Professional Documents
Culture Documents
3, MAY 2005
AbstractWe analyze a mobile multiple input multiple output can fluctuate due to user mobility, frequency hopping etc. In
wireless link with transmit and receive antennas operating practice, these fluctuations can be too rapid to allow reliable
in a spatially correlated Rayleigh flat fading environment. Only channel estimation. Multiple transmit and receive antennas
the correlations between the channel coefficients are assumed to
make channel estimation especially difficult due to the large
be known at the transmitter and the receiver. The channel coeffi-
cients are correlated in space and uncorrelated in time from one number of channel coefficients that need to be estimated. Since
coherence interval to another. These coefficients remain constant outdoor wireless systems strive to accommodate higher user
for a coherence interval of symbol periods after which they mobility and indoor wireless communication systems such
change to another independent realization according to the spatial as HomeRF and BlueTooth rely on frequency hopping, these
correlation model. For this system we characterize the structure issues necessitate further research into MIMO link capacity in
of the input signal that achieves capacity. The capacity achieving the absence of CSIT and CSIR. As a first step in this direction,
transmit signal is expressed as the product of an isotropically
distributed unitary matrix, an independent nonnegative diagonal it was shown by Marzetta and Hochwald in [4] that increasing
matrix and a unitary matrix whose columns are the eigenvectors the number of transmit antennas does not increase channel
of the transmit fade covariance matrix. For the case where the capacity in a fast fading scenario. Specifically, they showed
number of transmit antennas is larger than the channel coher- that with uncorrelated flat Rayleigh fading, increasing the
ence interval , we show that the channel capacity is independent number of transmit antennas beyond the channel coher-
of the smallest eigenvalues of the transmit fade covariance ence interval length does not increase capacity. Marzetta
matrix. In contrast to the previously reported results for the spa-
tially white fading model where adding more transmit antennas and Hochwalds work [4] underscores a fundamental property
beyond the coherence interval length ( ) does not increase of multiple antenna systems. The capacity benefits of using
capacity, we find that additional transmit antennas always increase multiple antennas depend dramatically on how well the channel
capacity as long as their channel fading coefficients are spatially variations can be tracked at the transmitter and the receiver.
correlated with the other antennas. We show that for fast hopping For users moving at high speeds or fast frequency hopping
or fast fading systems ( = 1) with only channel covariance systems the coherence interval is quite small and the channel
information available to the transmitter and receiver, transmit
fade correlations are beneficial. Mathematically, we prove this by
fluctuates rapidly. Based on the results of [4], it seems that there
showing that capacity is a Schur-convex function of the vector of is no capacity advantage from using multiple antennas in such
eigenvalues of the transmit fade correlation matrix. We also show scenarios.
that the maximum possible capacity gain due to transmitter fade Before resigning ourselves to the less than optimistic results
correlations is 10 log dB. of [4] we note that these results assume spatially white fading.
Index TermsAntenna correlation, channel capacity, channel Uncorrelated fading is a valid assumption for systems with
state information (CSI), multielement antenna arrays, wireless widely spaced antennas. In this paper we ask the following
communications. question: with no channel state information (CSI) is it beneficial
to place the transmit antennas closer together? Our motivation
for introducing spatial correlations comes from the observation
I. INTRODUCTION
that even if the channel variations are too fast, it is possible
much easier task than estimating the rapidly fluctuating instan- and is the additive noise vector. The transmit power
taneous channel. The estimated distribution can be made avail- constraint at each instant can be expressed as follows:
able to the transmitter through a feedback channel. Note that
because knowledge of the time-varying channel statistics con- (3)
stitutes some information about the channel that is available to
the transmitter and the receiver, this assumption has also been Next, we proceed to describe the spatial correlation model.
termed partial CSI [5], imperfect CSI or covariance feedback1
[6]. However, limiting our time horizon to the duration where A. Spatial Correlation Model
the channel is described by a given, fixed distribution, the instan-
taneous channel state becomes independent of the fixed channel In general the spatially correlated channel can be modeled by
statistics. Thus, following the definitions in [7] we observe that
covariance information falls under the no CSIT or CSIR as- (4)
sumption.
