186 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO.

1, JANUARY 2010

Transmit Diversity vs. Spatial Multiplexing in

Modern MIMO Systems
Angel Lozano, Senior Member, IEEE, and Nihar Jindal, Member, IEEE

Abstract—A contemporary perspective on transmit antenna ∙ Performance metrics related to the error probability.
diversity and spatial multiplexing is provided. It is argued that, Adding nuance to the term, more than one such metric
in the context of most modern wireless systems and for the can be defined (cf. Section IV).
operating points of interest, transmission techniques that utilize
∙ Transmission and/or reception techniques designed to
all available spatial degrees of freedom for multiplexing outper-
form techniques that explicitly sacrifice spatial multiplexing for improve the above metrics.
diversity. Reaching this conclusion, however, requires that the In this paper, we carefully discriminate these meanings.
channel and some key system features be adequately modeled
We use ’selectivity’ to refer to channel features, which are
and that suitable performance metrics be adopted; failure to do
so may bring about starkly different conclusions. As a specific determined by the environment (e.g., propagation and user
example, this contrast is illustrated using the 3GPP Long-Term mobility) and by basic system parameters (e.g., bandwidth
Evolution system design. and antenna spacing). In turn, the term ’diversity’ is reserved
Index Terms—Diversity, spatial multiplexing, OFDM, MIMO, for performance metrics and for specific transmit/receive
DMT, multiantenna communication. techniques, both of which have to do with the signal. Note
that channel selectivity is a necessary condition for diversity
strategies to yield an improvement in some diversity metric.
I. I NTRODUCTION

M ULTIPATH fading is one of the most fundamental fea-

tures of wireless channels. Because multiple received
replicas of the transmitted signal sometimes combine de-
A. Diversity over Time
Archaic electrical communication systems from a century
ago already featured primitive forms of diversity, where op-
structively, there is a significant probability of severe fades. erators manually selected the receiver with the best quality.
Without means of mitigating such fading, ensuring reasonable Automatic selection of the strongest among various receivers
reliability requires hefty power margins. was discussed as early as 1930 [1]. This naturally led to
Fortunately, fades are very localized in space and frequency: the suggestion of receive antenna combining, initially for
a change in the transmitter or receiver location (on the order of microwave links [2]–[5]. MRC (Maximum Ratio Combining),
a carrier wavelength) or in the frequency (on the order of the by far the most ubiquitous combining scheme, was first pro-
inverse of the propagation delay spread) leads to a roughly posed in 1954 [6]. In addition to receive antenna combining,
independent realization of the fading process. Motivated by other approaches such as repeating the signal on two or more
this selectivity, the concept of diversity is borne: rather than frequency channels were also considered for microwave links
making the success of a transmission entirely dependent on [7]. (Systems were still analog and thus coding and interleav-
a single fading realization, hedge the transmission’s success ing was not an option.) Given the cost of spectrum, though,
across multiple realizations in order to decrease the probability approaches that consume additional bandwidth were naturally
of failure. Hedging or diversifying are almost universal actions unattractive and thus the use of antennas quickly emerged
in the presence of uncertainty, instrumental not only in com- as the preferred diversity approach. Recognizing this point,
munications but also in other fields as disparate as economics receive antenna combining was debated extensively in the
or biology. 1950s [8]–[11] and has since been almost universally adopted
In communications specifically, the term ’diversity’ has, for use at base station sites. The industry, however, remained
over time, acquired different meanings, to the point of be- largely ambivalent about multiple antennas at mobile devices.
coming overloaded. It is used to signify: Although featured in early AMPS trials in the 1970s, and
∙ Variations of the underlying channel in time, frequency, despite repeated favorable studies (e.g., [12]), until recently
space, etc. its adoption had been resisted.1
Multiple base station antennas immediately allow for uplink
Manuscript received October 16, 2008; revised February 9, 2009 and June receive diversity. It is less clear, on the other hand, how to
23, 2009; accepted October 26, 2009. The associate editor coordinating the
review of this paper and approving it for publication was B. S. Rajan. achieve diversity in the downlink using only multiple transmit
A. Lozano is with the Universitat Pompeu Fabra, Barcelona 08005, Spain antennas. In Rayleigh fading, transmitting each symbol from
(e-mail: [email protected]). His work is partially supported by the project every antenna simultaneously is equivalent to using a single
CONSOLIDER-INGENIO 2010 CSD2008-00010 “COMONSENS”.
N. Jindal is with the University of Minnesota, Minneapolis, MN55455, transmit antenna [14, Section 7.3.2]. Suboptimal schemes were
USA (e-mail: [email protected]). His work was partially conducted during a
visit to UPF under the sponsorship of Project TEC2006-01428. 1 The sole exception was the Japanese PDC system [13], which supported
Digital Object Identifier 10.1109/TWC.2010.01.081381 dual-antenna terminals since the early 1990s.
1536-1276/10$25.00 ⃝
c 2010 IEEE

Authorized licensed use limited to: University of Minnesota. Downloaded on January 15, 2010 at 10:02 from IEEE Xplore. Restrictions apply.
LOZANO and JINDAL: TRANSMIT DIVERSITY VS. SPATIAL MULTIPLEXING IN MODERN MIMO SYSTEMS 187

formulated that convert the spatial selectivity across the trans- of different arguments that lead to this conclusion, and which
mit antennas into effective time or frequency selectivity. In will be elaborated upon:
these schemes, multiple copies of each symbol are transmitted ∙ Modern systems use link adaptation to maintain a target
from the various antennas, each subject to either a phase shift error probability and there is essentially no benefit in
[15] or a time delay [16]. From the standpoint of the receiver, operating below this target. This makes diversity metrics,
then, the effective channel that the signal has passed through which quantify the speed at which error probability is
displays enhanced time or frequency selectivity and thus a driven to zero with the signal-to-noise ratio, beside the
diversity advantage can be reaped via coding and interleaving point.
(cf. Section I-C). ∙ Wireless channels in modern systems generally exhibit a
More refined transmit diversity techniques did not develop notable amount of time and frequency selectivity, which
until the 1990s. Pioneered in [17], these techniques blossomed is naturally converted into diversity benefits through
into OSTBC (orthogonal space-time block codes) [18] and, coding and interleaving. This renders additional transmit
subsequently, onto space-time codes at large. Albeit first pro- diversity superfluous.
posed for single-antenna receivers, OSTBCs can also be used ∙ Block error probability is the relevant measure of relia-
in MIMO (multiple-input multiple-output) communication, bility. Since the channel codes featured in contemporary
i.e., when both transmitter and receiver have a multiplicity systems allow for operation close to information-theoretic
of antennas. This allows for additional diversity, and thus limits, such block error probability is well approximated
reliability, but no increases in the number of information by the mutual information outages. Although uncoded er-
symbols per MIMO symbol. ror probability is often quantified, this is only an indirect
Concurrently with space-time coding, the principles of performance measure and incorrect conclusions can be
spatial multiplexing were also formulated in the 1990s [19]– reached by considering only uncoded performance.
[22]. The tenet in spatial multiplexing is to transmit different It is also imperative to recognize that the notion of diver-
symbols from each antenna and have the receiver discriminate sity is indelibly associated with channel uncertainty. If the
these symbols by taking advantage of the fact that, due to transmitter has instantaneous CSI (channel-state information),
spatial selectivity, each transmit antenna has a different spatial then it can match the rate to the channel rendering the error
signature at the receiver. This does allow for an increased num- probability dependent only on the noise. Diversity techniques,
ber of information symbols per MIMO symbol; depending on which aim precisely at mitigating the effects of channel un-
the particular transmission technique used, reliability benefits certainty, are then beside the point. Although perhaps evident,
may or may not be reaped. this point is often neglected. In some models traditionally used
Altogether, the powerful thrust promised by MIMO is to evaluate diversity techniques, for instance, the channel fades
finally bringing multiantenna devices to the marketplace. very slowly yet there is no transmitter adaptation. As we shall
Indeed, MIMO is an integral feature of emerging wireless see, these models do not reflect the operating conditions of
systems such as 3GPP LTE (Long-Term Evolution) [23], most current systems.
3GPP2 Ultra Mobile Broadband, and IEEE 802.16 WiMAX
[24].
C. Time/Frequency Diversity v. Antenna Diversity
Perhaps the simplest manifestation of the efficacy of di-
B. Overview of Work versity in the aforementioned traditional models is receive
With the advent of MIMO, a choice needs to be made antenna combining: if two receive antennas are sufficiently
between transmit diversity techniques, which increase relia- spaced, the same signal is received over independently faded
bility (decrease probability of error), and spatial multiplexing paths. Even with simple selection combining, this squares the
techniques, which increase rate but not necessarily reliability. probability of error; MRC performs even better.
Applications requiring extremely high reliability seem well Based upon the specifics of receive antenna combining,
suited for transmit diversity techniques whereas applications it may appear that multiple, independently faded copies of
that can smoothly handle loss appear better suited for spatial the same signal are required to mitigate fading. Although
multiplexing. It may further appear that the SNR (signal-to- this is an accurate description of receive combining, it is an
noise ratio) and the degree of channel selectivity should also overly stringent requirement in general. This point is clearly
affect this decision. illustrated if one considers a frequency-selective channel. One
Our findings, however, differ strikingly from the above simple but naı̈ve method of mitigating fading in such a
intuitions. The main conclusion is that techniques utilizing channel is to repeat the same signal on two sufficiently spaced
all available spatial degrees of freedom for multiplexing frequency channels. Unlike receive combining, this technique
outperform, at the operating points of interest for modern doubles the number of symbols transmitted and therefore the
wireless systems, techniques that explicitly sacrifice spatial necessary bandwidth. Is repetition, which seems inefficient,
multiplexing for transmit diversity. Thus, from a performance the only way to take advantage of frequency diversity? It
perspective there essentially is no decision that need be made is not—if coding is taken into consideration. By applying
between transmit diversity and multiplexing in contemporary a channel code to a sequence of information bits, the same
MIMO systems (cf. [25]). This conclusion is established on benefit is gained by transmitting different portions of the
the basis of a suboptimal multiplexing technique, and it is only coded block over different frequency channels. No repetition is
strengthened with optimal multiplexing. There are a number necessary; rather, information bits are coded and interleaved,

