Spawc 2018 8446023
Spawc 2018 8446023
Spawc 2018 8446023
The primary purpose of this paper is to propose FDD- To model a realistic system, we assume that the differences
based cell-free massive MIMO systems that do not require between uplink and downlink multipath components, θ̃i and
any feedback from the user. The key idea behind the pro- β̃i , are i.i.d random variables with zero mean and variance
posed scheme is to extract the multipath components used σθ2 , σβ2 1 [10].
for the basestation cooperation from the uplink pilot signal.
To this end, we propose a computationally efficient multipath
component estimation technique based on the gradient descent III. FDD-BASED C ELL -F REE M ASSIVE MIMO S YSTEM
method. Furthermore, we propose the precoding and the power In this section, we present the FDD-based cell-free massive
allocation schemes that employ the multipath components MIMO systems that directly acquire multipath components
exclusively. To be specific, by using the property that the angle from the uplink pilot signal and use them for the bases-
information represents the channel direction, the proposed tation cooperation. We first explain the proposed multipath
angle-based precoding scheme can effectively reduce ICI and component estimation technique based on the gradient de-
also improves the spectral efficiency of the user. Moreover, scent method. We next discuss the angle-based zero-forcing
since each propagation path has a different path gain, a proper (ZF) precoding that only requires the angle information in a
power allocation to the propagation paths can improve the precoding design. In our analysis, we obtain a closed-form
energy efficiency. Simulation results show that the proposed rate expression that takes into account the imperfect angle
multipath component estimation technique outperforms the reciprocity. Finally, we discuss an efficient way to allocate
conventional subspace-based technique [7]. Moreover, the pro- the transmission power into each propagation path.
posed FDD-based cell-free massive MIMO systems save 19%
A. Gradient Descent-Based Multipath Component Estimation
of transmission power over the conventional cellular systems.
There are various techniques to estimate the multipath
II. C HANNEL M ODEL components. Among these, subspace-based techniques (MU-
SIC and ESPRIT) [7] are widely used due to the computa-
We consider M basestations with uniform linear array
tional benefit. However, in the massive MIMO regime, even
(ULA) of N antennas and K users with single antenna. We
subspace-based techniques which are known to be simple to
use a flat fading, one-ring scattering channel model with P
implement are infeasible due to the huge computational over-
propagation paths. Owing to the angle reciprocity, we can
head caused by the N -dimensional eigenvalue decomposition
safely assume that the uplink angle-of-arrival (AOA) and
(EVD) operation. Whereas, the proposed technique employs
downlink angle-of-departure (AOD) are similar. Moreover,
the gradient descent method that requires only P -dimensional
since the large-scale fading coefficient is not a function of
EVD operation. Since the number of antennas N is much
frequency, we also assume that the uplink and downlink
larger than the number of paths P in the massive MIMO
large-scale fading coefficients are similar. Whereas, uplink
regime, the proposed scheme is computationally efficient.
and downlink small-scale fading coefficients are distinct since
Let φk ∈ C1×τ be the uplink pilot signal with τ symbol
these are functions of frequency [8]. In the sequel, AOAs
and unit norm sent by the user k. In order to distinguish
(or AODs) and large-scale fading coefficients are collectively
the propagation paths of distinct users, we design the uplink
called the multipath components of the channel. In this setting,
training time of users in an non-overlapping fashion. In this
the uplink and downlink channels are given by [9]
setting, the received signal Ymk ∈ CN ×τ at the basestation
P
m sent by the user k is given by
hl = βil gil a(θil , λl ), l ∈ {UL, DL}, (1) √
i=1 Ymk = ρr hUL mk φk + Nmk , (6)
where βil is the large-scale fading coefficient which accounts where ρr is the uplink transmission power and Nmk ∈ CN ×τ
for the path loss and the shadowing effect, gil ∼ CN (0, 1) is the Gaussian noise with variance σn2 IN . We assume that
is the small-scale fading coefficient, and λl is the channel ρr is known to the basestations. Multiplying (6) by φH k and
wavelength. We assume that the small-scale fading coefficients collecting T samples, we have
are i.i.d. random variables. In addition, a(θ, λ) is the array √
Y(t)φH = ρr AUL BUL sUL (t) + NφH
steering vector given by √
= ρr AUL dUL (t) + n̄, t = 1, · · · , T, (7)
1 2πd sin θ 2πd sin θ
a(θ, λ) = √ [1, e−j λ , · · · , e−j(N −1) λ ]T , (2)
N where dUL = BUL sUL and n̄ = NφH ∼ CN (0, σn2 IN ). For
brevity, we henceforth omit the basestation and the user indices
where d is the antenna spacing. Note that the channel vector
for brevity. Multiple samples of (7) are combined into a matrix
in (1) can be represented as hl = Al Bl sl where
form as
√
Al = a(θ1l , λl ), · · · , a(θPl , λl ) , (3) Ȳ = ρr AUL DUL + N̄, (8)
Bl = diag( β1l , · · · , βPl ), (4) where Ȳ = [Y(1)φH , · · · , Y(T )φH ], DUL =
[d (1), · · · , d (T )], and N̄ = [n̄(1), · · · , n̄(T )].
