Codebook Adaptation for Quantized MIMO

Beamforming Systems
Roopsha Samanta and Robert W. Heath, Jr.
roopsha,[email protected]
Wireless Networking and Communications Group
Dept. of Electrical and Computer Engineering
The University of Texas at Austin
AbstractQuantized multiple-input multiple-output (MIMO)
beamforming systems use predesigned codebooks for the quan-
tization of transmit beamforming vectors. The quantized vector,
which is conveyed to the transmitter using a low-rate feedback
channel, is used for transmission to provide signicant diversity
and array gain. The codebook for quantization is a function of the
channel distribution, and is typically designed for xed channel
distributions. In this paper, we propose a channel adaptive
feedback strategy for arbitrary channel distributions, and present
a simple codebook design methodology based on the channel
statistics. The codebook for quantization is dynamically chosen
from a structured set of pre-designed codebooks, called the
codeset, wherein all codebooks are derived from one mother
codebook. Simulations illustrate that the proposed method can
improve error rate performance in correlated and/or channels
with strongly line-of-sight components.
I. INTRODUCTION
Multiple-input multiple-output (MIMO) wireless systems
can achieve signicant diversity and array gain by using simple
techniques like transmit beamforming and receive combining
[1], [2]. Optimal beamforming, however, requires perfect
channel knowledge at the transmitter. This can be obtained
using the reciprocity of the channel in certain scenarios. In
the most general case, one needs to rely on the use of a
feedback channel for conveying channel state information to
the transmitter. One approach for narrowband channels [3]
[5] is based on quantization of the instantaneous channel
information at the receiver followed by its conveyance to the
transmitter using a low-rate feedback channel. This is the
limited feedback approach. There are other approaches [6][8]
which are based on feedback of partial (or statistical) channel
information, e.g., the channel mean or the the channel covari-
ance matrices. These methods, in general, do not perform as
well as the the ones using instantaneous feedback since they
do not track the rapid uctuations of the channel. There are
intermediate schemes as well, e.g. [9], which sends back the
signal covariance matrix using a limited number of bits.
In [3][5], the transmit beamforming vector is quantized at
the receiver using a xed codebook available at both the trans-
mitter and the receiver. In the current literature, codebooks
are predominantly designed and optimized for uncorrelated
Rayleigh fading channels with various codeword selection
This material is based in part upon work supported by a UT Research Grant
award and the National Science Foundation under grant CCF-514194.
criteria based on the instantaneous channel knowledge. These
quantization strategies lack the ability to exploit knowledge of
the channel correlation, and hence result in a degradation of
the receive signal-to-noise ratio (SNR) when the channel is not
uncorrelated Rayleigh fading. There are other codebook design
schemes, which tailor the codebooks for spatially correlated
Rayleigh fading channels using knowledge of the channel
correlation matrix [10], [11]. There are more general schemes
[12][14] that dynamically adapt to the distribution of the
channel. Unfortunately, none of these papers design structured
codebooks exploiting the channel statistics.
In this paper, a channel adaptive strategy based on switched
codebook quantization is proposed for dealing with arbitrary
channel distributions. In our proposed scheme, the codebook
for quantization is dynamically chosen from a structured set
of pre-designed codebooks, called the codeset. The codebook
is chosen in response to the average behavior of the channel
over a temporal window while the optimal codeword in the
given codebook is chosen based on instantaneous channel
knowledge. The codebook is switched periodically, at a rate
on the order of coherence times. The optimal
codeword is updated at a more frequent rate on the order of
a fraction of the coherence time. Thus with this approach the
codebook is determined in part by the channel statistics but
the codeword is still determined by the instantaneous channel.
The effectiveness of a switched codebook approach requires a
exible codeset design that can deal with a variety of channel
statistics.
The novelty of this paper lies in the design of a structured
codeset. We create our codeset by starting with a Grassman-
nian codebook [4], [5] and augmenting with several additional
codebooks derived from different scalings and rotations of
a single mother codebook with limited support. The idea
is that the Grassmannian codebook will be good for more
isotropic distributions like the Rayleigh channel, e.g. Fig. 1,
while the other codebooks will take advantage of clustering
in the beamforming vectors due to effects like correlation
or strong rank-one line-of-sight components, e.g., Fig. 2. To
create these codebooks we dene rotation and local scaling
transformations on the space of the transmit beamforming vec-
tors - the complex Grassmann manifold. These are analogous
to translation and scaling in Euclidean spaces. The intuitive
idea is to rotate and scale the root codebook in an attempt
to match the support of the new codebook with that of the
376 1424401321/05/$20.002005IEEE
