for meaat selected values of are known as “totalsuring izers” or “level distribution recorders” and exist in various forms. suppose one is interested in the average value of Then (56) i. one uses a specific mathematical model distribution (e. This class is nonempty for positive distances. using whichever notation seems most convenient for the expression involved.e. One such fact. to discuss parent populations for -distributions. The reason for this is that a parent population in the classical sense does not usually exist in this environment. No theory or representation or mathematical model whatever can predict the particular distribution function of a particular waveform in a particular interval exactly. and whose half-hour medians (ipso facto unique) had yearly distribution functions within a specified distance of a fixed log-normal distribution function. and there are many such cases. More generally. In the power when is a voltage or current and or or or can be and is written light of (57). so that . It would be a simple matter to provide a more formal framework for this material by defining suitable classes of functions . and the probability that is simply the fracin which . ample. extent it can be predicted by whatever means. One may. But the approximation may be very close.. It is important to understand the sense in which (57) is applicable.g. as in Section XIII. it would be possible.. The associated density function [18] . howand count the fraction of sampled values that are ever. no more. tion of some interval is determined. but to the BRENNAN: LINEAR DIVERSITY COMBINING TECHNIQUES Fig. no less. what is important. discuss a “distribution” of distributions. in general.instrumentation for an equal-gain combiner is completely independent of what one chooses to think of as a SNR. the time average of th moment of the distribution: (57) a result that is especially useful in computing the average . much used in the body of this paper) for predicting facts such as averages of the form (57).g. One could similarly replace our heuristic language about approximate constants with a more formal treatment that was liberally seasoned with epsilons and deltas and rigorous inequalities. Thus. Of course. 19) were all within a specified distance (in the sense of Lévy’s metric) of some Rayleigh distribution function. for any given value of . A particularly simple special case of this arises when the random variable in question is some voltage or current waveform given as a function of time. simply by adding up the lengths of the intervals for which and dividing their sum by the total duration of the observation. one has available a method of predicting many important facts about the situation. a distribution funcis the probability that (some random variable) is tion less than or equal to . is given by the for any value of . APPENDIX I CERTAIN FACTS ABOUT PROBABILITY THEORY It will be recalled [18] that. In cases where the nature of the possible departures from the model distribution can be estimated. e. In the notation of Section XIII. Another method of obtaining the distribution function of some random variable is to sample it at discrete intervals . is the first moment of the distribution. in circumstances where the distribution function of some waveform can be approximately predicted from either theoretical or empirical grounds. In this case. However. 19. 19 Definition of P (x) = fraction of the time of observation that f (t) x. The importance of this fact is considerable. however. but not necessarily desirable. say . as indicated in Fig. but. The central purpose in using such distribution functions in radio engineering stems from the fact that many different individual waveforms have approximately the same distribution function. as is done in engineering language in that section. or at least have distribution functions that differ in describable ways. all those whose half-hour local distribution functions (in the sense of Fig. [Some statisticians and noise theorists may be bothered by the absence from this discussion of any reference to the classical notions of sample and population. and would suffice for most of the purposes of Sections II–VI. interchangeably for such averages. but for suitable (not present time 351 . is that all time avare given by erages of in the interval .) No fixed distribution can serve as a population distribution for any nonstationary process. of the first importance. 1 and their corresponding multidimensional distribution functions. In many applications. the possible departures in the corresponding time averages can similarly be estimated via (57). such distributions will generally depend on the and the duration . (57) is applicable. For exthe moments [18] of the distribution function . the Rayleigh distribution. Thus.] These considerations above extend directly to several random variables given as in Fig..