where is the spatially white MIMO channel with
We use the following notation throughout the paper:
is the column vector obtained by stacking the columns of on i.i.d. zero mean unit variance circularly symmetric complex
top of each other; for integer , is defined to be one when Gaussian components, and is the covariance
and zero otherwise; and are the th row and the matrix
th column of the matrix , respectively; is the
(5)
principal submatrix of ; is the th element of the matrix
; is the trace of the matrix , and denotes Kronecker Though the model described above is capable of representing
product. any correlation effects between the elements of , we use a
simpler model
II. SYSTEM MODEL AND PROBLEM STATEMENT
We focus on a point-to-point wireless communication system (6)
using transmit antennas and receive antennas over a nar-
Although less general than (4), this model is a commonly used
rowband flat Rayleigh fading channel. Following the notation
[8], [9] spatial correlation model. While the model has been
and system model of [4], we have the following mathematical
found to be satisfactory based on field measurements [10], we
representation:
would also like to point out that recent work by Ozcelik et al.
[11] has shown that the model may not render the multipath
structure correctly, leading to pessimistic capacity estimates in
some cases. and are called the transmit and receive fade
(1) covariance matrices respectively. It is easily seen that , , and
are related by
where is the signal received at th receive antenna at time
, is the fade coefficient between transmit antenna and (7)
receive antenna , is the signal transmitted from the th
transmit antenna at time and is the additive noise at the th
receive antenna at time . The noise components are assumed III. CONDITIONAL PROBABILITY DENSITY
to be independent identically distributed (i.i.d.) zero mean unit The starting point of the analysis is the conditional probability
variance circularly symmetric complex Gaussian. The channel density . Conditioned on the transmitted signal, the re-
coefficients remain fixed for the coherence time after which ceived signal has zero mean complex jointly Gaussian compo-
they change to a new set of values generated according to nents because the channel and the white noise are zero mean
the spatial correlation model described in Section II-A. The complex jointly Gaussian. The probability density is
channel coefficients are assumed to be independent from one completely described by the second moments, obtained as fol-
coherence interval to another. The underlying assumptions be- lows:
hind this block-fading model and its applicability to fast fading
or frequency hopping systems are discussed by Abou-Faycal,
Shamai, and Trott for the single antenna case in [3].
In matrix notation, the system model admits the following
representation:
(8)
(2)
which gives us the conditional probability density as
where is the received signal matrix, is the
transmitted signal matrix, is the channel matrix,
1The term feedback comes from the underlying model that the channel statis-
(9)
tics are provided by the receiver to the transmitter through a feedback link.
992 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 4, NO. 3, MAY 2005
Comparing this to the density expression in [4] for independent IV. PROPERTIES OF CAPACITY-ACHIEVING
fades TRANSMITTED SIGNALS
Subject to the average power constraint (3) the link capacity
is given by the following expression [4]:
(10)
(11) (14)
(15)
In this case, the matrix is a sufficient statistic.
Property 4: For uncorrelated transmit antennas, i.e., which, after a suitable change of variables, becomes
, the pdf of (9) simplifies only slightly as
(16)
Here
(12)
Also, in this case for any unitary matrix , (17)
. (18)
The special cases of uncorrelated transmit antennas
(19)
or uncorrelated receive antennas are of interest
because depending on the antenna separation and the density of (20)
scatterers around the antenna array the fades corresponding to
either the transmitter or the receiver may be correlated or un- It is easy to see that and have the same statistics as
correlated. It is a common scenario that the mobile unit is sur- and , respectively, and satisfies the power constraint
rounded by a rich scattering environment while the base station (3) if and only if satisfies it. Thus, the capacity optimization
is relatively unobstructed. This makes the decorrelating antenna problem with transmitter fade covariance matrix and receiver
separation much smaller at the mobile unit (centimeters) than at fade covariance matrix is equivalent to the capacity opti-
the base station (meters). But at the same time, the mobile unit mization problem with transmit fade covariance matrix and
is more size-constrained than the base station. For a particular receiver fade covariance matrix .