Authorized licensed use limited to: University of Minnesota. Downloaded on January 15, 2010 at 10:02 from IEEE Xplore. Restrictions apply.
188 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 1, JANUARY 2010

and then the first half of the coded block is transmitted on time/frequency locations, and possibly beamforming and
the first frequency and the other half on the second frequency. precoding.
The information bits can be correctly decoded as long as both ∙ The channels of high-velocity users vary too quickly
frequencies are not badly faded. The same principle applies in time to allow for feedback of CSI or even of the
to time selectivity: instead of repeating the same signal at supportable rates. Thus, the signals of such users are
different time instants, transmit a coded and interleaved block dispersed over the entire available bandwidth thereby
over an appropriate time period [26].2 taking advantage of extensive frequency selectivity. In
addition, time selectivity is naturally available because
II. M ODERN W IRELESS S YSTEMS of the high velocity.
Wireless systems have experienced dramatic changes as The above points evidence the disparity between the low-
they evolved from their initial analog forms to today’s ad- and high-velocity regimes and hence it is necessary to distin-
vanced digital formats. Besides MIMO, features of modern guish between them.
systems—that in many cases were completely absent in earlier
designs—include: III. T HE L OW-V ELOCITY R EGIME
∙ Wideband channelizations and OFDM. At low velocities, timely feedback regarding the current
∙ Packet switching, complemented with time- and state of the channel becomes feasible. This fundamentally
frequency-domain scheduling for low-velocity users. changes the nature of the communication problem: all uncer-
∙ Powerful channel codes [27]–[29]. tainty is removed except for the noise. With powerful coding
∙ Link adaptation, and specifically rate control via variable handling that remaining uncertainty, outages are essentially
modulation and coding [30]. eliminated.3 Transmit diversity techniques, whose goal is
∙ ARQ (automatic repeat request) and H-ARQ (hybrid- precisely to reduce outages, are then beside the point. Rate
ARQ) [31]. maximization becomes the overriding transmission design
These features have had a major impact on the operational principle, and the optimum strategy in this known-channel
conditions: setting is spatial waterfilling [21].
∙ There is a target block error probability, on the order Although the above consideration posited perfect CSI at the
of 1%, at the output of the decoder. (When H-ARQ is in transmitter, it also extends to imperfect-CSI settings (caused
place, this target applies at termination.) A link adaptation by limited rate and/or delay in the link adaptation loop). At
loop is then tasked with selecting the rate in order to a minimum, the supportable rate can be fed back; this still
maintain the performance tightly around this operating removes outages. Additional CSI feedback enables adaptive
point. The rationale for this is two-fold: techniques such as scheduling, power control, beamforming
i) There is little point in spending resources and precoding [34].4
pushing the error probability on the traffic chan- In multiuser settings, furthermore, CSI feedback is collected
nels much below the error probability on the from many users and time- and frequency-domain scheduling
control plane, which, by its very nature (short offers additional degrees of freedom. In this case, trans-
messages and tight latency requirements), cannot mit diversity techniques can actually be detrimental because
be made arbitrarily small [32]. they harden the possible transmission rates to different users
ii) Lower error probabilities often do not im- thereby reducing potential multiuser scheduling gains [35],
prove end-to-end performance: in some applica- [36].
tions (e.g., voice) there is simply no perceivable These conclusions apply almost universally to indoor sys-
improvement in the user experience while, in tems, which conform to this low-velocity regime, as long
others (e.g., data communication requiring very as their medium-access control features the necessary func-
high reliability), it is more cost effective to let tionalities. In outdoor systems, they apply to stationary and
the upper protocol layers handle the losses [33]. pedestrian users.
∙ The fading of low-velocity users can be tracked and fed
back to the transmitter thereby allowing for link adap- IV. T HE H IGH -V ELOCITY R EGIME
tation to the supportable rate, scheduling on favorable Having established that diversity is not an appropriate
2 When explaining the exploitation of selectivity through coding and in-
perspective in the low-velocity regime, we henceforth focus
terleaving, it is important to dispel the misconception that channel coding exclusively on the high-velocity regime. This is the regime
incurs a bandwidth penalty. If the constellation is kept fixed, then coding of interest for vehicular users in outdoor systems. At high
does reduce the rate relative to an uncoded system. However, there is no velocities, the fading (and therefore the time-varying mutual
rate penalty if the constellation size is flexible as in modern systems. For
instance, a system using QPSK with a rate-1/2 binary code and an uncoded information) is too rapid to be tracked. The link adaptation
BPSK system both have an information rate of 1 bit/symbol. For a reasonably
3 The rate supported by the channel may be essentially zero at some
strong code, though, the coded system will achieve a considerably smaller
bit error probability than the uncoded one. More importantly, the advantage time/frequency points, but with proper link adaptation this does not constitute
of the coded system in terms of block error probability is even larger and an outage in the sense that data is not lost (cf. Eq. 4).
this advantage increases with blocklength: the block error probability of a 4 Feedback mechanisms are sometimes studied under the assumption that
coded system decreases with the blocklength whereas, without coding, it they convey information regarding the transmit strategy, e.g., which beam-
actually increases with the blocklength. As will be emphasized throughout former or precoder to use, but not regarding rate selection, in which case
the paper, modern wireless systems cannot be conceived without powerful outages still occur. This, however, is not well aligned with modern system
channel coding. designs in which rate control is paramount.

Authorized licensed use limited to: University of Minnesota. Downloaded on January 15, 2010 at 10:02 from IEEE Xplore. Restrictions apply.
LOZANO and JINDAL: TRANSMIT DIVERSITY VS. SPATIAL MULTIPLEXING IN MODERN MIMO SYSTEMS 189

