Performance Comparison of Limited Feedback

Codebook-Based Downlink Beamforming Schemes

for Distributed Antenna Systems
Vinay Uday Prabhu, Sotiris Karachontzitis and Dimitris Toumpakaris
Wireless Telecommunications Laboratory
Department of Electrical and Computer Engineering, University of Patras
Email: vinay@tenet.res.in,{karaxodg, dtouba}@upatras.gr
AbstractIn this paper, we consider limited feedback
codebook-based downlink beamforming involving two Remote
Antenna Units with two antennas each transmitting to a single-
antenna Mobile Terminal. A Central Unit to which the two
Remote Antenna Units are connected enables centralized joint
beamforming. We consider different practical channel scenarios
that may arise in Distributed Antenna Systems and gauge the
performance of the Grassmannian, Kerdock and Unitary-DFT
codebooks. We then compare the performance when modied
code books that exploit information such as transmit antenna
correlation are used. We conclude the paper by listing the
strengths and drawbacks of each of the three codebooks.
I. INTRODUCTION
Codebook-based limited feedback is widely being accepted
as an elegant way to learn the partial channel state information
at the transmitter (CSIT) in practical precoded multiple-input
single/multiple-output (MISO/MIMO) wireless systems [1-5].
Lately, there has been an evolution from vector quantization-
based strategies, such as [5], towards xed pre-known code-
books constructed on the principles of Grassmannian Line
Packing (GLP) [2], [3], maximizing minimum correlation be-
tween columns of truncated Fourier matrices [6], Equiangular
frames [7] and most recently, mutually unbiased bases (MUB)
[8]- [9]. We examine the possibility of using these off-the-
shelf codebooks (that are publicly available through online
repositories such as [10], [11] and which are constructed on the
assumption of the i.i.d Rayleigh channel) under unique channel
scenarios that emerge in Distributed Antenna Systems (DAS).
We consider such a distributed antenna system involving two
Remote Antenna Units (RAUs) which are connected to a
Central Unit (CU) through an optical back-haul. The RAUs
with 2 transmit antennas each beamform jointly to a single-
antenna mobile terminal (MT). In this paper, we compare
the performance of the off-the-shelf Grassmannian, Kerdock
and Unitary-DFT codebooks and benchmark them with the
ideal case (of perfect CSIT) for the DAS under consideration.
In Section II we provide the system model. In Section III,
we briey explain the principle of construction of the three
codebooks. In Section IV, we motivate a few idiosyncratic
scenarios that may arise in a DAS and present simulation
This work was supported by the FUTON research project FP7 ICT-2007-
2015533 of the European Commission.
results comparing the performance of these codebooks as
well as the modied codebooks which take into account the
correlation between transmitting antennas [4]. Harnessing the
results of Section IV, we present the observations in Section
V. We conclude the paper in Section VI.
II. SYSTEM MODEL
Fig. 1 provides a diagrammatic representation of the DAS
under consideration. Considering at fading for all wireless
Fig. 1. The system model of the DAS under consideration.
links between the RAUs and the MT and N
t
= 2 antennas
per RAU, the received symbol at a given time instant can be
written as:
y= h

vs + w, (1)
where s C is the data symbol to be communicated to the MT,
v C
41
is the chosen beamforming vector, h = [h
T
1
h
T
2
]
C
41
is the concatenated MISO channel of the individual
channels (h
1
, h
2
C
21
are the downlink channels from
the rst and the second RAU to the MT respectively) and
w N
C
(0,
2
) is additive white Gaussian receiver noise.
At the receiver, we assume perfect Channel State Information
(CSI) and the existence of error-free and zero-delay feedback
paths to the RAUs. The beamforming vector is chosen from a
xed pre-known codebook T = {v
1
, ..., v
N
} C
4N
. In this
work, we consider the Grassmannian [2], [3], Kerdock [8]- [9]
and Discrete Fourier matrix based Unitary-DFT codebooks [6]
which are under consideration in standards such as 3GPP LTE
[12]. The optimal beamforming vector is chosen as:
v = arg max
i[1,N]
{h