beamforming vector distribution.The design methodology is
exible enough to be extended to any number of codebooks
and greatly reduces storage requirements of the codeset. Sim-
ulation results demonstrate a reduction in the required SNR
at a given symbol error rate of a quantized beamforming-
combining system with the proposed adaptation algorithm,
compared to non-adaptive quantization strategies designed for
uncorrelated Rayleigh fading channels.
Fig. 1. Example of a set of realizations of beamforming vectors on the unit
sphere taken from an isotropic distribution.
Fig. 2. Example of a set of realizations of beamforming vectors on the unit
sphere taken from a correlated distribution. The vectors realizations are more
clustered than the isotropic case.
II. SYSTEM OVERVIEW
Consider a MIMO system with M
t
transmit antennas and
M
r
receive antennas employing transmit beamforming and re-
ceive combining. Assuming a narrowband channel, the discrete
time baseband signal model at time instant k can be written
as
1
x k z
H
kH kw ks k z
H
kn k, (1)
where s kis a complex scalar which denotes the transmitted
symbol at instant k, H kis an M
r
M
t
matrix with complex
1
We use (.)
H
to denote conjugate transposition, | | to denote absolute
value, to denote Frobenius norm, E{} to denote expectation, CN(, )
to denote the complex normal distribution with mean and variance , C
t
to denote the t-dimensional complex vector space, card() to denote the
cardinality of a set.
entries that represents the channel transfer function at time
k and x k is the combined observation at the receiver. The
vectors w k and z k are the beamforming and combining
vectors, respectively at time instant k. The additive white
Gaussian noise (AWGN) at the receiver is denoted by n k
with i.i.d. entries distributed according to (^( , N
0
) . Further,
the symbol energy is given by c
t
E
_
[s[
2
_
.
We assume that H kis perfectly known to the receiver, and
that the system employs maximal ratio transmission (MRT)
and maximal ratio combining (MRC) [2] to maximize the
overall signal-to-noise ratio (SNR). The optimal combining
vector is given by the unit vector
z
opt
k
H kw k
|H kw k|
. (2)
The corresponding optimal beamforming vector is given by
w
opt
k
w[k]C
M
t , w[k]=1
|H w |, (3)
where |w k| reects the power constraint at the
transmitter. It follows from (3) that the optimal beamforming
vector is the right singular vector corresponding to the largest
singular value of H k. Note that this optimal vector is
not unique as it is invariant to transformations of the form
w
opt
k e
j
w
opt
k, where R. The nonuniqueness
can be incorporated by considering beamforming vectors to
be points on G(M
t
, ) , the Grassmann manifold that consists
of the set of one-dimensional subspaces of C
Mt
. If
1
is the
largest eigenvalue of H k
H
H k, the resulting SNR can be
written as
1
Et
N0
.
To realize this SNR, it is evident that the transmitter must
have knowledge of either the channel H k or the optimal
beamforming vector w
opt
kat any time instant k. This moti-
vates the general approach of quantized MIMO beamforming
[4]. We assume that a low-bandwidth, error-free, zero-delay
feedback channel exists between the transmitter and the re-
ceiver and is used to convey quantized channel information to
the transmitter. Codebook-based feedback schemes proposed
earlier in [4], [5] use a xed codebook available at the
transmitter and the receiver to enable feedback of the optimal
quantized beamforming vector. Specically, the index of the
codeword that maximizes a performance metric like the output
SNR or the outage probability is fed back to the transmitter.
We consider a block-stationary channel distribution model
in which the channel distribution remains constant within a
block of time. The receiver uses a fraction of this time interval
to estimate specic statistical information about the channel
using training data. In the discrete-time signal model, we
assume that the training interval length is T
train
symbols and
the total block length, inclusive of the training interval is T
stat
symbols.
III. CODESET DESIGN
We propose a switched codebook-based feedback strategy,
where the codebook for quantization of beamforming vectors
is dynamically chosen at the receiver from a set of pre-
designed codebooks, called the codeset. The codeset (
(
0
, (
1
, . . . , (
K
consists of K different codebooks, and is
377
available at both the transmitter and the receiver. The receiver
selects a codebook from the codeset ( based on statistical
information about the channel obtained during the training
phase, and sends back its index to the transmitter. The receiver
then uses instantaneous channel information to choose the
optimal codeword from the selected codebook, and conveys
its index to the transmitter through the feedback channel. Note
that the receiver sends back the optimal codeword index for
every channel realization, while the side information about the
codebook of choice is sent back only once every block. Thus,
the feedback requirement for transmission of the codeword
index is log
2
(N) | bits for every channel realization, and the
feedback overhead corresponding to codebook identication is
log
2
(K ) | bits for every block. Note that it is reasonable
to assume a block-stationary channel distribution model as the
channel statistics typically vary at a much slower rate than the
channel realizations, and this helps us minimize the codebook
identication overhead.
Let us denote the unit sphere in C
Mt
as
Mt
, dened
as
Mt