loop can therefore only match the rate to the average channel B. The Outage-Rate Tradeoff and the DMT
conditions. The scheduler, likewise, can only respond to Eq. (4) fully specifies the tradeoff between outage and
average conditions and thus it is not possible to transmit only rate at any particular SNR, but closed forms do not exist in
to users with favorable instantaneous channels; we thus need general for (4). This led to the introduction of metrics whose
not distinguish between single-user and multiuser settings. tradeoff can be more succinctly characterized. In particular,
the diversity order was introduced as a proxy for the outage
A. Channel Model and Performance Metrics probability. The traditional notion of diversity order equals the
Let 𝑛T and 𝑛R denote, respectively, the number of trans- asymptotic slope of the outage-SNR curve (in log-log scale)
mit and receive antennas. Assuming that OFDM (orthogo- for a fixed 𝑅. Although meaningful in early wireless systems,
nal frequency division multiplexing), the prevalent signalling where 𝑅 was indeed fixed, this is not particularly indicative
technique in contemporary systems, is used to decompose of contemporary systems in which 𝑅 is increased with SNR,
a possibly frequency-selective channel into 𝑁 parallel, non- the average signal-to-noise ratio. A more general formulation
interfering tones, the received signal on the 𝑖th tone is was introduced in [41], where 𝑅 depends on SNR according
to some function 𝑅 = 𝑓 (SNR). The diversity order
y𝑖 = H𝑖 x𝑖 + n𝑖 (1)
log 𝑃out (SNR, 𝑓 (SNR))
where H𝑖 is the 𝑛R × 𝑛T channel matrix on that tone, y𝑖 is 𝑑 = − lim (6)
SNR→∞ log SNR
the 𝑛R × 1 received signal, n𝑖 is the 𝑛R × 1 thermal noise,
IID (Independent Identically Distributed) circularly symmetric still captures the asymptotic slope of the outage-SNR curve (in
complex Gaussian with unit variance, and x𝑖 is the 𝑛T × 1 log-log scale), albeit now for increasing 𝑅. A proxy for rate,
transmitted signal subject to a power constraint SNR, i.e., termed the multiplexing gain, was defined as
𝐸[∣∣x𝑖 ∣∣2 ] ≤ SNR. The receiver has perfect knowledge of the 𝑓 (SNR)
𝑟 = lim , (7)
𝑁 channel matrices, the joint distribution of which is specified SNR→∞ log SNR
later.
which is the asymptotic slope, in bits/s/Hz/(3 dB), of the rate-
For a particular realization of H1 , . . . , H𝑁 , the average
SNR curve.
mutual information thereon is
The DMT (diversity-multiplexing tradeoff) specifies the
𝑁
1 ∑ (𝑟, 𝑑) pairs that are achievable for SNR → ∞, and thus
ℐ(SNR) = 𝐼(x𝑖 ; y𝑖 ) (2) characterizes the tradeoff between 𝑟 and 𝑑 [41]. For a quasi-
𝑁 𝑖=1
static channel model where each coded block is subject to
in bits/s/Hz. The mutual information on each tone is deter- a single realization of the fading process, the DMT specifies
mined by the chosen signal distribution. If the signals are IID that min(𝑛T , 𝑛R ) + 1 distinct DMT points are feasible, each
complex Gaussian5 with 𝐸[x𝑖 x†𝑖 ] = SNR
𝑛T I, then corresponding to a multiplexing gain 0 ≤ 𝑟 ≤ min(𝑛T , 𝑛R )
( ) and a diversity order
SNR †
𝐼(x𝑖 ; y𝑖 ) = log2 det I + H𝑖 H𝑖 . (3)
𝑛T 𝑑(𝑟) = (𝑛T − 𝑟)(𝑛R − 𝑟). (8)
Since approaching this mutual information may entail high The full DMT frontier of achievable (𝑟, 𝑑) pairs is obtained
complexity, simpler MIMO strategies with different (lower) by connecting these points with straight lines.6 More precisely,
mutual informations are often used. Expressions for these are this is the optimum DMT frontier, corresponding to the mu-
given later in this section. tual information in (3). Each transmit-receive architecture is
Once a transmission strategy has been specified, the corre- associated with a DMT frontier, determined by the outage-rate
sponding outage probability for rate 𝑅 (bits/s/Hz) is then relationship for that architecture, which may be smaller than
𝑃out (SNR, 𝑅) = Pr{ℐ(SNR) < 𝑅}. (4) this optimum. For example, for 𝑛T = 𝑛R = 2 the optimum
tradeoff connects the points (0, 4), (1, 1), and (2, 0), while
With suitably powerful codes, the error probability when not in Alamouti transmission yields 𝑑(𝑟) = 4(1 − 𝑟) and spatial
outage is very small and therefore the outage probability is an multiplexing with a MMSE-SIC receiver (cf. Section IV.D)
accurate approximation for the actual block error probability gives 𝑑(𝑟) = 1 − 𝑟/2.
[38]–[40]. We shall therefore use both notions interchangeably The DMT governs the speed at which outage decreases with
henceforth. SNR: if rate is increased as 𝑟 log SNR, then outage decreases
As justified in Section II, modern systems operate at a target (ignoring sub-polynomial terms) as SNR−𝑑(𝑟) . The DMT is
error probability. Hence, the primary performance metric is thus a coarse description, through the proxies 𝑑 and 𝑟, of the
the maximum rate, at each SNR, such that this target is not fundamental tradeoff between outage and rate. This coarseness
exceeded, i.e., arises from the definitions of 𝑟 and 𝑑, which:
𝑅𝜖 (SNR) = max{𝜁 : 𝑃out (SNR, 𝜁) ≤ 𝜖} (5) 1) are asymptotic, thereby restricting the validity of the
𝜁
insights to the high-power regime,7 and
where 𝜖 is the target.
6 If the coded block spans several fading realizations, then this additional
5 Actual systems use discrete constellations, for which counterparts to (3) time/frequency selectivity leads to larger diversity orders but does not increase
exist in integral form [37]. As long as the cardinality of the constellation the maximum value of 𝑟 [41], [42].
is large enough relative to the SNR, the gap between the actual mutual 7 Non-asymptotic DMT formulations have been put forth but they lack the
information and (3) is small and is inconsequential to our conclusions. simplicity and generality of (8) [43], [44].

Authorized licensed use limited to: University of Minnesota. Downloaded on January 15, 2010 at 10:02 from IEEE Xplore. Restrictions apply.
190 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 1, JANUARY 2010

TABLE I
2) involve only the slopes of the outage-SNR and rate-SNR MIMO-OFDM S YSTEM PARAMETERS
curves, thereby ignoring constant offsets.8
Indeed, 𝑑 does not suffice to determine the outage probability Tone spacing 15 kHz
OFDM Symbol duration 71.5 𝜇s
at a given SNR but simply quantifies the speed at which the Bandwidth 10 MHz (600 tones, excluding guards)
outage falls with SNR. Similarly, 𝑟 does not suffice to determine Resource block 12 tones over 1 ms (168 symbols)
the rate, but it only quantifies how it grows with SNR. H-ARQ Incremental redundancy
H-ARQ round spacing 6 ms
Notice that 𝑑(0) corresponds to the traditional notion of Max. number H-ARQ rounds 6
diversity order, i.e., with a fixed rate. A multiplexing gain Power delay profile 12-ray TU
𝑟 = 0 signifies a rate that does not increase (polynomially) Doppler spectrum Clarke-Jakes
with the SNR while 𝑑 = 0 indicates an outage probability that Max. Doppler frequency 185 Hz
Antenna correlation None
does not decrease (polynomially) with the SNR.
It is important to bear in mind that 𝑟 need not coincide
with the number of information symbols per MIMO symbol,
equals the maximum multiplexing gain (i.e., the value of 𝑟
which is the intuitive notion of spatial multiplexing. Likewise,
when 𝑑 = 0) for the transmit-receive strategy being used. This
𝑑 is not solely determined by the channel selectivity. Rather,
quantity can, at most, equal min(𝑛T , 𝑛R ) and, as recognized
𝑟 measures how aggressively 𝑅 is increased with SNR while 𝑑
in [25], any strategy that does not attain 𝑟 = min(𝑛T , 𝑛R ) is
measures how rapidly the outage falls with SNR.
strictly suboptimal in terms of 𝑅𝜖 (SNR) for SNR → ∞. Hence,
we conclude that the optimum strategies are those that utilize
C. Transmit Diversity v. Spatial Multiplexing in Modern Sys- all spatial degrees of freedom for multiplexing. Although this
tems holds asymptotically in the SNR, the extent to which it holds
for SNR values of interest in a selective channel can only be
Armed with the notions furnished by the DMT framework,
determined through a more detailed (non-asymptotic) study.
for the sake of comparison we define a rate-sacrificing transmit
To shed light on this point, a case study is presented next.
diversity scheme as one that achieves, for 𝑟 = 0, the maximum
diversity order 𝑑(0) = 𝑛T 𝑛R , but for which the maximum
spatial multiplexing is strictly 𝑟 = 1. On the other hand, spatial D. Case Study: A Contemporary MIMO-OFDM System
multiplexing schemes are defined as those for which 𝑟 > 1. Let us consider the exemplary system described in Table I,
Nevertheless, to make our points as clean as possible we shall which is loosely based on the 3GPP LTE design [23]. (With
focus exclusively on spatial multiplexing schemes that operate only slight modifications, this system could be made to
at 𝑟 = min(𝑛T , 𝑛R ). conform with 3GPP2 UMB or with IEEE 802.16 WiMAX.)
In the high-velocity scenario considered in this section, Every feature relevant to the discussion at hand is modeled:
frequency-flat analyses are likely to indicate that dramatic re- ∙ A basic resource block spans 12 OFDM tones over 1 ms.
ductions in outage probability can be had by increasing 𝑑. On Since 1 ms corresponds to 14 OFDM symbols, a resource
these grounds, transmission strategies that operate efficiently block consists of 168 symbols. In the high-velocity
at the full-diversity DMT point have been developed. The regime being considered, the 12 tones are interspersed
value of these strategies for modern wireless systems, however, uniformly over 10 MHz of bandwidth. There are 600
is questionable because: usable tones on that bandwidth, guards excluded, and
1) The outage need not be reduced below the target error hence every 50th tone is allocated to the user at hand
probability. while the rest are available for other users.9
2) The channel is highly selective: ∙ Every coded block spans up to 6 H-ARQ transmission
i) By the same token that the fading is too rapid rounds, each corresponding to a basic resource block,
to be tracked, it offers time selectivity. with successive rounds spaced by 6 ms for a maximum
ii) Since, in this regime, modern systems dis- temporal span of 31 ms. (This is an acceptable delay for
perse the signals over large swaths of band- most applications, including Voice-over-IP.) The H-ARQ
width, there tends to be abundant frequency process terminates as soon as decoding is possible. An
selectivity. error is declared if decoding is not possible after 6 rounds.
∙ The channel exhibits continuous Rayleigh fading with a
The DMT describes (coarsely) the entire outage-rate fron-
tier, but modern wireless systems operate at a target outage Clarke-Jakes spectrum and a 180-Hz maximum Doppler
probability 𝜖 and the quantity of interest is 𝑅𝜖 (SNR). Within frequency. (This could correspond, for example, to a
the DMT framework, a fixed outage corresponds to the 𝑑 = 0 speed of 100 Km/h at 2 GHz.) The power delay profile is
point. As a result, all that one can infer about 𝑅𝜖 (SNR) on the given by the 12-ray TU (typical urban) channel detailed
basis of the DMT is that its asymptotic slope in Table II. The r.m.s. delay spread equals 1 𝜇s.
∙ The antennas are uncorrelated to underscore the roles of
𝑅𝜖 (SNR) both diversity and multiplexing. Comments on antenna
lim (9)
SNR→∞ log SNR correlation are put forth in Section IV.E.
8 The choice of specific signal combining schemes at the receiver or of 9 For low velocity users, in contrast, the 12 tones in a resource block are
specific signal covariances at the transmitter is inconsequential to the DMT contiguous so that their fading can be efficiently described and fed back for
[45]–[47] link adaptation and scheduling purposes as discussed in Section II.