. one can write (60) Hence. (62) and of the are Hence. at least as easily as a numerical evaluation of (66). The virtue of this geometric approach is that it often enables the expression of an event of practical interest in terms of such a region. which is well known among mathematicians. [18. the distri. More generally. if the distribution functions of two variables are given numerically and the distribution function of the sum is required. it is a semi-infinite rectangle. such extensive use of it is made in the body of the paper that it is advisable to mention a few consequences here. while in the case of (64). 619] Then (61) becomes if .necessarily long) values of the duration . (See Appendix IV. . if and are random variables and and are constants. in many—possibly most—practical applications. It was pointed out by Bennett [18] that if and are independent random variables with density functions and . a few words on computing the distribution functions of sums of independent random variables may be useful. as an immediate consequence of (56). Finally. First. PROCEEDINGS OF THE IEEE. some of which are by no means obvious. 91. In the case of (63). much less the explicit distribution of . Joint density functions of three or more variables can similarly be integrated over regions in space of three or more dimensions. This is clear when considered then as time averages and.e.” after which the probability of the event in question can be computed by integrating the joint density function over the region. (61) If the are independent. a procedure unnecessarily involved. i. respectively. and . FEBRUARY 2003 . evaluate (66) numerically. note the simple algebraic fact that can be written . the probability that by (63) 352 and the joint distribution function is simply (64) Notice that both (63) and (64) can be written in the form (65) over a certain region in the of an integral of plane. 2. also holds for the corresponding distribution averages.. and are constants and (58) then (59) In order to consider higher moments than the first. (This has sometimes been redensity functions ferred to as “combining” the and distributions.e. the region is an ordinary rectangle. and then sum the resulting approximate density function.) Hence. is given Thus. NO. and .) Although this fact is well known. p. then p. It is quite well known that averaging is a linear operation. denote the distribution functions of . is of bution function of . [32. not necessarily independent. if the first two moments known. 190] may be defined as the limit of approximating sums of the form (68) where the form a suitably fine partition of the range of interest. It will be seen in several sections of this paper that these simple facts can lead to interesting and important results. the probability that more interest than the density function. This does not require independence of and . VOL. This can be expressed in terms of the component distribution functions as (67) which can be seen from (66) by writing and integrating (66) on from to . this dependence may be considered to be negligible for certain engineering purposes. i. This and other numerical integration techniques may be used to evaluate (67) numerically. but fairly commonly used.) However. the density function of the sum is given by (66) which is called the “convolution” or “composition” of the and . (Let . the average square of can be computed without even knowing the distribution of the . there is no need to transform the given distribution functions into approximate density functions. . not necessarily a “rectangle. Another well-known fact is that a joint density function of two random variables can be integrated over [18] a region in the plane to obtain the probability of the region. if are random variables. This is used at several places in the body of the paper. The integral (67). respectively.. by (59). and write (56) three times.

The integral (73) is easily evaluated by contour integration and the residue theorem. such that . and which coincides with the desired distribution function for positive real values of . The characteristic function where of is then simply (72) valid for all values of . An additional check was provided. equality holds in (19) if and only if the ’s and ’s are proportional. for the term is the as can be seen from (67). computations from (41) that for even so that the density function of (78) is where (73) BRENNAN: LINEAR DIVERSITY COMBINING TECHNIQUES (79) 353 .. and ments of . APPENDIX III THE MAXIMAL-RATIO DISTRIBUTION The characteristic function [18] of the is was comwhere each integral over the range of length puted by a 16-point Gaussian quadrature formula. The result is APPENDIX II THE SCHWARZ INEQUALITY There are several ways of proving the inequality (74) (19a) of which the simplest is perhaps to notice the algebraic identity . using an IBM structed [35] from (75) for 704 computer. This is precisely the APPENDIX IV COMPUTATION OF THE EQUAL-GAIN DISTRIBUTION The function of (41) is given recursively by (69) (75) which can be verified at once simply by expanding both sides. Each value . if and only if (76) (70) and it is not difficult to see that this will happen if and only if there are constants and . For all values of were consistent to four decimals for the entire range of . Evidently there is equality holding in (19) if and only if every term in the double sum on the right in (69) vanishes.e. It is for all odd not especially difficult to show that and for . tables for the smaller lated) to six decimals. i. Tables of . . ranging from . Tables of structed by this method agreed with (39) (separately tabu. where have been conRayleigh density function. The values for this integration were obtained from the preof by a modified Tchebyviously constructed table of concheff-Everett interpolation formula. Tables were constructed for various incredown to . the material of Section IV does not justify this assertion. The Rayleigh density and distribufor the range tion functions were generated in the computation program by rational approximations accurate to six decimals. where . was then computed from is obviously nonnegative. It can be shown on the basis of extensive . It is this fact that accounts for the “and only if” assertion following (14). That is. The function defined by (41) for all complex values of is an entire function. not both zero. for . which therefore admits a power series. more precise information can be extracted from this. However. It is clear that (19) follows immediately from (69). while for result (29). (77) (71) is not an index.

