Antenna Grouping Based Feedback Compression For FDD-based Massive MIMO Systems

This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TCOMM.2015.2460743, IEEE Transactions on Communications
TO APPEAR IN IEEE TRANSACTIONS ON COMMUNICATIONS 1
Antenna Grouping based Feedback Compression for

FDD-based Massive MIMO Systems
Byungju Lee, Member, IEEE, Junil Choi, Member, IEEE, Ji-Yun Seol, Member, IEEE, David J.
Love, Fellow, IEEE, and Byonghyo Shim, Senior Member, IEEE,
AbstractRecent works on massive multiple-input multiple- of each terminal inversely proportional to the number of
output (MIMO) have shown that a potential breakthrough in basestation antennas for uplink [4].
capacity gains can be achieved by deploying a very large number Presently, standardization activity for massive MIMO has
of antennas at the basestation. In order to achieve the perfor-
mance that massive MIMO systems promise, accurate transmit- been initiated [5], [6], and there is on-going debate regarding
side channel state information (CSI) should be available at the the pros and cons of time division duplexing (TDD) and
basestation. While transmit-side CSI can be obtained by employ- frequency division duplexing (FDD). In obtaining the channel
ing channel reciprocity in time division duplexing (TDD) systems, state information (CSI), FDD requires the CSI to be fed back
explicit feedback of CSI from the user terminal to the basestation through the uplink [7] while no such procedure is required for
is needed for frequency division duplexing (FDD) systems. In this
paper, we propose an antenna grouping based feedback reduction TDD systems owing to the channel reciprocity [8]. In fact,
technique for FDD-based massive MIMO systems. The proposed under the assumption that RF chains are properly calibrated
algorithm, dubbed antenna group beamforming (AGB), maps [9], the CSI of the downlink can be estimated using the pilot
multiple correlated antenna elements to a single representative signal in the uplink. Due to this benefit, most of the massive
value using pre-designed patterns. The proposed method modifies MIMO works in the literature have focused on TDD [10]
the feedback packet by introducing the concept of a header
to select a suitable group pattern and a payload to quantize (possible exceptions are [11][19]). However, FDD dominates
the reduced dimension channel vector. Simulation results show current cellular networks and offers many benefits over TDD
that the proposed method achieves significant feedback overhead (e.g., small latency, continuous channel estimation, backward
reduction over conventional approach performing the vector compatibility), and it is important to identify and develop
quantization of whole channel vector under the same target sum solutions for potential issues arising from FDD-based massive
rate requirement.
MIMO techniques.
Index TermsMassive multiple-input multiple-output, antenna One well-known problem of FDD system is that the amount
group beamforming, feedback reduction, vector quantization, of CSI feedback must scale linearly with the number of anten-
Grassmannian subspace packing.
nas to control the quantization error [20][23]. Therefore, it is
not hard to convince oneself that the overhead of CSI feedback
I. I NTRODUCTION is a serious concern in the massive MIMO regime. Needless to
Multiple-input multiple-output (MIMO) systems with large- say, a technique that efficiently reduces the feedback overhead
scale transmit antenna arrays, often called massive MIMO, while affecting minimal impact on system performance is
have been of great interest in recent years because of their crucial to the success of FDD-based massive MIMO systems.
potential to dramatically improve spectral efficiency of future In this paper, we provide a novel framework for FDD-
wireless systems [3], [4]. By employing a simple linear based massive MIMO systems that achieves a reduction in the
precoding in the downlink and receive filtering in the uplink, CSI feedback overhead by exploiting the spatial correlation
massive MIMO systems can control intra-cell interference and among antennas. The proposed algorithm, henceforth dubbed
thermal noise [3]. Additionally, massive MIMO can improve antenna group beamforming (AGB), maps multiple correlated
the power efficiency by scaling down the transmit power antenna elements to a single representative value using prop-
erly designed grouping patterns. When the antenna elements
B. Lee and B. Shim are with Institute of New Media and Communications are correlated, the loss caused by grouping antenna elements is
and School of Electrical and Computer Engineering, Seoul National Univer- shown to be small, meaning that grouping of antenna elements
sity, Seoul, Korea (e-mail:[email protected],[email protected]).
J. Choi is with the University of Texas at Austin, Austin, TX, USA (e- with correlated channels is an effective means of reduced
mail:[email protected]). dimension channel vector generation. In fact, by allocating
D. J. Love is with School of Electrical and Computer Engineering, Purdue a small portion of the feedback resources to represent the
Univ., West Lafayette, IN, USA (email:[email protected]).
J. Seol is with Samsung Electronics Co., Ltd., Suwon, Korea grouping pattern, the number of bits required for channel
(email:[email protected]). vector quantization can be reduced substantially, resulting in a
This work was supported by the National Research Foundation of significant reduction in feedback overhead. In order to support
Korea (NRF) grant funded by the Korean government (MSIP) (No.
2014R1A5A1011478 and 2014049151) and MSIP (Ministry of Science, ICT the antenna grouping operation, the proposed AGB algorithm
& Future Planning), Korea in the ICT R&D Program 2013 (No. 1291101110- uses a new feedback packet structure that divides the feedback
130010100). resources into two parts: a header to indicate the antenna group
This paper was presented in part at the International Conference on Com-
munications (ICC), 2014 [1] and Vehicular Technology Conference (VTC), pattern and a payload to indicate the codebook index of the
2014 [2]. reduced dimension channel vector. At the user terminal, a pair
0090-6778 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
of group pattern and codeword minimizing the quantization

Basestation CSI
distortion is chosen. Using the information delivered from the User 1 quantization
1
user terminal, the basestation reconstructs the full-dimensional s1
channel vector and then performs transmit beamforming. 2
s2
In our analysis, we show that when the transmit antenna CSI
Beamforming x
elements are correlated, the proposed AGB algorithm exhibits W H
User 2
quantization
smaller quantization distortion than the conventional vector
Nt-1
quantization employing a channel statistic-based codebook. sK
This in turn implies that the number of quantization bits Nt
CSI
required to meet a certain level of the performance for the quantization
User K
AGB algorithm is smaller than that of conventional vector
quantization under the same level of quantization distortion. Codeword index
(B bits/user)
We also investigate an estimated required number of feedback
bits to maintain a constant gap with respect to the system Fig. 1. CSI feedback in the multi-user downlink system.
with perfect CSI. It is shown that the use of antenna grouping
in correlated channels enables to considerably reduce the
amount of feedback overhead. Moreover, due to the fact where Rt,k CNt Nt is the transmit correlation matrix of the
that dimension of the codeword being searched is reduced, k-th user [24] and gk, CNt is the innovation process whose
and hence the proposed AGB brings additional benefits in elements are independent and identically distributed according
search complexity over the conventional vector quantization. to gk, CN (0, INt ) and is a temporal correlation coeffi-
We confirm by simulation on realistic massive MIMO channels cient (0 1). We assume the block-fading channel has
that the proposed AGB algorithm achieves up to 20%70% a coherence time of L, which means that the channel is static
savings in feedback information over the conventional vector for L channel uses in each block and changes from block-to-
quantization under the same target sum rate requirement. block. In this setup, the received signal of the k-th user for
The remainder of this paper is organized as follows. In the n-th channel use in the -th fading block can be expressed
Section II, we briefly review the system model and the con- as
ventional beamforming technique. In Section III, we provide
yk, [n] = hH H
k, wk, sk, [n] + hk, wj, sj, [n] + zk, [n] (1)
a detailed description of the proposed AGB algorithm and
j=k
subspace packing based grouping pattern generation scheme.
We present the simulation results in Section IV and present where hk, C is the channel vector from the basestation
Nt
conclusions in Section V. antenna array to the k-th user, wi, CNt is the unit norm
Notations: Lower and upper boldface symbols are used beamforming vector (wi, 2 = 1), si, [n] C is the message
to denote vectors and matrices, respectively. The superscripts signal for the i-th user, and zk, [n] CN (0, 1) is normalized
()H , ()T , and () denote Hermitian transpose, transpose, and additive white Gaussian noise at the k-th user. Since the
conjugate, respectively. X and XF are used as the two- beamforming is performed separately per block, in the sequel
norm and the Frobenius norm of a matrix X, respectively. E[] we focus on the operation of a single block and drop the fading
denotes the expectation operation, and CN (m, 2 ) indicates a block index . The matrix-vector form of (1) is expressed as
complex Gaussian distribution with mean m and variance 2 .
y[n] = Hx[n] + z[n] (2)
tr() is the trace operation, and vec(X) is the vectorization of
matrix X. Let X C|||| denote a submatrix of X whose where H = [h1 h2 . . . hK ]H CKNt is the composite
(i, j)-th entry is X ((i), (j)) for i, j = 1, . . . , || ( is the channel matrix, z[n] = [z1 [n] z2 [n] . . . zK [n]]T CK is
set of partial indices and || is the cardinality of ). the complex Gaussian noise vector (z[n] CN (0, IK )), x[n]
is the transmit vector normalized with the power constraint
II. MIMO B EAMFORMING (E[x[n]2 ] = P ), and y[n] = [y1 [n] y2 [n] . . . , yK [n]]T is
A. System Model and Conventional Beamforming the vectorized received signal vector. In order to control the
We consider a multiuser multiple-input single-output inter-user interference, beamforming is applied using x[n] =
(MISO) downlink channel with Nt antennas at the basestation Ws[n] where W = [w1 w2 . . . wK ] CNt K and s[n] =
and K user terminals each with a single antenna1 (see Fig. [s1 [n] s2 [n] . . . sK [n]]T CK are the beamforming matrix
1). We assume spatially correlated and temporally correlated and the message vector, respectively.
block-fading channels where the channel vector hi, follows In generating the beamforming vectors, we consider zero-
the first-order Gauss-Markov model as forcing beamforming (ZFBF) ( [25][29]
) where the right pseudo
1
1/2 inverse2 Wzf = HH HHH of the quantized channel
hk,0 = Rt,k gk,0
matrix H = [h1 , h2 , , hK ] is applied to the message
H
1/2
hk, = hk,1 + 1 2 Rt,k gk, , 1 vector s[n] to alleviate the inter-user interference. In order to
1 Note that the proposed method can be easily extended to a MIMO scenario 2 The proposed AGB in this paper is aimed to generate h with reduced
k
by vectorizing the channel vector corresponding to each receive antenna. For feedback overhead and can be applied to any precoding method including
simplicity, we consider the MISO setup for the rest of this paper. ZFBF.
satisfy the transmit power constraint, the beamforming vector III. A NTENNA G ROUPING BASED F EEDBACK R EDUCTION
wk should be normalized as The key feature of the proposed scheme is to map mul-
k
Wzf tiple correlated antenna elements into a single representative
wk = (3) value using grouping patterns. As a result, the channel vector
Wzf
k
dimension is reduced and a codeword is chosen from the
k codebook generated by the reduced dimension vector. When
where Wzf is the k-th column of Wzf . Under the assumption
that the basestation allocates equal power for all users3 , the the channel is correlated (i.e., antenna elements in a group are
achievable rate of the k-th user is similar), the loss caused by the grouping of antenna elements
( ) is negligible and the target performance can be achieved with
K |hk wk |
P H 2
Rk = log2 1 + K (4) smaller number of feedback bits than a conventional scheme
j=1,j=k |hk wj |
P H 2
1+ K requires. In this section, we explain the overall procedure of
the proposed AGB algorithm and then discuss the pattern set

