Layered Space-Time Architecture for

Wireless Communication in a Fading

Environment When Using Multi-Element
Antennas
Gerard J. Foschini
This paper addresses digital communication in a Rayleigh fading environment when
the channel characteristic is unknown at the transmitter but is known (tracked) at
the receiver. Inventing a codec architecture that can realize a significant portion of
the great capacity promised by information theory is essential to a standout longterm position in highly competitive arenas like fixed and indoor wireless. Use (nT , nR )
to express the number of antenna elements at the transmitter and receiver. An (n, n)
analysis shows that despite the n received waves interfering randomly, capacity
grows linearly with n and is enormous. With n = 8 at 1% outage and 21-dB average
SNR at each receiving element, 42 b/s/Hz is achieved. The capacity is more than 40
times that of a (1, 1) system at the same total radiated transmitter power and bandwidth. Moreover, in some applications, n could be much larger than 8. In striving for
significant fractions of such huge capacities, the question arises: Can one construct
an (n, n) system whose capacity scales linearly with n, using as building blocks n separately coded one-dimensional (1-D) subsystems of equal capacity? With the aim of
leveraging the already highly developed 1-D codec technology, this paper reports
just such an invention. In this new architecture, signals are layered in space and time
as suggested by a tight capacity bound.

Introduction
This paper describes a new point-to-point communication architecture employing an equal number
of antenna array elements at both the transmitter and
receiver. The architecture is designed for a Rayleigh
fading environment in circumstances in which the
transmitter does not have knowledge of the channel
characteristic. This new communication structure,
termed the layered space-time architecture, targets application in future generations of fixed wireless systems,
bringing high bit rates to the office and home. The
architecture might also be used in future wireless local
area network (LAN) applications for which it promises
extraordinarily high bit rates.
The architecture is a method of presenting and

processing higher dimensional signals with the aim of

leveraging the already highly developed one-dimensional (1-D) codec technology. Note that in this context, higher dimensional refers to space. (Generally, a
bandwidth-efficient 1-D code involves many dimensions over the temporal domain. 1-D refers to a complex
alphabet which is, of course, 2-D in terms of reals.)
As the paper points out, the capacity that this
architecture enables is enormous. At first, the number
of bits per cycle might seem too great to be meaningful. The capacity is achieved, however, in terms of n
equal lower component capacities, one for each
antenna at the receiver (or transmitter). A form of the
new architecture attains a capacity equal to a tight

Bell Labs Technical Journal Autumn 1996

41

Panel 1. Abbreviations, Acronyms, and Terms

AWGNadditive white Gaussian noise
BERbit error rate
codeccoder/decoder
iidindependent identically distributed
LANlocal area network
MEAmulti-element array
MMSEminimum mean square error
SNRsignal-to-noise ratio
lower bound on the capacity attainable using multielement arrays (MEAs) with an equal number of elements at both ends of the link. The next section
describes this lower bound on capacity. Subsequently,
the layered space-time architecture is discussed.
Although additional background details are available,1 this paper provides a self-contained description
of the architecture. The perspective is one of a complex baseband view of signaling over a fixed linear
matrix channel with additive white Gaussian noise
(AWGN). Time proceeds in discrete steps, normalized
so that t = 0, 1, 2, ... . The following notation and basic
assumptions should be reviewed:
Number of antennas. The MEA at the transmitter has nT. The MEA at the receiver has nR. For
convenience, the pair (nT, nR) denotes a communication system with nT transmit elements
and nR receive elements. Figure 1 illustrates
the notation.
Transmitted signal s(t). This signal has fixed narrow bandwidth. The total power is constrained
$ regardless of n (which is the dimension
to P
T
of s(t)). The bandwidth is narrow enough that
the channel frequency characteristic can be
treated as flat across the band.
Noise at receiver (t). This is the complex nRdimensional AWGN. The components are statistically independent and of identical power N
at each of the nR antenna outputs.
Received signal r(t). At each point in time, this is
an nR-dimensional signal. There is one complex vector component per receive antenna.
With each transmit antenna transmitting
$ n , P denotes the average power at
power P/
the output of each receiving antenna, with

42

Bell Labs Technical Journal Autumn 1996

average meaning spatial average.

Average signal-to-noise ratio (SNR) at each receive
antenna. This is = P/N, independent of nT.
Matrix channel impulse response g(t). This matrix
has nT columns and nR rows. The notation h(t)
is used for the normalized form of g(t). The
normalization is such that each element of h(t)
has a spatial average power loss of unity.
The basic vector equation describing the channel
operating on the signal is
r = g * s + ,
(1)
where * means convolution. These three vectors are
complex nR-dimensional vectors (2nR real dimensions). Because of the narrowband assumption, the
channel Fourier transform G is treated as a matrix
constant over the band of interest. Thus, g is written
for the nonzero value of the channel impulse
response, thereby suppressing the time dependence of
g(t). The same is true for h and its Fourier transform
H. Thus, in normalized form, (1) becomes

r=

/ nT h s + .

(1a)

The random channel model we use is the Rayleigh

channel model. Assume that the MEA elements at
each end of the communication link are separated by
about half a wavelength. At 5 GHz, for example, half a
wavelength measures only about 3 cm, so many
antenna elements are often possible. (Additionally,
there are two states of polarization [see Figure 2]).
With a half-wavelength separation, the Rayleigh
model for the nR ! nT matrix H representing the channel in the frequency domain is approximated by a
matrix having the following independent identically
distributed (iid), complex, zero-mean, unit-variance
entries:
Hi j = Normal 0 ,1 / 2

+ 1 Normal 0,1 /

2 .

|H ij | 2 is a chi-squared variate with two degrees

2
of freedom denoted by 2 but normalized so
E|Hij | 2 = 1.

Capacity
The viewpoint assumed in treating capacity is discussed next. We stress that capacity is a limit to errorfree bit rate that is provided by information theory,

Transmitter

Receiver

(9, 12)

ssor

Proce

Figure 1.
Shorthand notation (nT , nR ) for the number of transmit antennas nT and receive antennas nR (dipole antennas shown).

and this limit can only be approached in practice with

the advance of technology: any working system can
only achieve a bit rate (at some desired small bit error
rate [BER]) that is only a fraction of capacity. In what
follows, the term capacity will often be used as an
indicator of some smaller deliverable bit rate.

