Paper 12 PDF
Paper 12 PDF
Paper 12 PDF
On the Performance of Channel-Statistics-Based Codebook [6],[7] and random vector quantization (RVQ) based codebook [8],[9]
for Massive MIMO Channel Feedback have been extensively investigated. For these codebooks, there is an
important conclusion that, to maintain a constant capacity degradation
Wenqian Shen, Linglong Dai, Yu Zhang, Jianjun Li, due to channel quantization error, the required number of feedback
and Zhaocheng Wang bits approximately scales linearly with the number of BS antennas [9].
Thus, for massive MIMO with a large number of BS antennas, the chan-
nel feedback overhead will be overwhelming. Therefore, several new
Abstract—The channel feedback overhead for massive multiple-input
multiple-output systems with a large number of base station (BS) antennas
codebooks have been proposed to reduce the channel feedback over-
is very high since the number of feedback bits of traditional codebooks head. Based on the assumption of high correlation among BS antennas,
scales linearly with the number of BS antennas. To reduce the feedback an antenna-grouping-based feedback scheme was proposed in [10].
overhead, an effective codebook based on channel statistics has been de- Another promising solution is the channel-statistics-based codebook
signed, where the required number of feedback bits only scales linearly [11] that is designed by multiplying each vector of an original code-
with the rank of the channel correlation matrix. However, this attractive
conclusion was only proved under a particular channel assumption in the book (e.g., RVQ-based codebook) by the channel correlation matrix.
literature. To provide a rigorous theoretical proof under a general channel It has been verified through extensive simulations that the number of
assumption, in this paper, we quantitatively analyze the performance of the feedback bits required by this channel-statistics-based codebook only
channel-statistics-based codebook. Specifically, we first introduce the rate scales linearly with the rank of the channel correlation matrix [12]. As
gap between the ideal case of perfect channel state information at the trans-
mitter and the practical case of limited channel feedback, where we find
the rank of the channel correlation matrix is much smaller than the
that the rate gap depends on the quantization error of the codebook. Then, number of BS antennas in massive MIMO [13], the feedback overhead
we derive an upper bound of the quantization error, based on which we can be reduced. However, Clerckx et al. [12] only provided a proof of
prove that the required number of feedback bits to ensure a constant rate this attractive conclusion under a particular channel assumption, where
gap only scales linearly with the rank of the channel correlation matrix. the singular values of the square root of the channel correlation matrix
Finally, numerical results are provided to verify this conclusion.
are the same except for the first one, whereas the rigorous proof under
Index Terms—Channel feedback, codebook, massive multiple-input a general channel assumption has not been provided in the literature.
multiple-output (MIMO), performance analysis. In this paper, we quantitatively analyze the performance of
the channel-statistics-based codebook under a general channel
assumption.1 Specifically, we first introduce the rate gap between the
I. INTRODUCTION ideal case of perfect CSIT and the practical case of limited channel
Massive multiple-input multiple-output (MIMO) using hundreds of feedback. We find that the rate gap depends on the quantization error
antennas at the base station (BS) is one of the key technologies for of the codebook. Although the quantization error is difficult to obtain,
future 5G wireless communications [1]. To achieve the expected high we are able to derive an upper bound of the quantization error by magni-
spectrum efficiency and energy efficiency of massive MIMO, accu- fying and shrinking the involved inequation. After that, by substituting
rate channel state information at the transmitter (CSIT) is crucial [2], the upper bound of quantization error into the rate gap expression, we
[3]. Utilizing the channel reciprocity, CSIT can be obtained from up- can obtain the upper bound of the rate gap. Finally, we prove that the
link channel estimation in time division duplexing (TDD) systems [4]. number of feedback bits required to ensure a constant rate gap scales
While most of the existing works on massive MIMO consider TDD linearly with the rank of the channel correlation matrix. To the best of
mode due to this reason, frequency division duplexing (FDD) has many our knowledge, this work is the first one to provide a rigorous proof of
benefits over TDD and thus still dominates current cellular networks this conclusion under a general channel assumption.
