1

Performance Analysis of ZF and MMSE

Equalizers for MIMO Systems: An In-Depth
Study of the High SNR Regime
Yi Jiang Mahesh K. Varanasi

Jian Li

Abstract
This paper presents an in-depth analysis of the zero forcing (ZF) and minimum mean squared error (MMSE)
equalizers applied to wireless multi-input multi-output (MIMO) systems with no fewer receive than transmit
antennas. In spite of much prior work on this subject, we reveal several new and surprising analytical results in
terms of the well-known performance metrics of output signal-to-noise ratio (SNR), uncoded error and outage
probabilities, diversity-multiplexing (D-M) gain tradeoff, and coding gain. Contrary to the common perception
that ZF and MMSE are asymptotically equivalent at high SNR, we show that the output SNR of the MMSE
equalizer (conditioned on the channel realization) is mmse = zf + snr , where zf is the output SNR of the
ZF equalizer, and that the gap snr is statistically independent of zf and is a non-decreasing function of input
SNR. Furthermore, as snr , snr converges with probability one to a scaled F random variable. It is also
shown that at the output of the MMSE equalizer, the interference-to-noise ratio (INR) is tightly upper bounded
by

snr
zf .

Using the decomposition of the output SNR of MMSE, we can approximate its uncoded error as well

as outage probabilities through a numerical integral which accurately reflects the respective SNR gains of the
MMSE equalizer relative to its ZF counterpart. The -outage capacities of the two equalizers, however, coincide
in the asymptotically high SNR regime, despite the non-vanishing gap snr . By analyzing a fictitious parallel
channel model with coding across the sub-channels in terms of the diversity-multiplexing (D-M) gain tradeoff,
we provide the solution to a long-standing open problem: applying optimal detection ordering does not improve
the D-M tradeoff of the V-BLAST (vertical Bell Labs layered Space-Time) architecture. However, by deriving
tight lower bounds to the outage probabilities of ZF and MMSE equalizers, we show that optimal ordering yields
a SNR gain of 10 log10 N dB in the ZF-V-BLAST architecture (where N is the number of transmit antennas)
whereas for the MMSE-V-BLAST architecture, the SNR gain due to ordered detection is even better, and
significantly so.
Keywords
This work was supported in part by the National Science Foundation Grant CCF-0423842 and CCF-0434410. This
paper was presented in part at Globecom 2005.
Y. Jiang and M. Varanasi are with the Department of Electrical and Computer Engineering, University of Colorado,
Boulder CO 80309, USA. Email: [email protected], [email protected]
J. Li is with the Department of Electrical and Computer Engineering, University of Florida, Gainesville, FL 32611-6130,
USA. Email: [email protected]

Zero forcing, minimum mean squared error, MIMO, error probability, V-BLAST, diversity gain, spatial
multiplexing gain, tradeoff, outage capacity, outage probability.

I. Introduction
Consider the complex baseband model for the wireless multi-input multi-output (MIMO) channel
with N transmit antennas and M receiver antennas
y = Hx + z,

(1)

where y CM 1 is the received signal and H CM N is a Rayleigh fading channel with independent,
identically distributed (i.i.d.), circularly symmetric standard complex Gaussian entries, denoted as
hij N (0, 1) for 1 i M, 1 j N . We assume that the number of receive antennas is no less
than the number of transmit antennas (M N ). We also assume that the N data substreams have
uniform power, i.e., x CN 1 has covariance matrix E[xx ] = x2 IN , where E[] stands for the expected
value, () is the conjugate transpose, and IN is an N N identity matrix. The white Gaussian noise
z N (0, z2 I) is also circularly symmetric. The input signal-to-noise ratio (SNR) is defined as
snr =

x2
.
z2

(2)

In this paper, we present an in-depth analysis of the performance of the zero forcing (ZF) and
minimum mean squared error (MMSE) equalizers applied to the channel given in (1). The linear
ZF and MMSE equalizers are classic functional blocks and are ubiquitous in digital communications
[1]. They are also the building blocks of more advanced communication schemes such as the decision
feedback equalizer (DFE), or equivalently, the V-BLAST (vertical Bell Labs layered Space-Time)
architecture [2][3], and various other MIMO transceiver designs (see, e.g., [4][5] and the references
therein). Despite their fundamental importance, however, the existing performance analyses of the ZF
and MMSE equalizers1 are far from complete. For instance, it is commonly understood that ZF is a
limiting form of MMSE as snr . But when the ZF and MMSE are applied to the MIMO fading
channel given in (1), one may observe through simulations that the error probabilities of MMSE and
ZF do not coincide even as snr . To the best of our knowledge, no rigorous account of such a
phenomenon is available in the literature. As another example, the problem of obtaining the exact
diversity-multiplexing (D-M) tradeoff [6] of V-BLAST with optimal detection ordering still remains
open, and so does the quantification of the gain due to optimal detection ordering. In this paper,
we attempt to provide an in-depth look at the classical ZF and MMSE equalizers with respect to the
well-known performance metrics of output SNR, uncoded error and outage probabilities, diversity1

In the sequel we refer to the ZF and MMSE equalizers as ZF and MMSE for simplicity.

multiplexing (D-M) gain tradeoff, and SNR gain.

The major findings of this paper are summarized in the following.
R1 A common perception about ZF and MMSE is that ZF is the limiting form of MMSE as snr .
Therefore, it is presumed that the two equalizers would share the same output SNRs, and consequently,
the same uncoded error or outage probability in the high SNR regime. We show, however, that the
output SNRs of the N data substreams using MMSE and ZF are related by
mmse,n = zf,n + snr,n ,

1 n N,

(3)

where zf,n and snr,n are statistically independent and snr,n is a nondecreasing function of snr. Moreover,
snr,n ,n

with probability one (w.p.1), as snr ,

(4)

M N +2
,n F2(N 1),2(M N +2) is of F-distribution.2 Further, the interference-to-noise
N 1
ratio (INR) of the nth substream at the output of MMSE (denoted as inrn ), is approximately upper
where

bounded as
inrn .

snr,n
.
zf,n

(5)

with the approximate upper bound being asymptotically tight for high SNR. Since

snr,n
zf,n

is inversely

proportional to the input SNR, (5) implies that the higher the input SNR, the smaller the leakage
from the interfering substreams.
R2 Using R1, we obtain tight approximations of the uncoded error and outage probabilities of MMSE
which can be evaluated via numerical integration rather than Monte-Carlo simulations. This analysis
also confirms that there is a non-vanishing SNR gain of MMSE over ZF as snr . Interestingly,
however, the -outage capacities of MMSE and ZF coincide in the asymptotically high SNR regime in
spite of the SNR gap between their outage probabilities.
R3 We obtain the following upper bounds of the output SNRs for the ZF and MMSE equalizers:
mmse,n

2N snr + 1
1
u

and zf,n

2N snr
,
u

(6)

where N is the smallest singular value of H and u is a Beta random variable that is independent of
N with a probability density function (pdf)
fu (x) = (N 1)(1 x)N 2 ,

0 x 1.

(7)

Given two independent Chi-square random variables a 2m and b 2n . The ratio c = a/m
is a random variable
b/n
n
m m 1
R
2
2
2
n
m
x
( m+n
z1 t
2 )
t
e dt. We denote c Fm,n [7].
with distribution fc (x) =
m+n , where (z) = 0
( m
n (n+mx) 2
2 ) (2)
2

Based on these upper bounds, we prove that for both ZF and MMSE, the D-M gain tradeoff of a
fictitious parallel channel (with N independent sub-channels) with coding across the N substreams is
the same as that for the ZF and MMSE equalizers applied to the MIMO channel with independent
coding over each individual substream, and this trade-off is given as

r
d(r) = (M N + 1) 1
.
N

(8)

That is, the SNR gain gap between the MMSE and ZF equalizers cannot be captured by the D-M
gain tradeoff analysis.
R4 As an important corollary of R3, we solve the well-known open problem on the diversity gain of
the V-BLAST architecture with optimal detection ordering [2]. Note that the V-BLAST architecture
can be regarded as employing ZF or MMSE equalizers combined with decision feedback [3], which in
the sequel are referred to simply as ZF-VB and MMSE-VB, respectively. We prove that with equal
rate for each substream and for any order of decoding, both ZF-VB and MMSE-VB have the D-M
gain tradeoff

r
dvb (r) = (M N + 1) 1
,
N

(9)

which means that the so-called V-BLAST order [2] does not yield an improvement in the D-M gain
tradeoff relative to unordered decoding.
R5 We also derive lower bounds on the outage probabilities of MIMO systems that use ZF and MMSE
(without decision feedback). The lower bounds are shown to be asymptotically tight for high SNR.
Based on these bounds, we prove that for ZF the strongest substream has a SNR gain of as much as
10 log10 N dB over an average one at high SNR. For MMSE, the SNR gain is even higher, and that too
by a significant margin. When applied to systems with decision feedback, as in V-BLAST, because
the overall outage probability is dominated by that of the first detected substream, this result also
quantifies the coding advantage of optimally ordered decoding over fixed order decoding.
The results R1 and R2 are on the distribution of the output SNR of the MMSE equalizer, the
asymptotic normality of interference-plus-noise at its output, and the coded (outage) and uncoded
error probability performance. Such problems are also investigated in [8][9] for the asymptotic property
of linear multiuser receivers. While their work focuses on large systems, we study finite systems with
asymptotically high SNR. The influence of non-Gaussian interference upon error probability in finite
CDMA systems is studied in [10] which shows that the larger an interfering users amplitude, the
smaller its effect on bit-error rate [10]. The (tight) upper bound of INR given in (5) yields more

insights into this observation. The output SINR decomposition (3) was proposed in the conference
version of this paper [11]. In the independent work [12], the authors show that such a decomposition
is possible even if the columns of H are correlated (but the rows need to be independent). In this
case the pdf of the output SINR mmse,n is very involved. The approach of [12] is to approximate
the first three asymptotic moments of snr,n as M, N , and then approximate it by a Gamma
(or generalized Gamma) random variable. Our strategy is to study the exact distribution of snr,n at
asymptotically high SNR, which leads to a more concise approximation.
The results R3, R4, and R5 are motivated by the problem of the D-M tradeoff of V-BLAST with
ordered decoding. Although this problem has inspired much research, previous attempts have only
achieved partial success and that too for the ZF-VB. For instance, it is shown in [13], [14] that optimal
ordering does not improve the diversity gain of ZF-VB but that it provides a 3 dB SNR gain when
there are two transmitting antennas (N = 2). The extension to the case of N 4 can be found in
[15]. It is also shown in [13] that a suboptimal column-norm ordering technique proposed in [16] does
not improve the diversity gain for arbitrary N . Note that a (loose) upper bound to the D-M tradeoff
of ZF-VB with optimal order detection is given in [6] to be

r
,
dvb (r) (M 1) 1
N

0 r N.