K
design problem. We also analyze the quantization distortion
and the corresponding sum rate becomes Rsum = Rk .
k=1
caused by the proposed AGB technique and show that the
quantization distortion indeed decreases with the transmit
correlation coefficient.
B. Conventional Limited Feedback
In order to feed back the CSI, a user quantizes its channel
direction hk = h hk
to a unit norm vector hk . Specifically, A. AGB Algorithm
k
the k-th user chooses the quantized vector (codeword) hk from As mentioned, the AGB algorithm reduces the dimension
a pre-defined B-bit codebook set C = {c1 , , c2B } that is of the channel vector from Nt to Ng (Nt > Ng ) by mapping
closest to its channel direction:4 multiple correlated antenna elements to a single representative
value (see Fig. 2). While a conventional scheme employs all
hk = arg max |hH 2
k c| . (5) feedback resources (B bits) to express the quantized channel
cC
vector, the proposed method uses a part of the feedback
Then, the index of the chosen codeword hk is fed back to the resources to quantize the (reduced dimension) channel vector
basestation. and the rest to express the grouping pattern. Both basestation
In general, the number of bits needed to express the and user terminal share a codebook of channel vector and
codeword should be scaled with the dimension of the channel grouping pattern matrices, and thus the receiver feeds back
vector to be quantized to control the distortion caused by the the index of these. In order to support this operation, we
quantization process. In particular, when there is no spatial divide the feedback resources into two parts: a header (Bp
correlation among antenna elements and the random vector bits) to indicate the antenna group pattern and a payload (B-
quantization (RVQ) codebook is used, the number of feedback Bp bits) to represent an index of the quantized channel vector
bits per user should be scaled with the number of transmit (see Fig. 3). Since antenna grouping based quantization is
antennas and SNR (in decibels) as [21], [33] performed separately for each user, in the sequel we focus
Nt 1 on the operation of a single user and drop the user index k.
Buser = (Nt 1) log2 P PdB (6) Suppose there are NP = 2Bp antenna group patterns, then
3
NP distinct reduced dimension channel vectors are generated.
to maintain a constant gap in terms of the sum rate from
Each pattern converts an Nt -dimensional channel vector into
the system with perfect CSI. Hence, when the number of
an Ng -dimensional vector by multiplying the channel vector
transmit antennas increases, the feedback overhead needs to
by a grouping matrix G(i) RNg Nt . The reduced dimension
be increased as well (e.g., Buser = 210 when Nt = 64, (i)
channel vector hr CNg of the group pattern i is
PdB = 10), let alone the computational burden caused by the
r = G h, i = 1, , NP .
h(i)
codebook selection. Therefore, a reduction in the number of (i)
(7)
channel vector dimensions would be beneficial in reducing the
feedback overhead of the FDD-based massive MIMO systems. In Fig. 4, we illustrate the concept of antenna group patterns.
One simple way to generate the reduced dimension channel
3 In this paper, we consider the equal power allocation scenario for simplic-
vector is to average the channel coefficients in an antenna
ity. In order to maximize the sum rate, one might consider more deliberate group. For example, if h = [h1 h2 h3 h4 ]T and two adjacent
power allocation strategies (e.g., waterfilling after the block diagonalization
[30]). channel coefficients (first and second, third and fourth) are
4 In practice, each user quantizes the estimated channel, which is obtained grouped, the mapping matrix is
using the observations of the pilot signals. Once the channel information [1 1 ]
corresponding to the pilot signals are estimated, the channel information for (i) 2 2 0 0
G = (8)
the data tones are generated via proper interpolation among pilot channels. 0 0 12 21
With an aim of reducing the pilot overhead of massive MIMO systems,
various approaches have been proposed in recent years. In [31], an algorithm (i)
exploiting data tones for channel estimation has been proposed and also a and hr = G(i) h = [ h1 +h 2
2 h3 +h4 T
2 ] .
pilot allocation strategy based on the sparsity of channel impulse response (i)
Once the reduced dimension channel vector hr is ob-
and compressive sensing (CS) principle [32]. Once the estimated channel (i)
information is obtained at the receiver, correlation matrix can be estimated tained, hr is quantized by a B Bp bit codebook C =
using samples of instantaneous channel information (i.e., R = E[hhH ]). {c1 , , c2BBp }. It is worth mentioning that a codebook
Pattern NP
Pattern 1
h1 + h2 + h3 + h4
hr,1 =
4
Vector
Antenna grouping Expansion
quantization
Reduced dimension
Original channel vector Quantized vector
channel vector Expanded quantized vector
h = [h1 h 2 h3 h 4 L h13 h14 h15 h16 ]T
hr = [hr,1 h r,2 h r,3 h r,4 ]T h r = [h r,1 h r,2 h r,3 h r,4 ]T ~
hr = [h r,1 h r,1 h r,1 h r,1 Lh r,4 h r,4 h r,4 h r,4 ]T
Fig. 2. Illustration of the AGB algorithm for Nt = 16, Ng = 4. The reduced dimension channel vector hr is obtained by mapping antenna elements of a
(i)
group as a representative value. Note that hr is the quantized version of hr and hr is expanded version of hr .
(i) Nt (i)
Feedback packet in hr to Ng elements in hr , is performed by multiplying
an expansion matrix E(i) RNt Ng to hr . The expanded
(i)

