Professional Documents
Culture Documents
Abstract—The theoretical analysis of downlink massive MIMO channels, both the path delay in the time domain and angle of
usually assumes the ideal Gaussian channel matrix with asymp- departure (AoD)/angle of arrival (AoA) in the angle domain
arXiv:1508.00070v2 [cs.IT] 27 Nov 2015
totic orthogonality of channel vectors associated with different are essential parameters to represent the channels. Therefore,
users, since it can provide the favorable propagation condition.
Meanwhile, recent experiments have shown that massive MIMO in this paper, we will jointly consider the SCS of massive
channels between a certain user and massive base station an- MIMO channels in both the time domain and angle domain.
tennas appear the spatial common sparsity (SCS) in both the Clearly, the actual massive MIMO channel experienced by
time domain and angle domain. This motivates us to investigate practical systems is different from the ideal Gaussian channel
whether realistic sparse massive MIMO channels could provide matrix for theoretical analysis. Such difference motivates us
the favorable propagation condition, and reveal the capacity gap
between massive MIMO systems over realistic sparse channels to investigate the propagation condition of realistic massive
and that under the ideal Gaussian channel matrix assumption. MIMO channels with SCS, and reveal the capacity gap be-
This paper theoretically proves that channel vectors associated tween massive MIMO systems over actual sparse channels
with different users in massive MIMO over realistic sparse and that under the ideal Gaussian channel matrix assumption.
channels satisfy the asymptotic orthogonality, which indicates Specifically, we first introduce the SCS of massive MIMO
that the favorable propagation condition can also be provided.
Moreover, the simulation results confirm the theoretical analysis. channels in both the time domain and angle domain, and
illustrate the connection between the sparse massive MIMO
channel model and the frequency-domain channel matrix
Index Terms—Massive MIMO, favorable propagation condi-
tion, spatial common sparsity (SCS), asymptotic orthogonality. model in massive MIMO-OFDM systems. Then, we prove
that channel vectors over sparse channels are asymptotically
orthogonal, where our analysis is based on the widely used
isotropic uniform linear array (ULA) without loss of generality
I. I NTRODUCTION
[3], [7], [8]. The asymptotic orthogonality of different channel
The theoretical performance of massive MIMO relies on the vectors indicates that massive MIMO channels with SCS could
favorable propagation condition, which is equivalent to the provide the favorable propagation condition, which is also
asymptotic orthogonality of channel vectors associated with confirmed by simulations.
different users [1]. For theoretical analysis, the channel is
usually modeled as the ideal Gaussian channel matrix with II. S YSTEM M ODEL
elements following the mutually independent Gaussian distri- In typical massive MIMO systems, the BS employs M
bution, so that channel vectors associated with different users antennas to simultaneously serve K single-antenna users.
are asymptotically orthogonal and can provide the favorable Usually, M is large, and much larger than K, e.g., M = 128
propagation condition for massive MIMO [1], [2]. and K = 16 [1]. In the downlink, the BS transmits data
Recent researches have shown that, due to the finite number x ∈ CM×1 to K single-antenna users, and the received signal
of significant scatterers and the compact antenna array at the y ∈ CK×1 for K users can be expressed as
base station (BS), realistic massive MIMO channels between a √ √
y = ρd Gx + n = ρd D1/2 Hx + n, (1)
certain user and massive base station (BS) antennas exhibit the
intrinsically spatial common sparsity (SCS) [3]–[8]. However, where ρd is the transmit power, G ∈ CK×M is the downlink
in [4]–[6], the proposed channel estimation schemes only channel matrix, n ∈ CK×1 is the additive white Gaussian
consider the massive MIMO channels with SCS in the time noise (AWGN) vector, the diagonal√ matrix D1/2 ∈ CK×K
domain, while [7] and [8] only consider the SCS of massive with the kth diagonal element βk represents the large-scale
MIMO channels in the angle domain. For massive MIMO channel fading due to shadow fading and path loss, while H ∈
CK×M denotes the small-scale channel fading matrix due to
Z. Gao, L. Dai, and Z. Wang are with Tsinghua National Laboratory channel variation. Note that when the widely used OFDM is
for Information Science and Technology (TNList), Department of Electronic
Engineering, Tsinghua University, Beijing 100084, China (E-mails: gao-
considered, (1) is valid for every frequency-domain subcarrier.