Long-Burst Perspective
Communication in long bursts means bursts having many symbolsso many that an infinite time
horizon information-theoretic description of communication portrays a meaningful idealization. Yet bursts
are assumed to be of short enough duration that a
channel is essentially unchanged during a burst. The
channel is assumed to be unknown to the transmitter
but learned (tracked) by the receiver. The channel
might change considerably from one burst to the next.
By a channel being unknown to the transmitter,
we mean that the realization of H during a burst is
unknown. Actually, the average SNR value and even
n R might not be known to the transmitter.
Nonetheless, for purposes of this discussion, these two
parameters are considered to be known. The reason
for this is that at the transmitter, one assumes that
communication is taking place with a user for which
at least a certain nR and average SNR are available.
These minimum values represent what the transmitter conservatively uses to determine a capacity value

that is nearly always available.

In a given system, not all communication bursts
are successful. As explained below, for some small percentage of instances of H, the transmitters assumed
capacity value may be too optimistic. In such cases,
delivering the bit rate at the desired BER required of a
successful burst may be impossible. When it is impossible, a channel outage is said to have occurred and the
channel is considered to be in the OUT state.
Outage is dealt with probabilistically because H is
random; thus, capacity is a random variable. The
channel is random even though the base and user in
an office LAN environment or the communication
sites in fixed wireless applications may be fixed. In
actuality, the reason for this is that such sites are only
nominally fixed because perturbations of the communicators and the communication medium are possible.
For indoor LANseven for a user at a desk
some motion in and around the workspace is likely.
Not only people but various (especially metal) objects
could be moving in the propagation path. For predominately outdoor fixed wireless links, weatherrelated motion of antenna structures occurs, as well as
significant channel changes due to, say, vehicles and
foliage. Half-wavelength movements can be important. Thus, assuming that the channel is fixed during
a burst, the channel may vary from burst to burst,

Bell Labs Technical Journal Autumn 1996

43

Detail shows monopole antenna

elements for two polarization states.

Section of available
surface of a base
or laptop

Dielectric
surface

/2
/2

/4

Metal
surface

/2

Dielectric
layer

/2

For example, at 5 GHz, /2 3 cm

Figure 2.
Section of casing paved with half-wavelength lattice.

and one might be interested in the capacity that can

be attained in nearly all transmissions (for example,
95 to 99% or even higher).
Complementary capacity distributions, discussed
below, focus on the high-probability tail. (In special
applications like very large file transfers, however,
maximum attainable throughput over long time durations may be a preferable figure of merit.) A companion article1 mentions that based on the results of
research,2,3 the transmitter can use a single code even
though the specific value of the H matrix is unknown.
The distribution of capacity is derived from an
ensemble of statistically independent Gaussian nR ! nT
H matrices (the aforementioned Rayleigh model). In
this paper, the system is considered to be either OUT
or NOT OUT for each realization of H. As mentioned
earlier, the OUT state corresponds to the event that a
prespecified capacity level (for example, X) cannot be
met. For instance, given a 1% outage level, one would
say a certain capacity can be assured at that level.
By employing MEAs, capacity tail probabilities can
be significantly improved. The subsection
Opportunity for Enormous Bit Rates below discusses
the great capacity available and how the tail probability improves with larger and larger n.
(For cases in which the time constant of channel
change is very large, extra receive antenna elements
may be needed to ensure that outage is minimal (see

44

Bell Labs Technical Journal Autumn 1996

the end of the Conclusion section). In very severe

situations, provision for movement of the receive
antenna may be desirable to avoid the risk of being
stuck with an undesirable H for excessive time.
Deployment of a relay site is yet another alternative.
Fading correlation time and its incorporation into
more refined performance criteria is an interesting
subject for future investigation in measurement and
analytical studies.)

Key Capacity Expressions

A generalized capacity formula and a capacity
lower-bound formula are referred to below.1 The generalized formula is derived from other basic formulas.4-6 Additionally, the capacity formula for optimum
ratio combining is needed.
The generalized capacity formula for the general
(nT, nR) case is

C = log 2 det I n + / n T HH b/s/Hz . (2)

In this equation, det means determinant, In is the

R
nR ! nR identity matrix and means transpose
conjugate.
The capacity lower bound for the (n, n) case in
terms of n independent chi-squared variates is
n

C>

log
k= 1

1 + ( / n) 2k b/s /Hz .

(3)

99 percentile (Prob[OUTAGE] = .01).

Average signal-to-noise ratio at each
receive antenna is 0, 6, 12, 18, and 24 dB.

300

10

24 dB

0 dB

200

Capacity (b/s/Hz)

8
18 dB

150
12 dB

100

0 dB

0
0

10

20

30
(a)

40

50

6
4
12 dB
18 dB
24 dB

60

10

20

24 dB
6

18 dB

12 dB
6 dB

0 dB
20

30

50

60

0.25
6 dB

0.20
0.15
0.10

12 dB

0.05
24 dB

0
10

40

99 percentile (Prob[OUTAGE] = .01).

Average signal-to-noise ratio at each
receive antenna is 6, 12, 18, and 24 dB.

0.30

30
(c)

99 percentile (Prob[OUTAGE] = .01).

Average signal-to-noise ratio at each
receive antenna is 0, 6, 12, 18, and 24 dB.

10

6 dB

6 dB

50

Capacity (b/s/Hz/dimension)

Capacity (b/s/Hz)

250

Capacity (b/s/Hz/dimension)

99 percentile (Prob[OUTAGE] = .01.

Average signal-to-noise ratio at each
receive antenna is 24, 18, 12, 6, and 0 dB.

40

50

60

10

20

18 dB
30

40

50

60

(d)

(b)

The number of transmit antennas equals the number of receive antennas.

Capacities
Capacity lower bounds

Figure 3(a).
Capacity in b/s/Hz versus the number of antenna elements at each site.
Figure 3(b).
Capacity in b/s/Hz/dimension versus the number of antenna elements at each site.
Figure 3(c).
Capacity in b/s/Hz versus the number of antenna elements at each site.
Figure 3(d).
Capacity in b/s/Hz/dimension versus the number of antenna elements at each site.