(i)
quantized vector hr is expressed as
r = E hr , i = 1, , NP
h(i) (i) (i)
(10)
Channel vector quantization (B bits) Nt
where E(i) = G(i)T and satisfies G(i) E(i) = INg ( = N g
).
(a) (i) (i)T
For
[ example, ]for the grouping matrix in (8), E = G =
Feedback packet T
1 1 0 0
and the expanded quantized vector is
0 0 1 1
[ ]T
(i) (i) (i) (i)
h(i) (i) (i)
r = E hr = G
(i)T (i)
hr = hr,1 hr,1 hr,2 hr,2 .
Pattern selection (Bp bits) Channel vector quantization (B-Bp bits) (11)
The group pattern index i minimizing the distortion between
(b) (i)
h and hr is
Fig. 3. Feedback packet structure: (a) conventional method and (b) proposed
method. i = arg min D(h, h(i)
r ). (12)
i=1, ,NP
Once the pattern index i is obtained, this index and the

designed for i.i.d channels is not a proper choice for correlated corresponding codeword index are sent to the basestation.
channels so that we use a channel statistic-based codebook for After receiving the pattern index and codeword index of
channel vector quantization [34] (see Section III.C for details). all user terminals, the basestation decompresses the reduced
(i) (i )
The codeword hr maximizing the absolute inner product with dimension channel vector via the expansion (h = E(i ) hr )
(i)
hr is chosen as and then performs the beamforming using the composite
channel matrix H. A block diagram of the proposed AGB
r = arg max |hr
h(i) c| , i = 1, , NP
(i)H 2
(9) algorithm is depicted in Fig. 5.
cC
(i) h(i)
where hr = (i) r
is the direction of the reduced dimension B. Antenna Group Pattern Generation
hr
channel vector for the i-th pattern. This process is repeated Since multiple correlated antenna elements are mapped to a
(i)
for each group pattern and Np candidate codewords hr , i = single representative value, the AGB algorithm is sensitive to
1, , NP , are chosen in total. the choice of the antenna group pattern. Apparently, selecting
Once NP candidate codewords are obtained, we need to se- the best pattern among all possible combinations would be an
lect the codeword that minimizes the distortion between h and ideal option but it is undesirable since the number of patterns
(i) (i)
hr . We note that the direct comparison between h and hr is increases exponentially with the number of transmit antennas.
(i)
not possible since the dimension of hr CNg is smaller than Without doubt, a simple yet effective pattern design is crucial
that of the original channel vector h CNt . In computing the to the success of the AGB algorithm.
(i) (i)
distortion defined as D(h, hr ) = E[h2 (1 |hH hr |2 )] One easy and intuitive way to construct an antenna group
caused by the grouping and quantization, therefore, we use pattern E(i) is to group highly correlated antenna elements
(i) (i)
hr CNt , an expanded version of hr . The expansion together. Typically, adjacent antenna elements are highly cor-
process, which essentially is done by copying each element related so that the grouping of nearby antenna elements would
Pattern 1 Pattern 2 Pattern 3

h1 h5 h9 h13
h2 h6 h10 h14
h3 h7 h11 h15
h4 h8 h12 h16
Fig. 4. Example of antenna group patterns (Nt = 16, Ng = 8, NP = 3). Antenna elements belonging to the same pattern are mapped to one representative
value.
be a desirable option in practice (see the example in Fig. randomly selected patterns (see Section IV.B). We summarize
4). Alternatively, one can consider Grassmannian subspace the antenna group pattern generation procedures in Table I.
packing in the design of the antenna group patterns [35]. The As a further means to lessen the computational burden of
main goal of Grassmannian subspace packing is, when the pattern generation process, we partition the antenna array into
subspace distance metric and the number of feedback bits multiple sub-arrays and then apply Grassmannian subspace
B are provided, to find a set of 2B subspaces in G(Nt , m) packing for each sub-array. In the partitioning process, we
that maximizes the minimum subspace distance between any basically divide the antenna array such that the divided sub-
pair of subspaces in the set.5 The chordal distance has been arrays are close to the square matrix. Specifically, when the
popularly used as a metric to measure the distance [36]. Our partitioning level is two, for the Nt1 Nt2 dimensional
task of generating the pattern set is similar in spirit to the antenna array (Nt = Nt1 Nt2 ), we divide the axis with larger
Grassmannian subspace packing based codebook generation dimension. That is, if Nt1 Nt2 , then the dimension of each
in the sense that we construct a pattern set (containing 2Bp sub-array becomes N2t1 Nt2 (see Fig. 6). When partitioning
patterns) from all possible pattern candidates (E(i) RNt Ng ) level is larger than two, we perform the same procedure for
using a distance metric exploiting the spatial correlation each partitioned sub-array. In doing so, we achieve significant
among antenna elements. reduction in the number of antenna group pattern candidates.
In the first step of the pattern set design, we compute the For example, if Nt = 16, Ng = 8, and M = 1 (no partition),
(i)
quasi-correlation matrix norm Rt F to measure the spatial then the total number of antenna group pattern candidates
proximity of the antenna elements in the antenna group. The Nmax is about 2 106 (see Table I). Whereas, if the antenna
(i) (i) 1/2
quasi-correlation matrix Rt , defined as Rt = Rt E(i) , M (M = 2), then Nmax will be expressed
array is partitioned
(i) as Nmax =
captures the actual influence of the pattern E on the trans- i=1 Nmax,i where Nmax,i is the total number
mit correlation matrix Rt . In general, a pattern generated of candidates for the i-th sub-array. Since Nmax,i = 105 in
by grouping closely spaced antenna elements tends to have this case, Nmax 104 . Note that since the pattern generation
a higher quasi-correlation matrix norm than that generated process is performed off the shelf, this process does not affect
by grouping antenna elements apart. Thus, one can deduce the real-time operation.
that a pattern with a large-correlation matrix norm exhibits
lower grouping loss than that with a small quasi-correlation C. Quantization Distortion Analysis
matrix norm. For patterns with high quasi-correlation matrix
norm, we perform subspace packing to generate 2Bp patterns We now turn to the performance analysis of the AGB
(expansion matrices) maximizing the minimum distance metric algorithm. In our analysis, we analyze the distortion D induced
h
between any pair of subspaces. In measuring the distance, we by the quantization of the channel direction vector h = h ,
use the correlation matrix distance dcorr (A, B) between two which is defined as
[ ]
matrices A and B [37] D = E h2 |hH hr |2
tr(AH B) [ ( )]
dcorr (A, B) = 1 . (13) = E h2 1 |hH hr |2 (14)
AF BF
Note that dcorr (A, B) measures the orthogonality between two where hr is the expanded version of the quantized vector hr
correlation matrices A and B. When the correlation matrices (see (10)).
are equal up to a scaling factor, dcorr is minimized (dcorr = 0). In the evaluation of the distortion D, we use the quanti-
Whereas, when the inner product between the vectorized zation cell upper bound (QUB) [38]. As mentioned, since a
correlation matrices is zero (i.e., vec(A) and vec(B) are codebook designed for the i.i.d channels is not the right choice
orthogonal), dcorr is maximized (dcorr = 1). In our numerical for correlated channels, we employ a channel statistic-based
simulations, we show that the proposed subspace packing codebook obtained by applying the transmit correlation matrix
1/2
approach achieves a substantial gain over an approach using Rt to the codebook generated from the Grassmannian line
packing. Let fi Cr be the i-th unit norm vector generated
5 G(N , m) is the set of m-dimensional subspaces in CNt (or RNt ). from the Grassmannian line packing, then the set of B-bit
t
Basestation
1 User k ~
s1 Pattern h r( i )
selection
s2 2
Antenna group
Antenna group Beamforming x i* pattern expansion
pattern expansion H
W
yk Antenna group h(ir ) Vector h (ri)
Nt-1 pattern mapping quantization
sK
Nt
i* c*
Pattern index Codeword index

(Bp bits/user) (B-Bp bits/user)
Fig. 5. Overall transceiver structure of the proposed AGB technique.
and hence hr = hA = [h1 h3 ]T (hB = [h2 h4 ]T ).

A-iii) Antenna elements in a group are highly correlated.
That is, E[|hHA hr | ] E[|hB hr | ] where hr is
2 H 2
the quantized vector of hr and generated from the

channel statistic-based codebook. Note that this as-
sumption justifies the use of antenna grouping pattern
in (17).
M=1 M=2
Fig. 6. Antenna group pattern of the partitioned sub-arrays when M = 1 (no It is worth mentioning that depending on the way of grouping
partition) and M = 2.
elements of antenna array, there are several ways to generate
subvectors hA and hB . Also, the group pattern used in (17)
codewords for the channel statistic-based codebook is [34] might be worse than the pattern generated by the subspace
{ }
1/2 1/2 packing or proposed in (8). Thus, assumptions in A-i) and
Rt f1 Rt f2B
C = {c1 , , c2B } = 1/2
, , 1/2
. (15) A-ii) would be clearly pessimistic, but makes our analysis
Rt f1 Rt f2B
tractable. The following proposition provides an approximate
When the channel statistic-based codebook is used, the nor- upper bound of the quantization distortion D under these
2 ] = 1 |h ci |2 between the channel
D H
malized distortion E[h assumptions.
direction vector h and codebook vector ci can be upper
bounded as [24] Proposition 1. The quantization distortion D of the AGB
Ri {h : 1 |hH ci |2 } (16) algorithm under the channel statistic-based codebook satisfies
22 r1
B
where = 12
2 (i is the i-th largest singular value of the
transmit correlation matrix Rt Crr ) and B is the number

of quantization bits. D . Nt + 2Nt (18)
In our analysis, we restrict our attention to the scenario
where two antenna elements are mapped to a single representa-
tive value for mathematical tractability. Nevertheless, since the
key factor affecting the quantization distortion is the transmit BBp
22
correlation coefficient (see (30)), our results can be readily where = 2 Ng 1
is an upper bound of the normalized
12
applied to the general scenario where more than two antenna distortion between hA and hr as defined in (16) (i is the i-th
elements are grouped together. The minimal set of assumptions largest singular value of Rt,A )6 and is the correlation coef-
used for the analytical tractability are as follows:
ficient between two random variables h2 and 1 |hH hr |2 .
A-i) The channel vector h is partitioned into two sub-
vectors hA and hB . hA and hB are composed of
odd and even entries in h (i.e., hA = [h1 h3 ]T ,
hB = [h2 h4 ]T ). Thus, Ng = N2t .
A-ii) The reduced dimension channel vector hr is de-
1 2 3
signed such that hr = hA . For example, if h = 1 2
6 For example, if Rt
= 2 and A = {1, 3}, then
[h1 h2 h3 h4 ]T , then the antenna grouping pattern is 1
[ ] [ ] 3 2 1
1 0 0 0 1 2
G= (17) Rt,A = .
0 0 1 0 2 1
TABLE I
S UMMARY OF THE ANTENNA GROUP PATTERN GENERATION .
Initialization Bp : the number of bits for the pattern set