(10)

The difficulty of this problem lies in the fact that the distribution of the layer gains becomes extremely complicated due to the channel-dependent detection ordering. We circumvent this difficulty
by identifying the sharp upper bound given in (6). Indeed, the result R5 is also related to the bound
(6).
The rate/capacity performance of ZF and MMSE receivers applied to the point-to-point fading
MIMO channel are addressed in [17], where the authors show that the average capacity loss due to
using the linear ZF or MMSE equalizers converges to a constant as SNR increases. Similar conclusions
with regard to the sum rate are made in [18] and [19] in the context of the multi-access channel
(MAC) and the broadcast channel (BC). Combined with these results on rate/capacity performance,
this paper provides a more detailed picture of the performance of ZF and MMSE applied to both
single user and multiuser MIMO fading channels, especially in the high SNR regime.
The remainder of this paper is organized as follows. Section II introduces some preliminary results
to be used in the paper. In Section III, we analyze the output SNR of MMSE. Section IV derives the
uncoded error and outage probabilities of MMSE at high SNR. The D-M gain tradeoffs of the system

using ZF and MMSE are derived in Section V. Based on a tight lower bound to the outage probabilities
of the N substreams, we derive the SNR gain of optimal detection ordering for V-BLAST in Section
VI. Section VII presents the numerical examples validating the theoretical analysis. Conclusions are
made in Section VIII.
II. Preliminaries
A. Basics of ZF and MMSE Equalizers
Consider the MIMO channel model given in (1) where the N data substreams are mixed by the
channel matrix. The ZF and MMSE equalizers can be applied to decouple the N substreams. The ZF
and MMSE equalization matrices are (see, e.g., [20])
1

Wzf = (H H)

H ,

and Wmmse

1
1

= H H+
I
H .
snr

(11)

Left multiplying the received signal vector y by Wzf and Wmmse , we obtain N decoupled substreams
with output SNRs
zf,n =

snr
[(H H)1 ]nn

1 n N,

(12)

and
mmse,n =

[(H H

snr
1,
1
+ snr
I)1 ]nn

1 n N,

(13)

respectively. Here []nn denotes the nth diagonal element. Denote hn the nth column of H and Hn
the submatrix obtained by striking hn out of H. It follows from (12) and the fact (see, e.g., [21])
[(H H)1 ]nn =

1
hn hn

(14)

hn Hn (Hn Hn )1 Hn hn

that

zf,n = hn hn hn Hn (Hn Hn )1 Hn hn snr = (hn P

Hn hn )snr,

(15)

where P
Hn = I Hn (Hn Hn ) Hn stands for the orthogonal projection onto the null space of Hn . In
2
the case of i.i.d. Rayleigh fading, hn P
Hn hn 2(M N +1) , with distribution [22]

fhn P

Hn hn

(x) =

1
xM N ex ,
(M N )!

x 0.

Similarly, we have an alternative expression for mmse,n [11]:

"
#

1
1
mmse,n = hn hn hn Hn Hn Hn +
I
Hn hn snr,
snr

(16)

1 n N.

(17)

B. Diversity-Multiplexing Gain Tradeoff

In [6], the authors established the framework of D-M gain tradeoff analysis in the asymptotically
high SNR regime. Denote R(snr) as the data rate of any communication scheme with input SNR snr.
The diversity gain and multiplexing gain are defined as follows [6].
Definition II.1: A scheme is said to have multiplexing gain r and diversity gain d if the data rate
R(snr) satisfies
R(snr)
= r,
snr log snr
lim

(18)

and the average error probability Pe (snr) satisfies

log Pe (snr)
= d.
(19)
snr
log snr
Because Pe (snr) and R(snr) are related, so are d and r. We denote d(r) the tradeoff between the
lim

diversity gain and multiplexing gain, which is always a non-increasing function.

C. Two Theorems
The following two theorems turn out to be very useful for the analysis in this paper. The first
theorem is a slight variation of [23, Lemma 2.6].
Theorem II.2: Let H be an M N Gaussian matrix, whose entries are i.i.d. complex Gaussian
random variables with zero-mean and unit variance. With H = UV being the singular value
decomposition (SVD) of H we have that both U and V are Haar matrices

and they are statistically

independent of .
The second theorem is implied in [6].
Theorem II.3: For an M N i.i.d. Rayleigh fading channel matrix H with ordered squared singular
values of H, 21 22 2N > 0,
lim

0+

log P(2n < )

= (M n + 1)(N n + 1),
log

1 n N.

(20)

In other words,
P(2n < ) = (M n+1)(N n+1)+o(1) ,

1 n N.

where o(1) stands for a vanishing term as 0.

3

A random matrix is a Haar matrix if it is uniformly distributed on the set of unitary matrices.

(21)

III. Analysis of the Output SNR of MMSE

Since the elements of the channel matrix H are i.i.d., the output SNRs of the N substreams are of
identical (but not independent) marginal distributions. Hence, to study the distribution of the output
SNRs of the N substreams, we only need to focus on one, say the nth substream. As shown in (15)
and (16), zf,n is equal to snr multiplied by a Chi-square random variable. However, the distribution of
the output SNR of MMSE is more complicated. We start with analyzing the gap between the output
SNRs of ZF and MMSE.
It follows from (15) and (17) that the difference between mmse,n and zf,n , which we denote as snr,n
is

"
snr,n , mmse,n zf,n =

snrhn Hn

(Hn Hn )1

1 #
1

Hn Hn +
I
Hn hn .
snr

(22)

Since mmse,n = zf,n + snr,n , the characterization of mmse,n is given by the following theorem.
Theorem III.1: The random variable snr,n is statistically independent of zf,n . Moreover, as snr
, it converges to a scaled F random variable w.p. 1. In particular,
w.p.1

snr,n ,n ,

(23)

where
M N +2
,n F2(N 1),2(M N +2) .
(24)
N 1
Proof: Let Hn = Un n Vn be the SVD, where Un CM (N 1) and n C(N 1)(N 1) . Then
"

1 #
1
2
2

n Un hn
I
(25)
snr,n = snrhn Un n n n +
snr

1
1
= hn Un 2n +
I
Un hn .
(26)
snr
It is readily seen from (26) that given Hn and hn , snr,n is a non-decreasing function of snr. According
to the i.i.d. Rayleigh fading assumption, E[hn hn ] = I, thus
h
i

E[P
h
h
U
]
=
E
E
[P
h
h
U
]
= EHn [P
n
n
n
n
H
h
|H
n
n
Hn
n
Hn
Hn Un ] = 0,
n
n

(27)

where the last equality follows by the fact that P

Hn Un = 0 for any instantiation of Hn . Since

both P
Hn hn and Un hn are zero-mean Gaussian random vectors, (27) implies that PHn hn is sta
tistically independent of Un hn . Note that P
Hn hn is also independent of n , because PHn hn =

(I Un Un )hn with both hn and Un independent of n (cf. Theorem II.2). Hence P

Hn hn is independent of hn Un (2n +

1
1
snr I) Un hn ,

since the latter is a function of Un hn and n . Consequently,

10
2

2
zf,n = kP
Hn hn k snr is also independent of snr,n = hn Un (n +

1
1
snr I) Un hn .

Here k k stands for

the Euclidean norm of a vector.

It follows from (26) and the fact that the diagonal elements of 2n are all nonzero with probability
one (w.p. 1) that

lim snr,n = hn Un 2
n Un hn .

(28)

snr

Defining ,n , hn Un 2
n Un hn , we have shown that
w.p.1

snr,n ,n ,

as snr .

(29)

Because snr,n is independent of zf,n , so is its limit ,n .

We now derive the distribution of ,n . Denoting g = Un hn C(N 1)1 , we have that ,n =

g 2
n g, where g N (0, I) since E[gg ] = Un Un = IN 1 . Moreover, g and n are independent

since the singular matrix and singular values are independent (cf. Theorem II.2). Consider a matrix
G CM (N 1) which has the same dimension and distribution of Hn and is independent of g. Using
, we have g (G G )1 g = g V 2 V g. It is seen that
the SVD of G = UG G VG
n
G G
G

VG
g g,

G n

g is independent of . Consequently, we have

and VG
G

,n g (G G)1 g.

(30)

Construct a unitary matrix Ug such that Ug g = [0T , kgk]T . Note that Ug is hence a Householder
matrix [24]. Then
,n = [0T , kgk](Ug G GUg )1 [0T , kgk]T .

(31)

Observe that G G is statistically invariant under unitary transformations. Hence,

,n [0T , kgk](G G)1 [0T , kgk]T = kgk2 (G G)1 (N 1),(N 1) .

(32)

It is clear that kgk2 is a Chi-Square random variable with 2(N 1) degrees of freedom, i.e., kgk2
22(N 1) . According to (14) and (16),
1
[(G G)1 ](N 1),(N 1)
4

22(M N +2) .

By a b, we mean that the random variables a and b have identical distribution.

(33)

11

Hence, we have
,n

X
Y

(34)

where X 22(N 1) and Y 22(M N +2) , or equivalently that

M N +2
,n F2(N 1),2(M N +2) ,
N 1

(35)

M!
xN 2
,
(N 2)!(M N + 1)! (1 + x)M +1

(36)

with the pdf of ,n given as

f,n (x) =

0 x < .