Notice that some nonstandard notations have been

later in an asymptotic sense, the bound in (3) for large

2
usedfor example, 2k to denote directly a chi-

and n is quite tight. While (3) was initially derived

squared variate with 2k degrees of freedom. Because

elsewhere,1 the subsection A (6, 6) Example of

the entries of H are zero mean unit variance complex

Processing at the Receiver below includes a rederiva-

Gaussians, the mean of this variate is k. As discussed

tion of (3) that is constructive. That is, the right-hand

Bell Labs Technical Journal Autumn 1996

45

Primitive
data stream
DEMUX
Equal
rates
Layer 1
(mod/code)

Layer 2
(mod/code)

Layer n
(mod/code)

Modulo-n shift of layer-antenna

correspondence every seconds

Antennas

Fading noisy matrix channel

Figure 4.
Transmission process using space-time layering.

side will be associated with the capacity of a limiting

form of a communication architecture using 1-D
codecs.
A special application of (2)important to this discussionis the case of optimum combining, which
corresponds to (1, n) receive diversity. OC(n) is written to convey a system employing optimum combining with n-fold receive diversity.
The capacity formula for optimum ratio combining or receive diversity (nT = 1, nR = n) is

C = log2 1 +

22n

b/s/Hz .

(4)

The lower-bound (3) suggests that in some sense, one

might be able to embed n OC(k) systems (with
k = 1, 2, ... n) in an (n, n) system. The argument of the
logarithm in (3) suggests that each of the n singletransmit antenna systems would have transmit power
P$ / n so that each receive array element has an average SNR of /n. As explained below, such embedding
is indeed possible.

Opportunity for Enormous Bit Rates

Before demonstrating how to do the embedding,
the great capacity at stake is worth reviewing.
Figure 3(a) shows the capacity 99 percentile (1% outage) for SNRs of 0 dB to 24 dB in steps of 6 dB. As

46

Bell Labs Technical Journal Autumn 1996

noted from the ordinate, which ranges to 300 b/s/Hz,

the capacities are enormous. For example, at a 12-dB
SNR and even for modest numbers of antenna elements like eight or 12, significant capacity is availablethat is, about 21 and 32 b/s/Hz, respectively.
Even at a 0-dB SNR, very significant capacity exists.
About 25 b/s/Hz is available for, say, n = 32.
From the standpoint of signal constellations, one
must be concerned with per-dimension capacities.
Figure 3(b) shows the same capacity results as Figure
3(a) but they are expressed in terms of the b/s/Hz/signal dimension. In preparing Figure 3(b), the cases
n = 1, 2, 4, 8, 16, 32, and 64 were actually computed
and the remaining cases interpolated. Note that even
when the overall capacities are great, the per-dimension capacities can be reasonable. Figures 3(a) and 3(b)
depict the capacity (bold curves), as well as the capacity lower bound (light curves). The capacity lower
bound is quite tight at the higher values.
As the antennas increase in number, saturation of
the lower bound on the per-dimension capacity
becomes evident as Figure 3(b) indicates. This asymptotic behavior can be explained by looking at the
right-hand side of the inequality in (3). For the righthand side divided by n, the large n asymptote is1
1

log (1 + x) dx
0

= 1 + 1 log 2 (1 + ) log 2 e.

(5)

Note that in the limit of large , the dominant term in

the last expression is log2 (/e). For example, as Figure
3(b) shows, for an SNR of 24 dB and n = 64, an
asymptotic value of about 6.5 b/s/Hz/dimension and
indeed log2 (102.4) ! 6.5. The Asymptotic Optimality
of the Layered Architecture subsection later on concludes that not just the lower bound but also the
capacity per dimension or C/n converges to log2 (/e)
as and n increase without bound.
Figures 3(c) and 3(d) correspond to Figures 3(a)
and 3(b) but only for SNRs that are negative when
expressed in dB. Note that even at negative SNRs,
interesting capacity levels are possible. The tightness of
the lower bound deteriorates more and more, however, as is lowered. The Related Options section
later in the paper further discusses these aspects.

Time

Associated transmitter element

Each of the above rectangles corresponds to a

finite sequence of noisy received 6-D vectors as
illustrated by the right-hand figure for the eighth
rectangle. Six identical copies of these rectangles
Space are stacked on top of each other as shown below.
6
5
Rectangles in
the same column
are distinguished
by how the
6-D vectors are
processed.

4
3

Time

2
1
7

Nominal processing timeline through

space-time (shown as thin solid directed line)

Figure 5.
Flow of nominal processing time for a received signal.

Layered Space-Time Architecture

This section provides a high-level description of a
form of the new architecture having an equal number
of antenna elements at both ends of the link. Its capacity is associated with the lower bound (3) on the
capacity achievable with MEAs for the higher SNR values indicated in Figure 3(a). A simple description of
the architecture is provided following some brief
mathematical background information, which is
needed to establish that the architecture indeed offers
tremendous capacity. The previously mentioned
Related Options section provides some advantageous
variations on the architectural theme.

Mathematical Background
The following linear algebra helps clarify the
architecture. Let H j with 1 j n denote the n
columns of the H matrix ordered left to right so
that H = H1, H2, ...Hn . For each k such that

1 k n + 1, let H[k, n] denote the vector space

spanned by the column vectors H j satisfying
k j n. Because no such column vectors exist when
k = n + 1, the space H[n +1, n] is simply the null space.
Note that the joint density of the entries of H is a
spherically symmetric (complex) n2-dimensional
Gaussian. This makes it possible to state that, with
probability one, H[k, n] is of dimension n k + 1.
Furthermore, with probability one, the space of vectors perpendicular to H[k, n] denoted as H[k, n] is k 1
dimensional. For j = 1, 2, ... n, j is defined as the pro
jection of H j into the subspace H[ j+1,n] .
As explained next, with probability one, each j is
essentially a complex j-dimensional vector with iid
N(0,1) components (n is just Hn ). Strictly speaking,
an j is n dimensional. When viewing j using an
orthonormal basiswith the first basis vectors being
those spanning H[ j+1,n] and the remaining vectors

Bell Labs Technical Journal Autumn 1996

47

(6, 6) example: layers labeled a, b, c, d, e, and f;

detection of first complete a layer.
Already
detected

Space
(associated
transmitter
element)

Detect
now

6
5
4
3
2
1

f
e
d
c
b
a
0

Detect
later

a
f
e
d

b
a
f
e

c
b
a
f

Canceled

d
e
f
Nulled
c
d
e
b
c
d
a
b
c
a
b
a
5

a
f
e
d
c
b
7

b
a
f
e
d
c
8

c
b
a
f
e
d
9

10

Time
Reconstruct signals for five underlying layers
and then subtract off interference from the
previously detected bits.
Avoid interference from five overlying layers
by forming decision statistics that avoid
interference from those overlying layers.