_
x C
Mt
|x|
_
. Since |w| , the
beamforming vectors lie in
Mt
. Further, the invariance of
the optimal beamforming vector to transformations of the
form w
opt
e
j
w
opt
induces a special distance metric
d (, ) in
Mt
, dened as d (x
1
, x
2
)
_
[x
H
1
x
2
[
2
, where
x
1
, x
2

Mt
. Using this distance metric, we can dene a
spherical cap in
Mt
with center o and radius as the open set
O(o, )
_
x C
Mt
d (x, o) <
_
. These are effectively
spherical caps on G(M
t
, ) .
Each codebook (
i
in ( has N codewords in
Mt
, i.e., (
i

w
i1
, w
i2
, . . . , w
iN
, where w
ij

Mt
for all i , . . . , K
and j , . . . , N. The rst codebook (
0
in ( consists of
uniformly spaced codevectors in the entire
Mt
space. This
codebook can be designed using VQ quantization (VQ) [3],
Grassmannian line packings (GLP) [4], [5] or equiangular
frames (EF) [15].
The rest of the codebooks are conned to a range of
spherical caps spanning different localized regions in
Mt
. We
constrain the set of possible centers of the spherical caps to be
the same as a uniform codebook in
Mt
with P codevectors,
again designed using VQ, GLPs or EFs. We denote this set
of centers by O o
1
, o
2
, . . . , o
P
. We further constrain the
radii of the spherical caps to take values from a nite set of L
values given by
1
,
2
, . . . ,
L
. Finally, for the sake of
simplicity, we impose an order on the sequence of codebooks
(
1
, (
2
, . . . , (
K
such that the codebooks corresponding to the
same center have adjacent indices, i.e., (
1
O(o
1
,
1
) , (
2