S: the set of patterns to be selected
Main operation 1) Initialize the index set (= {1, . .). , Nmax } where
Ng 1 Nt n
n=0
Nmax = Ng !
.
Nt
= Ng
is the number of elements in an antenna group.
2) For each pattern i , calculate the Frobenius norm of the quasi-correlation matrix
(i)
ri = Rt F .
Without loss of generality, assume r1 r2 rNmax .
3) Choose J( 2Bp ) patterns T = {r1 , . . . , rJ }.
4) Apply the subspace packing to T to generate the pattern set S.
( J ) k,1 k,2 k,2Bp
Construct Nc = 2B p candidate sets {Sk }Nk=1 where Sk = {Rt , Rt , , Rt
c
}.
Rk,i
t is the i-th quasi-correlation matrix from the k-th candidate set.
5) Calculate the minimum dcorr of a Sk
dk,min (Sk ) = min1mn2Bp dcorr (Rk,m t , Rk,n
t ).
Decide the pattern set S
S = arg maxk=1, ,Nc dk,min (Sk ).
Proof. Using (14), we have where (a) is because |hH hr |2 = |hH A hr + hB hr | , (b)
H 2
[ ( )] H
follows from E[Re((hA hr ) (hB hr ))] 0 (see Appendix
H
D = E h2 1 |hH hr |2 (19) A), (c) follows from A-iii), and (d) follows from the
[ ] ( [ ])
QUB in (16). That is, by plugging h = h hA
, C =
= E h2 1 E |hH hr |2 + Cov(h2 , 1 |hH hr |2 ) { 1/2 1/2
} A
BBp
Rt,A f1 R f BBp
2
(20) , , t,A 2
, and = 22 2 Ng 1 into
( [ ])
1/2 1/2
Rt,A f1 Rt,A f BBp
[ ] [ 2 ] [
1
]
= E h2 1 E |hH hr |2 (16), we get E |hH h | 2
(1 )E h 2
. Finally, (e)
A r A
( ) [ ]
follows from E hA 2 = Nt = 2 .
N g 1
+ V ar (h ) V ar 1 |hH hr |2
2 (21)
( [ ]) Plugging (23) into (22), we have
[ ]
E h2 1 E |hH hr |2
[ ]
( [ ])2 D E h2 + V ar [h2 ] 1 (1 )2 (24)
+ V ar [h2 ] 1 E |hH hr |2 (22)
= Nt + Nt (1 (1 )2 ) (25)

where (20) is because Cov(X, Y )= E[XY ] E[X]E[Y ], Nt + 2Nt (26)
(21) is because Cov(X, ( Y ) = V
) ar(X) [V ar(Y ) ]and
(22) is because V ar 1 |h hr | H 2
= E |hH hr |4 where (25) is because E[h2 ] = Nt and V ar[h2 ] = Nt ,
( [ ])2 ( [ ])2
E |hH hr |2 1 E |hH hr |2 . and (26) is because (2 ) 2 where 2, which is the
[ ] desired result.
The normalized distortion term 1 E |hH hr |2 in the
right-hand side of (22) is approximately upper bounded as
[ ] We note that the relationship between the quantization
1 E |hH hr |2
distortion D and the transmit antenna correlation is not clearly
[ ]
(a)
= 1 E |hH h + h H
h | 2 shown in (18). When a specific correlation model is used,
A r B r
however, we can observe the relationship between the two. For

= 1 E[|hH
A hr | + |hB hr | + (hA hr ) (hB hr )
2 H 2 H H
example, if the exponential correlation model is employed, the
+ (hH H transmit correlation matrix Rt is expressed as [39]
B hr ) (hA hr )]
[ ] [ ]
= 1 E |hH A hr |
2
E |hH
B hr |
2
[ ] 1 Nt 1
2E Re(hH H 1 Nt 2
A hr ) (hB hr )
Rt = .. .. .. .. (27)
(b) [ ] [ ] . . . .
1 E |hH A hr |
2
E |hH
B hr |
2
( )Nt 1 ( )Nt 2 1
(c)
[ ]
= 1 2E |hH A h r | 2
(d) [ ] where = ej is the transmit correlation coefficient, and

. 1 2(1 )E hA 2 is the magnitude of correlation coefficient, and is the phase
(e) of the coefficient. When the number of transmit antennas Nt
= (23) is large, (non-ordered) singular value i of Rt approximately
behaves as [40]

Nt 1 0.6
2ik
i |k| ej Nt
k=(Nt 1) 0.5
Normalized quantization distortion

1 2
, i = 1, . . . , Nt . (28)
1 + 2 2 cos( 2i
Nt )
0.4
Using the first and second largest singular values of7 (28), we 0.3
have
0.2
2 1 + 2 2
t 1
. (29)
1 1 + 2 2 cos(2 NN t
) 0.1
Analysis upper bound (AGB algorithm)
Numerical evaluation (AGB algorithm)
Analysis upper bound (Conventional vector quantization)
BBp Numerical evalution (Conventioanl vector quantization)
22 Ng 1 0
Using this together with = 12
2 in Proposition 3.1, 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Correlation coefficient ||
we have
BBp
(1+2 2)2
D . Nt Nt 1 2 2 Ng 1 Fig. 7. Normalized quantization distortion as a function of the correlation
(1+2 2 cos(2 N ) ) coefficient (Nt = 16, Ng = 8, B = 16, Bp = 8).
t
1+2 2
p
2(N 1)
BB
+ 2Nt Nt 1 2
g . (30)
1+2 2 cos(2 N ) Jensens inequality, R(P ) can be upper bounded as [21]
t

In (30), we observe that the quantization distortion D de- P
K
creases with the correlation coefficient . Fig. 7 plots the R(P ) E log2 1 + k wj |
|hH 2
D
K
normalized quantization distortion E[h j=1,j=k
2 ] as a function of the

2
correlation coefficient . Note that = 22 2 r1 is obtained
B K
P
1 log2 1 + E |hHk wj |
2
. (32)
from the assumption that all non-zero singular values except K
j=1,j=k
the dominant one (i.e., 1 ) are the same (2 = 3 = ).
Note also that Ri is tight in a regime where transmit antennas Using orthogonality between wj and hr,k ,
are highly correlated since 21 (i.e., ) decreases with the
correlation coefficient. Readers are referred to [23], [24] for k wj | + hk |hk hr,k | ,
hk 2 |hH 2 2 H 2
(33)
more details. We observe that if || > 0.3, the quantization then (32) becomes
distortion D of the AGB algorithm is better (smaller) than that
of conventional vector quantization. We can also observe that P
K
the analysis matches well with the simulation results when the R(P ) log2 1 + E hk 2 (1 |hH 2
k hr,k | )
K
transmit antennas are highly correlated (|| > 0.6). However, j=1,j=k
( )
when the magnitude of is small, the assumption in A-iii) is (K 1)
= log2 1 + P D (34)
violated so that the proposed bound is invalid. K
The following proposition provides the upper bound of the
where (34) is due to D = E[hk 2 (1 |hH
k hr,k | )]. Using
2
sum rate gap R(P ).
(18) and (34), we get the desired result.
Proposition 2. When an equal power allocation per user is
Next proposition specifies the number of feedback bits
applied, the sum rate gap (per user) between the ZFBF with
needed to maintain a constant rate gap from the system with
perfect CSI and the proposed method satisfies
perfect CSI.
( )
K 1 Proposition 3. In order to maintain a rate gap (between the
R(P ) . log2 1 + P (Nt + 2Nt ) (31)
K ZFBF with perfect CSI and the proposed method) within log2
bps/Hz per user, it is sufficient to scale the number of bits per
where R(P ) is the difference between the achievable user according to
rate achieved by (4) and wk =
Wzfk
where Wzf = ( )
( )1
Wzfk (1 + 2 2)2
B Bp + (Ng 1)[log2
HH HHH . (1 + 2 2 cos(2 NNt 1
))2

t
Proof. Note that R(P ) is given by R(P ) = E[log2 (1 + + 2 + 4( 1) P (K1) K

|hH w |2 ].
P
K |hk wk | )] E[log2 (1 + 1+ P
P H 2 KK k k H )]. Using 2 log2 (35)
|h w |2
K j=1,j=k k j 2 Nt
7 Due to the symmetric property of (i.e.,

i Nt > Nt 1 = 1 >
Proof. In order to maintain a rate loss of R(P ) log2
Nt 2 = 2 > ), 1 = Nt and 2 = Nt 1 = 1 . bps/Hz per user, we set the rate gap upper bound given in
the pattern generation process. The number of required flops