Demod a
waveform on:
t < 2
2 t < 3
3 t < 4
4 t < 5
5 t < 6
6 t < 7

Avoiding interference
from transmitter antenna:

6
5 and 6
4, 5, and 6
3, 4, 5, and 6
2, 3, 4, 5, and 6

Figure 6.
Space-time layering in reception.

being those spanning H[ j+1,n] the first j components

of j are iid complex Gaussians while the remaining
components are all zero. Looking at these projections
in the order n, n1, ... 1 shows that the totality of
the n ! (n+1)/2 nonzero components are all iid standard complex Gaussians. Consequently, the ordered
sequence of squared lengths are statistically independent chi-squared variates having 2n, 2(n 1), ... 2
degrees of freedom, respectively. With the choice of
normalization presented in this paper, the mean of the
squared length of j is j.

Transmission
In a spectrally economical system, the layered
space-time architecture described here would be
employed in conjunction with an efficient 1-D code.
The form of the code employed in a specific instance of
the architecture is not within the scope of this paper.
For expositional simplicity, however, it is best to begin
the description by considering some nonspecific block
code rather than a convolutional code implementation.
Figure 4 illustrates the transmission process. A

48

Bell Labs Technical Journal Autumn 1996

primitive data stream is demultiplexed into n data

streams of equal rate. Each data stream is encoded in
some unspecified way except to say that the encoders
can proceed without sharing any information with
each other. Rather than committing each of the nencoded streams to an antenna, the bitstream/antenna association is periodically cycled. The
dwelling time on any association is seconds so that a
full cycle takes n ! seconds. The n-encoded substreams, then, share a balanced presence over all n
paths to the receiver. Therefore, none of the individual
substreams is hostage to the worst of the n paths.
With communication structured in this balanced
way, each subchannel has the same capacity. This
setup serves to uniformize the multiplexing/demultiplexing and coding/decoding processesthat is, all n
constituents are rendered virtually identical in structure. Of course, because the balance makes it possible
to use the same constellation for each subchannel, the
lowest maximum number of constellation points per
subchannel is obtained. Each channel is essentially the
same regarding the opportunity for coding.

Set layer index to initial

Detect bits in current layer

Reconstruct ideal received

version of signal in current layer

Remove all interference stemming

from signal in current layer

Is
layer index
equal to
final?

No

Increment
layer index

Yes
End

Figure 7.
Temporal view of the processing of successive
space-time layers.

In the next subsection, it will be seen that within

each substream, good symbols (those with a high
SNR) can compensate for bad symbols (those with a
low SNR) through coding. The subsection will help
clarify that in a certain sense, the capacity obtained is
the sum given by the right-hand side of (3).
As pointed out in the Related Options section
below, advantages can be achieved by allowing an
optional second stage of multiplexing/demultiplexing.
For now, however, this discussion assumes only one
stage of multiplexing/demultiplexing as indicated in
Figure 4.

A (6, 6) Example of Processing at the Receiver

In describing processing at the receiver, a (6, 6)
example is used. The extension to arbitrary (n, n) is
immediate. A training phase (not described here) is
assumed to be already completed. During this start-up
phase, known signals were transmitted and processed
at the receiver to expedite the H matrix becoming
accurately known to the receiver. The transmitter,
however, does not know the channel.
The top of Figure 5 shows the first eight of a finite

sequence of rectangles. Each rectangle in this linear

sequence symbolizes a sequence of 6D received vectors. The right side of the figure illustrates the succession of 6D vectors (with complex components)
corresponding to the eighth rectangle. These are the
vectors arriving on the time interval [7, 8). The
heavy dots in the planes (circular sections shown) represent the complex received signal components that
can be seen for the first and last of this sequence of
vectors. Each of these vectors includes noise plus ninterfering transmitted signals from the transmit
antennas.
For clarification, visualize constructing a stack of
six identical copies of this aforementioned sequence of
rectangles, one atop the other as shown in the figure.
This stack is a visual aid for explaining how the
received signals are preprocessed. A spatial element is
associated with each rectangle on this wall of rectanglesspecifically, a transmitter antennadepending
on the ordinate value. These elements are numbered
1, 2, ... 6 as are the ordinate values to which they correspond. The resulting rectangular partition of spacetime must be understood figuratively as both space
and time are discrete. The duration can span any
number of time units and, as previously mentioned,
each space element is associated with the single transmitter element as indicated on the left of the stack.
Note that for the same rectangle base interval of
duration , the very same -consecutive vector signals with complex components are associated with
each of the six vertically stacked rectangles. The six
rectangles having the same -duration base interval
will be distinguished by the preprocessing applied to
the vector signals. The transmitter antenna associations were made because a rectangle at ordinate i will
play a key role in the process of extracting the signal
radiated by the ith transmitter. As explained later,
besides relating to the nature of information to be
extracted, the ordinate also determines the way in
which interference must be handled in the course of
extracting information.
Processing time is distinct from signal reception
time. Figure 5 illustrates the flow of time in processing
the received signal. As processing time passes, processing proceeds top to bottom along a succession of con-

Bell Labs Technical Journal Autumn 1996

49

Sampled nD vector signal

from receiver front end

First: top of stack

Second: next to top

Last: bottom of stack

Seeking signal from

transmitter n

Seeking signal from

transmitter n1

Seeking signal from

transmitter 1

Linear
combination
avoiding no
interference

Linear
combination
avoiding
interference
from transmit
antenna n

Linear
combination
avoiding
interference
from transmit
antennas n, n1 2

Sum of
n1 unavoided
interferences
+

Sum of
n2 unavoided
interferences
+

n scalar signals for

further processing
n spatial coordinates for a
fixed time interval of duration
(unavoided interferences can be subtracted only
after the signals from which they stem have been detected)

Figure 8.
Spatial view of receiver processing.

secutive space-time layers (diagonals), moving left to

right as indicated by the thin solid directed line (really
an ordered sequence of directed lines). The time flow
in Figure 5 is only nominal for two reasons. First, with
block coding where we assume that a layer is synonymous with a code block, no time arrow need be associated with the processing of the symbols of a block.
Second, for convolutional coding, which has a definite
time direction, a significant modification of the obvious time progression within each full layer might have
an advantage as explained later.
The central theme of the architecture is interference
avoidance, and this discussion assumes that interfering
signals will be nulled out. (As discussed later, instead
of nulling, SNR could be maximized. In such a case,
noise means including not just AWGN but all interferers not yet subtracted out.) Fewer interferers must
be nulled at the higher stack levels. The interferences