[5]. Therefore, it is important to study the CSIT acquisition problem Notation: Lower case and upper case boldface letters denote vec-
for FDD massive MIMO systems. tors and matrices, respectively, (·)H and (·)−1 denote the conjugate
However, the channel reciprocity does not exist in FDD systems, transpose and inverse of a matrix, respectively, IK denotes the iden-
so accurate channel feedback from users to the BS is required. Tradi- tity matrix of size K × K, E[·] denotes the expectation operator, and
tional codebooks for channel feedback such as Grassmannian codebook Pr{A} denotes the probability of A.
0018-9545 © 2017 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.
7554 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 66, NO. 8, AUGUST 2017
where Rk ∈ C M ×M is the square root of the kth user’s chan- C. Downlink Precoding and Per-User Rate
1/ 2
nel correlation matrix2 Rk ∈ C M ×M , and hw , k ∈ C M ×1 is a vec- We can utilize the widely used zero-forcing (ZF) precoding at the
tor whose elements are independent identically distributed complex BS based on the feedback channel matrix Ĥ to eliminate interferences
Gaussian distributed with zero mean and unit variance. Furthermore, among multiple users. The transmitted signal x ∈ C M ×1 after precod-
1/ 2 1/ 2
Rk can be decomposed as Uk Λk UH k , where U k is a unitary ma- ing at the BS is given by
1/ 2
trix, and Λk is a diagonal matrix with the diagonal elements de-
noted by {σ1 , σ2 , . . . , σr , 0, . . .}, where r is the rank of correlation γ
x= Vs (4)
matrix. The concatenation of K channel vectors can be denoted by K
H = [h1 , h2 , . . . , hK ] ∈ C M ×K .
where γ is the transmit power, s = [s1 , s2 , . . . , sK ] ∈ C K ×1 is the sym-
bol vector intended for K users with the normalized power E[|si |2 ] = 1,
B. Limited Channel Feedback and V = [v1 , v2 , . . . , vK ] ∈ C M ×K is the precoding matrix con-
Although the training overhead to obtain the downlink channel vec- sisting of K different M -dimensional unit-norm precoding vector
tor at the user side is increased in massive MIMO systems, there are vi ∈ C M ×1 . We denote U = Ĥ(ĤH Ĥ)−1 , then the precoding vec-
some recently proposed effective downlink training methods [13],[14] tors vi can be obtained as the normalized ith column of U, i.e.,
U (:, i )
proposed with reduced training overhead. Thus, in this paper, each user vi = U (:, i )
.
is assumed to know its channel vector hk . Such information is required The received signal yk at the kth user can be described as
at the BS via limited feedback channel. Each user quantizes its channel
yk = hH
k x + nk
to B bits and then feeds them back to the BS. Quantization is realized
by using a quantization codebook, which is known to the BS and users. γ H γ
K
For the traditional RVQ-based codebook [9] W = = h k v k sk + hH
k v i si + n k (5)
K K
{w1 , w2 , . . . , w2B }, the unit-norm column vector wi ∈ C M ×1 i = 1, i = k
is randomly generated by selecting vector independently from the
where nk is the complex Gaussian noise at the kth user with zero
uniform distribution on the complex unit sphere. The required number
mean and unit variance. Thus, the signal-to-interference-plus-noise
of feedback bits scales linearly with the number of BS antennas
ratio (SINR) at the kth user is [9]
to ensure a constant capacity degradation [9]. Thus, the channel
feedback overhead becomes overwhelming for massive MIMO with γ
|hH k vk |
2
at the BS and the user location, which is the second-order channel statistics that Rpractical = E log2 1 + K
K . (9)
can be usually assumed to be static due to the moderate user velocities [11]. 1 + Kγ i = 1, i = k |hk vi |
H 2