z11@mails.tsinghua.edu.cn; {daill, zcwang}@tsinghua.edu.cn). To fully exploit advantages of massive MIMO, the BS
C. Yuen is with Singapore University of Technology and Design, Singapore requires the downlink channel information for precoding,
138682 (e-mail: yuenchau@sutd.edu.sg). user scheduling, etc. The propagation condition of downlink
This work was supported in part by the National Key Basic Research
Program of China (Grant No. 2013CB329203), Singapore A*STAR SERC channel matrix determines the achievable capacity of downlink
Project (Grant No. 142 02 00043), National Natural Science Foundation of massive MIMO, which can be expressed as [1]
China (Grant Nos. 61201185 and 61271266), and Beijing Natural Science
Foundation (Grant No. 4142027). C = log2 det (IK + ρd GG∗ ) , (2)
2
where IK ∈ CK×K is the unit matrix, and (·)∗ is the Furthermore, for a certain multipath from one scatterer, the
Hermitian transpose operation. It can be observed from (2) AoD seen from M antennas at the BS are very similar, which
that the achievable capacity heavily depends on the property is referred to as the SCS in the angle domain of massive
of channel matrix G. MIMO channels [3], [7], [8]. Due to such SCS in the angle
For the ideal Gaussian channel matrix, G provides the domain, the path gains associated with M antennas via the
favorable propagation condition, where elements of the small- sk th scatterer can be expressed as the steering vector, i.e.,
scale channel fading matrix H follow the independent and h sk(M −1)
i
αsk0k0 , αsk1k1 , · · · , αk(M −1) =
identically distributed (i.i.d.) circularly symmetric complex s s s
2πd sin(θ k ) 2π2d sin(θ k ) 2π(M −1)d sin(θ k )
Gaussian distribution CN (0, 1). It has been shown that row s
αkk 1, ej λ
k
, ej λ
k
, · · · , ej λ
k
,
vectors of G are asymptotically orthogonal [1], i.e., (6)
lim GG∗ /M = lim D1/2 HH∗ D1/2 /M = D. (3) where the isotropic ULA with the antenna spacing d is
sk0
M→∞ M→∞
considered without loss of generality, αk0 = αskk is the path
Under such ideal Gaussian channel matrix assumption, the sk
gain of the first (m = 0) BS antenna, θk is the AoD of the
available capacity has the simple asymptotic form as follows sk th propagation path, and λ denotes the wavelength.
K The channel model with SCS in both the time domain and
M≫K X
log2 det (IK + ρd GG∗ ) ≈ log2 (1 + ρd M βk ). (4) angle domain is different from the ideal Gaussian channel ma-
k=1 trix with the inherent favorable propagation condition widely
Such favorable propagation condition can ensure that the inter- used for theoretical analysis. This motivates us to investigate
user interference vanishes when M → ∞ [1]. Accordingly, whether the realistic sparse massive MIMO channels meet the
the superiority of massive MIMO in spectral efficiency can be favorable propagation condition, and reveal the capacity gap
guaranteed. For practical massive MIMO with a finite M , the between massive MIMO over realistic sparse channels and that
favorable propagation condition indicates that the condition under the ideal Gaussian channel matrix assumption.
number of G should be as small as possible [1].
B. Asymptotic Orthogonality Analysis
III. A SYMPTOTIC O RTHOGONALITY A NALYSIS FOR
According to (5), the channel frequency response over the
M ASSIVE MIMO C HANNELS WITH SCS nth subcarrier, denoted by gkm [n], can be expressed as
SPkm
skm
j2πnfs τkm
We first illustrate the realistic massive MIMO channels gkm [n] = αskm
km
exp − N
skm =1
with SCS in both the time domain and angle domain. Then SPkm
(7)
we theoretically prove that channel vectors are asymptotically = αskm
km
exp (−j2πγn τkmskm
), 1 ≤ n ≤ N,
skm =1
orthogonal, where the first and second-order moments of the
inner product of different channel vectors are derived. where N is the size of the OFDM symbol and γn = nfs /N .