UL UL
sl = [g1l , · · · , gPl ]T . (5) Then, the log-likelihood function of (8) is
2018 IEEE 19th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC)
θ̂ = argmin tr P⊥ AUL Ȳ Ȳ
H
, (12) ŵmk = γmk,i ĝmk,i = Ĝmk γ mk , (17)
i=1
where P⊥ AUL = IN − A (A
UL UL,H UL −1 UL,H
A ) A is the or-
thogonal projection matrix onto the orthogonal complement of where ĝmk,i is the ((k − 1)P + i)-th column of Ĝm =
−1
AUL . For simplicity, we let f (θ UL ) = tr P⊥ Ȳ Ȳ H
in the Âm (ÂHm Âm ) which corresponds to a(θmk,i UL
, λDL ) and
AUL
sequel. Note that (12) is a non-linear P -dimensional optimiza- Ĝmk = [ĝmk,1 , · · · , ĝmk,P ]. Also, γmk,i is the normalized
tion problem. Since it is very difficult to find out the closed- complex weight for the i-th propagation path that satisfies
P 2
i=1 |γmk,i | = 1 and γ mk = [γmk,1 , · · · , γmk,P ] . The
T
form solution of (12), we use the gradient descent method that
finds out the local minimum of the multi-dimensional function. transmit signal at the basestation m is given by
The update equation to find out the angle vector θ UL is
K
√
θ UL = θ UL − αn ∇θUL f (θ UL xm = ρd ŵmk sk , (18)
(n) (n−1) (n−1) ), (13)
k=1
where θ UL
(n) is the n-th estimate of θ , αn is the n-th step
UL
where sk is the data symbol for the user k satisfying
size determined by the Armijo’s rule, and ∇θUL f is E |sk |2 = 1 and ρd is the maximum transmit power sat-
∇θUL f = Im diag CUL AUL,H EP⊥ H UL
AUL Ȳ Ȳ A (A
UL,H UL −1
A ) , isfying
(14) K
2
where C UL
= diag cos θ1 , · · · , cos θP
UL UL
and E = E x
m 2 = ρd Ĝmk γ mk 22 ≤ ρd . (19)
√2πd diag (0, · · · , N − 1). k=1
N λUL
2018 IEEE 19th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC)
Then, the received signal at the user k is given by it can be relaxed into a convex semi-definite program (SDP)
using the semi-definite relaxation (SDR) [12]. By introducing
M
rk = hDL,H a new variable Γmk = γ mk γ H mk , the quadratic constraints can
mk xm +nk
m=1 be transformed into the linear constraints on SP . Since the
rank constraint of Γmk 1 is non-convex, we need to relax the
√ √ DL,H
M M K
= ρd hDL,H
mk ŵmk sk + ρd hmk ŵmj sj +nk , rank constraint to obtain the feasible SDP formulation of (22)
m=1 m=1j=k as
S I
M
K
min tr ĜH
mk Ĝmk Γmk
where nk ∼ CN (0, 1) is the additive noise. Note that the {Γmk }
m=1 k=1
received signal can be decomposed into three parts: 1) desired
M
2 H
signal part S, 2) ICI part I, and 3) noise wk . If the angle s.t. tr (BUL,2 UL,2 UL,2
mk + σθ Q̂mkk Cmk Bmk Q̂mkk )Γmk −
reciprocity is perfect (σθ2 = σβ2 = 0), then one can easily m=1
show that M K ξk
ξk σθ2 tr Q̂H UL,2 UL,2
mkj Cmk Bmk Q̂mkj Γmj ≥ , ∀k
ρ
sDL,H
mk Bmk γ mj if k = j
DL m=1j=k
d
hDL,H
mk ŵmj = (20)
0 otherwise.