(i)
34
Perfect CSIT for computing the quasi-correlation matrix norm Rt F
AGB algorithm
Conventional vector quantization and the correlation matrix distance dcorr can be obtained as
30
2Nt Ng and 2Ng2 Nt + 4Nt Ng , respectively. Then, accord-
ing to Table I,(the)(total) computational complexity becomes
2Nmax Nt Ng + NJP N2P (2Ng2 Nt + 4Nt Ng ). Note that Nmax
26
Sum rate (bps/Hz)
can be reduced significantly by applying the partitioning

22
approach discussed in Section III.B. As mentioned, the pattern
generation process does not affect the real-time operation
18
since this process is performed off the shelf. In contrast
14
to the operations we just described, the codeword search
complexity is quantified by O() notation. Note that codeword
10
search complexity grows exponentially with the dimension of
the vector to be quantized and the codeword search com-
6 plexity for the (conventional
) approach
( and proposed
) method
0 5 10
SNR (dB)
15 20
is given by O Nt 2B and NP O Ng 2BBp , respectively.
Note also that the complexity of additional operations (i.e.,
Fig. 8. Sum rate as a function of SNR when (Nt = 32, Ng = 16, K =
NP (Cp + Cg + Ce )) is much( smaller than
) that of the codeword
4, = 0.9, = 0.05, Bp = 8). search complexity NP O Ng 2BBp . Overall, the proposed
method brings additional benefits in search complexity over
Proposition 3.2 equal to the maximum allowable gap of log2 the conventional approach due to the fact that dimension of
as the codeword being searched is reduced.
( )
K 1
R(P ) . log2 1 + P (Nt + 2Nt ) , log2 . IV. S IMULATION R ESULTS AND D ISCUSSIONS
K
(36) A. Simulation Setup
By inverting (36) and solving for B, we get the desired result. In this section, we compare the sum rate performance of the
conventional vector quantization using the channel statistic-
based codebook [34] and the proposed AGB algorithm. While
Fig. 8 plots the sum rate as a function of SNR when B in all the feedback resources (B-bit) are used to quantize the
(35) is applied. We fix = 2 in order to maintain a SNR channel vector hk in the conventional vector quantization
gap of 3 dB. We observe that by using a proper scaling of approach, B-bit feedback resource is divided into Bq (channel
feedback bits, we can limit the rate loss within 3 dB (in fact vector quantization) and Bp (pattern selection) in the proposed
around 2 dB), as desired. method. To express the feedback allocation, we use the nota-
tion B = (Bq , Bp ) in the sequel. As a pattern set, we use
D. Comments on Complexity the combination of patterns for each sub-array. Let Bp,sub be
B
In this subsection, we discuss the complexity of the AGB the number of pattern bits of each sub-array (Bp,sub = Mp ),
Bp,sub
algorithm and conventional vector quantization. While the then 2 patterns are generated by applying the proposed
major operation of the conventional approach is to search the subspace packing approach. As a transmit antenna model, we
codeword index, computations associated with pattern index consider the exponential correlation model in (27) [42] and
selection, grouping and expansion process, pattern generation two-dimensional uniform planar array (UPA) model [43]. We
are additionally required for the proposed method. We first use Jakes model [44] for the temporal correlation coefficient
analyze the computational complexity of the pattern index = J0 (2fD ) where J0 () is the 0-th order Bessel function
selection, grouping process, and the expansion process, which of the first kind, fD = vfc /c denotes the maximum Doppler
are performed on the fly. Denoting the complexity associated frequency, and = 5ms is the channel instantiation interval.
with pattern index selection, grouping process, and the expan- With the user speed v = 3km/h, the carrier frequency
sion process as Cp , Cg , Ce , respectively, then the number of fc = 2.5GHz, and the speed of light c = 3 108 m/s,
required floating-point operations (flops) for each step is as the temporal correlation coefficient becomes = 0.9881.
follows [41]: Assuming a 5ms coherence time and frame structure of 3GPP
LTE FDD systems [5], each fading block consists of L 10
Cp requires 4Nt flops for computing the distortion D in
static channel uses.
(12).
Cg requires (2Nt 1)Ng flops for the matrix multiplica-
tion in (7). B. Simulation Results
Ce requires (2Ng 1)Nt flops for the matrix multiplica-
We first consider the exponential channel model
tion in (10). {
|ji|
Note that these operations need to be computed for NP k ij
rij = |ji| (37)
times. We next measure the computational complexity of (k )H i>j
11 20
Perfect CSIT (N =32)
t
Perfect CSIT (N =N , case1)
t g
Perfect CSIT (N =N , case2)
18 t g
AGB algorithm (N =32)
t
10 Conventional vector quantization (N =32)

t
AGB algorithm (N =N ,case2)
Sum rate (bps/Hz)
16 t g
Conventional vector quantization (N =N ,case1)
Sum rate (bps/Hz)

t g
14
9
12
8
10
Conventional beamforming
AGB algorithm (subspace packing)
AGB algorithm (random codebook generation)
8
AGB algorithm (grouping of adjacent antenna elements)
7
2 3 4 5 6 7 8 9 10 12 16 20 24 28 32
Number of pattern bits Bp Number of feedback bits B
Fig. 9. Sum rate as a function of the number of pattern bits (Nt = 16, M = Fig. 11. Sum rate as a function of the number of feedback bits B when
2, K = 1, Ng = 8, Bq = 16, B = Bq + Bp ), SN R=10dB. SNR= 10 dB and B = (Bq , 8) (Nt = 32, M = 4, K = 4, Ng = 16, =
0.8)
24 sum rate gain over the approach using randomly generated

22
patterns, AGB with grouping of adjacent antenna elements as
well as the conventional vector quantization technique. For
20 example, to achieve 9 bps/hz, AGB with subspace packing
18
requires B = 18 bits while AGB with grouping of adjacent
Sum rate (bps/Hz)
antenna elements, AGB with random patterns and conventional

16 vector quantization require 20, 21 and 24 bits, respectively.
14
We next measure the sum rate as a function of SNR. In this
case, we set Nt = 32, M = 4, K = 4, = 0.8 and investigate
12 the performance for two scenarios (B = Nt and 2Nt ). In
10
addition, we plot the system with perfect CSIT as an upper
Perfect CSIT
AGB algorithm (B=64)
bound. As shown in Fig. 10, the AGB algorithm achieves
8 Conventional vector quantization (B=64) significant gain over the conventional vector quantization tech-
AGB algorithm (B=32)
6
Conventional vector quantization (B=32) nique, bringing in more than 3 dB gain at mid SNR regime.
0 5 10 15 20
In particular, with B = 2Nt , the AGB algorithm performs
SNR(dB)
within about 2 dB of perfect CSIT system until SNR = 7 dB
while others suffer from more than 5 dB loss compared to the
Fig. 10. Sum rate as a function of SNR when B = (Bq , 8) (Nt = 32, M =
4, K = 4, Ng = 16, = 0.8). perfect CSIT system.
In Fig. 11, we plot the sum rate as a function of the number
of feedback bits. In this case, we set K = 4, = 0.8, Ng = 16
where rij is the (i, j)-th element of Rt,k and k = ejk is and compare the performance of the AGB algorithm when
a transmit correlation coefficient for the k-th user where is Nt = 32 with the following two scenarios; 1) a system
the magnitude of correlation coefficient and k is the phase having a reduced number of transmit antennas (Nt = Ng )
of the k-th user. Note that the phase of each user is randomly and 2) a system where one antenna per group is selected and
generated from to and independent among each user. all remaining antennas per group is shut down (Nt = Ng ).
Note also that all users have the same transmit correlation Interestingly, by taking advantage of high correlation among
coefficient |k | = since is determined by the antenna the antennas in a group, we observe that the performance of
spacing at the basestation. case 2) is better than that of the conventional system and case
In order to observe the effectiveness of the subspace packing 1). Nevertheless, due to the number of active antennas, the
approach discussed in Section III.B, we compare the proposed proposed algorithm when Nt = 32 still achieves significant
approach to the random pattern generation and grouping of feedback overhead reduction over the case 2). We observe that
adjacent antenna elements. In our simulations, we set Bq = the proposed AGB algorithm with Nt = 32 requires smaller
16, Nt = 16, Ng = 8, M = 2, K = 1 and measure the sum number of bits to achieve the same level of performance. For
rate as a function of the number of pattern bits Bp . To set the example, the proposed approach achieves significant gain over
same level of feedback, we set B = Bq + Bp bit for the con- the conventional vector quantization techniques, resulting in
ventional vector quantization. Overall, we observe from Fig. more than 60% feedback overhead reduction.
9 that the subspace packing approach provides a considerable In Fig. 12, we consider the two-dimensional UPA (NV
TABLE II
S IMULATION PARAMETERS FOR UPA MODEL .
Variables Simulation parameters