Eq. (7) builds the connection between time-domain massive
A. Massive MIMO Channels with SCS MIMO channels and frequency-domain channel matrix. For
multi-user MIMO-OFDM systems, gkm [n] is the element of
Due to the finite number of significant scatterers, for the the kth row and the mth column in the channel matrix
time-domain channel impulse response (CIR), the majority of G[n] over the nth subcarrier (see G in Eq. (1)). Here
channel gains are concentrated on a small number of channel T
G[n] = (g1 [n])T , (g2 [n])T , · · · , (gK [n])T
and gk [n] =
taps although the delay spread can be large in typical wireless
[gk1 [n], gk2 [n], · · · , gkM [n]]. Based on (7), we will investigate
communication systems [3]. The sparse downlink time-domain
the asymptotic orthogonality of row vectors of the frequency-
CIR between the kth user and the mth BS antenna can be
domain channel matrix over sparse massive MIMO channels.
expressed as [4]
Due to the SCS in the angle domain as shown in (6), for
XSkm
hkm (t) = αskm
km skm
δ (t−τkm ), Skm ≪ τmax fs , (5) the sk th propagation path, the path gain associated with the
skm =1 mth BS antenna αskm km
can be expressed as
where αskmkm skm
and τkm are the skm th path gain and path delay, sk
αskm
km
= αskk ej2πmd sin(θk )/λ
, 0 ≤ m ≤ M − 1. (8)
respectively, Skm is the number of resolvable multipaths, τmax
is the maximum delay spread, fs is the system bandwidth, and Furthermore, due to the SCS in the time domain, under the
τmax fs is the normalized maximum delay spread. far field assumption, the corresponding path delay of the mth
Moreover, there are measurements showing that time- BS antenna via the sk th scatterer can be expressed as [4]
domain CIRs between different co-located BS antennas and a skm
τkm = τksk + md sin(θksk )/c, 0 ≤ m ≤ M − 1, (9)
certain user share very similar path delays [1], i.e., skm = sk
and Skm = Sk , for 0 ≤ m ≤ M − 1. This is because the scale where τksk
is the path delay of the first (m = 0) BS antenna,
of the compact antenna array at the BS is small compared with and c is the velocity of electromagnetic waves.
the large distance of signal transmission, which leads to the Consequently, by substituting (8) and (9) into (7), we have
s
Sk md sin(θ k )
fact that channels of different transmit-receive antenna pairs k λγ s
(1− cn )−j2πγn τk k
αskk ej2π
P
gkm [n] = λ
share very similar scatterers. The inherent sparsity of time- sk =1
Sk s
md sin(θ k )
(10)
domain massive MIMO channels discussed above is referred s k λγ
(1− cn ) s
αkk ej2π
P
= λ µnk
k
,
to as the SCS in the time domain. sk =1
3
sk √
where µsnkk
= e−j2πγn τk . proportional to 1/ x when x > 0 increases [10], we have
For massive MIMO systems, since the distance between M −1
|Eα {υpqn }| J0 (am)2
P
different users simultaneously communicating with the BS
n o
gp [n]gq [n]∗ m=0
lim E = lim
M M
is usually large, their respective CIRs are usually mutually M →∞
MP
−1
M →∞
MP
−1 (15)
independent [1]. Meanwhile, we assume that their AoDs Eα {|υpqn |} J0 (am)2 p J0 (am)2
m=0
≤ lim ≤ Sp Sq lim m=0 M = 0,
follow the i.i.d. uniform distribution U [0, 2π) [9]. Due to M →∞ M M →∞
the limited dominant scatterers, the limited AoDs appear the where the first inequality is based on Jensen’s inequality.