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 66, NO. 8, AUGUST 2017 7555
We define the rate gap ΔR(γ) as the difference between per-user Using (1) and (2), we can rewrite |h̃H ci |2 as
rate achieved by ideal CSIT and limited channel feedback using the
channel-statistics-based codebook |(R1/ 2 hw )H R1/ 2 wi |2
|h̃H ci |2 =
R1/ 2 hw 2 R1/ 2 wi 2
ΔR(γ) = Rideal − Rpractical . (10)
(a ) w U(Λ
|hH ) Λ1/ 2 UH wi |2
1/ 2 H
= (19)
Following the results from [9] and[10] and using Jensen’s inequality, R hw 2 R1/ 2 wi 2
1/ 2
normalized channel vector h̃k can be decomposed along two orthogonal Since Λ1/ 2 is a diagonal matrix with r nonzero diagonal elements
direction, one is the direction of quantization vector ck , F k , and the other {σ1 , σ2 , . . . , σr }, we rewrite (21) as
is in the null space of ck , F k . Mathematically H H
|g Γ Γv|2
√ √ Pr |h̃H ci |2 < z = Pr < z (22)
h̃k = 1 − Xck , F k + Xs (13) Γg2 Γv2
where s is a unit vector distributed in the null space of ck , F k [9]. where g ∈ C r ×1 with g(j) = hw (j), v ∈ C r ×1 with v(j) = wi (j),
Combining hk = hk h̃k and (13), we have and Γ ∈ C r ×r with Γ(j, j) = Λ1/ 2 (j, j), j = 1, 2, . . . , r. The nondi-
agonal elements Γ(i, j) = 0. Actually, g and v are random vectors
k vi | = hk
|hH 2 2
(1 − X)|cHk , F k vi | + X|s vi |
2 H 2
. (14) in the r-dimensional hypersphere, whereas Γg and Γv are randomly
distributed in the r-dimensional hyperellipse, which is obtained by
Since the precoding vector vi is obtained as the normalized ith column
stretching the hypersphere according to the diagonal elements of Γ.
of Ĥ(ĤH Ĥ)−1 , vi is orthogonal to the kth user’s feedback channel
As we know, g|g2 v
H v |2
is the squared cosine of the angle between
vector ĥk = hk ck , F k , i.e., cH
k , F k vi = 0. Thus, we have |hk vi | =
H 2 2
two vectors g and v in hypersphere, whose cumulative distribution
hk X|s vi | . Since |s vi | = cos ((s, vi )) ≤ 1, we have
2 H 2 H 2 2
function (CDF) is given by Pr{ g|g2 v
H v |2
2
≤ z} = 1 − (1 − z)r −1 for
k vi | ≤hk X.
|hH 2 2
(15) z ∈ [0, 1] [8]. In (22), |g Γ Γv| H H 2
is the squared cosine of the angle
Γ g 2 Γ v 2
As the norm of a vector is independent of its direction [9], hk 2 is between two vectors Γg and Γv in the hyperellipse, whose CDF is
independent of X. Therefore, we obtain (12). very difficult be obtained. However, we can prove in the Appendix that
Pr{ |g
H Γ H Γ v |2
≤ z} < Pr{ g|g2 v
H v |2
Combining (11) and Lemma 1, we can further obtain Γ g 2 Γ v 2 2
≤ z}. Thus, we have
γ
ΔR(γ) ≤ log2 1+ (K −1)E hk 2 E sin2 ((h̃k , ĥk )) . |g H ΓH Γv|2
K Pr <z < 1 − (1 − z)r −1 . (23)
Γg2 Γv2
(16)
Then, by combining (18), (22), and (23), we can obtain the upper
B. Quantization Error bound of Z’s CDF as
In this section, we discuss the quantization error E[sin2 ((h̃k , ĥk ))] Pr {Z < z} < (1 − (1 − z)r −1 )2 . (24)
B
Fig. 1. Per-user rate against the SNR at the receiver. Fig. 2. Number of feedback bits against the rank of correlation matrix.
where (a) is obtained by setting s = 1 − z, and (b) and (c) are obtained perfect CSIT and the practical case using RVQ-based codebook to
from [9, Appendixes I and II], respectively. Finally, we can obtain the realize CSIT becomes greater when SNR increases since the required
upper bound of quantization error as number of feedback bits for RVQ-based codebook should scale linearly
with the number of BS antennas, which is much larger than the rank
B
E sin2 ((h̃, ĥ)) = 1 − E [Z] < 2− r −1 . (27) of the channel correlation matrix. Moreover, we also show the per-user
rate of the principle eigenvector-based method [15] for comparison,
which is overperformed by the channel-statistics-based codebook.