v
i

2
|v
i
T} , (2)
and the received SNR is equal to
SNR =
h

v
2
2
P

2
, (3)
where P = E
_
|s|
2
_
. The instantaneous capacity in bits/s/Hz
of this beamforming scheme is simply
C = log
2
(1 + SNR). (4)
III. KERDOCK, GRASSMANIAN AND UNITARY-DFT
CODEBOOKS - A BRIEF REVIEW
A. Grassmannian Codebooks
Given (1), it is well known that the optimal beamforming
vector is the principal right singular vector of the channel [13].
By exploiting the distribution of this vector in independent
identically distributed (i.i.d) Rayleigh fading matrix channels,
the problem of a size N codebook design was shown to
be equivalent to solving the problem of Grassmannian line
packing (GLP) [3]. The resultant codebooks, known as Grass-
mannian codebooks (T
grass
), satisfy the following design
criterion:
The codebook vectors in T
grass
= [v
1,grass
, ..., v
N,grass
] are
chosen such that they maximize the sine of the smallest angle
between any pairs of lines. That is:
T
grass
= arg max
TC
N
t
N
(T), where,
(T) = sin(
min
) = min
1k<lN
_
1

v
H
k
v
l

2
(5)
B. Unitary-DFT codebooks
The Unitary-DFT codebook used is as designed in Sec. IV-
C of [6]. The codebooks vectors T
dft
= [v
1,dft
, ..., v
N,dft
]
are the columns of the codebook matrix generated as:
T
dft
=
_
D
0
dft
F, D
1
dft
F, ..., D
Ni
dft
F
_
, where
F(k, l) =
1

N
t
e
{
2kl
N
t
}
; F C
NtNt
and D
dft
C
NtNt
.
(6)
Here, N
i
is chosen such that the codebook size, N = N
i
N
t
.
D
dft
is a diagonal generator matrix that is derived as shown
in [6]. As seen, the generation of this codebook requires only
a diagonal matrix and the Fourier matrix which has resulted
in this codebook being proposed for the 3GPP LTE standard
[12].
C. Kerdock Codebooks
Recently, Inoue and Heath [8]- [9] have produced a qua-
ternary alphabet beamforming codebook based on Kerdock
codes that has attractive properties such as systematic con-
struction, minimal storage and online search computability.
This is exciting, as the traditional codebook design often
involved computationally intensive non-linear searches which
would be performed ofine before writing the codebooks at
both the transmitter and the receiver. The special structure
in the codebook has been shown to facilitate derivation of
spatial multiplexing codebooks from multiple columns of the
beamforming codebook resulting in further storage reduction.
Furthermore, as revealed in [8]- [9], this codebook performs
as well as the Grassmannian codebook for the i.i.d. Rayleigh
channel. The codebook T
ker
is generated as,
T
ker
= [D
ker,0
H
Nt
, D
ker,1
H
Nt
, ..., D
ker,Ni1
H
Nt
] .
(7)
Here H
Nt
C
NtNt
is the Sylvester-Hadamard Matrix [15]
of order N
t
and [D
ker,0
, ..., D
ker,Ni1
] C
NtNt
are the
diagonal generator matrices designed as shown in [9]. N
i
is
chosen such that N = N
i
N
t
.
IV. CONSIDERED SCENARIOS AND SIMULATION RESULTS
We compare the three codebooks and benchmark their
performance with respect to the ideal case, which involves
full Channel State Information at the Transmitter (CSIT) at
the RAUs for different channel scenarios. For fair comparison,
the codebooks in all the cases are chosen to be of xed size
(N = 16), which entails a 4-bit feedback. On account of the
beamforming strategy being optimal only at lower SNRs, we
x P/
2
in our simulations to be 5 dB.
A. Rayleigh-I.I.D Case - h N
C
(0, 1)
We consider this case only as a theoretical benchmark.
It is highly unlikely that this scenario occurs in practice.
As predicted in [8], the Kerdock codebook performs very
well (nearly the same as Grassmannian and Unitary-DFT
codebooks) in spite of having codebook symbols drawn from
a restricted quaternary alphabet. The Capacity CDF (CCDF)
comparison is as shown in Fig. 2.
B. Mixed variance Rayleigh fading - (h
1
N
C
(0,
2
1
), h
2

N
C
(0,
2
2
))
Here, the downlink channels are Rayleigh i.i.d. with differ-
ent variances. Fig.3 indicates the CCDF performance compar-
ison. In the gure, Grassmann-Corr denotes the correlated-
Grassmannian codebook derived as described in [4]. For
the model under consideration, the correlated-Grassmannian
codebook T
corr,grass
is generated using
T
corr,grass
=
_
Dv
1,grass
Dv
1,grass