system, whether the fades corresponding to the transmitter or While the eigenvectors of the transmitter and receiver fade
the receiver are independent or correlated depends on the rela- covariance matrices do not affect capacity, they could still con-
tive values of these parameters. stitute relevant information required to achieve the capacity. In
In the next section, we use these special properties of the other words, the proof provided above does not imply that the
conditional pdf to derive some properties of the ca- transmitter does not need to know or that the receiver does
pacity-achieving input signal for the MIMO channel with cor- not need to know . We explain this observation further as
related Rayleigh fading. follows. Consider the system described by (14) and the system
JAFAR AND GOLDSMITH: MULTIPLE-ANTENNA CAPACITY IN CORRELATED RAYLEIGH FADING 993
described by (16). The input, output, noise, transmit-fade co- Theorem 1: For any coherence interval and any number
variance matrix, and receive-fade covariance matrix for system of receiver antennas, the capacity obtained with trans-
(14) are , , , and , respectively. For system (16), the mitter antennas in spatially correlated fading depends on the
corresponding quantities are , , , , and , respectively. transmit fade correlation matrix through only the
These systems are equivalent because starting from system (14) largest eigenvalues of and is not a function of the eigen-
the transmitter can rotate the input by and the receiver can vectors of or .
rotate the output by and the resulting system is (16). If the Note that if the transmitter fades are uncorrelated ,
transmitter does not know or the receiver does not know , all eigenmodes of are equally strong and using the largest
then these transformations can not be made and therefore we can eigenmodes amounts to using any of the transmit antennas, as
not claim that these two systems have the same capacity. Since found in [4]. This is true even with correlated receiver antennas
knowledge of and can only increase capacity, without .
the knowledge of and , the capacity of (16) will only be
an upperbound on the capacity of (14). Thus, the knowledge of B. Capacity Dependence on the Number of Transmit Antennas
is required at the transmitter in order to rotate to yield . One of the key results proved by Marzetta and Hochwald in
Similarly, the receiver needs to know to obtain from . [4] for spatially white fading is that regardless of the number of
Interestingly enough, the transmitter does not need to know receiver antennas, increasing the number of transmit antennas
and the receiver does not need to know . The receiver may beyond the coherence interval duration does not increase ca-
need to estimate just to provide it to the transmitter through pacity. This is in sharp contrast to the perfect CSIR case where
a feedback link, but other than that, we notice that the trans- capacity is known to increase linearly with even
mitter and receiver need to know only their corresponding fade in the absence of CSIT [1], [2]. With perfect CSIR, the linear
covariance eigenvectors. increase in channel capacity was shown in [9] to be true even
Lemma 2: The link capacity (13) obtained with with spatially correlated fading although the rate of growth is
antennas depends only on the largest eigenvalues of . reduced relative to the spatially white fading case. The avail-
Proof: Suppose that a particular joint distribution of the ability of CSIR seems to be crucial as it makes all the difference
elements of achieves capacity. Let the singular value de- between a linear growth in capacity with transmit antennas (for
composition of be given by ) and no growth at all. To explore this surprising re-
sult further, in this section we investigate the channel capacity
(21) dependence on the number of transmit antennas in spatially cor-
related fading.
where is a unitary matrix, is a oblong matrix
The following theorem states our main result in this section.
containing the singular values along the main diagonal and zeros
Theorem 2: The capacity of a MIMO channel without CSI
elsewhere, and is an unitary matrix. Without loss of
increases almost surely (with probability 1) with the number of
generality, we assume that . Let be
antennas when the transmit antenna fades are spatially corre-
the leading principal submatrix of . Also, let be
lated.
the matrix obtained by padding with trailing rows of
Thus, Theorem 2 establishes that if we are able to introduce
zeros.
and track spatial correlations, then additional transmit antennas
Note that . Next, define a new
almostly surely increase capacity. An outline of the proof is as
matrix
follows. First, we show that adding a correlated transmit antenna
(22) almost surely increases the principal eigenvalue of the transmit
fade covariance matrix. Then we show that a larger principal
Clearly eigenvalue implies a higher capacity. Putting these results to-
gether we obtain Theorem 2. The full proof is presented next.
(23) Starting from a system with transmit antennas, suppose we
add another antenna at the transmitter, increasing the number of
But the left-hand side of (23) depends only on the largest
transmit antennas to . The new transmit fade correlation
eigenvalues of . Thus, if we use instead of with the corre-
sponding transformation of the input distribution, we still get the matrix can be represented as the following bordered matrix
same mutual information which is independent of the
smallest eigenvalues of . To complete the proof we need to
show that if satisfies the power constraint then so does .