(81). For example. rather than absolute. A closely related problem that sometimes arises in this connection is the assumption. The term wise integration from the first to second line of (84) is easily justified. [32.e.. several investigators. it is not difficult to give counter-examples. provided (82) is satisfied.In particular. This was used to for for construct a table of with a guaranteed accuracy of six decimals. 91. Since the terms alternate in sign. even when and the separate input distributions are known. was originally found The result essentially by accident and verified by induction on . 2. 289]The two-dimensional form of the central limit theorem suggests that this assumption would often be very reasonable. but it is far from clear that it is equally applicable to fading radio waves in general. [55] which was reproduced in a paper by Booker. Ratcliffe. then the correlation of two Rayleigh variables could not be negative. [21] and McNicol [36. if Uhlenbeck’s result always held. [42] However. i. which has not always been 354 we make the change of variable (84) which is the result (18). In other words. but should be recognized as an assumption.9 Another direct approach was also suggested by Stein. one uses and for It can be shown from (78) that for recognized as such. This table agreed in this range with that constructed from (76) to six decimals or better. APPENDIX V CERTAIN QUESTIONS RELATED CORRELATED FADING TO THE PROBLEM . becomes (86) 9The procedure used here was suggested by one of the IRE reviewers. that two random variables that are individually Gaussian or normal have a joint distribution that is a two-dimensional normal distribution. obtaining (83) Using the series for OF All of the presently published treatments of correlated fading known to the present writer rely on a result due to Uhlenbeck. such results should be understood as representative. 4] have found such negative correlation. The prevalence of the quoted statement probably stems in part from some insufficiently explicit language of Cramér. 16 was unhesitatingly included in Section VII. this need not be true. In connection with . the correlation coefficient does not uniquely determine diversity performance.. however. p. It seems very probable that Uhlenbeck’s result is satisfactory as a first approximation for engineering purposes. which has resisted strenuous attempts at simplification. who pointed out that (17) could be written (85) which. is not as useful for the explicit computation of coefficients as the recursion relation (81) which can also be established from (41). NO. however. integrating by parts times. including Grisdale.” As a mathematical matter. the coefficient . this means that the error in terminating (77) at the th term is less than the magnitude of the th term. (81). FEBRUARY 2003 . It is essentially certain that it is applicable to narrow-band random noise of the type originally studied by Rice and Uhlenbeck. For larger values of . and Shinn. and (82) are principally due to Michael Ginsburg. APPENDIX VI SELECTION DIVERSITY MEAN POWER RATIOS In the integral (17) for . [56] Uhlenbeck’s result rests in turn on the joint distribution of two Rayleigh variables given by Rice. of the leading term is (80) but this term alone is not sufficiently accurate for useful values of . (17a) (82) the terms of (77) are monotonically decreasing in magnitude. et al. This is why Fig. VOL. Fig. This assumption is the basis of the common statement that “uncorrelated normal variables are independent. PROCEEDINGS OF THE IEEE. The results (78). meaningful sufficient conditions under which this distribution is applicable to correlated Rayleigh fading do not appear to be known. (78).