Fig. 2 shows the required number of feedback bits to limit the rate
C. Feedback Bits gap between the ideal case of perfect CSIT and the practical case of
using the channel-statistics-based codebook within 0.09 bps/Hz. We
In this section, we discuss the required number of feedback bits B can observe that the required number of feedback bits B scales linearly
to ensure a constant rate gap ΔR(γ). By combining (16) and (27), we with the rank of the channel correlation matrix r, which is consistent
can easily obtain with the theoretical result (29) shown by the dash curve in Fig. 2.
γ
B
ΔR(γ) ≤ log2 1 + (K − 1)E h2 2− r −1 . (28) V. CONCLUSION
K
To ensure the rate gap ΔR(γ) ≤ log2 (b), based on (28) and letting It has been shown that the number of feedback bits scales linearly
B with the rank of the channel correlation matrix for the channel-statistics-
log2 (1 + Kγ (K − 1)E[h2 ]2− r −1 ) ≤ log2 (b) bps/Hz, we have
based codebook. However, this attractive conclusion was only proved
under a particular channel assumption in the literature. To this end,
r−1 K −1
B≥ SNR + (r − 1) log2 (29) in this paper we provided a rigorous proof under a general channel
3 b−1
assumption for the first time. At first, we found that the rate gap between
where the signal-to-noise ratio (SNR) at the receiver is defined as the ideal case of perfect CSIT and using the channel stastics-based
SNR = 10 log10 Kγ E[h2 ]. We can observe that the required number codebook depends on the channel quantization error. Then, we derived
of feedback bits only scales linearly with the rank of correlation matrix an upper bound of the quantization error, based on which we proved
r when SNR increases. that the required number of feedback bits scales linearly with the rank
of the channel correlation matrix. Moreover, the proof techniques in
IV. SIMULATION VERIFICATION this paper can be used to analyze the performance of other codebooks.
Fig. 1 shows the per-user rate against SNR. We can observe from 2
between two isotropic vectors g and v uniformly distributed in a
Fig. 1 that the rate gap between the ideal case of perfect CSIT and the
hypersphere. For z ∈ [0, 1], the CDF of g|g2 v
H v |2
For |g
H H
Γ Γv| 2
Γ g 2 Γ v 2
in (30), the elements of g and v are multiplied by the [5] H. Xie, F. Gao, S. Zhang, and S. Jin, “A unified transmission strategy for
TDD/FDD massive MIMO systems with spatial basis expansion model,”
diagonal elements {σ1 , σ2 , . . . , σr } of Γ. The uniformly distributed IEEE Trans. Veh. Technol., to be published.
vectors g and v in hypersphere are projected onto nonuniformly dis- [6] D. J. Love, R. W. Heath, and T. Strohmer, “Grassmannian beamforming
tributed vectors Γg and Γv in hyperellipse, which is generated by for multiple-input multiple-output wireless systems,” IEEE Trans. Inf.
stretching the hypersphere. That means the angle between two vectors Theory, vol. 49, no. 10, pp. 2735–2747, Oct. 2003.
[7] K. K. Mukkavilli, A. Sabharwal, E. Erkip, and B. Aazhang, “On beam-
in hyperellipse tends to be smaller, i.e., the squared cosine of the angle forming with finite rate feedback in multiple-antenna systems,” IEEE
between two vectors in hyperellipse tends to be larger. Thus, the CDF Trans. Inf. Theory, vol. 49, no. 10, pp. 2562–2579, Oct. 2003.
of |g
H Γ H Γ v |2
is smaller than that of g|g2 v
H v |2
, i.e., we have (30). [8] C. K. Au-yeung and D. J. Love, “On the performance of random vector
Γ g 2 Γ v 2 2
quantization limited feedback beamforming in a MISO system,” IEEE
Now we give an intuitive example of an extreme case where Trans. Wireless Commun., vol. 6, no. 2, pp. 458–462, Feb. 2007.