50

Bell Labs Technical Journal Autumn 1996

that need not be nulled are those that will be subtracted out. Of course, when nulling interferers, any
possible enhancement of the noise caused by the interference nulling process must be carefully assessed. As
explained later in reference to the mathematical setup
that has been carefully tailored for capacity analysis,
the noise assessment will be easy to do.
Figure 6 illustrates additional details of the steps
required for proceeding along iterated diagonal layers.
For expositional convenience, a repetitive abcdef labeling on the stack is included. Detection of the first complete diagonal a layer through which is drawn a
dashed diagonal line is described. Other layers, including boundary layers, are handled similarly. Boundary
layers are those layers involved with where a burst
starts or ends (those having fewer than six rectangles).
The first complete a layer comprises six parts,
aj(t) (j = 1, 2, ... 6), in which the subscript indicates at

Primitive
bit stream

n equal rate
substreams
1:n
DEMUX

Periodic t-varying
vector to avoid
interference from
detected bits*

Encoder

Periodic
t-varying
vector
Symbol
stream

Interference
vector from
periodically
varying set of
n1 antennas

AWGN
vector

+
n1 detected
substreams

<., .>

Compose vector
of interfering
detected symbols

Memory of
previously
detected symbols

Periodic t-varying
vector to avoid
interference from
undetected bits*

<., .>

Detected
bitstream

Interference-free
encoded stream

n:1
MUX

Decoder*

Notationally, <., .> means complex scalar product.

AWGN Additive white Gaussian noise
* Channel knowledge required.

Figure 9.
System diagram of the processing involved at the receiver (discrete time baseband perspective).

what point in time the part that lasts for time units
begins. All layers relatively disposed to be located partially underneath this a layer are assumed already successfully detected while all layers disposed to be
partially above the a layer are yet to be detected. The
capacity associated with this case will also be found,
and then the capacity associated with the (n, n) case
will be apparent. (With block coding, a full layer could
correspond to exactly one block, although as pointed
out later, associating more than one block within a
layer can sometimes be advantageous).
Next, before computing capacity, a connection is
made to the earlier Mathematical Background subsection by pointing out the relevance of projecting a

received signal vector into H[ k+1,n] . Assume that one

received signal vector within a rectangle of ordinate k
is being preprocessed. The purpose of the preprocessing stage is to help later determine the signal sent from
antenna k. The aim of the preprocessing is to yield a
vector free of interference from all signals that were

simultaneously transmitted from antennas other than

the kth. The interference stemming from signals that
were simultaneously transmitted from antennas
1, 2, ..., k1 are inconsequential because these signals
are assumed to have been already perfectly detected
and subtracted out. What is necessary is to null out the
interference from yet undetected signalsnamely,
those simultaneously transmitted from antennas
k+1, k+2, ... n. This is exactly what is accomplished by
projecting a received signal vector into H[ k+1,n]
because that space is the maximal subspace orthogonal
to the subspace spanned by the signals received from
transmitters k+1, ... n.
This discussion has stressed that on each of the six
time intervals for the a layer, a different number of
interferers to be nulled must be addressed. For each of
the six intervals in turn, the capacity of a corresponding hypothetical system is expressed in which the
additive interference situation holds for all time. For
the first time interval, the five layers below have

Bell Labs Technical Journal Autumn 1996

51

Time
Space
No processing is done above and to the right
of the layer currently being processed. The
corresponding interferences do impair the
layer currently being processed.

6
5
4
3
2

Processing is completed below and to the left

of the layer currently being processed. The
corresponding interferences do not impair
the layer currently being processed.

A layer in space-time can correspond to a

superblock. Illustrated above is a superblock
comprising three separate blocks, which can
be independently coded and decoded. Each
block comprises six sub-blocks.

Three blocks each comprising six sub-blocks.

Figure 10.
Blocks and sub-blocks in a space-time layer show how parallel processing can be used to advantage.

already been detected and all interference from the

signal components transmitted from antennas labeled
1 to 5 has been subtracted out. Therefore, no interferers are present. Consequently, for the first time interval, a sixfold receive diversity effectively exists. Under
such nonexistence of interference, the capacity would
be

C = log2 1 + ( / 6) 12

b/s/Hz.

One interferer is present during the next interval;

the other four have been subtracted out. For a system
in which this level of interference prevailed forever,
the capacity would be

C = log2 1 + ( / 6) 10

b/s/Hz.

The process of nulling one interferer is what

caused the reduction of the chi-squared subscript (giv2
2
ing 10 in place of 12 ). This process is repeated
until finally encountering the sixth interval. In this
case, all five signals from the other antennas interfere.
Therefore, they must be nulled out so that the corresponding capacity would be

52

Bell Labs Technical Journal Autumn 1996

C = log2 1 + ( / 6) 2

b/s/Hz.

Because each signal radiated by the six transmit2

ting elements multiply a different H j , the six
variates are statistically independent of each other for
the reason given in the previous section. Similar to
what was reported elsewhere,1 for a system cycling
among these six conditions with an equal amount of
time spent on each, the capacity would be

C = (1/6)

log

k =1

1+ ( / 6) 2k
2

b/s/Hz.

Assume that six such systems are running in par2

allel with the same realization of 2k (k = 1, 2, ... 6)
occurring in each one. The capacity would then be six
times that given by the previous sixfold sum, or

C=

log

k =1

1 + ( /6) 2k
2

b/s/Hz.

In the limit of infinitely many symbols in a layer

and because every sixth layer is an a layer, this last
expression gives the capacity of the layered architecture for the (6, 6) case. Obviously then, in the large

Code

Periodic
time-varying
channel

Decode

Code

Periodic
time-varying
channel

Decode

Periodic
time-varying
channel

Decode

Periodic
time-varying
channel

Decode

Bit rate x/6

To
multiplexer

Bit rate x/18

Code

To
sub-multiplexer

Bit rate x/18

Code
Bit rate x/18

Subsystem at the top is replaced by three subsystems of one-third the bit rate.

Figure 11.
In parallel receiver processing, three time-varying channels run in parallel.

number of symbols limit, the capacity for an (n, n) sys-

the major steps in the iterative detection of the n lay-

tem is given by

ers. In providing a spatial view, Figure 8 highlights

C=

k =1

how interferences are handled differently at the dis-

log2 1 + ( / n) 2k b/s/Hz .
2

(6)

Figure 7 provides a high-level temporal view of

tinct vertical levels for the same received vectors.

Figure 9 is a system diagram of the processing
involved. The way past and future decisions are han-

Bell Labs Technical Journal Autumn 1996

53

dled is reminiscent of zero forcing decision feedback.7

Factors corresponding to catastrophic error propagation in decision feedback systems are discussed next.

