Beruflich Dokumente
Kultur Dokumente
I. INTRODUCTION
HERO AND MARZETTA: CUTOFF RATE AND SIGNAL DESIGN FOR THE QUASI-STATIC RAYLEIGH-FADING SPACE-TIME-CHANNEL
2401
computing the cutoff rate. In Section VI, general characterizations for the feasible set are given. In Section VII, we specialize
to the case of finite constellations with dimension smaller than
, for which the cutoff rate is closed-form.
II. PRELIMINARIES
A. Rayleigh-Fading Spatio-Temporal Channel
We use substantially the same notation as in the papers of
transmitter
Hochwald and Marzetta [20], [14]. Let there be
receiver antennas and let the
channel
antennas and
fading coefficients be constant over an interval of length
time periods, called the coherent fade sampling interval. A
matrix having complex valued
transmitted signal is a
denote the set of all possible signal
entries. Let
is defined as
matrices. The norm of a matrix
where
denotes the Hermitian transpose of and
are the singular values of . Note that if
only
of
these singular values will be nonzero.
For receiver antennas and an observation time interval of
time periods, the received signal is a sequence
of
complex-valued
matrices
which
has the representation [20]
(1)
is the th transmitted signal,
is the norwhere
the expected SNR at each receiver per
malized SNR with
is an
matrix of complex channel
transmit antenna,
is a
matrix of complex noises. The
coefficients, and
quasi-static Rayleigh flat fading model corresponds to taking
elements of the matrices
and
the
to be independent and identically distributed (i.i.d.) complex
zero-mean (circularly symmetric) Gaussian random variables
with unit variance. Therefore, the joint conditional probability
density of the observations factors into a product of marginals
where
(2)
is the
identity matrix, and
denotes
the magnitude determinant of square matrix .
denote the probability of decoding error for a block
Let
code of rate (nats) with block length . It is well known [34],
[10], [9] that for below capacity , the minimum decoding
of the best code satisfies
error probability
(3)
2402
where
is a reliability function, called the random
coding error exponent, controlling the error rate of the optimal
decoder
nats/symbol
(4)
where the inner maximization is performed over a suitably constrained set of probability distributions defined over the set
of signal matrices . Additional constraints on are determined
by factors such as total power budget or envelope constraints
which atand are left implicit in (4). The distribution
gives an optimal signal distributains the maximum in
tion for which there exists a decoder achieving minimum probability of decoding error for sufficiently large block lengths. Genis not discrete and an optimal set of signal matrices
erally,
can only be constructed by a random coding procedure. The
was studied for the spatio-temporal Rayleigh
function
quasi-static fading model by AbouFaycal and Hochwald [1]
and Telatar [31] under a mean power constraint on transmitted
signal matrices. Unfortunately, the double maximization in (4)
is generally very difficult since the inner integral is raised to a
. The cutoff rate specified
fractional exponent when
which is
below is a lower bound on the error exponent
frequently simpler to analyze.
can be lowerThe random coding error exponent
having slope
:
bounded by the tangent line
, where
is this lines intercept on both the
and the axes [34]. The point of tangency
nats/symbol
(5)
where the
matrix and the
matrix
are mutually independent matrices of i.i.d. zero-mean complex Gaussian
random variables. Substituting the SVD of into the model we
obtain
HERO AND MARZETTA: CUTOFF RATE AND SIGNAL DESIGN FOR THE QUASI-STATIC RAYLEIGH-FADING SPACE-TIME-CHANNEL
2403
Note that, as
is unitary,
is a
matrix of i.i.d. zero-mean complex Gaussian random variables
, the
trans(r.v.s). Note also that, as
satisfies the same peak-power constraint as
formed signal
signal . Thus, the signal
measured at
the original
antennas
the receiver after transmission of the signal on
is
is statistically equivalent to received when the signal
transmitted on only antennas.
, the SVD of is identical to
When
above except that now
, where
is an
diagonal matrix. Therefore,
where
Readers familiar with [20, Theorems 1 and 2] might suspect that characterization of the statistical distribution of the
optimal cutoff achieving signal matrix can be obtained. Indeed, paralleling the arguments of [20], it can be shown that, as
is invariant to unitary premultiplication of
and as the maximization in the definition of
is over a concave function of , the peak-power constrained cutoff rate is
where is
attained by random matrices of the form
isotropically distributed matrix,
is a
a
diagonal matrix, and and are statistically
random
independent.