Antenna elements spacing D = 0.5
( )1
Propagation path loss = 1 + ( rs )pl
Path loss exponent pl = 3 ( )
Angular spread (vertical) V = 12 ( arctan( s+r
u
) arctan( sr
u )
)
1 s+r sr
Angle of arrival (vertical) V = 2 arctan( u ) + arctan( u )
Angular spread (horizontal) H = arctan( rs )
Angle of arrival (horizontal) H,k (, ]
Elevation of the transmit antenna u = 60m
Radius of the scattering ring for the receiver r = 30m
Distance from the transmitter s = 50m
NH array) model, which is more realistic antenna model for

massive MIMO scenarios. The UPA model can be obtained
by the Kronecker product of the vertical correlation matrix
RV CNV NV and the horizontal correlation matrix RH,k
CNH NH . The resulting transmit correlation matrix of the UPA
20
t
model is expressed as Rt,k = RV RH,k where is the

AGB algorithm (N t =64)
18
Conventional vector quantization (N t =64)
Perfect CSIT (Nt =32) Kronecker product operator and each of the spatial correlation
16
AGB algorithm (N t =32) matrices is defined by
Conventional vector quantization (N t =32)
14

Sum rate (bps/Hz)
q +q,k

12 [Rq,k ]m,p = ej2D(mp) sin() d (38)
2q q +q,k
10
where q {H, V }, denotes propagation path loss between
8
the transmitter and the receiver, q is the angular spread,
6 D is the antenna elements spacing, and q,k is the angle of
arrival (AoA) for the k-th user. We summarize the simulation
4
parameters for UPA model in Table II. In Fig. 12, we plot the
2 sum rate as a function of SNR and the number of feedback
0 5 10 15 20
SNR (dB)
bits for Nt = 32, 64 and K = 2. For the UPA model, we set
NV = 4, NH = 8 for Nt = 32, M = 4 and NV = 8, NH = 8
(a) for Nt = 64, M = 8, respectively. We observe from Fig.
12(a) that the proposed approach achieves better sum rate
12 than the conventional scheme produces in particular for high
t
AGB algorithm (N t =64)
SNR regime. We also observe from Fig. 12(b) that the AGB
Conventional vector quantization (N t =64) algorithm outperforms the conventional vector quantization
11
Perfect CSIT (Nt =32) technique with a large margin, resulting in more than 50%
t feedback overhead reduction.
Sum rate (bps/Hz)
10 Conventional vector quantization (N t =32)

So far, we have assumed that the receiver has knowledge
of full CSI. In Fig. 13, we investigate the performance of the
9 AGB algorithm when the estimated CSI is employed. Since
the mismatch between the actual CSI and the estimated CSI is
8 unavoidable in a real communication, and this might result in
degradation performance, it is of importance to investigate the
7
effect of channel estimation error. In our simulation, we use an
additive channel estimation model where hk,est = hk + hk,err
where hk,est , hk and hk,err represent the estimated channel
6
24 32 40 48 56 64 vector, the original channel vector and the estimated error
Number of feedback bits B vector, respectively. We assume that hk,err is uncorrelated with
(b) hk,est , and hk,err has i.i.d elements with zero mean and the
2
Fig. 12. Sum rate as a function of (a) SNR when B = (24, 8) for Nt = 32 estimation error variance e,h . We observe from Fig. 13 that
and B = (48, 16) for Nt = 64 and (b) number of feedback bits B for the AGB algorithm is more robust to the estimation errors than
UPA correlation model when SNR= 10 dB, B = (Bq , 8) for Nt = 32, and the conventional vector quantization. For example, the sum
B = (Bq , 16) for Nt = 64.
rate gain at 7 bps/Hz of the AGB algorithm is about 5 dB
2
over the conventional vector quantization when e,h = 0.05,
2
while the gain is around 3 dB for e,h = 0.01.
conventional vector quantization technique is 30% for Nt = 32

12
AGB algorithm (2 =0)
e,h
and 25% for Nt = 64, respectively.
2
11 AGB algorithm (e,h =0.01)
AGB algorithm (2 =0.05)
e,h
10 Conventional vector quantization (2 =0)
e,h
V. C ONCLUSIONS
Conventional vector quantization (2 =0.01)
e,h
9
Conventional vector quantization (2 =0.05)
e,h
In this paper, we proposed an efficient feedback reduction
8 algorithm for FDD-based massive MIMO systems. Our work
Sum rate (bps/Hz)
7
is motivated by the observation that the CSI feedback overhead
must scale linearly with the number of transmit antennas so
6
that conventional vector quantization approach performing the
5 quantization of the whole channel vector is not an appropriate
4 option for the massive MIMO regime. The key feature of the
3
antenna group beamforming (AGB) algorithm is to control
relentless growth of the CSI feedback information in the
2
massive MIMO regime by mapping multiple correlated an-
1
0 5 10 15 20 tenna elements into a single representative value with grouping
SNR (dB) patterns and then choosing the codeword from the codebook
generated from the reduced dimension channel vector. It has
2
Fig. 13. Sum rate performance with various e,h for UPA correlation model been shown by distortion analysis and simulation results
when B = (24, 8) (Nt = 32, M = 4, K = 4, Ng = 16). that the proposed AGB algorithm is effective in achieving a
substantial reduction in the feedback overhead in the realistic
28
massive MIMO channels.
Although our study in this work focused on the single-cell
26 Perfect CSIT (N =64)
t scenario, we expect that the effectiveness of the proposed
24
t
Conventional vector quantization (N =64)
method can be readily extended to multi-cell scenario. In
t
Perfect CSIT (N =32) fact, in the multi-cell scenario, more aggressive feedback
t
22 AGB algorithm (N =32) compression is required since the channel information of the
Sum rate (bps/Hz)
20
Conventional vector quantization (N =32)
t interfering cells as well as the desired cell may be needed at
the basestation to properly control inter-cell interference. In
18 this scenario, the proposed AGB algorithm can be used as an
effective means to achieve reduction in the feedback informa-
16
tion. Also, investigation of nonlinear transmitter techniques
14 with user scheduling [46] would be interesting direction to
be investigated. Finally, we note that the proposed method
12
can be nicely integrated into the dual codebooks structure in
10 LTE-Advanced [47], [48] by feeding back the pattern index
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
for long-term basis and the codebook index for short-term
Transmit antenna correlation coefficient basis. Since the main target of the massive MIMO system is
slowly varying or static channels, dual codebook based AGB
Fig. 14. Sum rate as a function of the transmit antenna correlation coefficient algorithm will bring further reduction in feedback overhead.
when SNR= 10 dB.
A PPENDIX A
Finally, in Fig. 14, we plot the sum rate as a function of D ERIVATION OF (23)

for system with Nt = 32, 64 and K = 4. In the AGB Denoting (hH A hr ) and hH B hr as r1 (cos 1 + j sin 1 )

algorithm, we assign one bit per antenna elements on average. and r2 (cos 2 + j sin 2 ), Re[(hH H
A hr ) (hB hr )] becomes
Specifically, for Nt = 32, we set B = (24, 8) and for Nt = 64, r1 r2 (cos 1 cos 2 sin 1 sin 2 ). Under the assumption that
we set B = (48, 16), respectively. In this case, Ng = 16 for sufficient number of bits is used and hence the distortion
Nt = 32 and Ng = 32 for Nt = 64, respectively. It is worth [ [ hA and hr is]] small (i.e, 1 0, r1 1),
between