, ...,
Dv
N,grass
Dv
N,grass

_
C
4N
D =
_
I
2
0
0
1
2
I
2
_
; h
1
N
C
(0,
2
1
), h
2
N
C
(0,
2
2
).
(8)
3
Fig. 2. Capacity CDF comparison for the Rayleigh i.i.d. case.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Capacity (Bps/Hz)
P
r
o
b
a
b
i
l
i
t
y
Ideal
Grassmann-Corr
DFT
Kerdock
Grassmann
Ideal
Grassmann-Corr
DFT/ Kerdock
Grassmann
Fig. 3. Capacity CDF comparison for the mixed variance Rayleigh fading
case.
As seen, we get to within 0.2 bps/Hz of the ideal capacity
curve when the correlated-Grassmannian code book is used.
However, this comes at the cost of additional complexity
involved in performing (8). Moreover, use of this codebook
requires the exact knowledge of the correlation matrix.
C. Correlated fading with different transmit correlation ma-
trices
We use the Kronecker MISO model [14]. The individual
transmit correlation matrices are assumed to be different
to accommodate for the different levels of mutual coupling
arising due to dissimilar antenna array-topologies used on the
two RAUs. This makes the simulations more realistic.
h
1
= R
1
2
t1
h
w
; h
2
= R
1
2
t2
h
w
; h
w
N
C
(0, 1);
(9)
Here, R
t1
and R
t2
C
22
are the transmit correlation
matrices of the two RAUs, modeled as
R
t
=
_
1 t
t 1
_
, t 1. (10)
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
0
0.2
0.4
0.6
0.8
1
Capacity (Bps/Hz)
P
r
o
b
a
b
i
l
i
t
y
Ideal
Grassmann
DFT
Kerdock
0 1 2 3 4 5 6
0
0.2
0.4
0.6
0.8
1
Capacity (Bps/Hz)
P
r
o
b
a
b
i
l
i
t
y
Ideal
Grassmann
DFT
Kerdock
Ideal
DFT /Kerdock
Grassmann
Grassmann
Ideal
DFT /Kerdock
Fig. 4. Capacity CDF comparison for Correlated fading with different
transmit correlation matrices.
With reference to Fig.4, we have set the transmit correlation
factors (also termed transmitter cross-correlation coefcients)
high with t
1
= 0.9 and t
2
= 0.8 in the top sub-plot. A
drop in the degree of deterioration in the performance of the
Grassmannian codebook is seen when the transmit correlation
factors are reduced to t
1
= 0.45 and t
2
= 0.4 respectively
in the bottom sub-plot. Here, we assume that the MT does
not re-compute its Grassmannian codebook as in (8) using the
correlation matrix:
D =
_
R
t1
0
0 R
t2
_
. (11)
The computational cost apart, precise knowledge of the trans-
mit correlation factors (t
1
, t
2
) is required and whose estimation
in practice is non-trivial. The Kerdock and Unitary-DFT
codebooks are preferable in this case and they lag the ideal
CCDF curve by about 0.4 bps/Hz.
D. The generic case: Correlated fading, different correlation
matrices with Rayleigh and Ricean links.
This scenario arguably is the most generic and quite realistic
for the DAS under consideration. Referring to Fig. 5, we see
that the user in the scenario depicted has an NLOS link with
RAU-B (h
1
) and a LOS link with RAU-A (h
2
). We model
h
1
as a purely Rayleigh fading channel and h
2
as a Ricean
channel with the Ricean factor: k = 4. That is,
h
w1
N
C
(0,
2
1
); h
w2
N
C
(0,
2
2
); h
LOS
= [1 1]
T
h
1
= R
1
2
t1
h
w1
; h
22
= R
1
2
t2
h
w2
; h
2
=
_
k
k + 1
h
LOS
+
_
1
k + 1
h
22
Ricean Factor : k = 4; t
1
= 0.9; t
2
= 0.8.
(12)
h
1
h
2
Fig. 5. Scenario for the generic channel model case where the MT has a
LOS link with only one of the two RAUs.
The results for this scenario are as shown in Fig. 6. In the
rst sub-plot, we see that the Grassmannian codebook, like
in Fig.3, performs better than the Kerdock and Unitary-DFT
codebooks. This leads to a clear conclusion that the Kerdock
and Unitary-DFT codebooks are sensitive to the Rayleigh
i.i.d. assumption. The second sub-plot clearly highlights the
problem with the standard Grassmannian codebooks. High
antenna correlation affects the performance substantially and
the Grassmannian codebook lags the Kerdock and Unitary-
DFT codebooks by about 0.2 bps/Hz. The related average
ergodic capacity versus SNR curves are as depicted in Fig.
7. The average ergodic capacity is computed as:
C
ave
= E
h
_
log
2
_
1 +
h