K
tr ĜH
mk Ĝmk Γmk ≤ 1, ∀m
One can observe that the ICI is removed. However, in a real- k=1
istic system where the angle reciprocity is imperfect (σθ = 0 Γmk 0, ∀m, k. (23)
and σβ = 0), the ICI still remains in the received signal and
The relaxed problem (23) is a convex SDP and can be solved
hence there is a loss in the data rate. In the following theorem,
by the convex optimization tool such as CVX [13]. Once the
we provide an approximated closed-form expression for the
optimal variable {Γ̂mk } are obtained, we can find the rank
achievable data rate when the angle reciprocity is imperfect.
one approximations of {Γ̂mk } which are feasible for the
Theorem III.1. The approximated closed-form expression of original problem (22) using the rank reduction methods [14].
achievable rate for the user k is
⎛ ⎞
M 2
2 2
IV. S IMULATION R ESULTS
⎜ m=1 BUL
mkγ mk +σθ BmkCmkQ̂mkkγ mk
UL UL
⎟
2 2
Rk ≈ log2⎜
⎝1+ M K 2
⎟,
⎠ In this section, we evaluate the performances of the pro-
σθ2 m=1 j=k BUL
mk Cmk Q̂mkj γ mj
UL
+ ρ1d posed FDD-based cell-free massive MIMO systems. We con-
2
(21) sider M basestations with N antennas and K users with single
where Q̂mkj ∈ CP ×P is a submatrix of Q̂m = antenna randomly distributed in a square area of 1 × 1km2 . In
−1
m EÂm (Âm Âm ) whose position is [(k − 1)P + 1 :
ÂH H our simulations, the QoS requirements of the users are fixed
kP, (j − 1)P + 1 : jP ]. into 2 bit/symbol. The system bandwidth is set to 10 MHz and
the carrier frequencies of the uplink and downlink channels are
f UL = 49.8GHz and f DL = 50GHz, respectively. The users
C. Power Allocation for Each Propagation Path transmit OFDM pilot signals of τ = 200 symbols with uplink
training power ρr = 200 mW and the number of pilot samples
In this subsection, we discuss how to solve the power
are T = 16. For the multipath channel model, the number of
allocation problem that optimizes the power coefficients γmk,i
propagation paths of the channel is P , the angle spread is 10◦ ,
for each propagation path. Since each propagation path has dif-
DL and the large-scale fading coefficients of LOS and NLOS paths
ferent path gain βmk,i , proper allocation of transmission power
are modeled similarly to the 3GPP LTE standard as [15]
can sufficiently improve the energy efficiency. To this end, we
formulate the power allocation problem that minimizes the −148 − 35 log10 (dmk ) + zmk,p , LOS
total transmission power of the basestations while satisfying 10 log10 (βmk,p ) =
−158 − 35 log10 (dmk ) + zmk,p , NLOS
the quality of service (QoS) requirements of the users as
M
K where dmk is the distance between the basestation m and user
min Ĝmk γ mk 22 k in kilometers and zmk,P ∼ N (0, σz2 ) is the shadow fading
{γ mk }
m=1 k=1 variable with σz = 8 dB.
M UL 2 2 2
Fig. 2 shows the root mean square error (RMSE) per-
m=1 Bmk γ mk 2 +σθ B Cmk Q̂mkk γ mk 2
UL UL
s.t. K 1
≥ ξk , ∀k formance of the proposed gradient descent-based multipath
σθ2 M 2
j=k Bmk Cmk Q̂mkj γ mj 2 + ρd
UL UL
m=1 component estimation technique for N = 32 and T = 16.
K
For comparison, we use the ESPRIT algorithm, a well-known
Ĝmk γ mk 22 ≤ 1, ∀m, (22) subspace-based multipath component estimation technique [7].
k=1
One can observe that the proposed gradient descent-based
where ξk is the QoS requirement of the user k. (22) is a technique outperforms the subspace-based technique in the
non-convex separable quadratic constraint quadratic program
(QCQP). While the non-convex QCQP is a NP-hard problem, 1Γ is rank one symmetric positive semi-definite matrix (PSD)
mk
2018 IEEE 19th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC)
Fig. 2: The RMSE performance of the multipath component Fig. 3: The total transmission power versus the number of
estimation versus SNR for N = 32 and T = 16 basestations for N = 512, P = 4, and T = 16
angle estimation. Also, the normalized RMSE performance R EFERENCES
of the large-scale fading coefficient estimation stays around [1] H. Ji, Y. Kim, J. Lee, E. Onggosanusi, Y. Nam, J. Zhang, B. Lee, and
−10 dB in every SNR region. B. Shim, “Overview of full-dimension MIMO in LTE-advanced pro,”
Fig. 3 shows the total transmission power of the basestations IEEE Comm. Mag., vol. 55, no. 2, pp. 176–184, 2017.