mentioning that correlated fading tends to decrease the size E Re (hH H
A hr ) (hB hr ) is expressed as
of space that channel vectors span and hence is beneficial [ [ ]]
to reduce the quantization distortion by employing a user
E Re (hH H
A hr ) (hB hr )
dependent channel statistic-based codebook [45]. As a result,
the sum rate of multiuser MIMO systems increases with the = E [r1 r2 (cos 1 cos 2 sin 1 sin 2 )]
transmit correlation coefficient. As shown in Fig. 14, when = E [r1 r2 ] E [cos 1 cos 2 sin 1 sin 2 ]
the transmit correlation coefficient increases, the antenna = E [r1 r2 ] E [cos(1 + 2 )]
grouping operation becomes effective and thus the sum rate
E [r2 ] E [cos 2 ] (39)
of the AGB algorithm improves drastically. For example, when
= 0.8, the sum rate gains of the AGB algorithm over the =0 (40)
where (39) follows from the fact that 1 0, r1 1, and (40) [22] C. K. Au-Yeung and D. J. Love, On the performance of random vector
is because E [cos 2 ] = 0 since 2 is uniformly distributed quantization limited feedback beamforming in a MISO system, IEEE
Trans. Wireless Commun., vol. 6, no. 2, pp. 458-462, Feb. 2007.
between and . [23] K. K. Mukkavilli, A. Sabharwal, E. Erkip, and B. Aazhang, On
beamforming with finite rate feedback in multiple-antenna systems, IEEE
Trans. Inf. Theory, vol. 49, no. 10, pp. 2562-2579, Oct. 2003.
R EFERENCES [24] B. Clerckx, G. Kim, and S. J. Kim, MU-MIMO with channel statistics
based codebooks in spatially correlated channels, Proc. IEEE Global
[1] B. Lee, J. Choi, D. Love, J. Seol, and B. Shim, Antenna Grouping based Telecommun. Conf., Dec. 2008.
Feedback Reduction Technique for FDD-based Massive MIMO Systems, [25] G. Caire and S. Shamai, On the achievable throughput of a multi-
Proc. IEEE Int. Conf. Commun., Jun. 2014. antenna Gaussian broadcast channel, IEEE Trans. Inf. Theory, vol. 49,
[2] B. Lee and B. Shim, An efficient feedback compression for large-scale no. 7, pp. 1691-1706, Jul. 2003.
MIMO systems, Proc. IEEE Veh. Technol. Conf., May 2014. [26] T. Yoo and A. Goldsmith, On the optimality of multiantenna broadcast
[3] T. L. Marzetta, Noncooperative cellular wireless with unlimited numbers scheduling using zero-forcing beamforming, IEEE J. Sel. Areas Com-
of base station antennas, IEEE Trans. Wireless Commun., vol. 9, no. 11, mun., vol. 24, no. 3, pp. 528-541, Mar. 2006.
pp. 3590-3600, Nov. 2010. [27] J. Park, B. Lee, and B. Shim, A MMSE vector precoding with block
[4] J. Hoydis, S. T. Brink, and M. Debbah, Massive MIMO in the UL/DL diagonalization for multiuser MIMO downlink, IEEE Trans. Commun.,
of cellular networks: How many antennas do we need?, IEEE J. Sel. vol. 60, no. 2, pp. 569-577, Feb. 2012.
Areas Commun., vol. 31, no. 2, pp. 160-171, Feb. 2013. [28] K. Huang, J. G. Andrews, R. W. Heath Jr., Performance of orthog-
[5] Evolved Universal Terrestrial Radio Access (E-UTRA); Physical Chan- onal beamforming for SDMA with limited feedback, IEEE Tran. Veh.
nels and Modulation, 3GPP TS 36.211 v11.0.0 Std. Technol., vol. 58, no. 1, pp. 152-164, Jan. 2009.
[6] Y. Nam, B. L. Ng, K. Sayana, Y. Li, J. Zhang, Y. Kim, and J. Lee, Full- [29] N. Ravindran and N. Jindal, Limited feedback-based block diagonal-
dimension MIMO (FD-MIMO) for next generation cellular technology, ization for the MIMO broadcast channel, IEEE J. Sel. Areas Commun.,
IEEE Commun. Mag., vol. 51, no. 6, pp. 172-179, Jun. 2013. vol. 26, no. 8, pp. 1473-1482, Oct. 2008.
[7] D. J. Love, R. W. Heath Jr., V. K. N. Lau, D. Gesbert, B. D. Rao, and M. [30] Q. H. Spencer, A. L. Swindlehurst, and M. Haardt, Zero-forcing
Andrews., An overview of limited feedback in wireless communication methods for downlink spatial multiplexing in multiuser MIMO channels,
systems, IEEE J. Sel. Area Commun., vol. 26, no. 8, pp. 1341-1365, Oct. IEEE Trans. Signal Process., vol. 52, no. 2, pp. 461-471, Feb. 2004.
2008. [31] S. Park, B. Shim, and J. Choi, Iterative channel estimation using virtual
[8] J. Jose, A. Ashikhmin, T. L. Marzetta, and S. Vishwanath, Pilot contam- pilot signals for MIMO-OFDM systems, IEEE Trans. Signal Process.,
ination and precoding in multi-cell TDD systems, IEEE Trans. Wireless vol. 63, no. 12, pp. 3032-3045, Jun, 2015.
Commun., vol. 10, no. 8, pp. 2640-2651, Aug. 2011. [32] J. Choi and B. Shim, Compressive sensing based pilot reduction
[9] E. Bjornson, J. Hoydis, M. Kountouris, M. Debbah, Massive MIMO technique for massive MIMO systems, Proc. IEEE Inf. Theory Appl.
systems with non-ideal hardware: energy efficiency, estimation, and Workshop, Feb. 2015.
capacity limits, IEEE Trans. Inf. Theory, vol. 60, no. 11, pp. 7112-7139, [33] P. Ding, D. J. Love, and M. D. Zoltowski, Multiple antenna broadcast
Nov. 2014. channels with shape feedback and limited feedback, IEEE Trans. Signal
[10] C. Shepard, H. Yu, N. Anand, L. E. Li, T. L. Marzetta, R. Yang, and Process., vol. 55, no. 7, pp. 3417-3428, Jul. 2007.
L. Zhong, Argos: Practical many-antenna base stations, Proc. ACM [34] D. J. Love and R. W. Heath Jr., Limited feedback diversity techniques
Mobicom Netw. Conf., Aug. 2012. for correlated channels, IEEE Trans. Veh. Technol., vol. 55, no. 2, pp.
[11] J. Choi, Z. Chance, D. J. Love, and U. Madhow, Noncoherent trellis 718-722, Mar. 2006.
coded quantization: A practical limited feedback technique for massive [35] D. J. Love and R. W. Heath Jr., Limited feedback unitary precoding
MIMO systems, IEEE Trans. Commun., vol. 61, no. 12, pp. 5016-5029, for spatial multiplexing systems, IEEE Trans. Inf. Theory, vol. 51, no.
Dec. 2013. 8, pp. 2967-2976, Aug. 2005.
[12] J. Nam, J. Ahn, A. Adhikary, and G. Caire, Joint spatial division and [36] D. J. Love, R. W. Heath Jr., and T. Strohmer, Grassmannian beamform-
multiplexing: opportunistic beamforming, user grouping and simplified ing for multiple-input multiple-output wireless systems, IEEE Trans. Inf.
downlink scheduling, IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, Theory, vol. 49, no. 10, pp. 2735-2747, Oct. 2003.
pp. 876-890, Oct. 2014. [37] M. Herdin, N. Czink, H. Ozcelik, and E. Bonek, Correlation matrix
[13] P. Kuo, H. T. Kung, and P. Ting, Compressive sensing based channel diatance, a meaningful measure for evaluation of non-stationary MIMO
feedback protocols for spatially correlated massive antenna arrays, Proc. channels, Proc. IEEE Veh. Technol. Conf., Jun. 2005.
IEEE Wireless Commun. Netw. Conf., Apr. 2012. [38] T. Yoo, N. Jindal, and A. Goldsmith, Multi-antenna downlink channels
[14] C. K. Au-Yeung, D. J. Love, and S. Sanayei, Trellis coded line pack- with limited feedback and users selection, IEEE J. Sel. Areas Commun.,
ing: large dimensional beamforming vector quantization and feedback vol. 25, no. 7, pp. 1478-1491, Sept. 2007.
transmission, IEEE Trans. Commun., vol. 10, no. 6, pp. 1844-1853 Jun. [39] C. Martin and B. Ottersten, Asymptotic eigenvalue distributions and
2011. capacity for MIMO channels under correlated fading, IEEE Trans.
[15] K. T. Truong and R. W. Heath. Jr., Effects of channel aging in massive Wireless Commun., vol. 3, no. 4, pp. 1350-1359, Jul. 2004.
MIMO systems, IEEE J. Commun. Netw., vol. 15, no. 4, pp. 338-351, [40] R. Gray, On the asymptotic eigenvalue distribution of Toeplitz matri-
Aug. 2013. ces, IEEE Trans. Inf. Theory, vol. 18, no. 6, pp. 725-730, Nov. 1972.
[16] D. Ying, F. W. Vook, T. A. Thomas, D. J. Love, and A. Ghosh, [41] G. H. Golub and C. F. V. Loan, Matrix Computations, 3rd edition. The
Kronecker product correlation model and limited feedback codebook Johns Hopkins University Press, 1989.
design in a 3D channel model, Proc. IEEE Int. Conf. Commun., Jun. [42] J. Choi and D. J. Love, Bounds on eigenvalues of a spatial correlation
2014. matrix, IEEE Commun. Lett., vol. 18, no. 8, pp. 1391-1394, Aug. 2014.
[17] J. Choi, D. J. Love, and P. Bidigare, Downlink training techniques for [43] A. Adhikary, J. Nam, J.-Y. Ahn, and G. Caire, Joint spatial division
FDD massive MIMO systems: open-loop and closed-loop training with and multiplexing: The large-scale array regime, IEEE Trans. Inf. Theory,
memory, IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 802-814, vol. 59, no. 10, pp. 6441-6463, Oct. 2013.
Oct. 2014. [44] J. G. Proakis, Digital Communication, 4th ed. New York: McGraw-Hill,
[18] S. Noh, M. D. Zoltowski, Y. Sung, and D. J. Love, Pilot beam pattern 2000.
design for channel estimation in massive MIMO systems, IEEE J. Sel. [45] B. Clerckx, G. Kim, and S. J. Kim, Correlated fading in broadcast
Topics Signal Process., vol. 8, no. 5, pp. 787-801, Oct. 2014. MIMO channels: curse or blessing?, Proc. IEEE Global Telecommun.
[19] M. Joham, P. M. Castro, W. Utschick, L. Castedo, Robust precoding Conf., Dec. 2008.
with limited feedback design based on precoding MSE for MU-MISO [46] B. Lee and B. Shim, A vector perturbation with user selection for
systems, IEEE Trans. Signal Process., vol. 60, no. 6, pp. 3101-3111, multiuser MIMO downlink, IEEE Trans. Commun., vol. 60, no. 11, pp.
Jun. 2012. 3322-3331, Nov. 2012.
[20] W. Santipach and M. L. Honig, Asymptotic performance of MIMO [47] 3GPP TSG RAN WG1 #62, R1-105011, Way Forward on 8Tx
wireless channels with limited feedback, Proc. IEEE Military Commun. Codebook for Rel.10 DL MIMO, Aug. 2010.
Conf., Oct. 2003. [48] C. Lim, T. Yoo, B. Clerckx, B. Lee, and B. Shim, Recent trend of
[21] N. Jindal, MIMO broadcast channels with finite rate feedback, IEEE multiuser MIMO in LTE-Advanced, IEEE Commun. Mag., vol. 51, no.
Trans. Inf. Theory, vol. 52, no. 11, pp. 5045-5059, Nov. 2006. 3, pp. 127-135, Mar. 2013.
Byungju Lee (M 15) received the B.S. and Ph.D. degrees in the School
of Information and Communication, Korea University, Seoul, Korea, in 2008
and 2014, respectively. He was a visiting scholar with Purdue University,
West Lafayette, IN, USA, in 2013. He is presently a postdoctoral fellow at
Seoul National University. His research interests include information theory
and signal processing for wireless communications.
Junil Choi (M 15) received the B.S. (with honors) and M.S. degrees in
electrical engineering from Seoul National University in 2005 and 2007,
respectively, and received the Ph.D. degree in electrical and computer en-
gineering from Purdue University in 2015. He is now a post-doctorate fellow
at The University of Texas at Austin. From 2007 to 2011, he was a member
of technical staff at Samsung Advanced Institute of Technology (SAIT) and
Samsung Electronics Co. Ltd. in Korea, where he contributed to advanced
codebook and feedback framework designs for the 3GPP LTE/LTE-Advanced
and IEEE 802.16m standards. His research interests are in the design and
analysis of massive MIMO, distributed communication, and vehicular com-
munication systems.
Ji-Yun Seol (M 05) received the B.S., M.S., and Ph.D. degrees in Electrical
Engineering from Seoul National University, Seoul, Korea, in 1997, 1999,
and 2005, respectively. He has been with Samsung Electronics Co., LTD.,
Suwon, Korea, since 2004. He has years of experience in development of
modem algorithms and standardization for Mobile WiMAX. He is currently a
Director at Advanced Communications Lab., Communications Research Team
at Samsung Electronics in Korea. He has been in charge of research for the
next generation (B4G/5G) mobile communications since 2011. His current
fields of interest include research/development of next generation mobile
communication system and advanced PHY algorithms.
David J. Love (F 15) received the B.S. (with highest honors), M.S.E.,
and Ph.D. degrees in electrical engineering from the University of Texas
at Austin in 2000, 2002, and 2004, respectively. During the summers of
2000 and 2002, he was with Texas Instruments, Dallas, TX. Since August
2004, he has been with the School of Electrical and Computer Engineering,
Purdue University, West Lafayette, IN, where he is now a Professor and
recognized as a University Faculty Scholar. He has served as an Editor for
the IEEE Transactions on Communications, an Associate Editor for the IEEE
Transactions on Signal Processing, and a guest editor for special issues of
the IEEE Journal on Selected Areas in Communications and the EURASIP
Journal on Wireless Communications and Networking. He is recognized as a
Thomson Reuters Highly Cited Researcher and holds 23 issued US patents.
Dr. Love is a Fellow of the Royal Statistical Society, and he has been
inducted into Tau Beta Pi and Eta Kappa Nu. Along with co-authors, he
was awarded the 2009 IEEE Transactions on Vehicular Technology Jack
Neubauer Memorial Award for the best systems paper published in the IEEE
Transactions on Vehicular Technology in that year and multiple Globecom best
paper awards. He was the recipient of the Fall 2010 Purdue HKN Outstanding
Teacher Award, Fall 2013 Purdue ECE Graduate Student Association Out-
standing Faculty Award, and Spring 2015 Purdue HKN Outstanding Professor
Award.
Byonghyo Shim (SM 09) received the B.S. and M.S. degrees in control and
instrumentation engineering from Seoul National University, Korea, in 1995
and 1997, respectively. He received the M.S. degree in mathematics and the
Ph.D. degree in electrical and computer engineering from the University of
Illinois at Urbana-Champaign (UIUC), USA, in 2004 and 2005, respectively.
From 1997 and 2000, he was with the Department of Electronics Engi-
neering, Korean Air Force Academy as an Officer (First Lieutenant) and an
Academic Full-time Instructor. From 2005 to 2007, he was with Qualcomm
Inc., San Diego, CA, USA, as a Staff Engineer. From 2007 to 2014, he
was with the School of Information and Communication, Korea University,
Seoul, as an Associate Professor. Since September 2014, he has been with
the Department of Electrical and Computer Engineering, Seoul National
University, where he is presently an Associate Professor. His research interests
include wireless communications, statistical signal processing, estimation and
detection, compressive sensing, and information theory.
Dr. Shim was the recipient of the 2005 M. E. Van Valkenburg Research
Award from the Electrical and Computer Engineering Department of the
University of Illinois and 2010 Hadong Young Engineer Award from IEIE.
He is currently an Associate Editor of the IEEE Wireless Communications
Letters, Journal of Communications and Networks, and a Guest Editor of the
IEEE Journal on Selected Areas in Communications (JSAC).