O(o
1
,
2
) , (
3
O(o
1
,
3
) , . . . , (
1+L
O(o
2
,
1
) , (
2+L

O(o
2
,
2
) , (
3+L
O(o
2
,
3
) , and so on. Henceworth, we will
index the codebooks using indices from the sets , , . . . , K
and (o
1
,
1
) , (o
1
,
2
) , . . . , (o
2
,
1
) , . . . , (o
P
,
L
)
interchangeably.
One can use VQ to generate a uniform codebook conned
to a specied spherical cap in
Mt
. Generating K codebooks
using VQ is a computationally intensive task which becomes
increasingly difcult as the dimension of the ambient space M
t
and the number of codebooks K increase. In this paper, we
propose an alternate scheme for the design of these codebooks,
wherein, only one codebook is designed using VQ, and all
Fig. 3. Illustration of how the mother codebook is rotated and scaled to form
a new codebook.
other codebooks are derived from this root codebook. The
basic idea is illustrated in Fig. 3. To enable this, we rst
need to dene rotation and local scaling transformations on
the space of the beamforming vectors. Note that this requires
some care due to the constraints of the space. For instance, a
scaling function of the form x x, where x
Mt
and
is a real scalar yields a vector that does not belong to
Mt
as
it does not satisfy the unit norm constraint. We henceforth
denote the root codebook as (
root
, and the corresponding
spherical cap as O(o
root
,
root
) . Any other codebook derived
from (
root
is denoted by (
deri
, with its spherical cap denoted
by O(o
deri
,
deri
) .
Denition 1: The rotation map r (
root
(
deri
is dened
as
w
deri
r (w
root
) U
root,deri
w
root
, (4)
where w
deri
(
deri
, w
root
(
root
, and U
root,deri
is a
unitary matrix that satises o
deri
U
root,deri
o
root
.
This map has the property that d (o
root
, r(w
root
) )
d (o
root
, w
root
) .
Denition 2: [16] The scaling map s (
root
(
deri
is
dened as
g
__
r
1
e
j1
, r
2
e
j2
, . . . , r
Mt
e
jM
t
_

_
_

2
( r
2
1
) e
j1
, r
2
e
j2
, . . . , r
Mt
e
jM
t
_
(5)
where w
root

_
r
1
e
j1
, r
2
e
j2
, . . . , r
Mt
e
jM
t

(
root
, and
< .
This map has the properties
1) s (o
root
) o
root
, and,
2) d (o
root
, s(w
root
) ) d (o
root
, w
root
) .
Both the rotation and scaling maps operate on all codewords
of (
root
supported by the spherical cap O(o
root
,
root
) to
generate the corresponding codewords of (
deri
supported by
the specied spherical cap O(o
deri
,
deri
) . Note that in case
of rotation, the center of the root codebooks support is rotated
378
to a new center, while the radius remains the same. In case
of scaling, the center of the root codebooks support remains
the same, while the radius is scaled to a new value. Thus,
given a desired support of a codebook, parameterized by o
deri
and
root
, one can rotate and scale the root codebook using
appropriate U
root,deri
and to generate the new codebook.
This design methodology is exible enough to be extended
to any number of codebooks and greatly reduces storage
requirements of the codeset. Only two codebooks in the
codeset need to be stored - (
0
and (
root
. The drawback of
using these transformations is that the codebooks may no
longer be uniform - it is not known if the scaling map preserves
the distribution of the codewords in the codebook.
IV. SELECTION CRITERIA
We now present the criteria for selection of the best code-
book and codeword for quantization of the optimal beam-
forming vector. As mentioned in Section II, the codebook
for quantization is selected from ( using information about
the long-term behavior of the channel in the block obtained
during the training phase. Once chosen, this codebook is xed
for a block duration equal to T
stat
symbol periods. Ergodicity
arguments under a block stationary assumption reveal that this
approach incorporates some information about the statistics of
the channel. Further analysis is beyond the scope of this paper.
The center of the spherical cap dening the optimal code-
book is chosen as the one that maximizes the average output
SNR over all possible centers during the training phase. Thus,
we have
o
opt