[2] H. Q. Ngo, A. Ashikhmin, H. Yang, E. G. Larsson, and T. L. Marzetta,
in the FDD-based cell-free massive MIMO systems for a “Cell-free massive MIMO versus small cells,” IEEE Trans. Wireless
various number of basestations and users with N = 512, Commun., vol. 16, no. 3, pp. 1834–1850, 2017.
P = 4, and T = 16. For comparison, we use the cellular [3] E. Nayebi, A. Ashikhmin, T. L. Marzetta, H. Yang, and B. D. Rao, “Pre-
coding and power optimization in cell-free massive MIMO systems,”
systems where each user is served by only one basestation. IEEE Trans. Wireless Commun., 2017.
The basestation mk serving the user k is selected as mk = [4] H. Q. Ngo, L.-N. Tran, T. Q. Duong, M. Matthaiou, and E. G. Larsson,
P “On the total energy efficiency of cell-free massive mimo,” IEEE Trans.
argmax p=1 βmk,p . To make a fair comparison, the angle-
m Green Commun. Netw., vol. 2, no. 1, pp. 25–39, 2018.
based ZF precoding is also applied in the cellular systems. [5] B. Lee, J. Choi, J.-Y. Seol, D. J. Love, and B. Shim, “Antenna grouping
based feedback compression for FDD-based massive MIMO systems,”
From Fig. 3, we observe that the energy efficiency of the IEEE Trans. Commun., vol. 63, no. 9, pp. 3261–3274, 2015.
FDD-based cell-free massive MIMO systems is considerably [6] H. Xie, F. Gao, S. Zhang, and S. Jin, “A unified transmission strategy for
better than the conventional cellular systems. For instance, TDD/FDD massive MIMO systems with spatial basis expansion model,”
IEEE Trans. Vehicle. Tech., vol. 66, no. 4, pp. 3170–3184, 2017.
when M = 16 and K = 4, the FDD-based cell-free massive [7] R. Roy and T. Kailath, “Esprit-estimation of signal parameters via
MIMO systems consume the total transmission power of 16.1 rotational invariance techniques,” IEEE Trans. Acoust. Speech. Signal.
W while the cellular systems consume the total transmission Process., vol. 37, no. 7, pp. 984–995, 1989.
[8] F. Fernandes, A. Ashikhmin, and T. L. Marzetta, “Inter-cell interference
power of 19 W. Thus, the FDD-based cell-free massive MIMO in noncooperative TDD large scale antenna systems,” IEEE J. Sel. Areas
systems save approximately 15% of the transmission power. in Commun., vol. 31, no. 2, pp. 192–201, 2013.
[9] P. Kyosti, “Winner II channel models,” IST, Tech. Rep. IST-4-027756
V. C ONCLUSION WINNER II D1. 1.2 V1. 2, 2007.
[10] K. Hugl, K. Kalliola, and J. Laurila, “Spatial reciprocity of uplink and
In this paper, we presented the FDD-based cell-free massive downlink radio channels in FDD systems,” Proc. COST 273 Technical
MIMO systems that directly acquire multipath components Document TD (02), vol. 66, p. 7, 2002.
[11] J. Park, B. Lee, and B. Shim, “A MMSE vector precoding with block
from the uplink pilot signal and use them for the basestation diagonalization for multiuser MIMO downlink,” IEEE Trans. Commun.,
cooperation. To be specific, we proposed the computation- vol. 60, no. 2, pp. 569–577, 2012.
ally efficient multipath component estimation technique based [12] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge
university press, 2004.
on the gradient descent method and basestation cooperation [13] M. Grant, S. Boyd, and Y. Ye, “CVX: Matlab software for disciplined
scheme that utilizes the multipath components. By exploiting convex programming,” 2008.
the angle reciprocity, no downlink CSI feedback from the [14] Z.-Q. Luo, W.-K. Ma, A. M.-C. So, Y. Ye, and S. Zhang, “Semidefinite
relaxation of quadratic optimization problems,” IEEE Signal Process.
user is required for the CSI acquisition. From the simulation Mag., vol. 27, no. 3, pp. 20–34, 2010.
results, we showed that the proposed multipath component [15] E. U. T. R. Access, “Further advancements for E-UTRA physical layer
estimation technique outperforms the conventional subspace- aspects,” 3GPP Technical Specification TR, vol. 36, p. V2, 2010.
based technique. Furthermore, we showed that the energy
efficiency of the proposed FDD-based cell-free massive MIMO
systems is much better than that of the cellular systems.