Antenna Grouping Based Feedback Compression For FDD-based Massive MIMO Systems

Uploaded by

Copyright:

Available Formats

Antenna Grouping Based Feedback Compression For FDD-based Massive MIMO Systems

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Antenna Grouping Based Feedback Compression For FDD-based Massive MIMO Systems

Uploaded by

Copyright:

Available Formats

This article has been accepted for publication in a future issue of this journal, but has not been

Antenna Grouping based Feedback Compression for

of group pattern and codeword minimizing the quantization

Once the pattern index i is obtained, this index and the

Pattern 1 Pattern 2 Pattern 3

Pattern index Codeword index

Fig. 5. Overall transceiver structure of the proposed AGB technique.

and hence hr = hA = [h1 h3 ]T (hB = [h2 h4 ]T ).

the quantized vector of hr and generated from the

Initialization Bp : the number of bits for the pattern set

(d) [ ] where = ej is the transmit correlation coefficient, and

Normalized quantization distortion

Proof. Note that R(P ) is given by R(P ) = E[log2 (1 + + 2 + 4( 1) P (K1) K

7 Due to the symmetric property of (i.e.,

the pattern generation process. The number of required flops

can be reduced significantly by applying the partitioning

10 Conventional vector quantization (N =32)

Sum rate (bps/Hz)

24 sum rate gain over the approach using randomly generated

antenna elements, AGB with random patterns and conventional

Variables Simulation parameters

NH array) model, which is more realistic antenna model for

model is expressed as Rt,k = RV RH,k where is the

10 Conventional vector quantization (N t =32)

conventional vector quantization technique is 30% for Nt = 32

You might also like