SCS. Then, based on (10) and according to the definition of Moreover, we investigate the second-order moment of
asymptotic orthogonality (3), we have ∗
gp [n]gq [n] /M , namely,
Sp sp
M −1
|gp [n]gq [n]∗ |2
md sin(θp )
o2
gp [n]gq [n]∗ s λγ
n o n
1 s
p j2π (1− cn ) gp [n]gq [n]∗ gp [n]gq [n]∗
αpp µnp
P P
M
= M
( e λ var M
=E M2
−E M
, (16)
m=0 sp =1
Sq sq
md sin(θq )
(11)
×
P s
(αqq µnq
s
q ∗ −j2π
) e λ
λγ
(1− cn )
), where var {·} is the variance of a RV. Furthermore, we have
sq =1 Sp Sq
MP s s
−1 sp jam sin(θpp ) sq jam sin(θq q ) ∗
P P
2 znp e znq e
where p and q denote two arbitrary users with 1 ≤ p < q ≤ |gp [n]gq [n]∗ | m=0 sp =1 sq =1
M2
= M
K. Based on (11), we will derive the first and second-order
∗
s′ s′
MP
−1 Sp Sq s′q jam′ sin(θ q ) s′p jam′ sin(θ p )
q p
P P
∗ znq e znp e
moments of gp [n]gq [n] /M . Taking the statistical expectation m′ =0 s′p =1 s′q =1
× ,
of (11) with respect to AoDs, we obtain M
(17)
Eθ {gp [n]gq [n]∗ /M } and we can obtain its expectation for AoDs as follows
Sp sp
M −1 P 2π md sin(θp ) λγ
s sp 1 R j2π (1− cn ) s
(
1
αpp µnp dθpp Sp Sq
P
= M e |gp [n]gq [n]∗ |2
M −1 M −1 P
λ P sp 2 sq 2
1 2
2π
znp znq J0(a(m−m′ ))
P P
m=0 sp =1 0 Eθ M2
=M2
Sq 2π sq
md sin(θq )
m=0 m′ =0 sp =1sq =1
s sq ∗ 1 (1− λγ n) s
αqq µnq e−j2π
R
dθq q
P
× λ c Sq
Sp P Sq ′
sq ∗ sq
2π 2 znq
sp
znq J0 (a(m−m′ ))J0 (am)J0 (am′ )
P P
sq =1 0 + znp
Sp Sq M −1 sp =1sq =1 s′q =1,s′q 6=sq
sp s 1 λγn
2
(znqq )∗ M J0 2π md
P P P
= znp λ
(1 − c
) Sq Sp Sp ′ ∗
sp
sp =1 sq =1 m=0 +
P P P znq 2 znp
sq sp
znp J0 (a(m′ −m))J0 (am)J0 (am′ )
M −1 sq =1sp =1 s′p =1,s′p 6=sp
υpqn P
= M
J0 (am)2 , Sp Sp Sq Sq
′∗ ′
m=0
P P P P sp sp sq∗ sq 2 ′ 2
+
(12) sp =1s′ =1,s′ 6=s sq =1s′ =1,s′ 6=s znp znp znq znqJ0 (am) J0 (am ) .
p p p q q q
where Eθ {·} represents the expectation operator for random (18)
sp Sp sq Sq
variables (RVs) θp sp =1 and θq sq =1 subjected to the Similar to (14), we can get the following inequality
i.i.d. U [0, 2π), J0 (·) is the zero-order Bessel function of the
|gp [n]gq [n]∗|2
|gp [n]gq [n]∗|2
M −1M −1n
R 2π ±jx 1 2
1 sin θ J0 (a(m−m′ ))
P P
first kind with the definition J0 (x) = 2π e dθ [10], E M 2 =E α E θ M 2 ≤ M 2
0 m=0 ′ =0
sp sp sp sq sq sq λγ m
znp =αp µnp , znq =αq µnq , a = 2π λd (1 − cn ), and υpqn = +(Sq +Sp −2)J0 (a(m−m′ )) J0 (am)oJ0 (am′ )
P SP P Sq sp sp sq sq ∗
sq =1 αp µnp αq µnq . Besides, due to fc ≫ fs in
2
sp =1 +(Sp −1)(Sq − 1)J0 (am)2 J0 (am′ ) .