Intuitively, ,n represents the power of the signal component hiding in the range space of Hn that
is recovered by the MMSE equalizer. In contrast, the ZF equalizer nulls out that signal component.
snr,n
For any full rank channel matrix,
0 as snr . Therefore, the interference from the other
zf,n
data substreams is negligible compared to the channel noise as snr . Consequently, for any full
rank channel realization, the ratio of the output SNR gains (in dB) of the MMSE to ZF equalizers
goes to unity or
10 log10

snr,n
mmse,n
= 10 log10 1 +
0,
zf,n
zf,n

as snr

In spite of the diminishing relative output SNR gain, the MMSE is shown to have remarkable SNR
gain over ZF even as snr owing to the fact that the limit of their difference is an F random
variable.
In the next section, we will provide applications of our analysis of mmse,n = zf,n + snr,n . It is
noted here that Theorem III.1 was originally presented in the conference version of this paper [11].
In the independent work of [12], the authors show that zf,n and snr,n are independent even if the
columns of H are correlated but with the rows of H being independent. However, in this case, the
exact distribution of ,n is unknown.
A. Interference-to-noise ratio (INR)
In recovering the signal xn in the range space of Hn , the MMSE equalizer admits some leakage
from the other interfering data substreams. It is shown in [10] that the leakage diminishes as input
power increases. A more careful study detailed in Appendix A shows that the INR at the output of
the MMSE equalizer is in fact inversely proportional to the input SNR.

12

Lemma III.2: The INR of the nth substream obtained using MMSE equalizer is upper bounded by
inrn .

snr,n
.
zf,n

(37)

This upper bound is asymptotically tight at high SNR.

Proof: See Appendix A.
IV. Applications of Theorem III.1
In this section, we apply Theorem III.1, i.e. the relationship mmse = zf,n + snr , to analyze the
uncoded error probability, outage probability, and -outage capacity of the MMSE equalizer. We shall
see that the gap snr brings about a remarkable difference in performance between the MMSE and
ZF with respect to the uncoded error probability and outage probability as snr . Interestingly
however, their -outage capacities coincide in the high SNR regime because this performance metric
depends only on the fact that the ratio of the output SNRs of the MMSE and ZF equalizers approaches
unity with increasing SNR.
A. Uncoded Error Probability Analysis
The uncoded error probability of the ZF equalizer is well known but we state it here for the sake of
completeness. Consider the input of binary phase-shift keying (BPSK). The error probability of the
nth substream obtained by ZF is (cf. (16))
Z
Pb,zf =
where the Q-function is Q(x) =

Q( 2snrx)

1
2

R
x

1
xM N ex dx,
(M N )!

(38)

t2

e 2 dt. The exact closed-form expression of Pb,zf is known

(see, e.g., [20]).

Pb,zf

snr
r
M N +1 M N

+1
1
+
X
M

N
+
n
1+snr
snr
1

.
1
=
2
1 + snr
2
n
n=0

(39)

We now consider the problem of analyzing the uncoded error probability for the MMSE equalizer.
Because the output SNRs of all the N substreams are of identical distribution, we only need to focus
on one substream.
For the error probability of MMSE, we assume that the sum of the perturbations due to the interference from the other data substreams and the channel noise can be well approximated as being
Gaussian. Consequently, the approximate error probability of MMSE equalizer can be calculated

13

through the Q-function (this Gaussian approximation is remarked on later):

p

Pb,mmse ' Emmse,n Q

2mmse,n .

(40)

It follows from (22) and (29) that

w.p.1

mmse,n zf,n + ,n ,

as snr .

(41)

p
2mmse,n around zf,n + ,n , we obtain

q

2
Q
2mmse,n = Q
2(zf,n + ,n ) +
(,n snr,n ) ,
2
2
where (mmse,n , zf,n + ,n ) and (x) = 12 exp x2 . Recall that [25]
Applying the Taylor expansion to Q

1 (x)
(x)
1 2
Q(x)
.
x
x
x
Therefore

( 2 )

(42)

(43)

Q( 2) at high SNR. Also note that zf,n + ,n and ,n snr,n 0 w.p.1

as snr . We can see from (42) that

q

Q
2mmse,n = Q
2(zf,n + ,n ) (1 + o(1))

w.p. 1.

(44)

Hence, at high SNR, we have that

p
E Q
2mmse,n
p
= 1,
lim
snr E Q
2(zf,n + ,n )
so that the error probability of MMSE can be further approximated as
q

Pb,mmse ' E Q
2(zf,n + ,n ) .

(45)

(46)

Since the distributions of zf,n and ,n are given in (16) and (36), respectively, (46) can be obtained
via numerical integration rather than Monte Carlo simulations. Invoking the alternative expression of
Q(x) [26], namely that
1
Q(x) =

Z
0

/2

exp

x2
2 sin2

d,

(47)

we have
Pb,mmse

" Z Z
#

2(snrx + ,n )
1 /2
1
M N x
' E,n
exp
x
e dx
d
0
(M N )!
2 sin2
0
#
"
Z Z /2
snrx
1
1
exp 2
d
xM N ex dx
< E,n e,n
0
(M N )!
sin
0

= E e,n Pb,zf

(48)
(49)
(50)

14
,n

where to obtain (49), we have used the fact that e,n e sin2 . Note that E [e,n ] is a constant
number strictly less than unity. The equation (50) shows the non-vanishing error probability gap
between ZF and MMSE even in the high SNR regime.
Calculating the error probabilities of a general quadrature amplitude modulation (QAM) is straightforward using the error probability expression in Q-function [27].
The representation of the error probability using the Q-function is based on the Gaussian approximation of the perturbation due to interference-plus-noise, and hence is not exact. In [10], Poor, et.
al. show that in some scenarios the error probability calculated based on the Gaussian approximation
is indiscernible from the exact one. It is explained essentially by observing that (i) the leakage from
the interfering substreams diminishes at high SNR and (ii) the interference term is dominated by the
noise at low SNR. In either case, the perturbation can be well-approximated by Gaussian noise. But
their work focused on non-fading channel. In fading channels, however, a rigorous justification for the
Gaussian approximation is still missing. Indeed, we have observed through extensive simulations that
for rank-deficient channel realizations there is a non-negligible discrepancy between the Q-function
approximation and the actual one, especially for channels with low dimensionality (say, M = N = 2).
Despite this phenomenon, the Gaussian approximation is still quite accurate in terms of average error
probability for channels that are full rank with probability one, a fact that is verified in a numerical
example given in Section VII.
B. Outage Probability and -Outage Capacity
Consider employing independent codes of rate R each over the N antennas. The nth antenna
transmission is in outage if the output SNR cannot support the target rate R. With the ZF equalizer
the outage probability of the nth substream is

zf
Pout,n

= P (log(1 + zf,n ) < R) = F2

2(M N +1)

2R 1
snr

(51)

15

where F2

2(M N +1)

(x) = 1 ex

outage probability of MMSE is

M
N
X
k=0

xk
2
is the cumulative density function (cdf) of 2(M
N +1) . The
k!

mmse
Pout,n
= P (log(1 + zf,n + snr,n ) < R)
R

Z 2R 1
2 1
=
F2
fsnr,n ()d
2(M N +1)
snr
0
R

Z 2R 1
2 1
'
F2
f,n ()d.
2(M N +1)
snr
0

We impose the upper limit 2R 1 on the integration because F2

2(M N +1)

(52)

(x) = 0 when x < 0. Inserting

(36) into (52), we calculate the outage probability of MMSE using numerical integration. Note that
F2

2(M N +1)

(x) = e

X
k=M N +1

xk
x(M N +1)
=
+ o x(M N +1)
k!
(M N + 1)!

around the origin. It follows from (52) that for high SNR (snr 2R 1),
R
Z 2R 1
M N +1
2 1

mmse
Pout,n F2
1 R
f,n ()d
2(M N +1)
snr
2 1
0
M N +1
Z 2R 1

zf
1 R
f,n ()d.
= Pout,n
2 1
0

(53)

mmse and P zf
Given a fixed rate R, there is a non-vanishing gap between Pout,n
out,n even as snr .

Moreover, as we can observe from (53), the gap would become smaller as R increases. This phenomenon
is validated in Section VII on numerical results.
The -outage capacity is the maximum supportable rate under the restriction that the outage probability is no greater than , and is defined as
Czf () , sup {R : P(log(1 + zf,n ) < R) }

(54)

and
Cmmse () , sup {R : P(log(1 + mmse,n ) < R) }
= sup {R : P(log(1 + zf,n + snr,n ) < R) } ,

(55)

respectively. Since the cdfs of both zf,n and mmse,n are continuous. It is easy to show that the cdfs
of log(1 + zf,n ) and log(1 + mmse,n ) are also continuous. Therefore Czf and Cmmse are the solutions
to P(log(1 + zf,n ) < R) = and P(log(1 + mmse,n ) < R) = , respectively. Now we obtain that

1
() ,
(56)
Czf () = log 1 + snrF2
2(M N +1)

16

where F1
2

2(M N +1)

is the inverse function of the cdf of 22(M N +1) thus satisfying

Z

F 1
2

()

2(M N +1)

1
xM N ex dx = .
(M N )!

(57)

For Cmmse (), we examine the relationship:

P (log(1 + mmse,n ) < Czf ())

1
()
= P zf,n + snr,n < snrF2
2(M N +1)

snr,n
1
2
()
= P 2(M N +1) < F2
snr
2(M N +1)

(58)

By the continuity of the cdf,

lim P 22(M N +1) < F1

2

snr

snr,n
()
snr
2(M N +1)

= P 22(M N +1) < F1

2

() = ,

w.p.1.