Robustness
In case the layered architecture described earlier
seems fragile, an explanation of why it can be quite
robust is included. At first, the architecture might seem
fragile. After all, the successful detection of each layer
relies on the successful detection of the underlying layers. Thus, any failure in any layer but the last will
likely cause the detection of all subsequent layers to
fail. A quantitative discussion is included in this subsection to illustrate that fragility generally is not a significant problem, especially when huge capacity is
available. As depicted in Figure 3(a), a huge capacity
can be a very reasonable assumption.
In practical implementations, the huge capacity
available can be invested in selecting a code that provides the required bit rate with very substantial error
protection. Let ERROR denote the event that a packet
(= long burst) contains at least one error for whatever
reason. Decomposing the ERROR event into two disjoint events gives

ERROR = ERRORnonsupp ERRORsupp

.
ERRORnonsupp denotes the event that channel realization simply does not support the required BER even
if receiver processing could be enhanced magically by
a genie removing interference entirely from all underlying layers.
ERRORsupp denotes the remaining ERROR events.
Assume that the required outage is 1%, packet size
(payload) is 10,000 bits, and a BER of 107 is required.
The extra capacity can be used instead to provide a
BER at least one order of magnitude lower. Because
10 4 10 8 = 10 4 , roughly one packet in 104 contains an error. Inflate the bit-error occurrences by
labeling all bits in such a packet in error. Such a
drastic inflation in the accounting of errors is a
harmless exaggeration. The reason for this is those
packets containing errors can be ascribed to outage
because they carry insignificant probability compared to Probability[ERRORnonsupp]. In effect, the
huge capacity available allows the luxury of taking
the perspective that ERROR = OUT. When the sys54

Bell Labs Technical Journal Autumn 1996

tem is not out, essentially error-free transmission is

provided.
Despite the robustness just described, providing
error-free transmission over a burst for an
extremely high fraction of the bursts can erode the
bit rate so that codes must be carefully selected for
any application.

Related Options
This section discusses some modifications of the
communication architecture previously described.
Suggestions are provided as to what might be gained
or lost by these changes. Some of these items are preliminary ideas that are included as possibilities for
future research.

Asymptotic Optimality of the Layered Architecture

The layered space-time approach to communication was based on the premise that the channel was
not known at the transmitter. Suppose instead that the
channel is known at the transmitter and that this
knowledge is used to transmit n noninterfering signal
components of equal power. Typically, this is done by
using the eigenmodes of HH to derive what amount
to n uncoupled subchannels. Other research has been
published on how these eigenmodes arise as the natural modes to drive when the channel is known to the
transmitter.8
For the large n and large asymptote, the capacity
benefit of communicating in the way just described is
compared below with using the layered space-time
architecture. It will be seen later in this paper that in
an asymptotic sense, the per-dimension capacity is not
improved by knowing the channel at the transmitter.
Under the assumption of a Rayleigh fading chan9
nel, research has shown that as n increases, the density of the eigenvalues of HH approaches

1
n

n 1
d on 0 4 n (0 elsewhere).
4

For the purpose of computing the per-dimension

capacity, convenience motivates renormalization in
the following capacity-invariant ways:
1
Channel H/ n 2 in place of H,
Transmit signal power per dimension P$
instead of P$ /n, and
Noise power per dimension unity.

Previously, an n 2 multiplier was attached to

each scalar transmit signal component to keep the
transmitters total radiated power constant at P$ inde1
pendent of n. This n 2 multiplier has been moved off
the transmit signal components and onto the Hijs. This
action fixes the limiting set of eigenvalue support of
HH/n to [0, 4].
Consider that for each n, the matrix representation for each random channel stems from an infinite
random matrix with indices ij (i 1, j 1), where for
each n the infinite matrix is projected into its northwest n ! n corner submatrix to obtain the random
n ! n channel matrix. For any fixed x on [0, 4] and a
corresponding integer (n) on [0, n], it can be written
that, with probability one,
1

Lim n
Number of Channel Eigenvalues
n
~

(7)

1 1
d .
4

Obviously, the per-dimension capacity Cch_knwn/n is

now expressed by

log (1 + 4x)
0

= log2 (1 + 4 )

ln 2

1
0

1 x
dx
x

x (1 x) + sin
1 + 4 x

(8)

dx .

The right-hand side of (8) follows from integration

by parts. In the limit of large , the last integral simplifies, enabling one to conclude that
C ch _ knwn / n log 2 ( / e) b/s/Hz/signal
dimension

(n, both large).

(9a)

This is the same asymptotic behavior as that for

the layered space-time architecture for which the
assumption was that the transmitter does not know the
channel. In the large- large-n asymptote, capacity
per dimension is not lost by lack of channel knowledge. Furthermore, in light of (9a), one can now conclude that the equations right-hand side also
expresses the limiting behavior of C/n for the capacity

C given by (2).
The large asymptote is anticipated to be of interest in some applications. However, the following is
worth mentioning: One can derive that the advantage
of knowing the channel for large n but vanishingly
small is a factor of two in capacity

C ch _ knwn / n
/ l n 2 b/s/Hz /signal dimension
= 2C / n (as n , 0).

(9b)

The next subsection provides additional information

about the small- realm.
The following question, related to the subject of
this subsection, is left for possible future research: Does
the lack of channel knowledge significantly diminish
the per-dimension capacity if x is allowed dependence
in the power distribution at the transmitter (optimal
(x) in place of constant )? So far, the constraint of
equal power out of all nT transmit elements has been
tacitly imposed. It would be worthwhile to explore the
aforementioned question while relaxing this restriction
even when H is unknown to the transmitter.
Despite the distributional spherical symmetry of
the elements of H, the receiver can break from symmetry in the reception process in an H-independent
manner that is known to the transmitter. Indeed, the
layered space-time reception process involves a symmetry breaking. The example associated with Figure 6
shows that the signal from transmit antenna six is
attended to first, then the signal from five, and so on.
The capacity advantage that comes from using information on reception asymmetry to distribute power
judiciously among the transmit antennas is an area
currently being researched.

Slowing Processing With Parallelism

Parallel processing can be used to advantage as
shown in Figure 10. The example involves slowing
processors by a factor of three. (In theory, one could
slow processing by any factor). To do so, each of the six
streams is further demultiplexed at the transmitter into
three demultiplexed substreams. At the receiver and
for each of the six streams, three separate processors
operate in parallel on the three distinctly encoded substreams. For block codes, each of the three sublayers
could constitute separate blocks. In effect, Figure 11
Bell Labs Technical Journal Autumn 1996

55

stresses that three time-varying channels run in parallel. Even a sublayer could be decomposed into blocks
for the purpose of block coding, especially if n is large.