(9)
(10)
Thus, the cutoff rate can be equivalently expressed as
III.
nats/symbol
(7)
where the maximization is performed over the set of distributions of suitably constrained signal matrices.
Proof of Lemma 1: For notational convenience define the
matrix
. Then the conditional probability density function (p.d.f.) (2) takes the form
As
is nonnegative definite
and, therefore,
. Thus, by Fubini we can interchange
order of integration in the triple integral to obtain
(11)
(8)
For practical coding schemes, it is of interest to restrict attention to finite sets of signal matrices. Let be a prespecified
finite positive integer. The cutoff rate (11) restricted to discrete
2404
(12)
where
(14)
(15)
is a suitably constrained set of signal matrices
and where
and signal probabilities
(see the following
subsection for examples).
the signal is never transmitted which motiWhen
vates a natural definition of a signal constellation.
in
Definition 1: A set of matrices
constellation if all assigned signal probabilities
.
positive,
and
(16)
In particular, under the constraint (14), the optimization (12)
decomposes into two nested minimizations
is a signal
are strictly
(17)
,
is the
matrix
..
.
Note that
where
is the average transmitter power budget. Thus, the
average power constraint introduces additional dependency between signal matrices and signal probabilities which complicates the optimization problem.
..
..
..
..
..
.
(19)
HERO AND MARZETTA: CUTOFF RATE AND SIGNAL DESIGN FOR THE QUASI-STATIC RAYLEIGH-FADING SPACE-TIME-CHANNEL
2405
Thus
Proposition 2: Let
-dimensional cutoff rate
where
where
is a multivariate circular complex Gaussian denand
sity with zero mean and covariance matrix
where
attains
is
As
(20)
As
that
(a.e.)
are identically zero. Hence,
is positive defimplies
inite.
As for assertion ii) of the Lemma we have by Cauchy
Schwarz
2406
where
.
We have thus shown that rate-optimal finite-dimensional constellations, defined in Definition 1, must have equalizer distributions. These results are summarized in the following.
,
is a -elewhere
is an undetermined multiplier
ment vector of ones, and
.
that must be chosen to enforce the equality constraint
is convex cup, the KuhnTucker conditions [26] assert
As
that the minimum exists and must satisfy
achieves
Proposition 3: A constellation of dimension
-dimensional peak constrained cutoff rate (12) only
the
over signal matrices
if the optimal distribution
in the constellation is an equalizer distribution of the form
.
An equivalent condition to (24) is that there exists a vector
, not identically zero, lying in the positive
which satisfies
orthant
if
if
or, equivalently,
(28)
if
if
(22)
(23)
is the th column of
. By reordering the entries
where
and the rows and columns of
we can assume, without
of
s satisfying (22) are the first
loss of generality, that the
columns of
with the remainder satisfying (23),
.
Making this assumption, (22) and (23) imply that
and
(24)
are distinct (see
Furthermore, we can assume that
is positive definite.
Lemma 9) and, therefore, by Lemma 2,
Therefore, from (22)(24), the strictly positive component of
is given by
the minimizer
(29)
is attained by a constellation of diBy Proposition 3,
is nonempty. Substitution of the form
mension only if
specified in Proposition 3
of the optimal probability vector
into (13) we obtain the following alternative characterization of
.
Proposition 4: Let be a positive integer. The peak-power
constrained -dimensional cutoff rate is
(30)
(25)
where is determined via the constraint
or
,
(26)
Note that
since
since
and
as the elements of
are
. Thus, in view
satisfy the equalof (24), the strictly positive elements of
ization condition
(27)
in Proposition
Observe that by taking the limit
4, we obtain the cutoff rate of constellations of countable, but
possibly infinite, dimension. The objective function
maximized in (30) is similar to the criterion used in Capons
method, also known as minimum variance distortionless
response (MVDR), for adapting the weights of a beamforming
array of antenna elements and for high-resolution spectral
estimation [19].