ojO

T
train
Ttrain

k=1
|H ko
j
|
2
. (6)
Once o
opt
is chosen, we use instantaneous channel infor-
mation to select the optimal codewords from all codebooks
centered at o
opt
, as shown in (7),
w
i,opt
k
w
i
[k]Co
opt
,
i
|H kw
i
k|
2
. (7)
Finally, we choose the average loss in SNR due to quan-
tization during the training phase as the cost function for
selecting the radius of the optimal codebook. Specically,
we pick the codebook centered at o
opt
, which maximizes the
average output SNR over all codebooks centered at o
opt
. This
is obtained by

opt

i

T
train
Ttrain

k=1
|H kw
i,opt
k|
2
. (8)
We do not search over all codebooks in the codeset to select
the optimal codebook. Once we pick o
opt
, we only need to
search over all codebooks centered at o
opt
.
Though we described the algorithm with a specic training
phase at the beginning of the block, it should be clear that it is
possible to relax this requirement. For example, a windowed
average of prior channels could be used to determine the
current channel state. Thus an explicit training phase would
not be required. Additional details of this method are beyond
the scope of this work.
V. SIMULATION RESULTS
In this section, we present the symbol error rate (SER)
performance of the proposed quantization scheme in uncor-
related Rayleigh fading, spatially correlated Rayleigh fad-
ing, and Ricean channels. We also compare the proposed
algorithm with previously proposed xed codebook quantiza-
tion schemes [3][5]. The three xed codebook quantization
schemes have been shown to have similar performance, and
we choose VQ [3] as a comparison reference.
All the simulations are done for the case of M
t
M
r

with QPSK modulation. The number of bits for feedback of
the codeword index is set to 2, i.e., N=4. Further, the number
of possible centers for the spherical caps, P , and the
number of radii, L . Thus, the total number of codebooks
in ( is K and the overhead for our scheme due to
codebook switching is 5 bits.
The distribution of the channel is assumed to be block-
stationary, and in this paper, we only look at the perfor-
mance within one such block. The transmit and receive
correlation matrices for the spatially correlated channel H
R
1/2
r
H
w
R
1/2
t
are generated using the procedure described in
[17]. The parameters used for R
t
and R
r
are - mean angle

0
,
0
; standard deviation 0.2, 0.2; and normalized antenna
spacing 0.25, 025. In the Rician model H
_
K
K+1

H
_
1
K+1
H
w
, the Rician factor K is chosen to be 2.

H is the
xed component dened as

H a(
r
) b
H
(
t
) , where
a(
r
) e
j2
d

r
e
j(Mr1)2
d

T
(9)
b(
t
) e
j2
d

t
e
j(Mt1)2
d

T
(10)
a(
r
) and b(
t
) are the array response vectors for the receiver
and transmitter antenna arrays. The wavelength normalized
antenna spacing
d