Maximizing SNR Instead of Nulling

Assume that interference from underlying layers is
subtracted out. In the previous Layered Space-Time
Architecture section, the linear combinations formed
to avoid remaining interferences were those combinations that null out all the remaining interferers. In the
(6, 6) example, zero to five such interferers existed
depending on the transmit antenna. The following
option can surpass the capacity denoted by the righthand side of (3): In processing the signal component
radiated from a specified transmit array element, proceed as before except choose the linear combination
that gives the maximum SNR instead of nulling. Note
that here, the meaning of noise includes all noncanceled interference along with thermal noise.
The maximum SNR alternative to nulling is reminiscent of minimum mean square error (MMSE) decision feedback 10-13 as an alternative to the zero
forcing7 approach. For the maximum SNR method,
capacity could be assessed assuming the code is so
advanced that it produces essentially white Gaussian
interference signals. From the curves (essentially lines)
in Figure 3(a), the capacity improvement over nulling
offered by maximizing SNR is marginal for the higher
SNR values. Not much improvement can be expected
for a of more than about 12 dB. This conclusion is
reached even though the curves for maximum SNR
are not depicted in Figure 3(a); such curves would
occur between the corresponding bold and light lines
shown.
Given the marginality and the Gauss-like
requirement on the interference, nulling could be
preferred over maximizing SNR. For the lower SNR
values in Figure 3(a), howeverand especially for all
the low SNR values of Figure 3(c)the lower bound
has lost its tightness. Thus, maximizing SNR looms as
an alternative.
Does the maximum SNR method perform well
when is small? While it is beyond the scope of this
paper to report performance curves, the maximum
SNR method does perform well in an important limiting sense. This conclusion is reached after reviewing

56

Bell Labs Technical Journal Autumn 1996

(2) in the context of the form of the expressions for

the n-maximum SNRs. One can establish that for fixed
nas tends toward zerothe capacity constrained
to use the maximum SNRs divided by the true capacity converges to one. The next paragraph delineates
the two steps required to show this effect.
First, consider each one of the n-maximum SNRs,
taking care to note that the so-called noise power in
the denominators involves thermal noise plus interference from yet unsubtracted signal components. When
expanding each of the n SNRs in a power series in ,
the interference terms in the denominator do not register in the terms of first-order importance. Second,
derive the linear- term in the argument of the logarithm in (2). Some of these terms clearly can be
neglected. From this derivation, one learns that the
capacity using the maximum SNR on each subchannel
is precisely the same as the capacity given by (2) when
contributions of only first-order importance are
retained. This derivation for the small- realm results
in a capacity tending toward the total capacity of n
parallel systems with n-fold optimal combining.
Namely, as tends to zero,
n

log
k =1
n

1+ ( / n) Hk Hk b/s/Hz

log2 1 + ( / n) 2n , k b/s/Hz .
2

(10)

k =1

The dot between vectors is the complex n-D scalar

2
product. As before, the extra k subscript on 2n , k is
employed to index over independent chi-squared variates. Because the limit of small has been carefully
analyzed insofar as retaining terms that are linear in ,
one can write
1
C (ln 2) ( / n) trace HH b/s/ Hz
(11)

(as 0)

Trace in this context means the sum of the diagonal

elements.

Coding
We previously stressed block coding for expositional simplicity. In practice, convolutional codes or
a form of trellis-coded modulation for more bandwidth efficiency might have a role. (The block-convolutional distinction is blurred if a block code is

formed by blocking off a convolutional code.)

Comments about parallelism are equally true for
both convolutional and block codes. With convolutional codes, adjoining layers can be decoded simultaneously at times as long as a decision depth
requirement is met. The depth constraint stipulates
that the detection processes of the different processors
be staggered in such a way to ensure that each layer is
decoded after the interferences below them have been
detected and subtracted off. Such subtraction is essential for approaching the high capacities expressed by
(2) and (3).
A preliminary idea related to convolutional coding is mentioned for addressing the nonstandard
context in which the AWGN variance is a periodically changing value. For illustrative purposes,
assume the existence of five symbols per sub-block
and elaborate the sequence of transmit antenna elements used for any of the three sub-blocks over
time. The 6 ! 5 = 30 consecutive time intervals
result in 666665555544444333332222211111. A 6
symbol (meaning a symbol transmitted from antenna
6) tends to need the least error protection (no interferences). A 5 symbol tends to need more protection
(one interference)and so on down to a 1 symbol,
which tends to need the most protection (five interferences). Tend is used because noise, interference
level, and channel realization are all random variables.
In decoding convolutional codes, one could pair
protector and protected symbols in a
more sensible way if doing so significantly expedites bit decisions. For example,
616161616125252525254343434343 would be an
improvement. The encoding (decoding) to accomplish
this involves nothing more than a straightforward permutation (inverse permutation) in the encoding
(decoding) process. Careful study is needed to quantify
the resulting benefit of this idea for promoting timeliness in making bit decisions.
The actual choice of codes remains an important
open issue, one that is best addressed in the context of
specific applications, However, it is worth mentioning
that a transformation of the architecture that renders
the coding context much more standard does exist
albeit at a price of about one-half the available capacity.

In explaining the transformation, n is assumed to

be even; n odd is a trivial extension. At the transmitter, the primitive bit stream is demultiplexed to n/2
streams rather than n streams. However, n transmit
(and receive) antennas are still used. During each
interval of duration , each of the n/2 demultiplexed
signals is now associated with a distinct pair of transmit antennas. The same coded and modulated signal
is transmitted out of the array elements in each pair
but at different times. The pairing is as follows: the
best with the worst symbols, the next to best with the
next to worst, and so on. (To be precise, one should
not simply say best and worst. Tending to be the
best and tending to be the worst are more definitive because of the random power-transfer characteristics.)
The motivation for pairing transmit antennas is to
have each demultiplexed signal component possess
the same optimum combining diversity level. Say, for
explanatory purposes, n = 6. As far as time of transmission is concerned, refer also to the (6, 6) example
in Figure 6. A signal transmitted out of antenna 6 during [, 2 ) is the same as that transmitted out of
antenna 1 during [6 , 7 ). Antenna pairs 5 and 2 and
4 and 3 are configured the same way, the latter pair
being associated with the contiguous time intervals,
the union of which is [3 , 5 ).
Evidently, with this pairing, n/2 optimal combiners exist in effect for arbitrary n, each of which has
n+1-fold diversity. Therefore, subject to the constraint
to communicate in this way, the formula for capacity
differs from (6). Specifically, instead of chi-squared
variates having the arithmetically progressing indices
2, 4, ..., 2n, all n/2 indices are 2n+2 so that the capacity is given by

C=

n /2

log
k =1

1 + ( / n) 2(n +1),k b/s /Hz . (12)

2

The subscript k indexes statistically independent

22n + 2 variates. Thus, a good part of the cyclic volatility of the received signal SNR has been removed.
Clearly, as n increases, the arguments of the n/2 base
two logarithms in (12) converge to 1+. The price in
lost capacity for the large n asymptote is easy to see.
That is, the capacity now increases linearly with only

Bell Labs Technical Journal Autumn 1996

57

an n/2 slope rather than an n slope. Considerable

capacity, however, is still possible.