2(M N +1)

Hence we have that limsnr P (log(1 + mmse,n ) < Czf ()) = w.p. 1, and consequently,
lim Cmmse () Czf () = 0,

snr

w.p.1,

(59)

i.e., the MMSE equalizer has the same -outage capacity as ZF at asymptotically high SNR.
The non-vanishing SNR gap between the outage probabilities of the zero-forcing and MMSE equalizers and the result in (59) may seem contradictory at first. The explanation for this apparent contradiction is that the difference between the outage probabilities of the two equalizers vanishes as the
rate increases such that the difference between the maximum rates achievable with ZF and MMSE
such that outage probability is less than a fixed threshold shrinks to zero with increasing SNR, a
phenomenon we illustrate again in the section on numerical results.
V. Analysis of D-M Gain Tradeoff
In this section, we obtain the exact D-M gain tradeoffs of the linear ZF and MMSE receivers
when independent, equal rate (and equal power) SISO Gaussian codebooks are employed over the N
antennas. As a by-product of this analysis, we further infer that no channel dependent ordering of
substream decoding can improve the D-M tradeoff of V-BLAST.
A. The Linear ZF Equalizer
Consider the MIMO system that employs independent coding for each substream and the ZF equalizer at the receiver. Each substream effectively experiences a scalar channel whose gain is of 22(M N +1)

17

distribution. With the overall multiplexing gain r, each substream has a multiplexing gain

r
N.

The

system is in outage if and only if at least one substream is in outage. Hence, the overall system outage
probability is given as

Pout,zf (snr) = P log 1 +

snr

[(H H)1 ]nn

r
<
log snr, for some n ,
N

(60)

where P(E) denotes the probability of event E. We can bound (60) by

P log 1 +

snr
[(H H)1 ]11

r
<
log snr
N

N
X
Pout,zf (snr)
P log 1 +
n=1

snr
[(H H)1 ]nn

r
<
log snr .
N
(61)

Because the output SNRs of the N substreams have identical distributions, the rightmost expression
of (61) is

Since

N P log 1 +

snr

[(H H)1 ]11

< r log snr .

log N snr
0, according to (19), the diversity gain
log snr

snr
log
P
log
1
+
<
log Pout,zf (snr)
[(H H)1 ]11
dzf (r) = lim
= lim
snr
snr
log snr
log snr

r
1
1
N
<
snr
log P [(H H)
1 ]
11
= lim
snr
log snr

Using the fact that

(62)

r
N

log snr

(63)
(64)

1
has pdf given in (16), one can easily derive from (64) that (see [20])
[(H H)1 ]11

r
dzf (r) = (M N + 1) 1
.
(65)
N

B. The Linear MMSE Equalizer

Analogously to (60), the outage probability is given as

!
!
snr
r
Pout,mmse (snr) = P log
log snr, for some n ,
<
1
N
[(H H + snr
I)1 ]nn

(66)

Since the MMSE equalizer has a higher output SNR than the ZF equalizer (by snr,n , a non-negative
random variable), we have that Pout,mmse Pout,zf . Therefore, it must be true that the D-M gain
tradeoff for MMSE is better than that for ZF, i.e.,

log Pout,mmse (snr)

r
(M N + 1) 1
.
snr
log snr
N

dmmse (r) , lim

Interestingly, equality in the above inequality holds, as we shall see next.

(67)

18

We first prove the following lemma on the distribution of the elements of a column of a Haar matrix
and their minimum value.
1
T
Lemma V.1: The joint pdf of {|vn |2 }N
n=1 , where v , [v1 v2 vN ] is a column vector of a Haar

matrix V CN N , is

for 0 |xn |2 1 and 0

f|v1 |2 , ,|vN 1 |2 (x1 , . . . , xN 1 ) = (N 1)!,

N
1
X

|vn |2 1.

(68)

n=1

The marginal pdf of |vn |2 for each n is

f|vn |2 (x) = (N 1)(1 x)N 2 ,

0 x 1,

(69)

Define u , min{|vn |2 , n = 1, , N.}. Then u has pdf

fu (x) = (N 2 N )(1 N x)N 2 ,

0x

1
.
N

(70)

Consequently, for any fixed n, |vn |2 has the same distribution as N u, i.e., |vn |2 N u.
Proof: As v is a column of a Haar matrix,
v

,
kk

where N (0, IN ) is a complex-valued circularly symmetric Gaussian vector. Denote Xi , |i |2 .

Then Xi s are i.i.d. with an exponential distribution, i.e., fXi (x) = ex for x 0. Consider the
conditional joint distribution of Xi for 1 i N 1 given Y , kk2 . Using the fact that Y =
PN
2
i=1 Xi 2N , we obtain
fX1 ,...,XN 1 |Y (x1 , . . . , xN 1 |y) =
=
As |vi | =

|i |2
kk2

Xi
Y ,

fY |X1 ,...,XN 1 (y|x1 , . . . , xN 1 )f (x1 , . . . , xN 1 )

fY (y)
e

PN

i=1

xi

1
N 1 ey
(N 1)! y

= (N 1)!y (N 1) .

(71)

it follows from (71) that

f|v1 |2 ,...,|vN 1 |2 |Y (x1 , . . . , xN 1 |y) = (N 1)!.

(72)

1
Since the joint pdf of {|vn |2 }N
n=1 is independent of Y , we have proven (68). Note that the random

vector [|v1 |2 , . . . , |vN 1 |2 ] has a uniform distribution over the simplex

)
(
N
1
X
1
2
|vn |2 1 ,
S = {|vn |2 }N
n=1 : 0 |vn | 1,

(73)

n=1
5

1
2 N
Here we present the joint pdf of {|vn |2 }N
n=1 rather than {|vn | }n=1 because the latter is a degenerated function since

PN

i=1

|vi |2 = 1.

19

which has volume Vol(S) =

1
(N 1)! .

By the property of symmetry, |vn |2 s have identical distribution. Note that |vn |2 has the same
distribution as

Z1
Z1 +Z2

where the two independent Chi-square random variables Z1 22 and Z2

22N 2 . Hence |vn |2 is of Beta distribution with parameter (1, N 1) (see, e.g., [28, pp. 60]):
f|vn |2 (x) = (N 1)(1 x)N 2 ,

0 x 1.

(74)

According to the definition u , min{|vn |2 , n = 1, , N.}, we have

P(u > x) = P(|vn |2 > x, 1 n N ).

(75)

It is not difficult to verify that the set {|vn |2 > x, 1 n N } corresponds to a subset of S:
(
)
N
1
X
1
1
2
{|vn |2 }N
|vn |2 > x , 0 x ,
n=1 : |vn | > x, , 1
N

(76)

n=1

which is a smaller simplex with volume (To see this point, note that the side length of this smaller
simplex is 1 N x.)
(1 N x)N 1 Vol(S).

(77)

P(u > x) = (1 N x)N 1 .

(78)

It follows from (75) and (77) that

Thus the pdf of u is

fu (x) =

dP(u > x)
= N (N 1)(1 N x)N 2 ,
dx

0x

1
.
N

(79)

Comparing (74) and (79), we see that |vn |2 has the same distribution as N u for all n.
Besides the MIMO system considered in the beginning of this section, which employs independent
coding for each substream, we also consider a system where a single SISO Gaussian code is applied
across N substreams. With a linear MMSE equalizer, the D-M gain tradeoff of the latter system is

P
N
log(1
+

)
<
r
log
snr
log P
mmse,n
n=1
dmmse (r) , lim
,
(80)
snr
log snr
while D-M tradeoff of the former system is dmmse (r) as defined in (67).
The following theorem establishes that the equality holds in (67).
Theorem V.2:

r
dmmse (r) = dmmse (r) = (M N + 1) 1
N

(81)

20

Proof: It is easy to see that (cf. (66))

Pout,mmse (snr) P

N
X

!
log(1 + mmse,n ) < r log snr .

(82)

n=1

Consequently,
dmmse (r) dmmse (r).

(83)

Let us denote the sum of the mutual informations between channel input and the output over the N
substreams of an MMSE equalizer as
Immse =

N
X

log(1 + mmse,n ) =

n=1

N
X

log

n=1

snr
1

[(H H + snr
I)1 ]nn

!
.

(84)

Let H H = V2 V be its SVD with 21 2N as the ordered diagonal entries of 2 . We can

rewrite (84) as
Immse =

N
X

log

n=1

snr

vn 2 +

1
1
vn
snr I

(85)

where vn is the nth column V . Recall from Theorem II.2 that V is a Haar matrix and is independent
of . Since
vn

1
N
X
1
2
+
I
vn =
|vni |2 (2i + snr1 )1 |vnN |2 (2N + snr1 )1 ,
snr

(86)

i=1

where vni is the ith element of vn , the mutual information of the nth substream can be bounded as
!

1 + 2N snr
snr
log
, n = 1, , N,
(87)
log
1

1
|vnN |2
vn 2 + snr
I
vn
and hence
Immse

N
X

log

n=1

1 + 2N snr
|vnN |2

Hence, we can upper bound the D-M gain tradeoff

P
2 < r log snr
log
|v
|
log P N log(1 + 2N snr) N
nN
n=1
dmmse (r) lim
.
snr
log snr

(88)

(89)

Since [v1N v2N vN N ]T is a column of the Haar matrix V , |vnN |2 has the pdf given in (69), we
can obtain after some routine calculations that
E[log |vnN |2 ] =

N
1
X
n=1

Because |vnN |s have the same pdf,

"
E

N
X

n=1

#
log |vnN |2 = N

1
.
n

N
1
X
n=1

1
, E,
n

(90)

21

which is a negative constant. Next, consider the following set inequality which allows us to further
lower bound Pout,mmse (snr),
(
2
N , {|vnN |2 }N
n=1 : N log(1 + N snr)

N
X

)
log |vnN |2 < r log snr

n=1

2
N , {|vnN |2 }N
n=1 : N log(1 + N snr) < r log snr + E

N
X

)
2

log |vnN | > E

(91)

n=1

Using the statistical independence between N and {|vnN |2 }N

n=1 we have

!
N
X
P N log(1 + 2N snr)
log |vnN |2 < r log snr
n=1

N
!
X

2
2
> P N log(1 + N snr) < r log snr + E P
log |vnN | > E .