is assumed to be 0.5, and the line of sight

3
.
In the all gures, we plot SER versus the input or transmit
SNR. When the channel is uncorrelated Rayleigh fading, we
observe from Figure 4 that the proposed algorithm matches the
performance of quantized beamforming using VQ. The SER
improvement obtained with our algorithm becomes signicant
in spatially correlated and Ricean channels as is evident from
Figures 5 and 6. We obtain array gains of nearly . dB over
xed codebook quantization without any loss of diversity.
VI. CONCLUSIONS
In this paper, we presented a quantized beamforming strat-
egy for arbitrary channel distributions, wherein the codebook
for quantization is selected on the basis of channel statistics.
A novel codebook design methodology was also proposed to
simplify the design and storage of the codeset. The simulation
results demonstrate substantial improvements over xed code-
book quantization schemes in spatially correlated and Ricean
channels, and no loss of performance in uncorrelated Rayleigh
channels.
VII. ACKNOWLEDGEMENT
The rst author would like to thank Bishwarup Mondal for
many insightful discussions, and in particular, for pointing out
his result for scaling vectors in
Mt
[16].
379
REFERENCES
[1] T. K. Y. Lo, Maximum ratio transmission, IEEE Trans. Commun.,
vol. 47, no. 10, pp. 14581461, Oct. 1999.
[2] A. Paulraj, R. Nabar, and D. Gore, Introduction to Space-Time Wireless
Communications. New York: Cambridge University Press, 2003.
[3] A. Narula, M. J. Lopez, M. D. Trott, and G. W. Wornell, Efcient use
of side information in mulitiple-antenna data transmission over fading
channels, IEEE Jour. Select. Areas in Commun., vol. 16, no. 8, pp.
14231436, Oct. 1998.
[4] D. J. Love, R. W. Heath, Jr., and T. Strohmer, Grassmannian beamform-
ing for multiple-input multiple-output wireless systems, IEEE Trans.
Inf. Th., vol. 49, no. 10, pp. 27352747, Oct. 2003.
[5] K. K. Mukkavilli, A. Sabharwal, E. Erkip, and B. Aazhang, On
beamforming with nite rate feedback in multiple antenna systems,
IEEE Trans. Inf. Th., vol. 49, no. 10, pp. 2562 2579, Oct. 2003.
[6] E. Visotsky and U. Madhow, Space-time transmit precoding with
imperfect feedback, IEEE Trans. Inf. Th., vol. 47, no. 6, pp. 2632
2639, Sept. 2001.
[7] S. Zhou and G. B. Giannakis, Optimal transmitter eigen-beamforming
and space-time block coding based on channel mean feedback, IEEE
Trans. on Sig. Proc., vol. 50, no. 10, pp. 25992613, Oct. 2002.
[8] , Optimal transmitter eigen-beamforming and space-time block
coding based on channel correlations, IEEE Trans. Inf. Th., vol. 49,
no. 8, pp. 16731690, Jul. 2003.
[9] R. S. Blum, MIMO with limited feedback of channel state information,
in Proc. Int. Conf. Acoust., Speech and Sig. Proc., vol. 4, April 2003,
pp. 8992.
[10] P. Xia and G. B. Giannakis, Design and analysis of transmit-
beamforming based on limited-rate feedback, in Proc. Veh. Technol.
Conf., vol. 3, Sep. 2004, pp. 1653 1657.
[11] D. J. Love and R. W. Heath, Jr., Grassmannian beamforming on
correlated MIMO channels, 2004, to appear in Proc. Glob. Telecom.
Conf.
[12] B. C. Banister and J. R. Zeidler, A simple gradient sign algorithm for
transmit antenna weight adaptation with feedback, IEEE Trans. on Sig.
Proc., vol. 51, no. 5, pp. 11561171, May 2003.
[13] J. C. Roh and B. D. Rao, An efcient feedback method for MIMO
systems with slowly time-varying channels, in Wireless Comm. and
Net. Conf., vol. 2, Mar. 2004, pp. 760 764.
[14] B. Mondal and R. W. Heath, Jr., Channel adaptive feedback for limited
feedback MIMO beamforming systems, accepted to IEEE Trans. Sig.
Proc., May 2005.
[15] B. Mondal, R. Samanta, and R. W. Heath, Jr., Frame-theoretic quan-
tization for limited feedback MIMO beamforming systems, in WIRE-
LESSCOM, 2005.
[16] B. Mondal and R. W. Heath, Jr., Asymptotic performance of block
quantizers on the Grassmann manifold, to be submitted to IEEE Trans.
Sig. Proc., 2005.
[17] A. Forenza, D. J. Love, and R. W. Heath, Jr., A low complexity
algorithm to simulate the spatial covariance matrix for clustered MIMO
channel models, in Proc. Veh. Technol. Conf., vol. 2, May 2004, pp.
889893.
Fig. 4. Uncorrelated Rayleigh fading channel.
Fig. 5. Spatially correlated Rayleigh channel.
Fig. 6. Ricean channel.
380