No Cycling
As Figure 4 illustrates, cycling the substream to
antenna association was required at the transmitter. Is
this cycling really necessary? A straightforward but
tedious asymptotic argument shows that, in the limit
of large n, the receive diversity compensates for any
inferior Hijs. Consequently, the asymptotic linear
capacity growth with n also occurs even without
cycling.

Discussion and Conclusion

With growing multiuser applications, efficient use
of spectral resources is especially important to avoid
highly contentious channel demand in a limited frequency band. Bit-rate delivery issues can be difficult to
decide solely from a fundamental standpoint. Indeed,
determination of implementation complexity can be
influenced by the legacy of past technology choices
limiting what is readily available for the short term.
Nonetheless, the results discussed in this paper inform
the evolution toward meeting future demand for
greater bit rates.
When the transmit volume is sufficient to allow
driving transmit antenna elements separated by onehalf a wavelength, the results presented in this paper
suggest considering doing so. When the receive volume (also assumed to be amply sized) is radiated by
waves involving distinct spatial degrees of freedom
all in the same frequency bandreceive antenna elements with half-wavelength spacing can serve to
capture that energy. Thus, transmit/receive volume
can be used to improve capacity dramatically over that
of systems in which the spatial dimension is not
exploited.
The layered space-time architecture is designed
largely to undo the coupling between distinct spatial
modes, yielding a system in which capacity increases
linearly with n for both fixed bandwidth and fixed
total radiated power. This n-D architecture can be
viewed in terms of n 1-D architectures of equal capacities. In future theoretical studies, a comparison of
extreme approaches would be informativefor example, a narrowband system using MEAs at both the

58

Bell Labs Technical Journal Autumn 1996

transmitter and receiver with wideband alternatives to

using MEAs to meet required capacity demands.
Understanding the relative merits of extreme
approaches could help clarify how the spatial and frequency domains should be combined to provide channels in various applications.
When adding more receiver elements than n to an
(n, n) system, the excess nR n can be used to improve
performance simply by adding twice the excess to the
degrees of freedom of the chi-squared variates appearing on the right-hand side of (6). If other users of the
same frequency spectrum are identified, the excess
receiver elements are effective for reducing co-channel
interference. The results of research have been published concerning handling co-channel interferers in a
Rayleigh fading environment.14

Acknowledgments
Figures 1 and 2 were largely the creation of
M. J. Gans, who also advised the author on antenna
theory. Valuable discussions with I. Bar-David,
J. Mazo, A. Saleh, J. Salz, and L-F. Wei are also
gratefully acknowledged.
References
1. G. J. Foschini and M. J. Gans, On Limits of
Wireless Communication in a Fading
Environment When Using Multiple Antennas,
Wireless Personal Communications, accepted for
publication.
2. E. Csiszar and J. Korner, Information Theory:
Coding Theorems for Discrete Memoryless Systems,
Academic Press, New York, 1981.
3. S. Z. Stambler, Shannons Theorems for a
Complete Class of Discrete Channels Whose State
Is Known at the Output, Problems of Information
Transmission, No. 11, Plenum Press, New York,
November 1976, pp. 263-270.
4. S. Kullback, Information Theory and Statistics,
John Wiley and Sons, New York, 1959.
5. D. B. Osteyee and I. J. Good, Information Weight
of Evidence: the Singularity Between Probability
Measures and Signal Detection, Springer-Verlag,
New York, 1970.
6. M. S. Pinsker, Information and Information
Stability of Random Processes, Holden Bay,
San Francisco, 1964, Chapter 10.
7. R. Price, Nonlinearly Feedback-Equalized PAM
vs. Capacity for Noisy Filter Channels,
Proceedings of the IEEE International Conference on
Communications, June 1972, IEEE Publishing

Company, New York, Sessions 22-12 to 22-17.

8. G. J. Foschini and R. K. Mueller, The Capacity
of Linear Channels with Additive Gaussian
Noise, Bell System Technical Journal, January
1970, pp. 81-94.
9. V. Grenander and J. W. Silverstein, Spectral
Analysis of Networks with Random Topologies,
SIAM Journal of Applied Mathematics, Vol. 32,
No. 2, March 1997, pp. 499-519.
10. C. A. Belfiore and J. H. Park Jr., Decision
Feedback Equalization, Proceedings of the IEEE,
67(8):1143-1156, August 1979.
11. J. Salz, Optimum Mean-Square Decision
Feedback Equalization, Bell System Technical
Journal, October 1973, pp. 1341-1372.
12. D. D. Falconer and G. J. Foschini, Theory of
Minimum Mean-Square-Error QAM Systems
Employing Decision Feedback Equalization,
Bell System Technical Journal, December 1973,
pp. 1821-1849.
13. R. D. Gitlin, J. F. Hayes, and S. Weinstein, Data
Communication Principles, Plenum Press, New
York, 1992, Chapters 5 and 7.
14. R. D. Gitlin, J. Salz, and J. H. Winters, The
Impact of Antenna Diversity on the Capacity of
Wireless Communication Systems, IEEE
Transactions on Communications, Vol. 42, No. 4,
April 1994, pp. 1740-1751.

GERARD J. FOSCHINI is a distinguished member of technical staff in the Wireless Communications

Research Department at Bell Labs in
Holmdel, New Jersey. He is conducting communication and information theory investigations for wireless communications at
both the point-to-point and systems levels. Mr. Foschini
is an IEEE Fellow and has taught at three New Jersey
universities. He holds a B.S.E.E. from the New Jersey
Institute of Technology in Newark, an M.E.E. from New
York University, and a Ph.D. in mathematics from the
Stevens Institute of Technology in Hoboken,
New Jersey.

(Manuscript approved October 1996)

Bell Labs Technical Journal Autumn 1996

59