σ1 = σ2 =, . . . , = σr −1 and σr σ2 . The r-dimensional hyperellipse [9] N. Jindal, “MIMO broadcast channels with finite-rate feedback,” IEEE
is generated by compressing the last dimension of a r-dimensional Trans. Inf. Theory, vol. 52, no. 11, pp. 5045–5060, Nov. 2006.
hypersphere. Thus, the vectors in hypersphere are projected onto a [10] B. Lee, J. Choi, J. Y. Seol, D. J. Love, and B. Shim, “Antenna grouping
based feedback compression for FDD-based massive MIMO systems,”
compressed hypersphere, i.e., the hyperellipse. The angle between two IEEE Trans. Commun., vol. 63, no. 9, pp. 3261–3274, Sep. 2015.
vectors in hyperellipse becomes smaller. Mathematically, the numera- [11] D. J. Love and R. W. Heath, “Limited feedback diversity techniques for
tor of the squared cosine can be approximately expressed as correlated channels,” IEEE Trans. Veh. Technol., vol. 55, no. 2, pp. 718–
722, Mar. 2006.
|g H ΓH Γv|2 ≈ |σ1 |4 |Σri =−11 gi∗ vi |2 . (32) [12] B. Clerckx, G. Kim, and S. Kim, “MU-MIMO with channel statistics-
based codebooks in spatially correlated channels,” in Proc. IEEE Global
The dominator of the squared cosine can be approximately ex- Commun. Conf., Nov. 2008, pp. 1–5.
[13] W. Shen, L. Dai, B. Shim, S. Mumtaz, and Z. Wang, “Joint CSIT ac-
pressed as Γg2 Γv2 ≈ |σ1 |4 (Σri =−11 |gi |2 )(Σri =−11 |vi |2 ). Thus, we quisition based on low-rank matrix completion for FDD massive MIMO
−1 g ∗ v |2
|Σ ri =
|g H Γ H Γ v |2
have Γ g 2 Γ v 2
≈ 1 i i
−1 |g |2 )(Σ r −1 |v |2 )
(Σ ri =
. Considering (31), we can get systems,” IEEE Commun. Lett., vol. 19, no. 12, pp. 2178–2181, Dec. 2015.
1 i i= 1 i [14] Z. Gao, L. Dai, Z. Wang, and S. Chen, “Spatially common sparsity based
adaptive channel estimation and feedback for FDD massive MIMO,” IEEE
|g H ΓH Γv|2 Trans. Signal Process., vol. 63, no. 23, pp. 6169–6183, Dec. 2015.
Pr ≤z ≈ 1 − (1 − z)r −2 . (33)
Γg2 Γv2 [15] S. Zhou and G. B. Giannakis, “Optimal transmitter eigen-beamforming
and space-time block coding based on channel mean feedback,” IEEE
Utilizing the fact that 1 − (1 − z)r −2 < 1 − (1 − z)r −1 , we can Trans. Signal Process., vol. 50, no. 10, pp. 2599–2613, Oct. 2002.
obtain (30).
REFERENCES
[1] F. Rusek et al., “Scaling up MIMO: Opportunities and challenges with
very large arrays,” IEEE Signal Process. Mag., vol. 30, no. 1, pp. 40–60,
Jan. 2013.
[2] C. Qi, G. Yue, L. Wu, and A. Nallanathan, “Pilot design for sparse channel
estimation in OFDM-based cognitive radio systems,” IEEE Trans. Veh.
Technol., vol. 63, no. 2, pp. 982–987, Feb. 2014.
[3] C. Huang, L. Liu, C. Yuen, and S. Sun, “Iterative channel estimation
using LSE and sparse message passing for mmWave MIMO systems,”
arXiv:1611.05653, 2016.
[4] X. Gao, L. Dai, S. Han, C.-L. I, and R. W. Heath, “Energy-efficient hybrid
analog and digital precoding for mmWave MIMO systems with large
antenna arrays,” IEEE J. Sel. Areas Commun., vol. 34, no. 4, pp. 998–
1009, Apr. 2016.