(92)

n=1

Hence, (89) can be further bounded as

dmmse (r) lim

N
2 >E
log
|v
|
log P N log(1 + 2N snr) < r log snr + E + log P
nN
n=1
log snr

snr

Since P

(93)

2 > E is a positive constant independent of snr, we have

log
|v
|
nN
n=1

P
N

log P N log(1 + 2N snr) < r log snr + E

dmmse (r) lim

snr
log snr

log P N log(1 + 2N snr) < r log snr

= lim
snr
log snr

r
2
log P N < snr N 1
.
= lim
snr
log snr

(94)

By Theorem II.3, we obtain the upper bound

r
dmmse (r) (M N + 1) 1
,
N

1 r N.

(95)

Combining this with (67) and the bound (83), we have that the D-M tradeoff of MMSE is

r
dmmse (r) = dmmse (r) = (M N + 1) 1
N

(96)

and the theorem is proved.

At first glance, the conclusion that even considering the lower bound on outage probability in (82)
does not improve system diversity gain is rather surprising. Since the N substreams usually have
distinct output SNRs (even if they have identical marginal distributions), it seems unlikely that all the
N substreams are in outage simultaneously. Note that Immse is the mutual information rate realized

22

in a parallel channel that results from fixing the receiver front-end to be the linear MMSE equalizer
with the transmitter using this knowledge to code across the N substreams. One might expect to
achieve an N -fold diversity gain with such coding compared to using independent coding across the
antennas. Theorem V.2 however implies that the output SNRs of the N substreams are actually highly
correlated. This point can be seen from (87); if N snr1 , then the mutual informations in all the
N substreams tend to be small.
Corollary V.3: For both ZF and MMSE, the N substreams, ranking from the strongest to the
weakest, have diversity gain of their individual outage probabilities to be all equal to M N + 1.
Proof: We first consider the MMSE case. Suppose that the strongest substream, with output
SNR max , has diversity gain dmax > M N + 1. According to (96), dmmse (0) = M N + 1. Denote
P
the mutual information Immse (snr) , N
n=1 log(1 + mmse,n ) log(1 + max ). Then
lim

snr

log P (Immse (snr) < const)

log P (log(1 + max ) < const)
lim
= dmax > M N + 1,
snr
log snr
log snr

where const stands for a finite constant. Note that the leftmost term of the above equation is equal to
dmmse (0) = M N + 1, which leads to a contradiction. Thus dmax M N + 1. On the other hand,
suppose the weakest substream has diversity gain dmin < M N + 1. With the same but independent
coding applied to all the substreams, the overall system outage probability would be dominated by
that of the weakest substream, which is of order snrdmin . It implies that dmmse (0) < M N + 1, which
also leads to a contradiction. Therefore dmin M N + 1, and we conclude that all the substreams
must have the same diversity gain M N + 1 for the linear MMSE equalizer. The argument for the
case of ZF is straightforward given the above.
C. D-M Gain Tradeoff of V-BLAST with Channel-Dependent Ordered Detection
Based on Corollary V.3, we are ready to answer the long standing open question as to what really
is the D-M tradeoff of V-BLAST with channel-dependent ordered decoding.
In contrast to the linear equalizers which decode the N substreams simultaneously, the V-BLAST
equalizer applies successive nulling and interference cancellation to recover the substreams one by one.
At each step, the V-BLAST estimates one substream according to the criteria of ZF or MMSE and then
eliminates the estimated component from the received data. Hence at the next step, the substream
to detect is subject to one less interferer [29]. We refer to the V-BLAST based on the ZF and MMSE
criteria as the ZF-VB and MMSE-VB, respectively. The output SNRs of the substreams estimated by

23

the ZF-VB or MMSE-VB are closely related to applying the QR decomposition to the channel matrix.
In particular, denoting the QR decomposition H = QR, the ZF-VB yields N substreams with output
SNRs (cf. [3], [13])
2
zfvb,n = rnn
snr

for

where rnn , n = 1, . . . , N are the diagonal entries of R.

1 n N,

(97)

Similarly, the MMSE-VB yields N substreams

with [29], [30]

2
mmsevb,n = rnn
snr 1 for

1 n N,

(98)

yielded by the QR decomposition

where rnn , n = 1, . . . , N are the diagonal entries of R

H
snr

12

=
I

R.
Because the N substreams employ independent SISO codes, the substream corresponding to the
Q
lowest output SNR is the bottleneck of the overall system [31]. One remedy of this undesirable effect
is to apply channel-dependent ordering [32][2]. The channel dependent ordering can be represented
by a permutation matrix (this is actually a function of H but we dont write (H), for simplicity)
and the output SNRs of the substreams obtained via ordered V-BLAST are therefore dependent on

H
R.
The permutation matrix is chosen
= Q
the QR decompositions H = QR and
21
snr I
2 } (or min
2 } in the MMSE-VB case) is maximized among the N !
such that min1nN {rnn
rnn
1nN {
2
2 } and r 2
2
2
permutations. Define rmax
, max min1nN {rnn
max
and
min , min min1nN {rnn }. And r
2
2 replaced by r
2 .
rmin
are similarly defined with rnn
nn

Then with any ordered detection the D-M tradeoffs of ZF-VB and MMSE-VB are sandwiched by

2 snr) < r log snr

2 snr) < r log snr
log P log(1 + rmin
log P log(1 + rmax
N
N
lim
dzfvb (r) lim
snr
snr
log snr
log snr
(99)
and

2 snr) <
log P log(
rmin
lim
snr
log snr

r
N

log snr

2 snr) <
log P log(
rmax
dmmsevb (r) lim
snr
log snr

r
N

log snr

(100)
respectively. It is not difficult to show that with detection ordering
2
rN
N =
6

1
[(T H H)1 ]N N

2
and rN
N =

[(T H H

To make the QR decomposition unique, the diagonal entries of R are positive.

1
+

1
1
snr I) ]N N

(101)

24
2
2
By choosing different , rN
N
N and r
N can take on one of N different values:
2
rN
N =

1
[(H H)1 ]nn

2
and rN
N =

1
[(H H

,
1
1
snr I) ]nn

1 n N.

(102)

Comparing (97) (98) to (12) (13) and invoking (102), we see that the first detected substream using
ZF-VB (or MMSE-VB) has the output SNR taken from zf,n (or mmse,n ), for 1 n N . Define
2
rN
N,max , max1nN

1
[(H H)1 ]nn

2
and rN
N,min , min1nN

1
,
[(H H)1 ]nn

2
2
while rN
N
N,max and r
N,min

are defined similarly. We show next that

2
2
rmax
rN
N,max

2
2
and rmin
= rN
N,min .

(103)

By definition,
2
2
2
rmax
= max min {rnn
} max rN
N = max

1nN

1nN [(H H)1 ]nn

2
= rN
N,max .

(104)

Hence the first inequality of (103) is proven. Again by definition, we can prove that
2
2
2
2
rmin
= min min {rnn
} min rN
N = rN N,min .
1nN

(105)

2
2
2
2
Moreover, suppose rmin
< rN
N,min , which means that for some permutation , rmin = rnn for some

matrix such that the nth column of H is

n 6= N . Then we can always find a new permutation
and the QR decomposition H
= QR
yields r2 r2 = r2 <
moved to the N th column of H,
nn
min
NN
2
2
2
rN
N
N,min , where r
N rnn because moving the column to the right always reduces its corresponding

rii as it has more interference to suppress. Now we have reached a contradiction and hence proven
2
2
that rmin
= rN
N,min . Using the same argument, we can prove that
2
2
rmax
rN
N,max

2
2
and rmin
= rN
N,min .

(106)

Let us first focus on the MMSE-VB case. Recall that the V-BLAST applies the same but independent
coding. Therefore, with spatial multiplexing gain r, the outage probability of V-BLAST is [31]

r
2
P log( min rnn
snr) <
log snr .
1nN
N
According to (100) and (106), we obtain that

r
2
2
log P log(
rN
snr)
<
log
P
log(
rN
log
snr
N,min
N,max snr) <
N
lim
dmmsevb (r) lim
snr
snr
log snr
log snr

r
N

log snr

(107)

25

or equivalently,
lim

snr

r
1
2
N
log P rN

snr
N,min
log snr

dmmsevb (r) lim

r
1
2
N
log P rN

snr
N,max
log snr

snr

(108)

By Corollary V.3, we have

r
r
1
1
2
2
N
N

log P rN

snr
log
P
r

snr
N N,max
N,min
r
= lim
= (M N + 1) 1
.
lim
snr
snr
log snr
log snr
N
(109)
Hence, for any channel-dependent detection ordering

r
dmmsevb (r) = (M N + 1) 1
,
N

0 r N.

(110)

Clearly, in similar vein, we can obtain

r
dzfvb (r) = (M N + 1) 1
,
N

0 r N.

(111)

Now we have established the following theorem.