and the rate-achieving signal set can
The cutoff rate
be iteratively computed using (30) in the following steps: 1) sein
, 2) use
lect a candidate set of signal matrices
, speciquadratic programming to find the zero entries of
fied by condition (23), e.g., using slack variable or active set
methods [23]; 3) reorder the zero and nonzero entries of
and permute the rows and columns of
as described above;
for ; 5) compute inner product
and
4) solve
take its natural logarithm; 6) perturb the candidate set of signal
matrices and repeat 1)5). This iterative procedure is repeated
for each new set of candidate signal matrices until
HERO AND MARZETTA: CUTOFF RATE AND SIGNAL DESIGN FOR THE QUASI-STATIC RAYLEIGH-FADING SPACE-TIME-CHANNEL
2407
. Then, for
where
timal distribution defined in (18) is
(31)
the op-
(34)
(32)
where
is no
The significance of the above proposition is that if
, any peak constrained set of
greater than
is in the feasible set
and thus
signal matrices
. Furthermore,
has an equalizer distribution
,
converges to the equiprobable distribution
as
. Using [13, Proposition 2], the following bound on
for the distance-optimal constellation can be derived:
As
is positive definite
is invertible. Furthermore,
are strictly less than one and
from (33), the eigenvalues of
by elementary matrix manipulations
is
. Also,
and
Therefore,
are bounded
. When
the eigenvalues
of
(33)
From the proof of Proposition 5 we can identify a weaker condition on the number of receiver antennas necessary to ensure
2408
as
since
and
. Therefore,
we have
and assume
Then, as
. Then
(35)
both and
in as
OF
Here we give two conditions, one sufficient and one neces. The sufficient condition reduces
sary, to ensure
to specifying the largest inscribed right circular cone which fits
while the necessary condition is equivalent
inside
to specifying the smallest right circular cone which contains
. This sufficient condition is slightly weaker than the
sufficient condition presented in Lemma 4 of the next section
and obtained by recursive inscription of a polyhedral cone. On
the other hand, as contrasted with the condition of Lemma 4, the
sufficient condition in this section is not easy to express recursively in .
will denote the th column of
As in (22) and (23),
HERO AND MARZETTA: CUTOFF RATE AND SIGNAL DESIGN FOR THE QUASI-STATIC RAYLEIGH-FADING SPACE-TIME-CHANNEL
For
, define
the
by deleting its th column. Define
2409
matrix obtained
(36)
. A sufficient conis
(37)
to be contained in
is
(38)
the rays
of
,
, which are the extremals
. Therefore, the radius of this smallest sphere is
Finally, if
B. A Recursive Construction
Here the objective will be to specify conditions for which
lies in
when it is known that
lies in
. To proceed, we will need a recursive update for
in terms of
. For a set of signal matrices
let the
matrix
be
partitioned as
,
to
(39)
where
Since
is arbitrary, we obtain the inequality (37).
there exists a
As for the necessary condition, for any
through which all positive rays in
sphere centered at
must pass (see Fig. 1). The smallest such sphere has
radius equal to the maximum distance between and the faces of
. Points at maximum distance must occur along one of
We also recall the partitioned matrix inverse identity for the case
is positive definite
that
(41)
2410
where
partitioned form
. Let
where
have
is a scalar.
(45)
(42)
where
(43)
(46)
is positive
Proof of Proposition 7: By Lemma 2,
definite. Applying the partitioned matrix inverse identity (41)
to the right-hand side of
and
The set
yields (42).
Then the
-element vector
is in
iff a)
and b)
is in
. Furthermore,
.
condition b) implies that
Proof of Corollary 1: It only need be shown that condition
. If this condition holds
b) of the corollary implies that
as
has nonnegative elements.
then
Therefore,
(44)
Lemma 4: Let
tinct outerproducts
sume that
be a signal such that
Then the
if
-element vector
lies in
a)
and
(47)
b)
where
and
By definition (43)
(48)
where
Adding this latter equation to the inequality (44) gives
and, as by Lemma 2
Interpretation of the sufficient conditions in Lemma 4 is similar to the interpretation of Lemma 3. Condition a) is equivalent
to
(49)
lies close to the column span of
then
If
is small and condition a) restricts the inner product of
the previous unnormalized probability vector
and
to
HERO AND MARZETTA: CUTOFF RATE AND SIGNAL DESIGN FOR THE QUASI-STATIC RAYLEIGH-FADING SPACE-TIME-CHANNEL
2411
contains
it will suffice to
or, equivalently,
is in
, under the conditions of Lemma 4. Applying the ShermanMorrisonWoodbury identity (46) to
where
, and
(50)
(52)
is the orthogonal projector defined in (36).