Mathematically, we need to show that where is an vector representing the correlations between
the added antenna and the previous transmit antennas, and
(24) is a positive real number representing the fading power from the
added antenna. Also, let the singular value decomposition of the
But this can be proved proceeding along the lines of a similar transmit correlation matrix be given by
proof presented in [12], albeit in a different context. The com-
plete proof is provided in the Appendix.
We conclude this section by summarizing our results in the
following Theorem. (25)
994 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 4, NO. 3, MAY 2005
unitary matrix as in [4], by premultiplication with an isotropi- than the channel corresponding to the transmit covariance ma-
cally distributed unitary matrix to yield the new input signal trix . Majorization as a measure of correlations is also used
which generates the same mutual information. Note in [9].
that a unitary transformation on does not change capacity [by A function of vector is said to be Schur-convex if
inspection from (9)]. whenever vector majorizes vector . For our
Comparing with the results for uncorrelated fading, we notice purpose this can be interpreted as follows. If the capacity is
that the structure of the capacity-achieving signal under corre- a Schur-convex function of the eigenvalues of the channel
lated Rayleigh fading is similar, except for a rotation along the transmit fade covariance matrix, then more spread out eigen-
eigenvectors of the transmitter fade correlation matrix. While values lead to higher capacity.
the receive correlations do not appear directly in the statement Next, we use these concepts of majorization and Schur-con-
of Theorem 3, the distribution of the diagonal matrix can de- vexity, to mathematically state our previous observation that
pend on the eigenvalues of both and . transmit fade correlations help capacity.
The structure of the capacity-achieving signal is relevant for
several reasons. Capacity computation is a maximization of mu-
B. Capacity is Schur-Convex
tual information over the space of all valid input distributions.
By identifying the structure of the capacity-achieving input dis- The following Corollary of Theorem 1 shows that transmit
tribution we limit the size of the optimization space. The only fade correlations help capacity.
unknown in this case is the distribution of the diagonal matrix . Corollary 1: For fast Rayleigh fading (or fast frequency hop-
Since is also independent of the size of the search space is ping), i.e. for , the MIMO channel capacity under corre-
considerably reduced. The structure of capacity achieving codes lated fading is Schur-convex in the eigenvalues of .
also provides insights into the kind of practical codes that per- Proof: Let and be the vectors containing eigen-
form well on the channel. In this case, the diagonal matrix can values of two different transmit fade correlation matrices
be interpreted as a random power allocation across the eigenvec- and , respectively. If majorizes then the principal
tors of the transmit fade covariance matrix. eigenvalue of is at least as large as the principal eigen-
value of . From the result of Theorem 1 we know that for
, capacity depends only on the principal eigenvalue of
V. SPATIALLY CORRELATED FADING: GOOD OR BAD? the transmit fade correlation matrix. Thus, the larger principal
eigenvalue leads to a larger capacity.
In the perfect CSI case, it is well known that in order to To estimate the benefits of transmitter fade correlations, we
achieve a higher capacity we want the channel gains associated next present a simple example comparing the extreme cases of
with each transmit antenna to fade independently [9]. However, independent and perfectly correlated fading. This allows us to
our results in the previous sections indicate that correlated fades compute precisely the maximum possible capacity gains due to
at the transmitter may be desirable when CSI is not available. correlated transmitter fades.
We saw that unlike the spatially white fading case where the ca- Consider the case of fast Rayleigh fading (or fast frequency
pacity does not increase with the number of transmit antennas hopping), . Let us denote the capacity achieved with
beyond the coherence interval duration (as proved in [4]), for transmit antennas and a total transmit power as . The
spatially correlated fading at the transmitter antennas the ca- superscript or 1 depending on whether the transmitter
pacity increases with the number of transmit antennas. Thus, fades are perfectly uncorrelated or perfectly cor-
it seems that transmit fade correlations help capacity. To make related , respectively. First, let us assume
this statement precise we need a mathematical notion that al- that the channel components fade independently. As shown in
lows us to say when a correlation matrix is more correlated or [4], the capacity achieved with transmit antennas is the same
less correlated than another correlation matrix. Such a notion is as the capacity achieved with only one transmit antenna. We de-
provided by the concept of majorization from order statistics. note this as
Thus, with perfect correlation of the transmit fades there is a long as the channel fading coefficients are spatially correlated.