E.I. Cohn. L. 334–351. N. L. Bennett. 19.” Proc. CS-4.. and J. Labs. August 19. vol.” Proc.” Bell Sys. May 1956.” Proc. [31] E. and D. vol. 46. June 1954. Chisholm. for radiotelegraphy. pp.. Stein. have been published. [34] “Tables of the Normal Probability Functions. pp. J.. vol. Glynn of the Division 3 Computing Section of Lincoln Laboratory provided substantial assistance with the material in Appendix IV and prepared several of the curves. 1854–1873. “A simplified diversity communication system for beyond-the-horizon links.. J. Telecommun. 1530. vol. 23. 1317–1335. Stand. “The statistics of combiner diversity.. December 1953. pp. Jelonek. Washington. T. Mason. 93–94. H. Brennan.. “On the maximal signal-to-noise ratio realizable from several noisy signals. N. Grisdale. March 1956. [42] . M. pp. W. IRE. Roche. M. Inc.Stein remarked that this is (87) is the Beta function [substitute in where (85)]. N. “Mathematical analysis of random noise. 1946. Appl. vol. “Evaluation of polarization diversity performance. pp. [17] D. Tables of the Incomplete 0 Function. pp. Palmer. 944–947. March 1951. P. Princeton.” J. Eds. IRE.. 39.. G. and H. 43. B.” Bell Sys. Bedford. Private Communication. W. Elekt. H. January 1957. Math.” Wireless Engr. E.” (unpublished memorandum). Cambridge.. ACKNOWLEDGMENT I am indebted to the several individuals who either provided helpful comments on [17]. D. October 1955. [45] C. vol. 636–642. A. on Extended Range and Space Communications. 46. Center. R.: Princeton University Press. “An analysis of dual diversity receiving systems. no. Bedford. Van Wambeck and A. [32] H. L. 1958. “Diversity receiving system of RCA communications. Mindes. [23] F. [10] J. 27. Staras.. R. Henze. vol. “Diversity Improvement in Frequency-Shift Keying for Rayleigh Fading Conditions. H. Ser. 24. McNicol. on Communications Systems. pp. 163–184. Bennett of the Bell Telephone Laboratories.. M. and J. [20] J. “The effect of fading on communication circuits subject to interference. Price and P. T. D. H. pp. June 1958.” in Natl. [12] F. [11] K. [25] A. pp. Y. [30] K.and polarizationdiversity reception in the 6–18 mc range. C. Pierce. “Performance of diversity receiving systems. vol. Bur. 9–13 before the complete tables of Mason. L. 39–51. O. IRE. Rainville. Cramér. pp. Waterman Jr. H. February 1947. “Ratio squarer. et al. Granlund of the Department of Electrical Engineering.. IRE. Peterson. “Diversity reception in UHF long-range communications. Bloom and M. IEE. pt. on Antennas and Propagation. 42. Tech. H. vol. E. 50–55. 517–524. [13] E. Root. May 1958. Rice. Chisholm. Gordon. H. “Theoretische untersuchungen über einige di-versity-verahren. E. pp. C. Adams.. 37. H. 45. IRE.” Proc. Beverage.” Proc. November 1954. Sichak. Terman. H. Hycon Eastern. H. vol. 889–895. F. A coust. [15] J. E.” IRE Trans. 362–363. January 1958. [7] H. BRENNAN: LINEAR DIVERSITY COMBINING TECHNIQUES [14] K. C. carefully read and criticized the manuscript. 42. Pierce. Brennan. pp. vol. April 1943.” Proc. Private Communication. vol. Phys. vol. F. Meyer. 46. 65. Nichols. Cambridge. vol. January 1958. 1704. IRE. Moore.” Proc. [18] W. Center. October 1955. pp. March 1958. Green Jr. [37] J. IRE. Mass. [38] S.. Altman and W. Staras of the RCA Laboratories. October 1955.. [3] Z. P. 355 . pp. April 1931. vol. Ser. pp. 1057–1058. Morrow Jr. December 1958. 43. pp. E. or did one or more of the foregoing. pp. Eng. J. J.” J. [36] R. “The fading of radio waves of medium and high frequencies. pp. F. 531–561. F. Ed. 19. 555–570. Elektr. vol. J. Clark and J. Roche. Stein of Hycon Eastern.” Cornell Univ.” Hochfrequenz. Rep.Y. “Filtered thermal noise—fluctuation of energy as a function of interval length. T. 1281–1289. [40] K. G. Stand. “Methods of solving noise problems. Fed.” Proc. vol. P.. Tech. Private Communication. “Effect of correlation on combiner diversity. Mack Jr. vol. 41. C. These included W. Air Force Cambridge Res.. vol. Inc. Elec. Schmelovsky. E. Fitch. 183–194.” in Symp. Packard. Sichak. Middleton. 1958. October 1958. “Evaluation of IF and baseband diversity combining receivers. Bryant. 14. inc. May 1958.. Morris. 44. F. S..” Proc. J. IRE. B. Beverage and H. Ross. Eng. 44. Bond of the Signal Corps Engineering Laboratories. Math. 46. “Einfluss der korrelation zwischen empfangsfeldstarken bei diversity-empfang. 104. and J. O. vol. Ginsburg. IRE. S. REFERENCES [1] H. to be published. [5] L. Pearson. [2] H. Soc. and that higher moments of are given by successive derivatives of the same function. and J. 359. Adams and B. “Theoretical diversity improvement in frequency-shift keying. Faber. Sichak of the ITT Laboratories. G. Radio Engineers Handbook. furnished data or results for inclusion in the present paper. January 1945. 