Authorized licensed use limited to: University of Minnesota. Downloaded on January 15, 2010 at 10:02 from IEEE Xplore. Restrictions apply.
LOZANO and JINDAL: TRANSMIT DIVERSITY VS. SPATIAL MULTIPLEXING IN MODERN MIMO SYSTEMS 191

TABLE II 5
TU POWER DELAY PROFILE

Received Signal Power (dB)

Delay (𝜇s) Power (dB)
0 -4 0
0.1 -3 0 600
0.3 0
0.5 -2.6
0.8 -3 Tone
-5
1.1 -5
1.3 -7
1.7 -5
2.3 -6.5
3.1 -8.6 -10
(a)
3.2 -11
5 -10 3

Received Signal Power (dB)

0
0 6 12 18 24 30
The impulse response describing each of the 𝑛T 𝑛R entries -3 Time (ms)
of the channel matrix is Time (ms)
-6
12
∑ √
ℎ(𝑡, 𝜏 ) = 𝛼𝑗 𝑐𝑗 (𝑡)𝛿(𝑡 − 𝜏𝑗 ) (10) -9
𝑗=1
-12
where the delays {𝜏𝑗 }12 12
𝑗=1 and the powers {𝛼𝑗 }𝑗=1 are spec-
12
ified in Table II and {𝑐𝑗 (𝑡)}𝑗=1 are independent complex -15
(b)
Gaussian processes with a Clarke-Jakes spectrum. Although
time-varying, the channel is suitably constant for the duration
of an OFDM symbol such that it is meaningful to consider its Fig. 1. (a) TU channel fading realization over 600 tones. The circles indicate
the locations of the 12 tones that map to a given resource block. (b) TU
frequency response as in (1). channel fading realization, for a given tone, over 30 ms. The circles indicate
The variability of the channel response over the multiple the locations of the 6 H-ARQ rounds.
tones and H-ARQ rounds of a coded block is illustrated in
Fig. 1. Note the very high degree of frequency selectivity and
how the channel decorrelates during the 6 ms separating H- ∙ A rate-sacrificing transmit diversity strategy that converts
ARQ rounds. the MIMO channel into an effective scalar channel with
Without H-ARQ, rate and outage are related as per (4). signal-to-noise ratio
With H-ARQ, on the other hand, the length of each coded SNR
{ }
block becomes variable. With IR (incremental redundancy) Tr H𝑖 (𝑘)H†𝑖 (𝑘) (13)
𝑛T
specifically, mutual information is accumulated over succes-
sive H-ARQ rounds [48]. If we let ℳ𝑘 (SNR) denote the mutual where H𝑖 (𝑘) denotes the channel for the 𝑖th symbol on
information after 𝑘 rounds, then the number of rounds needed the 𝑘th H-ARQ round. By applying a strong outer code
to decode a particular block is the smallest integer 𝐾 such to this effective scalar channel, the mutual information
that after 𝑘 rounds is, at most [41]
ℳ𝐾 (SNR) > 6 𝑅𝜖 (SNR) (11) ∑𝑘 168 ( { })
1 ∑ SNR
ℳ𝑘 (SNR) = log 1 + Tr H𝑖 (ℓ)H†𝑖 (ℓ)
where 𝐾 ≤ 6. A one-bit notification of success/failure is fed 168 𝑖=1 𝑛T
ℓ=1
back after the receiver attempts to decode following each H- (14)
ARQ round. An outage is declared if ℳ6 (SNR) ≤ 6 𝑅𝜖 (SNR) This transmit diversity strategy provides full diversity
and the effective rate (long-term average transmitted rate) is order with reduced complexity, but its multiplexing gain
cannot exceed 𝑟 = 1, i.e., one information symbol for
6 𝑅𝜖 (SNR) every vector x𝑖 in (1); precisely, 𝑑(𝑟) = 𝑛T 𝑛R (1 − 𝑟).
ℛ𝜖 (SNR) = . (12)
𝐸[𝐾] Note that, when 𝑛T = 2, (14) is achieved by Alamouti
The initial rate is selected such that the outage at H-ARQ transmission [17].
termination is precisely 𝜖. This corresponds to choosing an ∙ A basic MMSE-SIC spatial multiplexing strategy where
initial rate of 6 𝑅𝜖 where 𝑅𝜖 corresponds to the quantity of a separate coded signal is transmitted from each antenna,
interest defined in (5) with the mutual information averaged all of them at the same rate [49]. The receiver attempts
over the 168 symbols within each H-ARQ round and then to decode the signal transmitted from the first antenna.
summed across the 6 rounds. An MMSE filter is applied to whiten the interference
In order to contrast the benefits of transmit diversity and from the other signals, which means that the first signal
spatial multiplexing, we shall evaluate two representative experiences a signal-to-noise ratio
transmission techniques that achieve the respective extremes ( 𝑛T )−1
of the optimal DMT frontier: h†𝑖,1 (𝑘) H𝑖,1 (𝑘)H†𝑖,1 (𝑘) + I h𝑖,1 (𝑘) (15)
SNR

Authorized licensed use limited to: University of Minnesota. Downloaded on January 15, 2010 at 10:02 from IEEE Xplore. Restrictions apply.
192 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 1, JANUARY 2010

8 15
ity MMSE-SIC structure is representative of the single-user
ers

Rate (bits/s/Hz)
7 10
it Div MIMO mode in LTE [23].
nT=4 nsm
Tra
6 5 MIMO
nR=4 IMO Let 𝑛T = 𝑛R = 4, the high-end configuration for LTE, and
n-M
Rate (bits/s/Hz)