db improvement in the capacity over the case of inde- This leads to the interesting conclusion that for fast fading or
pendent fading. fast frequency hopping channels where channel state
Thus with fast fading or fast frequency hopping , it information is hard to obtain, it is beneficial to place the trans-
is beneficial to place transmit antennas close together to obtain mitter antennas close to each other in order to generate highly
correlated fades instead of the traditional approach that seeks to correlated fades. Mathematically, we proved this by showing
place transmit antennas separated by the decorrelating distance that capacity is a Schur-convex function of the vector of eigen-
so that the fades corresponding to different transmit antennas values of the transmit fade correlation matrix. We also proved
are uncorrelated. In short, for fast fading channels transmitter that the capacity gains due to transmitter fade correlations are
fade correlation helps capacity, and the capacity gains from cor- bounded above by dB. This is in sharp contrast to the
related transmitter fades are bounded above by dB. well known capacity benefits of uncorrelated fades when perfect
Similar results have been found to apply to the case of mul- CSIR and perfect CSIT is assumed.
tiple input single output (MISO) channels with perfect CSIR
and no CSIT [14], [15]. Boche and Jorswieck show in [14] APPENDIX
and [15] that if the transmitter can adapt to the channel spatial
correlations then fade correlations improve capacity. Although A. Proof of the Inequality:
their channel model is significantly different from our model We proceed as follows:
in this paperwe assume no CSIR and multiple receive an-
tennas while [14], [15] assume perfect CSIR and a single re-
ceive antennait is interesting that the findings are similar. In
(35)
both cases, if we allow the transmitter to track the spatial corre-
lations then capacity is improved.
(36)
VI. CONCLUSION
Define . So we need to prove that
We analyzed a point to point mobile MIMO wireless link with
transmit and receive antennas operating in a spatially cor- (37)
related Rayleigh flat fading environment. While for our discus-
sions in this paper we focused on mobile communications when
users are travelling at high speeds, our results are equally appli- From this point onwards, the proof follows along the lines of
cable to frequency hopping systems. In either case the channel a similar proof in [12]. However, for the sake of completeness,
may not stay constant for long enough to allow enough time we provide of the rest of the proof in this paper. We need the
for reliable channel estimation. To address this case no channel following three lemmas.
state information was assumed at either the transmitter or the Lemma 5: For a Hermitian matrix the vector of diagonal
receiver. The channel coefficients were allowed to be correlated entries majorizes the vector of eigenvalues .
in space but constrained to be uncorrelated in time from one co- Lemma 6: For any two given positive real vectors ,
herence interval to another. The coefficients were assumed to the permutation that minimizes the sum
remain constant for a coherence interval of symbol periods is such that and are in the opposite
after which they change to another independent realization ac- order: i.e., , , if , then
cording to the spatial correlation model. For this system, we .
showed that the channel capacity is independent of the smallest Lemma 7: If , and are
eigenvalues of the transmit fade covariance matrix, as three vectors with components arranged in descending order,
well as the eigenvectors of the transmit and receive fade covari- i.e., if , and
ance matrices. We observed that the eigenvectors of the transmit , and if majorizes then the following
fade covariance matrix need to be known only to the transmitter is true:
while those of the receive fade covariance matrix need to be
known only to the receiver. We characterized the structure of (38)
the input signal that achieves capacity. The capacity achieving
transmit signal is expressed as the product of an isotropically
distributed unitary matrix, an independent nonnegative diagonal Lemma 5 is [13, Theorem .3.26]. Lemma 6 is Theorem 2 in
matrix and a unitary matrix whose columns are the eigenvectors [12]. Lemma 7 is Theorem 3 in [12].
of the transmit fade covariance matrix. Also, in contrast to the If we let be the vector of diagonal entries of arranged in
previously reported results for the spatially white fading model ascending order, then based on Lemma 6 we have
where adding more transmit antennas beyond the coherence in-
terval length does not increase capacity, we found (39)
that additional transmit antennas always increase capacity as
JAFAR AND GOLDSMITH: MULTIPLE-ANTENNA CAPACITY IN CORRELATED RAYLEIGH FADING 997
Now from Lemma 5 we have that the vector of diagonal entries [15] , Optimum power allocation and complete characterization of the
of , majorizes the vector of eigenvalues . Using impact of correlation on the capacity of MISO systems with different
CSI at the transmitter, in Proc. IEEE Int. Symp. Information Theory,
Lemma 7 we obtain 2003, p. 353.