v
2
2
P

2
__
. (13)
However, we see a marked improvement if we incorporate
the correlation matrix and modify the Grassmannian codebook.
Exact knowledge of the same will yield a near ideal perfor-
mance which is as shown in Fig. 8.
V. OBSERVATIONS FROM THE SIMULATION RESULTS
In the previous section, we saw that in all the cases, the DFT
and Kerdock codebooks performed almost the same. Kerdock
codes have the least storage requirement and computational
complexity. The restricted quaternary alphabet renders mul-
tiplication operations in to simple sign ipping or ipping
0 0.5 1 1.5 2 2.5 3 3.5 4
0
0.2
0.4
0.6
0.8
1
Capacity (Bps/Hz)
P
r
o
b
a
b
i
l
i
t
y
Ideal
Grassmann
DFT
Kerdock
0 1 2 3 4 5 6
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Capacity (Bps/Hz)
P
r
o
b
a
b
i
l
i
t
y
Ideal
Grassmann
DFT
Kerdock

1
2
=1

2
2
=0.125
t
1
=0.9
t
2
=0.8
k=4

1
2
=1

2
2
=1
t
1
=0.9
t
2
=0.8
k=4
Ideal
Grassmann
/
DFT Kerdock
/
DFT Kerdock
Grassmann
Ideal
Fig. 6. Capacity CDF comparison for the generic channel model.
real and imaginary components. This motivates their usage
in MTs with limited computational and memory capabilities.
The major drawback of this code is its non-availability for odd
number of antennas as well as its limited maximum codebook
size. As shown in [8], the maximum codebook size is limited
to:
N
max
= N
t
(N
t
+ 1) (14)
This is where the DFT codebook nd its utility. The relatively
simple construction and scalability coupled with its superior
resistance to antenna correlation (as seen in Fig. 4,6) makes it
an ideal candidate where the MT is antenna correlation blind
and has enough memory to store larger codebooks. In cases,
where the MT has access to the correlation information as well
as sufcient memory Grassmannian codebooks are the best
choice. As seen in Fig. 3,4 and Fig. 8, Grasmmanian code-
book outperforms the other codebooks in scenarios involving
correlated fading provided we modify the codebook-based
on the nature of the correlation matrix. In the case of xed
codebook size schemes (like the N = 16 case considered),
we deem it prudent to have both the standard Grassmann and
Kerdock codebooks residing on the RAU and the MT. We
can then shrewdly utilize either of the two depending on the
information that the MT possesses about both transmit antenna
correlation and its environment parameters such as the Ricean
k-factor. This of course requires that we use one bit more than
the general case to indicate which of the two codebooks is
being used in the initial phase of communication between the
CU and the MT. It is equally important that we consider the
existence of larger Grassmannian and Unitary-DFT codebooks
(N = 64). Figure 9 clearly highlights the superior performance
obtained by the Grassmannian and Unitary-DFT codebooks
5 0 5 10 15
0
1
2
3
4
5
6
SNR (in dB)
A
v
e
r
a
g
e

C
a
p
a
c
i
t
y

(
C
a
v
e
)

i
n

b
p
s
/
H
z
Ideal
Grassmann
DFT
Kerdock
5 0 5 10 15
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
SNR (in dB)
A
v
e
r
a
g
e

C
a
p
a
c
i
t
y

(
C
a
v
e
)