No g
consider first a simplistic model where the fading is frequency-
5 0
xin flat and there is no H-ARQ. Every coded block is therefore
ple
ulti
0 5 10 15 20 25 30 35
SNR (dB) M
4 atia
l subject to essentially a single realization of the Rayleigh
Sp
fading process. Under such model, the spectral efficiencies
3
achievable with 1% outage, 𝑅0.01 (SNR), are compared in Fig. 2
2 Frequency-Flat Fading alongside the corresponding efficiency for the non-MIMO
nT=1
No H-ARQ reference (𝑛T = 1, 𝑛R = 4). Transmit diversity is uniformly
1 1% Outage
nR=4 superior to spatial multiplexing in the SNR range of interest.
0 In fact, spatial multiplexing results in a loss even with respect
0 5 10 15 20 to non-MIMO transmission with the same number of receive
SNR (dB) antennas. The curves eventually cross, as the DMT predicts10
and the inset in Fig. 2 confirms, but this crossover does not
occur until beyond 30 dB.
Fig. 2. Main plot: MMSE-SIC spatial multiplexing v. transmit diversity with
𝑛T = 𝑛R = 4 in a frequency-flat channel with no H-ARQ. Also shown is the Still with 𝑛T = 𝑛R = 4, consider now the richer model
non-MIMO reference (𝑛T = 1, 𝑛R = 4). Inset: Same curves over a wider described in Tables I–II. The effective mutual information for
SNR range.
each block is averaged over tones and symbols and accu-
mulated over H-ARQ rounds. The corresponding comparison
15 ing is presented in Fig. 3. In this case, transmit diversity offers
TU Channel ex
ltipl
H-ARQ
al
Mu a negligible advantage whereas spatial multiplexing provides
1% Outage
pati ample gains with respect to non-MIMO.
S
The stark contrast between the behaviors observed under the
Rate (bits/s/Hz)

10 nT=4
MIMO
nR=4 ity
different models can only be explained by the abundant time
ivers
Trans
mit D and frequency selectivity neglected by the simple model and
actually present in the system. This renders transmit antenna
5 diversity superfluous, not only asymptotically but at every SNR.
nT=1 Non-MIMO Under the simple model, the signal from the first antenna does
nR=4
not benefit from any diversity and thus limits the overall rate.
Under the richer channel model, however, that first signal reaps
0 diversity from time/frequency selectivity and thus the lack of
0 5 10 15 20 spatial diversity is mostly inconsequential. To further highlight
SNR (dB) this effect, consider the outage 𝜖-ℛ𝜖 curves in Fig. 4 for the
simple and rich channel models. (Recall that, in the simple
model, ℛ𝜖 = 𝑅𝜖 because there is no H-ARQ). The enrichment
Fig. 3. MMSE-SIC spatial multiplexing v. transmit diversity with 𝑛T =
𝑛R = 4 in the channel described in Tables I–II. Also shown is the non- of the channel greatly increases the MMSE-SIC rate, but has
MIMO reference (𝑛T = 1, 𝑛R = 4). a much smaller effect on the transmit diversity rate due to the
diminishing returns of diversity.
Although the above results were for a highly selective
during the 𝑘th H-ARQ round. If successful, the effect channel, this behavioral contrast is robust. Even if the speed
of the first signal is subtracted from the received sam- is reduced down to where the low-velocity regime might start,
ples and decoding of the second signal is attempted, as in Fig. 5, the behaviors are hardly affected because there
and so forth. No optimistic assumption regarding error is still significant selectivity. Likewise, the performances are
propagation is made: an outage is declared if any of largely preserved if the bandwidth is diminished significantly
the 𝑛T coded signals cannot support the transmitted rate. below 10 MHz or the delay spread is reduced below 1 𝜇s.
The aggregate mutual information over the 𝑛T antennas
after 𝑘 H-ARQ rounds is shown at the bottom of the 10 For transmit diversity, the DMT frontier is a straight line connecting
page, where h𝑖,𝑚 (ℓ) is the 𝑚th column of H𝑖 (ℓ) and (𝑟 = 0, 𝑑 = 16) and (𝑟 = 1, 𝑑 = 0), while for MMSE-SIC it connects
H𝑖,𝑚 (ℓ) = [h𝑖,𝑚+1 (ℓ)h𝑖,𝑚+2 (ℓ) ⋅ ⋅ ⋅ h𝑖,𝑛T (ℓ)]. While de- (𝑟 = 0, 𝑑 = 1) and (𝑟 = 4, 𝑑 = 0). These frontiers intersect at (𝑟 =
ficient in terms of diversity order, this strategy yields full 20/21, 𝑑 = 16/21) and thus transmit diversity outperforms MMSE-SIC (as
SNR → ∞) for 𝑑 > 16/21. In our 𝑑 = 0 setting, transmit diversity only
multiplexing gain, 𝑟 = min(𝑛T , 𝑛R ), when 𝑑 = 0; more achieves a slope of 𝑟 = 1 bit/s/Hz/(3 dB) compared to 𝑟 = 4 for MMSE-SIC.
generally 𝑑(𝑟) = (𝑛R −𝑛T +1)(1−𝑟/ min(𝑛T , 𝑛R )). This This difference in slope explains the eventual crossover.

{ ( )}
𝑘
∑ 1 ∑
168 ( 𝑛T )−1
† †
ℳ𝑘 (SNR) = 𝑛T min log 1 + h𝑖,𝑚 (ℓ) H𝑖,𝑚 (ℓ)H𝑖,𝑚 (ℓ) + I h𝑖,𝑚 (ℓ) (16)
𝑚=1,⋅⋅⋅ ,𝑛T 168 𝑖=1 SNR
ℓ=1

Authorized licensed use limited to: University of Minnesota. Downloaded on January 15, 2010 at 10:02 from IEEE Xplore. Restrictions apply.
LOZANO and JINDAL: TRANSMIT DIVERSITY VS. SPATIAL MULTIPLEXING IN MODERN MIMO SYSTEMS 193

1 20
Spatial Multiplexing
Transmit Diversity TU Channel
H-ARQ
SNR =10 dB 1% Outage
15
nT=4

Rate (bits/s/Hz)
nT=4, nR=4
0.1 Ergodic nR=4
P out

10
Flat fading TU Channel
0.01 No H-ARQ H-ARQ

5 nT=2
nR=2

0.001 0
0 2 4 6 8 10 0 5 10 15 20
Rate (bits/s/Hz) SNR (dB)

Fig. 4. Outage probability as function of ℛ𝜖 for transmit diversity and for Fig. 6. In solid lines, 1%-outage rate achievable with MMSE-SIC spatial
MMSE-SIC spatial multiplexing at SNR = 10 dB, for both the frequency-flat multiplexing in the channel described in Tables I–II. In circles, corresponding
channel without H-ARQ and for the channel described in Tables I–II. ergodic rate for the same numbers of antennas.

16
Spatial Multiplexing
course, be taken under consideration when determining the
nT=4 appropriate transmission strategy.
12 MIMO
nR=4
Rate (bits/s/Hz)

Transmit Diversity
8 F. Optimal MIMO
While in the case study we considered the performance of
nT=1 TU Channel a low complexity but suboptimal detection scheme for spatial
4 Non-MIMO
nR=4 H-ARQ multiplexing, the continual increase in computational power
1% Outage
SNR = 20 dB is now rendering optimal or near-optimal MIMO feasible.
Low Rather than transmitting separate coded signals from the 𝑛T
~

0 Velocity
antennas, a single one can be interleaved over time, frequency
0 20 40 60 80 100
Velocity (Km/h)
and the transmit antennas. At the receiver side, each vector
symbol is then fed to a detector that derives soft estimates
of each coded bit—possibly by use of a sphere decoder—
Fig. 5. MMSE-SIC spatial multiplexing, transmit diversity and non-MIMO to a standard outer decoder (e.g., message-passing decoder),
transmission as function of velocity for the channel described in Tables I–II at with subsequent iterations between the MIMO detector and
SNR = 20 dB. (Below some point, the system transitions to the low-velocity
regime and thus the curves are no longer meaningful.) the decoder [52]. Such techniques, and others such as mutual
information lossless codes [53]–[55], can approach the mutual
information in (3).
E. Ergodic Modeling It is worthwhile to note that the mutual information in (3) is
greater than that in (13) for any channel matrix H. Denoting
As it turns out, the time/frequency selectivity in modern by 𝜆ℓ the ℓth eigenvalue of HH† ,
systems is so substantial as to justify the adoption of an er-
godic model altogether. Shown in Fig. 6 is the correspondence ( ) (𝑛 ( ))
SNR ∏R
SNR
†
between the exact rates achievable with 1% outage in the log det I + HH = log 1+ 𝜆ℓ (17)
channel described in Tables I–II and the respective ergodic 𝑛T 𝑛T
ℓ=1
( 𝑛R
)
rates. SNR ∑
From a computational standpoint, this match is welcome ≥ log 1 + 𝜆ℓ (18)
𝑛T
news because of the fact that convenient closed forms exist ℓ=1
( )
for the rates achievable in an ergodic Rayleigh-faded channel SNR { }
= log 1 + Tr HH†
[50]. Moreover, the optimum transmission strategies and the 𝑛T
impact upon capacity of more detailed channel features such (19)
as antenna correlation, Rice factors, colored out-of-cell inter-
ference, etc, can then be asserted by virtue of the extensive where (17) holds because the determinant equals the product
body of results available for the ergodic setting [45], [51]. of the eigenvalues, (18) comes from { dropping
} ∑terms in the
𝑛R
Antenna correlation, for example, leads to a disparity in the product, and (19) follows from Tr HH† = ℓ=1 𝜆ℓ .
distribution of the spatial eigenmodes that effectively reduces Hence, optimal MIMO is uniformly superior to rate-
the spatial multiplexing capability. Such effects should, of sacrificing transmit diversity in the sense that it achieves a