Theorem V.4: For both ZF-VB and MMSE-VB with any channel-dependent detection ordering, the
D-M gain tradeoff of the overall system is

r
dvb (r) , dmmsevb (r) = dzfvb (r) = (M N + 1) 1
, 0 r N.
(112)
N
Theorem V.4 stands for the final answer to the long-standing open problem on whether optimal
ordering in V-BLAST improves system diversity gain. Our answer is no.
The result that the maximal diversity gain of the ZF-VB is M N + 1 even with detection ordering
was first established in the conference version of this paper [11]. This result was also reached by Zhang
et. al. in [33], where they further predict that the whole diversity multiplexing tradeoff curve will
not be improved by optimal ordering. However, their result can not be extended to the MMSE-VB
case.
We conclude this section with the following corollary.
Corollary V.5: In the asymptotically high SNR regime, the overall outage probability of V-BLAST
is dominated by that of the first detected layer for any detection ordering, .
Proof: Suppose there exists an ordering technique T which yields the n-th (n 2) detected layer
with diversity gain D M N + 1, i.e., its outage probability Pout (snr) snrD in the high SNR
regime. With a random detection ordering, there is

1
N!

chance that the random ordering coincides

with T . Hence the outage probability of the n-th layer is no less than

1
N ! Pout (snr)

and therefore its

26

diversity gain is no greater than D M N + 1. However it is well known that the V-BLAST with
random detection ordering yields the n-th detected layer with diversity gain M N + n [14], which
is strictly greater than D, which leads to a contradiction. Hence for any ordering technique, the n-th
(n 2) detected layer has diversity gain strictly greater than M N + 1. Since the first detected
layer has diversity gain M N + 1 by Theorem V.4, the corollary has been proven
VI. A Closer Look: Outage Probability and Coding Gain
In this section, we consider the case where independent coding is applied to the N substreams. We
analyze the outage probability of the substreams. Despite the pessimistic conclusion of Section V with
respect to the diversity gain, we show that there is a remarkable SNR gain due to applying the optimal
detection ordering in the V-BLAST architecture which we quantify next.
To facilitate the analysis, we rewrite the output SNRs of ZF
zf,n =

snr
[(H H)1 ]nn

(113)

and MMSE
mmse,n =

snr
1.
[(H H + snrI)1 ]nn

(114)

According to (86) we have the upper bound

mmse,n

snr
1.
+ snr1 )1

|vnN |2 (2N

(115)

Similarly, we obtain
zf,n

2N snr
.
|vnN |2

(116)

Given the target rate R of the nth substream, its outage probability is
zf
Pout,n
(R, snr) = P (log(1 + zf,n ) < R)
!

snr
R
= P PN
<2 1
2 2
i=1 |vni | i
!

snr
R
zf
P
2 < 2 1 , Pout,n (R, snr).
2
|vnN | N

(117)
(118)

if ZF is used, and for MMSE

mmse
Pout,n
(R, snr) = P (log(1 + mmse,n ) < R)

!
snr
R
= P PN
<2
2 2
1 1
i=1 |vni | (i + snr )

snr
R
mmse
P
(R, snr),
<2
, Pout,n
|vnN |2 (2N + snr1 )1

(119)
(120)

27

With regard to the lower bounds (118) and (120) , we have the following lemma.
Lemma VI.1:
lim

zf
(R, snr)
Pout,n

= 1,

(121)

mmse (R, snr)

Pout,n
mmse (R, snr) = 1.
snr Pout,n

(122)

snr

zf
Pout,n
(R, snr)

and
lim

That is, the lower bounds in (118) and (120) are asymptotically tight at high SNR.
Proof: The proof is rather technical. We relegate it to Appendix B.
We are now ready to establish the following theorem which quantifies the SNR gain that accrues by
applying the optimal decoding ordering to the V-BLAST architecture.
zf
Theorem VI.2: Given the input SNR snr and the target rate R for each substream. Denote Pout,min
(R, snr)
mmse (R, snr) the outage probabilities of the substreams with the highest output SNR obtained
and Pout,min

using ZF and MMSE, respectively. Then in the high SNR regime,

and

snr
zf
zf
' Pout,n
(R, snr),
Pout,min
R,
N

n,

(123)

snr
mmse
mmse
Pout,min
R,
< Pout,n
(R, snr),
N

n.

(124)

where the approximation in (123) is asymptotically accurate as snr (in the sense that the limit
of the ratio of the two probabilities tends to unity with increasing SNR). Comparing to the fixed
decoding order, applying the optimal decoding order yields 10 log10 N dB SNR gain for ZF-VB, and
more than 10 log10 N dB SNR gain for MMSE-VB.
Proof: Due to (121) and (122), we can closely approximate the outage probability of the nth
substream by

zf
Pout,n
(R, snr)

'P

and

mmse
Pout,n
(R, snr)

'P

2N
2R 1
<
|vnN |2
snr

2N + snr1
2R
<
|vnN |2
snr

zf
, Pout,n
(R, snr),

mmse
, Pout,n
(R, snr),

at high SNR. Define u , min{|vnN |2 , 1 n N }. Then

2

N
2R 1
zf
zf
Pout,min (R, snr) ' P
<
, Pout,min
(R, snr),
u
snr
and

mmse
Pout,min
(R, snr)

'P

2N + snr1
2R
<
u
snr

(125)

(126)

(127)

mmse
, Pout,min
(R, snr).

(128)

28

Recall from Lemma V.1 that |vnN |2 N u, n, which implies that

2
2
N
N
2R 1
2R 1
P
<
=P
<
.
u
snr/N
|vnN |2
snr

(129)

It follows from (129), (127), and (125) that

and hence

snr
zf
zf
Pout,min
R,
= Pout,n
(R, snr),
N

(130)

snr
zf
zf
Pout,min
R,
' Pout,n
(R, snr).
N

(131)

Because the approximations in (127) and (125) are asymptotically accurate as snr , so is the
approximation in (131). Therefore with respect to ZF, we can conclude that the strongest substream
has 10 log10 N dB SNR gain over an average one. Because V-BLAST applies independent coding to
each layer, V-BLAST is in outage if and only if at least one layer is in outage. It is known from
Corollary V.5 that in the high SNR regime, the outage events of V-BLAST are dominated by those of
the first detected layer. Consequently, the ZF-VB with the optimal detection ordering has 10 log10 N
dB SNR gain over ZF-VB with fixed detection ordering.
As for MMSE, according to (128),

snr
snr
mmse
< 2R u
Pout,min
= P 1 + 2N
R,
N
N

2
= P N + N snr < 2R |vN n |2

mmse
< P 1 + 2N snr < 2R |vN n |2 = Pout,n
(R, snr).

(132)
(133)
(134)

As the approximations in (126) and (128) are asymptotically accurate at high SNR, we can see that
snr
mmse
mmse
Pout,min
R,
< Pout,n
(R, snr),
N

(135)

which implies that the SNR gain of strongest substream over an average substream is more than
10 log10 N dB. Moreover, it can be seen that the gap between (133) and (134) gets larger as N
increases.
Finally, we remark that the key fact used to determine the coding gain advantage of ordered
detection is that |vnN |2 N u, where vnN is the (n, N )th entry of the unitary matrix V , and
u = min{|vnN |2 , n = 1, , N }. According to Lemma V.1, |vnN |2 N u as long as V is a Haar
matrix. If the columns of H are independent, then the distribution of H is invariant under the right
multiplication of a unitary matrix. Hence V is a Haar matrix. We see that Theorem VI.2 still holds
even the rows of H are correlated but the columns of H are statistically independent.

29

VII. Numerical Examples

In this section, we present several numerical examples to validate the preceding theoretical analysis.
Figure 1 presents the quantile-quantile (qq) figures for the distribution of
M N +2
N 1

M N +2
N 1 snr

against

which is of F-distribution (see 35). We see that the F -distribution approximation works

very well for all cases in the high SNR regime. We also see that the approximation is less accurate
when N = M . The explanation is as follows. It is known that for M = N , the channel matrix tends
to be more ill-conditioned thus the smallest diagonal entry of n (see (26)) is close to zero. Therefore
the convergence of (26) to the limit (28) is slower. For the case of N = M/2, the F -distribution
approximation is very accurate even for a moderate SNR of 10 dB.
M = 4, N = 4
15

SNR = 30 dB
SNR = 20 dB
SNR = 10 dB
45 degree line

10

SNR = 30 dB
SNR = 20 dB
SNR = 10 dB
45 degree line

12
Empirical quantitles

Empirical quantitles

M = 16, N = 16
14

10
8
6
4
2

10
Fquantiles

15

5
10
Fquantiles

M = 4, N = 2

M = 16, N = 8
4

Empirical quantitles

Empirical quantitles

SNR = 20 dB
SNR = 10 dB
45 degree line

8
6
4
2
0

4
6
Fquantiles

Fig. 1. Quantile-quantile plots for

M N +2
N 1 snr .

SNR = 20 dB
SNR = 10 dB
45 degree line

2
Fquantiles

The range of the quantiles is 1% - 99%.

In the second example, we consider an i.i.d. Rayleigh channel with M = 5 and N = 4. We calculate
the INRs in the output of MMSE based on 104 Monte Carlo trials. Figure 2 shows the INRs at different
input SNRs, in which each curve represents the 104 INRs sorted in the non-decreasing order. This
simulation result agrees with Lemma III.2; the INR is inversely proportional to the input SNR.
Figure 3 compares the uncoded error probabilities of ZF and MMSE equalizers in a 44 channel with

30

M = 5, N = 4

10

20 dB
1

10

25 dB
30 dB

INR

10

35 dB
40 dB

10

10

10

10

2000

4000

6000

8000

10000

Fig. 2. INRs in the output of MMSE with input SNR equal to 20, 25, 30, 35, and 40 dB. The results are based
on 104 Monte Carlo trials of the channel matrix. M = 5, N = 4.

BPSK input. We see that the error probabilities of MMSE obtained via averaging over 105 Monte
Carlo simulations match extremely well with the high SNR approximation of (46) for a moderate
SNR (snr 10 dB). Moreover, as predicted in (50), there is a non-vanishing gap between the error
probability curves of MMSE and ZF.
Figure 4 compares the outage probabilities of ZF and MMSE equalizers in an i.i.d. Rayleigh channel
with M = 6 and N = 5. We consider the three cases where the target rates are one, two, and
four bps/Hz per substream. Similar to Figure 3, the outage probabilities of MMSE obtained via
averaging over 107 Monte Carlo simulations (represented by +) match exactly at high SNR with
the approximation in (52) which is obtained via numerical integration (represented in solid lines).
Although for fixed R MMSE has a non-vanishing SNR gain over ZF, the gap becomes smaller as
R increases. Figure 4 also illustrates that the -outage capacity is the same for MMSE and ZF at
asymptotically high SNR.
In the fifth example, we consider a channel with M = N = 3. Suppose N independent capacityachieving codes are applied to each substream, and the target rate is R = 3 bps/Hz for each substream.
mmse
mmse , P zf
zf
, Pout,n
The four solid lines in Figure 5, from top to bottom, represent Pout,n
out,min , and Pout,min ,

respectively. The four dashed lines underneath the solid lines are the corresponding lower bounds,
zf
mmse , P zf
mmse
i.e., Pout,n
, Pout,n
out,min , and Pout,min . We see from Figure 5 that (i) all the lower bounds are

31

M=N=4
1

10

10

Pb

10

10

MMSE (high SNR approx.)