λγn nfs
typical massive MIMO systems, we have c = N fc ≪ 1 (19)
for 1 ≤ n ≤ N and a > 0, where fc is the carrier frequency. Similar to (13) and (15), we can further obtain
∗
According to (12), the first-order moment of gp [n]gq [n] /M
|gp [n]gq [n]∗ |2
|gp [n]gq [n]∗ |2
can be expressed as lim E M 2 = lim E α E θ M 2 =0.
M →∞ M →∞
M −1 (20)
oo Eα {υpqn} P J0 (am)2 ∗
n
gp [n]gq [n]∗
o n n
gp [n]gq [n]∗ m=0 Then we have lim var gp [n]gq [n] /M = 0 by substituting
E M
= Eα Eθ M
= M
, M→∞
(13) (15) and (20) into (16). Consequently, the asymptotic orthog-
where E {·} is the expectation operator for all RVs, Eα {·} is onality of row vectors of the channel matrix G[n] is proven.
s Sp
the expectation operator for RVs αpp sp =1 and αqq sq =1 ,
s Sq It should be pointed out that we can further exploit
the statistical characteristics of path gains to simplify (13)
and path gains and AoDs are mutually independent.
and (16). Specifically, if we consider the Sk -path channels
Since the large-scale channel fading has no impact on the S
with {αskk }skk=1 following the i.i.d. CN (0, 1/Sk ), we have
asymptotic orthogonality of channel vectors, without loss of nP o s
Sk sk 2
generality, we consider channels with normalized energy, i.e., E sk =1 |α | = 1, Eα {υpqn } = 0, E znpp = 0,
PSp sp 2 PS s 2 s
sp =1 αp = 1 and sqq=1 αqq = 1. Then we have and E znqq = 0. Therefore, (13) and (16) can be further
expressed as
Sp Sq sq sq ∗ p
|υpqn | ≤
P P αspp µsnp
p
(αq µnq ) ≤ Sp Sq , (14) M −1
E {gp [n]gq [n]∗ /M } = J0 (am)2 Eα {υpqn }/M = 0,
P
sp =1 sq =1 (21)
m=0
where the second inequality utilizes the Cauchy-Schwarz n n
2
oo
inequality [10]. var {gp [n]gq [n]∗ /M } =Eα Eθ |gp [n]gq [n]∗ | /M 2
M −1 M −1 (22)
Consequently, based on the discussion above, as well as the =
P P 2
J0 (a(m−m′ )) /M 2 .
boundedness (|J0 (x)| < 1 for x > 0) and the delay of J0 (x) m=0 m′ =0
4
0 1
0.2 10
0.9 i.i.d. Gaussian
Simulated, d=0.2λ Simulated
Sparse, d=λ/2
0.15 Simulated, d=0.5λ Analytical 0.8
Sparse, d=λ
var{gp[n]gq[n] /M}
|E{g [n]g [n] /M}|
Prob( σ ≤ Abscissa)
Simulated, d=λ −1 0.7 Sparse, d=2λ
10
*
Smallest, 128x6
*
0.5
Largest, 128x6
0.05 0.4
p
−2
10
0.3
0 d=0.5λ Smallest, 6x6 Largest, 6x6
0.2
−3
d=λ 0.1
−0.05 10
1 2 3 1 2 3 0
10 10 10 10 10 10 −30 −25 −20 −15 −10 −5 0 5 10 15 20 25
Number of antennas M Number of antennas M
(a) (b) Eigenvalues of GG*
Fig. 1. |E {gp [n]gq [n]∗ /M } | and var {gp [n]gq [n]∗ /M } against M and d. Fig. 2. CDFs of smallest/largest eigenvalue of GG∗ in MIMO systems.
10
i.i.d. Gaussian, SNR = 10 dB
IV. N UMERICAL R ESULTS
(Bits/Channel Use)
i.i.d. Gaussian, SNR = 20 dB