Authorized licensed use limited to: University of Minnesota. Downloaded on January 15, 2010 at 10:02 from IEEE Xplore. Restrictions apply.
194 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 1, JANUARY 2010

10 1.0
1
n T =2, n R =2 al 8 bits/symbol
tim C
1% Outage Op -SI
SE

Symbol Error Probability

8 MM -1
0.1
10 Transmit
Rate (bits/s/Hz)

ti Diversity
mou
Ala
6 TU Channel 10-2
0.0 Sp
H-ARQ 4 bits/symbol atia
uti lM
l mo ulti
ima Ala Sp
4 Opt 10-3
0.0 atia ple
xin
lM g
ult
iple
2 -SIC xin
Frequency Flat MMSE 10-4
0.0 Transmit g
No H-ARQ Diversity

0 10-5
0.0 Frequency-Flat Fading
0 5 10 15 20
SNR (dB)
0.0
0 5 10 15 20 25 30
Fig. 7. Spectral efficiencies achievable with Alamouti transmission, with SNR (dB)
spatial multiplexing (MMSE-SIC) and with optimal MIMO for 𝑛T = 𝑛R =
2. The comparisons are shown for both a frequency-flat channel without H-
ARQ and for the channel described in Tables I–II.
Fig. 8. Uncoded symbol error probability for transmit diversity and spatial
multiplexing with 𝑛T = 𝑛R = 2. The comparisons are shown for both 4
bits/symbol and 8 bits/symbol.
smaller outage probability for any rate and SNR.11 Drawing
parallels with the discussion in Section I about the subopti-
mality of repeating the same signal on two frequency channels Consider 𝑛T = 𝑛R = 2 for the sake of specificity.
versus transmitting different portions of a coded block thereon, Comparisons must be conducted at equal SNR and rate, e.g.,
one could equate transmit diversity with the former and the Alamouti with 16-QAM v. spatial multiplexing with 4-QAM
optimal MIMO strategy with the latter. or Alamouti with 256-QAM v. spatial multiplexing with
Optimal MIMO allows achieving both the rate and the out- 16-QAM. These constitute different space-time modulation
age benefits of multiantenna communication, whereas transmit formats with 4 and 8 bits per MIMO symbol, respectively.
diversity and MMSE-SIC spatial multiplexing obtain only Fig. 8 presents the symbol error probabilities, averaged over
the outage or only the rate benefit, respectively. Nonetheless, the fading distribution, for a maximum-likelihood detector
optimal MIMO is still subject to the fundamental outage-rate with no outer coding. (Note that, in the absence of outer
tradeoff governed by the optimal DMT region specified in (8). coding, changing the modulation format is the only way to
In Fig. 7, the spectral efficiencies of Alamouti transmission coarsely perform link adaptation). Consistent with previous
and spatial multiplexing (for optimal and MMSE-SIC) are sections, one can compare the different schemes at a particular
shown for 𝑛T = 𝑛R = 2, for both the frequency-flat model and operating point, which in this case corresponds to a fixed
the richer model in Tables I–II. Optimal spatial multiplexing symbol error probability. From the two sets of curves in the
is superior to Alamouti with both models, as per the above figure, spatial multiplexing outperforms Alamouti if the error
derivation, but the difference is considerably larger when the probability is above 10−2 . For lower error probabilities, which
rich model is used. Consistent with the earlier case study, is where an uncoded system would likely have to operate,
MMSE-SIC performs well below Alamouti in the frequency- Alamouti outperforms spatial multiplexing at 4 bits/symbol but
flat setting but outperforms it under the rich model. Spatial not necessarily at 8 bits/symbol. In fact, the behavior of the
multiplexing with MMSE-SIC and optimal spatial multiplex- uncoded error probability as a function of SNR is qualitatively
ing have the same maximum multiplexing gain and thus the similar to that of the mutual information outages in a non-
same slope, 𝑟 = 2 bits/s/Hz/(3 dB), but the DMT cannot selective channel: from Fig. 4, at outage probabilities above
capture the constant offset between the two. roughly 0.2 spatial multiplexing outperforms transmit diversity
(in flat fading with no H-ARQ) whereas for lower outage
V. U NCODED E RROR P ROBABILITY: A P OTENTIALLY probabilities the opposite is true.12
M ISLEADING M ETRIC Hence, uncoded analysis does not always correctly predict
the behavior with strong outer coding in much the same
In the previous section, the superiority of spatial multiplex-
way that a mutual information analysis without selectivity
ing relative to rate-sacrificing transmit diversity was illustrated
does not (cf. Section IV). In a system such as the one
in the context of modern wireless systems, which exhibit
described in Tables I and II, the outer code makes use of
abundant time and frequency selectivity and utilize powerful
frequency selectivity across tones and time selectivity across
outer coding. However, a different conclusion is sometimes
H-ARQ rounds. Without an outer code, on the other hand,
reached if one compares the error probabilities of the two
schemes in the absence of outer coding. 12 The uncoded symbol error probability is the average of the uncoded
symbol error probability conditioned on H. This conditional error probability,
11 The mutual information in (3) is also greater than the MMSE-SIC mutual which is determined only by the noise, is essentially 0 or 1 for most channel
information in (16), because (3) equals the sum of the 𝑛T mutual informations realizations depending on whether the mutual information corresponding to
whereas (16) is 𝑛T times the minimum. One can make the two equal by H is larger or smaller, respectively, than the transmit rate. Intuitively then,
separately rate controlling each coded signal based on instantaneous channel the average symbol error probability roughly mirrors the mutual information
conditions [56], but this is infeasible in the high-velocity regime. outage probability in non-selective channels.

Authorized licensed use limited to: University of Minnesota. Downloaded on January 15, 2010 at 10:02 from IEEE Xplore. Restrictions apply.
LOZANO and JINDAL: TRANSMIT DIVERSITY VS. SPATIAL MULTIPLEXING IN MODERN MIMO SYSTEMS 195

this selectivity would not be exploited and thus averaging ∙Performance assessments are to be made at the correct
uncoded error probabilities does not have the same operational operating point, particularly in terms of error probability.
meaning of averaging mutual informations. Since modern ∙ The assumptions regarding transmit CSI must be consis-
communication systems rely on powerful channel codes, it is tent with the regime being considered. At low velocities,
the coded performance that is relevant. One should be careful adaptive rate control based on instantaneous CSI should
when predicting such performance on the basis of uncoded be incorporated; at high velocities, only adaptation to
error probabilities. average channel conditions should be allowed.
∙ Coded block error probabilities or mutual information