ZF
MMSE (simulation)

10

10

10

15

20
SNR

25

30

35

40

Fig. 3. Error probabilities of ZF given in (39) and MMSE (Monte Carlo trials and high SNR approximation
given in (46)). M = N = 4.
M = 6, N = 5

10

Outage Probability

10

R=4

10

R=2
3

10

R=1
4

10

10

10

10

10

20

30

40

SNR (dB)

Fig. 4.

Outage probabilities of ZF given in (51) (represented by dashed lines) and MMSE via Monte Carlo

trials (+) and high SNR approximation given in (52) (solid line). M = 6, N = 5.
zf
zf
is 10 log10 N = 4.77
asymptotically tight as SNR increases, (ii) the gap between Pout,min
and Pout,n
mmse over P mmse is significantly larger than 4.77 dB. This numerical
dB, (iii) the coding gain of Pout,min
out,n

32

example agrees with Theorem VI.2.

0

10

Pzf
outage,n
1

Pmmse
outage,n

Outage Probability

10

Pzf
outage,min
2

10

Pmmse
outage,min
4.77
dB

10

10

10

15

20
SNR (dB)

25

30

35

40

zf
mmse
zf
mmse
. The solid lines are the true values and the
, and Pout,min
, Pout,min
Fig. 5. Outage probabilities Pout,1
, Pout,1

dash lines are lower bounds. M = N = 3. The result is obtained via averaging over 106 Monte Carlo trials of
the channel matrix.

Figure 6 adopts the same simulation layout as that of Figure 5 except that the channel dimensionality
zf
zf
is changed to M = N = 5. As we can see from Figure 6, the gap between Pout,min
and Pout,1
is
mmse over P mmse is even more significant compared to
10 log10 N = 6.99 dB, and the SNR gain of Pout,min
out,1

Figure 5. This result agrees with our analysis of (133) and (134).
In the final example, we consider an i.i.d. Rayleigh channel with M = N = 4. We compare the outzf
mmse ), their lower bounds (P zf
mmse
age probabilities of the strongest substreams (Pout,min
, Pout,min
out,min , Pout,min ),

along with the outage probabilities of ZF-VB and MMSE-VB with optimal detection ordering. It is
zf
mmse are also tight lower bounds to the
seen from Figure 7 that the lower bounds to Pout,min
and Pout,min

outage probability of ZF-VB and MMSE-VB with optimal detection ordering, respectively. Hence
theorem VI.2 can be used to predict the SNR gain of ordered detection in V-BLAST architecture.
This example also validates Corollary V.4; the detection ordering cannot improve the diversity gain,
which is M N + 1 = 1 in this example, although MMSE-VB manifests higher diversity gain in the
low to moderate SNR regime (snr 15 dB).

33

M=N=5

10

zf
outage,n

P
1

10

Pmmse
Outage Probability

outage,n

zf

10

Poutage,min
6.99
dB

10

10

10

Pmmse
outage,min

10

10

15

20
SNR (dB)

25

30

35

40

zf
mmse
zf
mmse
. The solid lines are the true values and the
, and Pout,min
, Pout,min
Fig. 6. Outage probabilities Pout,1
, Pout,1

dash lines are lower bounds. M = N = 5.

M=N=4

10

ZFVB
Pzf

outage,min

Pzf

Outage Probability

10

outage,min

ZF

MMSEVB
Pmmse

outage,min

Pmmse

10

outage,min

10

MMSE

10

10

10

15

20

25

30

35

40

SNR (dB)
zf
mmse
Fig. 7. Outage probabilities Pout,min
, and Pout,min
, ZF-VB and MMSE-VB with optimal ordering. The dash
mmse
zf
and Pout,min
. M = N = 4. The result is obtained via averaging over 107
lines are the lower bounds Pout,min

Monte Carlo trials.

VIII. Conclusions
In this paper, we have analyzed the performances of the zero forcing (ZF) and minimal mean squared
error (MMSE) equalizers applied to an M N wireless multi-input multi-output (MIMO) systems,

34

in terms of output SNR, uncoded error and outage probabilities, diversity-multiplexing (D-M) gain
tradeoff, and SNR gain. We show that there is a gap between the output SNRs of ZF and MMSE
equalizers, which converges with probability one to a random variable of scaled F-distribution as
input SNR goes to infinity. Based on this result, we can accurately approximate the uncoded error
probability of MMSE equalizer via numerical integration rather than time-consuming Monte-Carlo
simulations. For coded systems, we show that although given fixed target rate MMSE has a nonvanishing SNR gain over ZF, the -outage capacities of MMSE and ZF coincide in the asymptotically
high SNR regime. We also prove that even with perfect coding across the N substreams, the D-M gain

r
. As
tradeoff of the MIMO system using either ZF or MMSE equalizer is d(r) = (M N + 1) 1
N
an important corollary, we prove that the V-BLAST equalizer (vertical Bell Labs layered Space-Time)
has a maximal diversity gain of M N + 1 even with optimal order detection. Moreover, we show
that for the ZF equalizer, the strongest substream has 10 log10 N dB SNR gain over an average one.
For MMSE, this SNR gain is much larger than even that. This analysis also quantifies the SNR gain
of applying ordered detection in V-BLAST architecture.
Appendix A
Proof of Lemma III.2
According to the matrix inverse lemma,
"
1
1 #

1
1
I
I + HH
H H +
= snr I H
H .
snr
snr
Therefore we can rewrite the MMSE filter matrix given in (11) by
"

1 #
1

Wmmse = snr I H
I + HH
H H
snr
1

1
I
.
= H HH +
snr

(136)

Denote wn the nth row of Wmmse , which is the MMSE nulling vector for the nth substream. Then

1
1
1 1

wn = HH +
I
hn (Hn Hn +
I) hn ,
snr
snr

(137)

where hn and Hn are defined as in (14). If we normalize wn such that wn hn = 1, then

wn =

1
(Hn Hn + snr
I)1 hn
.
1
hn (Hn Hn + snr
I)1 hn

(138)

35

Without loss of generality, we assume here that x2 = 1 and z2 =

1
snr .

Applying wn to the received

data vector y (cf. (1)), the power of noise is

Pns,mmse =

1
hn (Hn Hn + snr
I)2 hn
1
kwn k2 =
,
1
snr
snr|hn (Hn Hn + snr
I)1 hn |2

(139)

and the sum power of noise and interference from the other N 1 substreams is
Pns,mmse + Pintf,mmse =

wn

1
1

Hn Hn +
I w=

snr
hn (Hn Hn +

1
1
snr I) hn

(140)

From (140) and (139), we have the ratio

1
snrhn (Hn Hn + snr
I)1 hn
Pns,mmse + Pintf,mmse
=
.
1
Pns,mmse
hn (Hn Hn + snr
I)2 hn

(141)

Denote Hn = Un n Vn the SVD of Hn , where Un CM (N 1) and n R(N 1)(N 1) . Then the

SVD of Hn Hn +

1
snr I

is

2 + 1 I
.
1
n ] n snr N 1
I = [Un ..U
Hn Hn +
snr
0

0
1
snr IM N +1

.
[Un ..U
n] .

(142)

Now we can rewrite

1
1
1

2
hn ,
nU
Hn Hn +
I
hn = hn Un n +
I
Un hn + snrhn U
n
snr
snr

(143)

2
1
1

2
hn .
nU
Hn Hn +
I
hn = hn Un n +
I
Un hn + snr2 hn U
n
snr
snr

(144)

hn
and
hn

nU
= P , we obtain from (141) that
Applying (143) and (144) the fact that U
n
Hn
2
1
hn P
Pns,mmse + Pintf,mmse
Hn hn + snr hn Un n +
=

2
2
Pns,mmse
hn P
Hn hn + snr hn Un n +

1
1
Un hn
snr I

2
1
Un hn
snr I

(145)

and
Pintf,mmse
Pns,mmse

=
.

1
2
1
1
Un hn snr2 hn Un 2n + snr
Un hn
I
snr I

2
1
2
2
Un hn
hn P
Hn hn + snr hn Un n + snr I
1
hn Un (2n + snr
I)1 Un hn
snr hn P
Hn hn

snr1 hn Un 2n +

(146)
(147)

It is easy to see that the upper bound of (147) is asymptotically tight as snr . We note that the
numerator and the denominator in (147) are exactly snr,n (see (26)) and zf,n (see (15)), respectively.
The Lemma is proven.