. Let be the distance
Proof of Lemma 5: Fix a value
and the closest face of
(see
between the vector
Fig. 2). As each face is contained in the column span of
and as
is convex and contains
, is specified by
the projection theorem
(53)
and
is an idempotent matrix
Since
As the elements of
which is positive under conditions a) and b) of the lemma. Thus,
is in the positive orthant which establishes condition b) of Corollary 1.
Since
is positive definite and
and
are convex,
does not lie on the boundary of
and
such that
there always exists a value
for all
. As the set
is monotone decreasing in , the largest possible inscribed cone is obtained by using the minimum possible value of . This gives the
least restrictive sufficient condition b) in (47). The form of this
in the Lemma below.
minimum is specified for
with
Lemma 5: For a given set of signal matrices
assume that the -element
distinct outerproducts
lies in
. For
, the minimum value
vector
ensuring
is the value
of
given by
(51)
(54)
as long as
.
Evidently,
be the minimum corresponding value of
for
Let
. Equating (53) and (54) yields the following
which
quadratic equation for :
2412
where
. For
, the
quadratic equation has two nonnegative solutions only one of
. For
,
which is contained in
and there is one solution in
. In both cases this solution is
given by (51).
VII. LOW-RATE OPTIMALITY OF UNITARY ORTHOGONAL
CONSTELLATIONS
to be considered is
When the number of signal matrices
sufficiently small, significant simplification of the optimization
(17) is possible. In particular, one obtains optimality of a set of
scaled mutually orthogonal unitary signal matrices and a simple
.
form for
The first result specifies the solution to optimization of the
defined in (10).
dissimilarity measure
define the integer
For given , , and
(55)
is the rank of
We will see below that under some conditions
the signal matrices in the optimal -dimensional constellation.
. Then
Proposition 8: Let
and
(57)
unitaries, i.e.,
,
are
unitaries, and
are
real
and
,
diagonal matrices of singular values
respectively.
of
and
only depends
The function
and
through
. Thus, unitaries
and
on
can be arbitrarily chosen without affecting
and
. We denote this functional dehence we can choose them as
. Therependence by writing
fore,
(58)
where
are
where , are the slope and intercept of the least squares linear
fit to the function
(59)
The approximation is a lower bound and underestimates the
, given by (55), by at most over less than
exact value of
. If
0.5% of the SNR range shown in Fig. 3
for
the SNR is sufficiently large, e.g., (from Fig. 3)
and
,
and the optimal signal matrices
transmit antennas. On the other hand, for small
utilize all
,
and the optimal
SNR, i.e., (from Fig. 3)
HERO AND MARZETTA: CUTOFF RATE AND SIGNAL DESIGN FOR THE QUASI-STATIC RAYLEIGH-FADING SPACE-TIME-CHANNEL
that
where
and , defined in Lemma 6, have
and . Therefore, when
and
the same column spaces as
are mutually orthogonal
and the second term in (59)
. This establishes that the opis equal to its maximum value
and
have orthogonal singular vectimum signal matrices
and
and that
tors
2413
where the
s are rank
mutually orthogonal rectangular
unitaries as defined in the statement of the Proposition.
Proposition 8 implies that for low symbol rate, the distanceoptimal signal constellations are constellations of scaled mutu.
ally orthogonal unitary matrices of rank
(60)
where the maximization on the right-hand side is subject to
. Observe that as each summand is monotone
increasing in the inequality constraint in the maximization is
always met with equality. Consider the Lagrangian
where
is an undetermined multiplier. For a suitable value
is identical
of , the maximum of over unconstrained
to the maximum of the right-hand side of (60) over power con. The derivative of with respect to is
strained
where
. Therefore, for each
stationary points of given by either
is one of the roots
With respect to (61) there are three cases which must be conboth roots are real negative and for
sidered. For
both roots are complex. For
both roots are
real positive. Thus, only the latter case is relevant. It can be veris equal to one of these two positive roots,
ified that, when
the second derivative of is
and let
be as defined in (55).