VI. C ONCLUSION outages, rather than uncoded error probabilities, should

be used to gauge performance.
Since the 1970’s, antenna diversity had been a preferred
Proper modeling is essential in order to evaluate the behav-
weapon used by mobile wireless systems against the dele-
ior of transmission and reception techniques in contemporary
terious effect of fading. While narrowband channelizations
and future wireless systems. As our discussion on trans-
and non-adaptive links were the norm, antenna diversity was
mit diversity and spatial multiplexing demonstrates, improper
highly effective. In modern systems, however, this is no
modeling can lead to misguided perceptions and fictitious
longer the case. Link adaptivity and scheduling have rendered
gains.
transmit diversity undesirable for low-velocity users whereas
abundant time/frequency selectivity has rendered transmit
diversity superfluous for high-velocity users. Moreover, the VII. ACKNOWLEDGMENTS
prevalence of MIMO has opened the door for a much more The authors gratefully acknowledge comments, suggestions
effective use of antennas: spatial multiplexing. Indeed, the and discussions with Profs. Jeff Andrews and Robert Heath Jr.,
spatial degrees of freedom created by MIMO should be re- from the University of Texas at Austin, Prof. Uri Erez, from
garded as additional ’bandwidth’ and, for the same reason that Tel-Aviv University, Prof. Andrea Goldsmith, from Stanford
schemes based on time/frequency repetition waste bandwidth, University, Prof. Albert Guillen, from the University of Cam-
rate-sacrificing transmit diversity techniques (e.g., OSTBC) bridge, Prof. Daniel P. Palomar, from the Hong Kong Uni-
waste ’bandwidth’. versity of Science and Technology, and Mr. Chris Ramming
Of all possible DMT points, therefore, the zero-diversity from Intel Corp., as well as the comments of the anonymous
one stands out in importance. Techniques, even suboptimum reviewers and of the associate editor.
ones, that can provide full multiplexing are most appealing to
modern wireless systems whereas techniques that achieve full
diversity order but fall short on multiplexing gain are least R EFERENCES
appealing. Our findings further the conclusion in [25], where [1] W. C. Bohn, “Automatic selection of receiving channels,” U. S. Patent
a similar point is made solely on the basis of the multiplexing No. 1747218, Feb. 1930.
[2] H. H. Beverage and H. O. Peterson, “Diversity receiving system of RCA
gain for frequency-flat channels. Although our conclusion has for radiotelegraphy,” Proc. IRE, vol. 19, pp. 531–584, Apr. 1931.
been reached on the premise that the coded error probabilities [3] F. J. Altman and W. Sichak, “A simplified diversity communication
of discrete constellations are well approximated by the mutual system for beyond-the-horizon links,” IRE Trans. Commun. Systems,
vol. 4, pp. 50–55, Mar. 1956.
information outages of Gaussian codebooks, we expect it to [4] P. Bello and B. Nelin, “Predetection diversity combining with selectively
hold in any situation where the code operates at a (roughly) fading channels,” IRE Trans. Commun. Systems, vol. 9, pp. 32–42, Mar.
constant gap to the mutual information. 1963.
[5] H. Makino and K. Morita, “Design of space diversity receiving and
The trend for the foreseeable future is a sustained increase in transmitting systems for line-of-sight microwave links,” IRE Trans.
system bandwidth, which is bound to only shore up the above Commun. Tech., vol. 15, pp. 603–614, Aug. 1967.
conclusion. LTE, which for our case study was taken to use [6] L. R. Kahn, “Ratio squarer,” Proc. IRE, vol. 42, p. 1704, Nov. 1954.
[7] W. T. Barnett, “Microwave line-of-sight propagation with and without
10 MHz, is already moving towards 20 MHz channelizations. frequency diversity,” Bell Syst. Tech. J., vol. 49, no. 8, pp. 1827–1871,
At the same time, exceptions to the foregoing conclusion Oct. 1970.
do exist. These include, for example, control channels that [8] D. G. Brennan, “Linear diversity combining techniques,” Proc. IRE,
vol. 47, pp. 1075–1102, 1959.
convey short messages. Transmit diversity is fitting for these [9] M. Schwartz, W. R. Bennett, and S. Stein, Communication Systems and
channels, which do benefit from a lower error probability but Techniques. New York: McGraw-Hill, 1965.
lack significant time/frequency selectivity. Other exceptions [10] A. Vigants, “Space diversity performance as a function of antenna
separation,” IEEE Trans. Commun. Tech., vol. 16, no. 6, pp. 831–836,
may be found in applications such as sensor networks or others Dec. 1968.
where the medium access control is non-existent or does not [11] W. C. Jakes, Microwave Mobile Communications. New York: IEEE
have link adaptation and retransmission mechanisms. Press, 1974.
Our study has only required evaluating well-known tech- [12] D. C. Cox, “Antenna diversity performance in mitigating the effects of
portable radiotelephone orientation and multipath propagation,” IEEE
niques under realistic models and at the appropriate operating Trans. Commun., vol. 31, pp. 620–628, May 1983.
points. Indeed, a more general conclusion that can be drawn [13] N. N. K. Kinoshita and M. Kuramoto, “Development of a TDMA digital
from the discussion in this paper is that, over time, the evolu- cellular system based on a Japanese standard,” in Proc. IEEE Vehic.
Tech. Conf. (VTC’91), pp. 642–645, 1991.
tion of wireless systems has rendered some of the traditional [14] A. J. Goldsmith, Wireless Communications. Cambridge University
models and wisdoms obsolete. In particular: Press, 2005.
[15] A. Hiroike and F. Adachi, “Combined effects of phase sweeping
∙ Frequency and time selectivity should always be properly transmitter diversity and channel coding,” IEEE Trans. Veh. Technol.,
modeled. vol. 33, pp. 37–43, Feb. 1984.

Authorized licensed use limited to: University of Minnesota. Downloaded on January 15, 2010 at 10:02 from IEEE Xplore. Restrictions apply.
196 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 1, JANUARY 2010