36

Appendix B
Proof of Lemma VI.1
We first prove (121). According to (117), we can rewrite

2
1
zf
Pout,n
(R, snr) = P N 2
2
P

N 1
|vnN | 1 + N
|vnN

i=1

|2

|vni |2
2i

<

2R

1
.
snr

(148)

Because 2n s are in non-increasing order, we have

N
1
X
i=1

N
1
X
|vni |2
|vni |2
1

2 .
2
2
n

N 1
i=1 N 1

Combining (148) and (149) yields

zf
Pout,n
(R, snr) P

(149)

2N

|vnN

|2

1+

2N
1
|vnN |2 2N 1

2R

snr

<

(150)

zf
= Pout,n
(R, snr) + P

2N
|vnN

|2

>

1 2
, N 1 <
snr

2N
|vnN |2

2R

2N
snr
2 | 2R 1
|vnN

We now focus on the second term of the right hand side of (151)

2N
2
R1
2

2
|v
|

P N 2 >
, 2N 1 < 2 nN
snr
|vnN |
snr
N
1
2 |
|vnN

= P

2N

2R

1
<
snr
|vnN |2

+P

2N
|vnN

|2

> 1+

< 1+
1
log snr

2N
2R 1
<
<
snr
|vnN |2

1
log snr

1+

(152)

2R 1

1 2
, N 1 <
snr

1 2
, N 1 <
snr

1
log snr

2R 1
snr

2N
snr
2 | 2R 1
|vnN

2N
snr
2 | 2R 1
|vnN

P
Second, because

2N
2R 1
<
snr
|vnN |2

< 1+

1
log snr

2N
snr
2 | 2R 1
|vnN

(153)

2R 1
2
+ P N 1 < (1 + log snr)
(154)
snr

2R 1
snr

2N
|vnN |2

2N
|vnN |2

To obtain (154) from (153), we have used the following two facts. First,

R
2N
1
2 1 2
2R 1

<
< 1+
, N 1 < 2
P
2
snr
|vnN |
log snr
snr
N

2N
|vnN |2

2R

2R

(151)

snr
2 | 2R 1
|vnN

2N
|vnN |2

1
(155)

37

is a decreasing function of

2N
,
|vnN |2

for
2N
>
|vnN |2

we have

2N
|vnN |2
2N
2 |
|vnN

snr
2R 1

1+

<
1+
1

1+

1
log snr

1
log snr

1
log snr

2R 1
,
snr

2R 1
snr

2R 1 snr
snr 2R 1

= (1 + log snr)

2R 1
.
snr

Hence,

2
P N 2 >
|vnN |

1
log snr

2R 1 2
, N 1 <
snr

1+

2N
|vnN |2
2N
snr
2 | 2R 1
|vnN

1
2 1 2
2R 1
, N 1 < (1 + log snr)
log snr
snr
snr

R
2 1
< (1 + log snr)
.
snr

2N
>
|vnN |2

P 2N 1

1+

(156)

Therefore using the inequalities (155) and (156), we obtain (154) from (153).
Combining (151) and (154), we have

zf
Pout,n
(R, snr)

2N
<
|vnN |2

1+

1
log snr

2R 1
snr

2R 1
2
+ P N 1 < (1 + log snr)
snr

(157)

According to Theorem II.3, as snr , the second term of the above equation
R
2(M N +2)+o(1)

(2 1)(1 + log snr)

2R 1
=
,
P 2N 1 < (1 + log snr)
snr
snr

(158)

while the first term

2N
<
|vnN |2

1+

1
log snr

2R 1
snr

2
> P N < c 1 +

= K

where c is some finite positive constant, say c =

1
2N ,

2R 1
snr

1
log snr

1+

1
log snr

2R 1
P(|vnN |2 > c)
snr
M N +1+o(1)
,
(159)

and K = P(|vnN |2 > c) cM N +1 is also a finite

positive constant.
It follows from (158) and (159) that given fixed R,

R 1
P 2N 1 < (1 + log snr) 2 snr
2

= 0.
lim
N
snr
2R 1
P |vnN
<
2
snr
|

38

Hence
zf
Pout,n
(R, snr)
lim
zf
snr P
out,n (R, snr)

lim

< 1+

1
log snr

< 1+

1
log snr

snr

2N
|vnN |2

lim

2N
|vnN |2

snr

lim 1 +

snr

2N
|vnN |2

1
log snr

<

R 1
+ P 2N 1 < (1 + log snr) 2 snr
2

N
2R 1
P |vnN
<
2
snr
|
R

2R 1
snr

2R 1
snr

2 1
snr

M N +1

= 1.

(160)

On the other hand, according to (118)

lim

zf
Pout,n
(R, snr)

snr

zf
Pout,n
(R, snr)

1.

(161)

Combining (160) and (161), we have proven (121).

The techniques used above in the proof for (121) can be equally applied to prove (122). We omit it
for simplicity.
IX. Acknowledgement
The first author thanks Dr. X. Zheng for the helpful discussions at the early stage of this research
work.
References
[1]

J. G. Proakis, Digital Communications. McGraw-Hill Inc., Third Edition, 1995.

[2]

G. J. Foschini, G. D. Golden, R. A. Valenzuela, and P. W. Wolniansky, Simplified processing for high spectral
efficiency wireless communication employing multiple-element arrays, Wireless Personal Communications, vol. 6,
pp. 311335, March 1999.

[3]

G. Ginis and J. M. Cioffi, On the relationship between V-BLAST and the GDFE, IEEE Communications Letters,
vol. 5, pp. 364366, September 2001.

[4]

D. Palomar, J. Cioffi, and M. Lagunas, Joint Tx-Rx beamforming design for multicarrier MIMO channels: A unified
framework for convex optimization, IEEE Transactions on Signal Processing, vol. 51, pp. 23812401, September
2003.

[5]

Y. Jiang, J. Li, and W. Hager, Uniform channel decomposition for MIMO communications, IEEE Transactions
on Signal Processing, vol. 53, pp. 4283 4294, November 2005.

[6]

L. Zheng and D. Tse, Diversity and multiplexing: A fundamental tradeoff in multiple-antenna channels, IEEE
Transactions on Information Theory, vol. 49, pp. 10731096, May 2003.

[7]

G. Casella and R. L. Berger, Statistical Inference. Duxbury Press, 2001.

[8]

J. Zhang, E. K. P. Chong, and D. N. C. Tse, Output mai distribution of linear mmse multiuser receivers in ds-cdma
systems, IEEE Transactions on Information Theory, vol. 47, p. 2001, March 2001.

39

[9]

D. Guo, S. Verdu, and L. K. Rasmussen, Asymptotic normality of linear multiuser receiver outputs, IEEE Trans.
Inform. Theory, vol. 48, pp. 30803095, December 2002.

[10] H. V. Poor and S. Verdu, Probability of error in MMSE multiuser detection, IEEE Transactions on Information
Theory, vol. 43, pp. 858871, May 1997.
[11] Y. Jiang, X. Zheng, and J. Li, Asymptotic analysis of V-BLAST, Proceedings of IEEE GlobeCom, St. Louis, MO,
November 2005.
[12] P. Li, D. Paul, R. Narasimhan, and J. Cioffi, On the distribution of SINR for the MMSE MIMO receiver and
performance analysis, IEEE Transactions on Information Theory, vol. 52, pp. 271286, January 2006.
[13] N. Prasad and M. K. Varanasi, Analysis of decision feedback detection for MIMO rayleigh fading channels and
the optimization of rate and power allocations, IEEE Transactions on Information Theory, vol. 50, pp. 10091025,
June 2004.
[14] S. Loyka and F. Gagon, Performance analysis of the V-BLAST algorithm: an analytical approach, IEEE Transactions on Wireless Communications, vol. 3, pp. 13261337, July 2004.
[15] S. Loyka and F. Gagon, Analytical framework for outage and BER analysis of the V-BLAST algorithm, Int.
Zurich Seminar on Communications (IZS), pp. 120123, Feb. 18-20 2004.
[16] M. O. Damen, K. Abed-Meriam, and S. Burykh, Iterative QR detection for an uncoded space-time communication
architecture, Proc. Allerton Conf. Communications, Control, and Computing, October 2000.
[17] X. Zhang and S. Kung, Capacity analysis for parallel and sequential MIMO equalizers, IEEE Transactions on
Signal Processing, vol. 51, pp. 29893002, November 2003.
[18] B. Hochwald and S. Vishwanath, Space-time multiple access: linear growth in the sum rate, Proceedings: Allerton
Conference on Communications, Control, and Computing, Monticello, IL, October 2002.
[19] N. Jindal, High SNR analysis of MIMO broadcast channels, Proc. IEEE Int. Symp. Information Theory, Adelaide,
Australia, Sept. 2005.
[20] D. Tse and P. Viswanath, Fundamentals of Wireless Communications. Cambridge Press, 2005.
[21] R. A. Horn and C. R. Johnson, Matrix Analysis. Cambridge: Cambridge University Press, 1985.
[22] T. W. Anderson, An Introduction to Multivariate Statistical Analysis. second edition, John Wiley and Sons, Inc.,
1984.
[23] A. M. Tulino and S. Verdu, Random Matrix Theory and Wireless Communications. Hanover, MA 02339, USA: now
Publishers Inc., 2004.
[24] G. H. Golub and C. F. Van Loan, Matrix Computations. Baltimore, MD: Johns Hopkins University Press, 1983.
[25] H. L. Van Trees, Detection, Estimation, and Modulation Theory, Part I. New York, NY: John Wiley and Sons, Inc.,
1968.
[26] J. Craig, New, simple and exact result for calculating the probability of error for two-dimensional signal constellations, Proc. Milcom, 1991.
[27] K. Cho and D. Yoon, On the general BER expression of one and two-dimensional amplitude modulations, IEEE
Transactions on Communications, vol. 50, pp. 10741080, July 2002.
[28] S. M. Ross, Introduction to Probability Models, Seventh Edition. HarCour Academic Press, 2000.
[29] M. Varanasi and T. Guess, Optimum decision feedback multiuser equalization with successive decoding achieves

40

the total capacity of the Gaussian multiple-access channel, Proceedings of the Thirty-First Asilomar Conference on
Signals, Systems and Computers, vol. 2, pp. 1405 1409, Nov 2-5 1997.
[30] B. Hassibi, A fast square-root implementation for BLAST, Thirty-Fourth Asilomar Conf. Signals, Systems and
Computers, pp. 12551259, November 2000.
[31] N. Prasad and M. K. Varanasi, Outage capacities of space-time architecture, Information Theory Workshop, San
Antonio, Texas, October 24-29 2004.
[32] M. Varanasi, Group detection for synchronous gaussian code-division multiple-access channels, IEEE Transactions
on Information Theory, vol. 41, pp. 10831096, July 1995.
[33] H. Zhang, H. Dai, Q. Zhou, and B. L. Hughes, On the diversity-multiplexing tradeoff for ordered SIC receivers
over MIMO channels, IEEE International Conference on Communications (ICC), Istanbul, Turkey, June 2006.