i
n

b
p
s
/
H
z
Ideal
Grassmann
DFT
Kerdock
Fig. 7. Comparison of average capacity versus SNR for the generic channel
model.
with N = 64. As discussed earlier, we can only have a N = 16
Kerdock codebook (N = 20 if we add the identity matrix too,
which translates in a simple antenna sub-set selection strategy).
Therefore, when RTs support larger codebooks, the N = 64
Grassmannian and Unitary-DFT codebooks come in to picture.
We capture these arguments in Fig.10.
VI. CONCLUSION
In this paper, we have attempted to subject the standard
Grassmannian, Unitary-DFT and Kerdock codebooks to a
limited feedback downlink beamforming scenario in a Dis-
tributed Antenna System. Such systems, owing to the dis-
tributed antenna topology yield unique idiosyncratic channel
realizations which are deviant from the i.i.d. Rayleigh channel.
We endeavored to gauge the performance variation of the
three codebooks, while allowing variations in the form of
modied correlation-aware codebooks as well as codebooks
with different sizes. We then presented the observations listing
the strong and weak points of the three codebooks used.
REFERENCES
[1] A. Narula, M. Lopez, M. Trott and G. Wornell, Efcient Use of
Side Information in Multiple-Antenna Data Transmission Over Fading
Channels, IEEE Journal on Selected Areas of Communications, vol. 16,
no. 8, pp. 1423-1436, Aug. 1998.
[2] K. Mukkavilli, A. Sabharwal, E. Erkip and B. Aazhang, On Beam-
forming with Finite Rate Feedback in Multiple-Antenna Systems, IEEE
Transactions on Information Theory, vol. 49, no. 10, pp. 2562-2579, Oct.
2003.
[3] D. J. Love, R. W Heath Jr. and T. Strohmer, Grassmannian Beam-
forming for Multiple-Input Multiple-Output Wireless Systems, IEEE
Transactions on Information Theory, vol. 49, no. 10, pp. 2735-2747,
Oct. 2003.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Capacity (Bps/Hz)
P
r
o
b
a
b
i
l
i
t
y
Ideal
GrassmannCorr
DFT
Kerdock
Ideal
DFT /Kerdock
GrassmannCorr
Fig. 8. Capacity CDF comparison for the generic channel model using the
correlated Grassmann codebook.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Capacity (Bps/Hz)
P
r
o
b
a
b
i
l
i
t
y
Ideal
GrassmannN=64
DFTN=64
KerdockN=16
Ideal
GrassmannN=64
DFTN=64
KerdockN=16
Fig. 9. Capacity CDF comparison for the generic channel model using the
N = 64 Grassmannian and Unitary-DFT codebooks.(
2
1
=
2
2
= 1|k =
4|t
1
= 0.9, t
2
= 0.8.)
[4] D. J. Love and R. W. Heath Jr. , Limited Feedback Diversity Techniques
for Correlated Channels, IEEE Transactions on Vehicular Technology,
vol. 55, no. 2, pp. 718-722, March 2006.
[5] J. Roh and B. Rao, Transmit Beamforming in Multiple-Antenna
Systems with Finite Rate Feedback: a VQ-Based Approach, IEEE
Transactions on Information Theory, vol. 52, no. 3, pp. 1101-1112,
March 2006.
[6] D. J. Love and R. W. Heath Jr, Limited Feedback Unitary Pre-coding
for Orthogonal Space-Time Block Codes, IEEE Transactions on Signal
Processing, vol. 53, no. 1, pp. 64-73, Jan. 2005.
[7] B. Mondal, R. Samanta and R. W. Heath Jr. , Frame Theoretic
Quantization for Limited Feedback MIMO Beamforming Systems in
Proc. Int. Conf. on Wireless Networks, Communications and Mobile
Computing, pp. 1065-1070 vol. 2, 2005.
[8] T. Inoue and R. W. Heath Jr. , Kerdock Codes for Limited Feedback
MIMO Systems, in Proc. IEEE ICASSP (Las Vegas, NV), pp. 3113-3116,
March 30-April,4-2008.
[9] T. Inoue and R. W. Heath Jr, Kerdock Codes for Limited Feed-
RT has limited computational capabilities
Serious memory constraints
KERDOCK CODEBOOK
RT has no access to correlation information
Memory can support larger codebooks
DFT CODEBOOK with
larger N, say N=64
RT has access to correlation information
But limited memory
Grassmannian CODEBOOK
N=16
N=16
RT has access to correlation information
Grassmannian CODEBOOK
N=64
Memory can support larger codebooks
Fig. 10. Diagrammatic representation of the observations.
back Precoded MIMO Systems, accepted for publication in IEEE
Transactions on Signal Processing, also available at: http://arxiv.
org/PS cache/arxiv/pdf/0711/0711. 0557v2. pdf
[10] http://users. ece. utexas. edu/ inoue/codebook/index.html
[11] http://cobweb. ecn. purdue. edu/ djlove/grass. html
[12] 3GPP, Physical Layer Aspects of UTRA High Speed Downlink Packet
Access, Technical Report TR25. 814, 2006. [Online] available at:
http://www. 3gpp. org/ftp/Specs/html-info/Meetings-R1.htm
[13] D. Tse and P. Viswanath, Fundamentals of Wireless Communications,
Cambridge University Press, Cambridge UK, 2005.
[14] J. P. Kermoal, L. Schumacher, K. I. Pedersen, P. Mogensen and F. Fred-
eriksen, A Stochastic MIMO Radio Channel Model with Experimental
Validation, IEEE Journal on Selected Areas in Communications, vol.
20, no. 6, pp. 1211-1226, June 2002.
[15] K. J. Horadam, Hadamard matrices and their applications, Princeton
University Press, Princeton, NJ, 2007.