[16] A. Wittneben, “A new bandwidth efficient transmit antenna modulation [44] K. Azarian and H. E. Gamal, “The throughput-reliability tradeoff in
diversity scheme for linear digital modulation,” in Proc. IEEE Int’l Conf. block-fading MIMO channels,” IEEE Trans. Inf. Theory, vol. 53, no. 2,
on Commun. (ICC’93), vol. 3, pp. 1630–1634, 1993. pp. 488–501, Feb. 2007.
[17] S. M. Alamouti, “A simple transmit diversity technique for wireless [45] A. Lozano, A. M. Tulino, and S. Verdu, “High-SNR power offset in
communications,” IEEE J. Sel. Areas Commun., vol. 16, pp. 1451–1458, multiantenna communication,” IEEE Trans. Inf. Theory, vol. 51, no. 12,
Oct. 1998. pp. 4134–4151, Dec. 2005.
[18] V. Tarokh, N. Seshadri, and A. R. Calderbank, “Space-time codes for [46] N. Jindal, “High SNR analysis of MIMO broadcast channels,” in Proc.
high data rate wireless communications: performance criterion and code IEEE Int’l Symp. on Inform. Theory (ISIT’05), Sep. 2005.
construction,” IEEE Trans. Inf. Theory, vol. 44, pp. 744–765, Mar. 1998. [47] N. Prasad and M. K. Varanasi, “MIMO outage capacity in the high
[19] G. J. Foschini and M. J. Gans, “On the limits of wireless communica- SNR regime,” in Proc. Int’l Symp. on Inform. Theory (ISIT 2005), pp.
tions in a fading environment when using multiple antennas,” Wireless 656–660, Sep. 2005.
Personal Commun., pp. 315–335, 1998. [48] G. Caire and D. Tuninetti, “The throughput of hybrid-ARQ protocols
[20] G. J. Foschini, “Layered space-time architecture for wireless commu- for the Gaussian collision channel,” IEEE Trans. Inf. Theory, vol. 47,
nications in a fading environment when using multi-element antennas,” no. 5, pp. 1971–1988, July 2001.
Bell Labs Tech. J., pp. 41–59, 1996. [49] G. J. Foschini, G. D. Golden, R. A. Valenzuela, and P. W. Wolnianski,
[21] I. E. Telatar, “Capacity of multi-antenna Gaussian channels,” Eur. Trans. “Simplified processing for high spectral efficiency wireless communica-
Telecommun., vol. 10, pp. 585–595, Nov. 1999. tion employing multi-element arrays,” vol. 17, no. 11, pp. 1841–1852,
[22] G. Raleigh and J. M. Cioffi, “Spatio-temporal coding for wireless Nov. 1999.
communications,” IEEE Trans. Commun., vol. 46, no. 3, pp. 357–366, [50] H. Shin and J. H. Lee, “Capacity of multiple-antenna fading channels:
Mar. 1998. spatial fading correlation, double scattering and keyhole,” IEEE Trans.
[23] S. Sesia, I. Toufik, and M. B. (Editors), The UMTS Long Term Evolution: Inf. Theory, vol. 49, pp. 2636–2647, Oct. 2003.
From Theory to Practice. Wiley, 2009. [51] A. M. Tulino, A. Lozano, and S. Verdu, “Impact of correlation on the
[24] J. G. Andrews, A. Ghosh, and R. Muhamed, Fundamentals of WiMAX. capacity of multiantenna channels,” IEEE Trans. Inf. Theory, vol. 51,
Prentice Hall, 2007. pp. 2491–2509, July 2005.
[52] B. M. Hochwald and S. ten Brink, “Achieving near-capacity on a
[25] B. Varadarajan and J. R. Barry, “The outage capacity of linear space-time
multiple-antenna channel,” IEEE Trans. Commun., vol. 51, no. 3, pp.
codes,” IEEE Trans. Wireless Commun., vol. 6, no. 6, pp. 2642–2648,
389–399, Mar. 2003.
Nov. 2005.
[53] B. A. Sethuraman, B. S. Rajan, and V. Shashidhar, “Full-diversity, high-
[26] D. Tse and P. Viswanath, Fundamentals of Wireless Communication.
rate space-time block codes from division algebras,” IEEE Trans. Inf.
Cambridge University Press, 2005.
Theory, vol. 49, pp. 2596–2616, Oct. 2003.
[27] R. Gallager, “Low-density parity-check codes,” IEEE Trans. Inf. Theory, [54] F. Oggier, J.-C. Belfiore, and E. Viterbo, “Cyclic division algebras: a
vol. 8, no. 1, pp. 21–28, Jan. 1962.
tool for space-time coding,” Foundations and Trends in Commun. and
[28] C. Berrou, A. Glavieux, and P. Thitimajshima, “Near Shannon limit Inf. Theory, vol. 4, no. 1.
error correcting coding and decoding: Turbo-codes,” Proc. IEEE Int’l [55] V. Shashidhar, B. S. Rajan, and B. A. Sethuraman, “Information-lossless
Conf. in Communic. (ICC’93), pp. 1064–1070, May 1993. space-time block codes from crossed-product algebras,” IEEE Trans. Inf.
[29] G. D. Forney and D. J. Costello, “Channel coding: the road to channel Theory, vol. 52, pp. 3913–3935, Sep. 2006.
capacity,” Proc. IEEE, vol. 95, no. 6, pp. 1150–1177, June 2007. [56] S. T. Chung, A. Lozano, H. C. Huang, A. Sutivong, and J. M. Cioffi,
[30] A. J. Goldsmith, “The capacity of downlink fading channels with “Approaching the MIMO capacity with V-BLAST: theory and practice,”
variable rate and power,” vol. 46, pp. 569–580, Aug. 1997. EURASIP J. Applied Signal Process. (special issue on MIMO), vol.
[31] D. Costello, J. Hagenauer, H. Imai, and S. B. Wicker, “Applications 2002, pp. 762–771, May 2004.
of error-control coding,” IEEE Trans. Inf. Theory, vol. 44, no. 6, pp.
2531–2560, Oct. 1998.
Angel Lozano is a Professor of Information and
[32] M. Meyer, H. Wiemann, M. Sagfors, J. Torsner, and J.-F. Cheng, “ARQ Communication Technologies at UPF (Universi-
concept for the UMTS long-term evolution,” Proc. IEEE Vehic. Technol. tat Pompeu Fabra) in Barcelona, Spain, where he
Conf. (VTC’2006 Fall), Sep. 2006. teaches and conducts research on wireless commu-
[33] P. Wu and N. Jindal, “Coding versus ARQ in fading channels: how nications. Born in Manresa, Spain, Angel Lozano re-
reliable should the PHY be?,” GLOBECOM’09, Nov. 2009. ceived the Telecommunications Engineering degree
[34] D. J. Love, R. W. H. Jr., V. K. N. Lau, D. Gesbert, B. D. Rao, and from UPC (Universitat Politecnica de Catalunya),
M. Andrews, “An overview of limited feedback in wireless communi- Spain, in 1992 and Master of Science and Ph.D.
cation systems,” IEEE J. Sel. Areas Commun., vol. 26, pp. 1341–1365, degrees in Electrical Engineering from Stanford
Oct. 2008. University in 1994 and 1998, respectively. Contem-
[35] B. M. Hochwald, T. L. Marzetta, and V. Tarokh, “Multiple-antenna chan- porarily, between 1996 and 1998, he also worked for
nel hardening and its implications for rate feedback and scheduling,” Rockwell Communication Systems (now Conexant Systems) in San Diego,
IEEE Trans. Inf. Theory, vol. 50, no. 9, pp. 1893–1909, Sep. 2004. USA. In 1999 he joined Bell Labs (Lucent Technologies, now Alcatel-Lucent)
[36] H. Viswanathan and S. Venkatesan, “The impact of antenna diversity in in Holmdel, USA, where he was a member of the Wireless Communications
packet data systems with scheduling,” IEEE Trans. Commun., vol. 52, Department until 2008. Between 2005 and 2008 he was also an Adjunct
pp. 546–549, Apr. 2004. Associate Professor of Electrical Engineering at Columbia University. Dr.
[37] A. Lozano, A. M. Tulino, and S. Verdú, “Power allocation over parallel Lozano is a senior member of the IEEE since 1999. He served as associate
Gaussian channels with arbitrary input distributions,” IEEE Trans. Inf. editor for the IEEE T RANSACTIONS ON C OMMUNICATIONS between 1999
Theory, vol. 52, no. 7, pp. 3033–3051, July 2006. and 2009, has guest-edited various other IEEE and non-IEEE journal special
[38] L. Ozarow, S. Shamai, and A. D. Wyner, “Information theoretic consid- issues, and he is actively involved in committees and conference organization
erations for cellular mobile radio,” IEEE Trans. Veh. Technol., vol. 43, tasks for the IEEE Communications Society. He has further participated in
pp. 359–378, May 1994. standardization activities for 3GPP, 3GPP2, IEEE 802.20 and the IETF. Dr.
[39] G. Caire, G. Taricco, and E. Biglieri, “Optimum power control over Lozano has authored over 80 technical journal and conference papers, holds
fading channels,” IEEE Trans. Inf. Theory, vol. 45, pp. 1468–1489, July 15 patents, and has contributed to several books. His papers have received
1999. two awards: the best paper at the 2006 IEEE Int’l Symposium on Spread
[40] N. Prasad and M. K. Varanasi, “Outage theorems for MIMO block- Spectrum Techniques & Applications, and the Stephen O. Rice prize to the
fading channels,” IEEE Trans. Inf. Theory, vol. 52, no. 12, pp. 5284– best paper published in the IEEE T RANSACTIONS ON C OMMUNICATIONS in
5296, Dec. 2006. 2008.
[41] L. Zheng and D. Tse, “Diversity and multiplexing: a fundamental
tradeoff in multiple antenna channels,” IEEE Trans. Inf. Theory, vol. 49,
pp. 1073–1096, May 2003.
[42] P. Coronel and H. . Bolcskei, “Diversity-multiplexing tradeoff in
selective-fading MIMO channels,” in Proc. IEEE Int’l Symp. on Inform.
Theory (ISIT’07), June 2007.
[43] R. Narasimhan, “Finite-SNR diversity-multiplexing tradeoff for corre-
lated Rayleigh and Rician MIMO channels,” IEEE Trans. Inf. Theory,
vol. 52, no. 9, pp. 3956–3979, Sep. 2006.

Authorized licensed use limited to: University of Minnesota. Downloaded on January 15, 2010 at 10:02 from IEEE Xplore. Restrictions apply.
LOZANO and JINDAL: TRANSMIT DIVERSITY VS. SPATIAL MULTIPLEXING IN MODERN MIMO SYSTEMS 197

Nihar Jindal (S’99-M’04) received the B.S. de- issue of the EURASIP J OURNAL ON W IRELESS C OMMUNICATIONS AND
gree in Electrical Engineering and Computer Sci- N ETWORKING on the topic of multiuser communication. Dr. Jindal’s research
ence from U.C. Berkeley in 1999, and the M.S. spans the fields of information theory and wireless communication, with
and Ph.D. degrees in Electrical Engineering from specific interests in multiple-antenna/multi-user channels, dynamic resource
Stanford University in 2001 and 2004. He is an allocation, and sensor and ad-hoc networks. Dr. Jindal was the recipient of
assistant professor in the Department of Electrical the 2005 IEEE Communications Society and Information Theory Society
and Computer Engineering at the University of Min- Joint Paper Award and 2008 Best Paper Award for the IEEE J OURNAL ON
nesota. His industry experience includes internships S ELECTED A REAS IN C OMMUNICATION (IEEE Leonard G. Abraham Prize),
at Intel Corporation, Santa Clara, CA in 2000 and at and was also awarded the University of Minnesota McKnight Land-Grant
Lucent Bell Labs, Holmdel, NJ in 2002. Dr. Jindal Professorship Award in 2007 and the NSF CAREER award in 2008.
currently serves as an Associate Editor for IEEE
T RANSACTIONS ON C OMMUNICATIONS, and was a guest editor for a special

Authorized licensed use limited to: University of Minnesota. Downloaded on January 15, 2010 at 10:02 from IEEE Xplore. Restrictions apply.