8–13. pp. Princeton. “Diversity reception. and D. 1956. Slepian. IRE. E. B.T. 122. p. H. O. “Fluctuations of random noise power. [39] S. and W. Glaser and L. “Investigations of the angular scattering and multipath properties of tropospheric propagation of short radio waves beyond the horizon. December 1956. April 1949. Sichak. vol. P. 56-117. deBettencourt. Nutley. p.. “Single-sideband techniques in UHF long-range communications. O. March 1956. pp. 562–584. I am also indebted to my colleagues W. CS-6. vol. Bolgiano Jr. 609–638. 41. [27] W. Bur. IRE.” Proc. Hausman. [41] S. October 6–8.” Proc. October 1949. May 1957. inc. pp. S. 1946. 660–661. Pearson and H. 256–264.. January 1956. Portmann. May 1957. Kahn.” Proc.. p. pp. Biometrika Tables for Statisticians. [16] R. Private Communication. Mass.. Staras. N. [22] J. IRE. N. [44] D. R. “Diversity Reception in Scatter Communication with Emphasis on Angle Diversity.” Proc. pp. K. N. E.: Cambridge University Press..: Cambridge University Press. G..” in Natl.” Air Force Cambridge Res. Washington. no. 24. Tables of ’The Distribution Functions of Sums of Rayleigh Variables. on Communication Systems. Meyer of SCEL. N. 96. vol. H.” Proc. [9] H.. pp.” Proc. [29] W. Ithaca. N... 74–76. Leon-hard. “Diversity reception with correlated signals. vol. “Fading of longdistance radio signals and a comparison of space. vol. S.. The differentiation indicated in (86) is straightforward. Mass. G. pp. Brennan. 216–227. [19] R. “Angular diversity reception at 2290 mcps over a 188-mile path. vol. who contributed very generously to this paper under each of the indicated headings. Inc. pp.” Proc. 46–156. [28] R.. IRE. Peterson.” IRE Trans. Phys. J.” Arch. vol. IRE. “On theoretical signal-to-noise ratios in FM receivers: A comparison with amplitude modulation.. H. Baghdady and J. IRE. [8] F. [26] J. IRE. 43. 903–910. “A rapid beam-swinging experiment in transhorizon propagation. 1954. L. vol. Pierce. J. “Carrier-to-noise statistics for various carrier and interference characteristics. Chalk. J.” J. 54–62. [21] G. IRE. I should especially like to thank W.: McGraw-Hill Book Co. Bond and H. [4] A. and several colleagues at Lincoln Laboratory. vol. IEE. T. 20. “A communication technique for multipath channels. Appl. 1943. “Linear techniques in diversity communication. Mason and M. Appl. 338–340. New York.” IRE Trans. 11. pp. 1954. November 1956.” Proc. vol. Engrg. Res.. [43] D. [24] R. “Diversity telephone receiving system of RCA communications. 1774–1778. E. Mack. [35] W. H. [6] D. August 1956. Hartley. AP-6. Appl. Among these. Übertragung. Ginsburg for the use of the equal-gain distributions in Figs. 44. Amer. [33] “Tables of the Error Function and Its Derivative. C. statistical evaluation of possible gain. pp. 1955. 1953. pp. April 1931.” Proc. Cambridge.” Proc. Mathematical Methods of Statistics.

MU-156. June. J. Royal Soc. “Etude de systemes de radiocommunication troposphérique UHF a longue distance. p. 1956.” Proc. H. M.... “Topics in the Design of Antennas for Scatter.” Onde Elect. Glaser and S. Stein. Rad.[46] J. Cambridge.. [54] H. August 1957. Van Wambeck. L... IRE. N. 1956. N. Fort Monmouth. March 1951. Ratcliffe. [56] H. Signal Corps Eng. Shinn. Rep. “Tech. Booker. Lexington.T. Granlund. “Experimental evaluation of diversity receiving systems.. 37. G. Morrow Jr. pp. Meyer.” M. 2. 1957. 1946. [51] W. J. Telecommun. PROCEEDINGS OF THE IEEE. Fort Monmouth.” Phil. H. 242. 454. [47] S. Shimony. pp. Mass. 135. vol. Lab. 91.. N. September 1950.: Princeton University Press. E. 444–449. [52] J. Mass. Sichak. vol. E. May 1957. Uhlenbeck. Lincoln Lab. Princeton..” Evans Signal Lab. NO. 39. A.I. Labs. 619. Nutley. VOL. [49] A. Cramér.I. J. “Diffraction from an irregular screen with applications to ionospheric problems. Memo. Tech. 1943. [50] W. J. Memo. J. Private Communication.” Fed. 252–255. 43. (London) A.. FEBRUARY 2003 . 356 [53] F. IRE. and D. Mathematical Methods of Statistics.. “Final Rep. Trans. private communication. F. No. [55] G. vol. 190.T. October 1995. pp. [48] “Scatter propagation issue.. Bond and H. E. Labs. N. 579–607. MU-156. vol.. 619..” Proc.