Corollary 2: Let
and define the set of signal
Assume
by
where
are a
matrices
mutually orthogonal rectangular unitary matrices
set of
(
and
,
). This set of
of rank
signal matrices are equidistant in the sense
Let
be any set of signal matrices in
the same power constraint. Then
and thus
maximizes
satisfying
where
. Therefore,
2414
, condition b) reads:
, which is satisfied regardless of the value of .
The final result of this section is an expression for the cutoff
rate.
and let
be as defined in (55).
Proposition 9: Let
. Then the peak constrained
Suppose that
dimensional cutoff rate (12) is
Thus, the minimizer
tribution
(63) establishes that
of
is the uniform dis. Substitution of this solution back into
and
is given by (56). Furthermore, an optimal constellais the set of rank
mutually orthogtion attaining
onal matrices specified in Corollary 2 and the optimal proba,
.
bility assignment is uniform:
Any unitary transformation on the columns (spatial coordinates) of a set of signal matrices produces a set of signal ma. In particular, any set of
mututrices with identical
permutation matrices, specified in the
ally orthogonal
course of proof of Corollary 2, has optimal distance properties.
This simple set of signal matrices corresponds to transmitting
energy on a single antenna element at a time, among a total of
elements, in each of the available time slots. Since
is increasing in , the maximum cutoff rate achievable
using these mutually orthogonal unitary matrices is obtained by
.
using the maximum possible number of them:
Observe that the resulting optimal constellation may correspond
to a code of quite low symbol rate, e.g., for
the symbol rate is only one bit per symbol.
It is noteworthy that the optimal peak constrained signal constellation specified by Proposition 9 does not include the zero. Including zero in the signal convalued signal matrix
stellation would allow signaling using ONOFF keying. ONOFF
keying is often proposed for average power constrained signaling over low-SNR channels since it permits energy discrimination at the receiver. As contrasted with ON-OFF keying all signals in the optimal peak constrained signal set have equal power.
We conjecture that the zero signal would result from replacing
the peak-power constraint with an average power constraint in
Proposition 9.
which is
Proof of Proposition 9: Define
strictly less than one under the hypotheses of the proposition.
satisfying the assumpFor any set of signal matrices
tions of Proposition 9, Proposition 8 asserts that
,
, with equality when
consists of the specified mutually orthogonal matrices. Using this inequality and the
,
fact
where
(63)
is the
HERO AND MARZETTA: CUTOFF RATE AND SIGNAL DESIGN FOR THE QUASI-STATIC RAYLEIGH-FADING SPACE-TIME-CHANNEL
and
2415
Since
Lemma 7 is established.
Lemma 8: Let
the trace norm of
and
be
satisfies
(64)
Furthermore, denoting by
of , where
where
. Identifying
, and
,
finishes the proof of
Lemma 6.
and
are the diagonal elements of and
where
, respectively. Equality occurs in (64) iff the trace norm of the
is equal to zero.
antidiagonal matrix
have diagonal eleProof of Lemma 8: Let
and consider the decomposition
ments
into diagonal matrix
and antidiagonal matrix .
Then
As
B. Bounds on
Lemma 7: The dissimilarity measure
where
is a nonnegative function of
satisfies
such that
2416
(65)
where
Let
and
Since
[7] G. J. Foschini, Layered space-time architectures for wireless communication in a fading environment when using multi-element antennas,
Bell Labs Tech. J., vol. 1, no. 6, pp. 4149, 1996.
[8] G. J. Foschini and M. J. Gans, On limits of wireless communication in
a fading environment when using multiple antennas, Wireless Personal
Commun., vol. 6, no. 3, pp. 311335, 1998.
[9] R. G. Gallager, A simple derivation of the coding theorem and some applications, IEEE Trans. Inform. Theory, vol. IT-11, pp. 38, Jan. 1965.
, Information Theory and Reliable Communication. New York:
[10]
Wiley, 1968.
[11] J. Hagenauer, Iterative decoding of binary block and convolutional
codes, IEEE Trans. Inform. Theory, vol. 42, pp. 429445, Mar. 1996.
[12] B. Hassibi and B. M. Hochwald, How much training is needed in multiple-antenna wireless links, Bell Labs., Lucent Technol., Murray Hill,
NJ, Tech. Memo, 2000.
[13] A. O. Hero and T. L. Marzetta, On computational cutoff rate for space
time coding, Bell Labs., Lucent Technol., Murray Hill, NJ, Tech.
Memo, 2000.
[14] B. M. Hochwald and T. L. Marzetta, Unitary space-time modulation for
multiple-antenna communications in Rayleigh flat fading, Bell Labs.,
Lucent Technol., Murray Hill, NJ, Tech. Memo, 1999.
, Unitary space-time modulation for multiple-antenna communi[15]
cation, IEEE Trans. Inform. Theory, vol. 46, pp. 543564, Mar. 2000.
[16] B. M. Hochwald, T. L. Marzetta, T. J. Richardson, W. Sweldens, and
R. Urbanke, Systematic design of unitary space-time constellations,
IEEE Trans. Inform. Theory, vol. 46, pp. 19621973, Sept. 2000.
[17] I. A. Ibragimov and R. Z. Hasminskii, Statistical Estimation: Asymptotic Theory. New York: Springer-Verlag, 1981.
[18] S. Kallel and K. Li, Biderectional sequential decoding, IEEE Trans.
Inform. Theory, vol. 43, pp. 13191326, July 1997.
[19] S. M. Kay, Modern Spectral Estimation: Theory and Application. Englewood-Cliffs, NJ: Prentice-Hall, 1988.
[20] T. L. Marzetta and B. M. Hochwald, Capacity of a mobile multipleantenna communication link in Rayleigh fading, IEEE Trans. Inform.
Theory, vol. 45, pp. 139158, Jan. 1999.
[21] J. L. Massey, Variable length codes and the Fano metric, IEEE Trans.
Inform. Theory, vol. IT-19, pp. 196198, Jan. 1972.
[22]
, Coding and modulation in digital communications, in Proc.
1974 Int. Zurich Sem., Digital Communication, Mar. 1974, pp.
E2.1E2.4.
[23] J. J. Mor and S. J. Wright, Optimization Software Guide, ser. Frontiers
in Applied Mathematics. Philadelphia, PA: SIAM, 1993.
[24] P. Narayan and D. L. Snyder, The two-user cutoff rate region for an
asynchronous and a synchronous multiple access channel are the same,
IEEE Trans. Inform. Theory, vol. IT-27, pp. 414419, July 1981.
[25]
, Signal set design for band-limited memoryless multiple access
channels with soft decision demodulation, IEEE Trans. Inform. Theory,
vol. IT-33, pp. 539556, July 1987.
[26] T. Rockafellar, Convex Analysis. Princeton, NJ: Princeton Univ. Press,
1970.
[27] J. E. Savage, Sequential decodingThe computation problem, Bell
Syst. Tech. J., vol. 45, pp. 149175, Jan. 1966.
[28] D. L. Snyder and I. B. Rhodes, Some implications of the cutoff-rate criterion for coded direct detection optical communications, IEEE Trans.
Inform. Theory, vol. IT-26, pp. 327338, May 1980.
[29] V. Tarokh, N. Seshadri, and A. R. Calderbank, Space-time codes for
high data rate wireless communication: Performance criteria and code
construction, IEEE Trans. Inform. Theory, vol. 44, pp. 744765, Mar.
1998.
[30] I. E. Telatar, Capacity of multi-antenna Gaussian channels, AT&T Bell
Labs., Tech. Memo, 1995.
[31]
, Capacity of multi-antenna Gaussian channels, Europ. Trans.
Telecommun., vol. 10, no. 6, pp. 585595, Nov.Dec. 1999.
[32] F.-Q. Wang and D. J. Costello, Probabilistic construction of large
constraint length trellis codes for sequential decoding, IEEE Trans.
Commun., vol. 43, pp. 24392448, Sept. 1995.
[33] L. Wei and H. Qi, Near-optimal limited-search detection on ISI/CDMA
channels and decoding of long convolutional codes, IEEE Trans. Inform. Theory, vol. 46, pp. 14591482, July 2000.
[34] J. M. Wozencraft and R. S. Kennedy, Modulation and demodulation
for probabilistic coding, IEEE Trans. Inform. Theory, vol. IT-12, pp.
291297, July 1966.