Professional Documents
Culture Documents
3D Massive MIMO Systems: Modeling and Performance Analysis
3D Massive MIMO Systems: Modeling and Performance Analysis
Hachem et al derived a Central Limit Theorem (CLT) for a reduced degree of randomness in the channel matrix. This
the MI of Kronecker channel model and rigorously proved calls for the use of an approach that does not depend on the
that the MI converges to a standard Gaussian random variable more generally employed results on the distribution of Wishart
in the asymptotic limit in [19]. More asymptotic results on matrices. However, the systematic structure of the model can
the statistical distribution of MI of 2D channels have been be exploited to our aid in characterizing its MI. With this
provided in [14]–[16], [18]. However, the extension of these channel model at hand, we derive analytical expressions for
results to the 3D case has not appeared yet. the cumulative density function (CDF) of the MI in different
The efficient design and better understanding of the limits signal-to-interference plus noise ratio (SINR) regimes. An
of wireless systems require accurate modeling and character- exact closed-form expression is obtained for systems with
ization of the 3D channels. Given the available knowledge a single Rx and multiple Tx antennas in the general SINR
on certain channel parameters like angle of departure (AoD), regime. The result is extended to systems with a finite number
angle of arrival (AoA), delay, amplitude, number of Tx and of Rx antennas in the low SINR regime. We also present an
Rx antennas, finding the best way to model the channel asymptotic analysis of the statistical distribution of the MI and
consistent with the information available and to attribute a joint show that it is well approximated by a Gaussian distribution for
probability distribution to the entries of the channel matrix any number of Rx antennas, as the number of paths and Tx an-
is of vital importance. It has been shown in literature that tennas grow large. Finally numerical results are provided that
the choice of distribution with the greatest entropy creates a illustrate an excellent fit between the Monte-Carlo simulated
model out of the available information without making any CDF and the derived closed-form expressions. The simulation
arbitrary assumption on the information that was not available results provide a flavor of the performance gains realizable
[20], [21], [22]. In the context of wireless communications, through the meticulous selection of the downtilt angles. We
the authors in [23] addressed the question of MIMO chan- think our work contributes to cover new aspects that have not
nel modeling from an information-theoretic point of view yet been investigated by a research publication and addresses
and provided guidelines to determine the distribution of the some of the current challenges faced by researchers and
2D MIMO channel matrix through an extensive use of the industrials to fairly evaluate 3D beamforming techniques on
principle of maximum entropy together with the consistency the basis of achievable rates. The results have immediate
argument. Several channel models were derived consistent applications in the design of 3D 5G massive MIMO systems.
with the apriori knowledge on parameters like AoA, AoD, The rest of the paper is organized as follows. In section
number of scatterers. These results can be extended to the II, we introduce the 3D channel model based on the 3GPP
3D MIMO channel model, which will be one of the aims of standards and information-theoretic considerations. In Section
this work. As mentioned earlier, almost all existing channel III, we provide analytical expressions for the CDF of the MI
models are 2D. However, owing to the growing interest in 3D in the general and low SINR regimes for a finite number of
beamforming, extensions of these channel models to the 3D Tx antennas. In section IV, we provide an asymptotic analysis
case have started to emerge recently [24], [8]. An important that holds for any number of Rx antennas. The derived results
feature of these models is the vertical downtilt angle, that can are corroborated using simulations in section V and finally in
be optimized to change the vertical beam pattern dynamically section VI, some concluding remarks are drawn.
and yield the promised MI gains [25]–[28]. Studies on 3D
channel modeling and impact of antenna downtilt angle on
II. S YSTEM AND C HANNEL M ODEL
the achievable rates have started to appear in literature, but
to the best of authors’ knowledge, there has not been any Encouraged by the potential of elevation beamforming
comprehensive study to develop analytical expressions for the to enhance system performance, many standardized channel
MI of these channels, that could help evaluate the performance models have started to emerge that define the next generation
of the future 3D MIMO systems. 3D channels. We base the evaluation of our work on these
The aims of this paper are twofold. First is to provide an models while making some realistic assumptions on the chan-
information-theoretic channel model for 3D massive MIMO nel parameters.
systems and second is to predict and analyze the performance
of these systems by characterizing the distribution of the MI.
We follow the guidelines provided in [23] to present the A. System Model
entropy maximizing 3D channel model that is consistent with We consider a downlink 3D MIMO system, where the BS
the state of available knowledge of the channel parameters. is equipped with NBS directional Tx antenna ports and the
The 3D channel used for the MI analysis is inspired from mobile station (MS) has NM S Rx antenna ports. The antenna
the models presented in standards like 3GPP SCM [29], configuration is shown in Fig.1. Note that θtilt represents the
WINNER+ [24] and ITU [30]. These standardized MIMO elevation angle of the antenna boresight, a parameter that does
channel models assume single-bounce scattering between the not appear in the 2D channel models. Also, zero electrical
transmitter and receiver, which allows us to assume the AoDs downtilt corresponds to θtilt = π2 . In LTE, the radio resource
and AoAs to be fixed and known apriori during the modeling is organized on the basis of antenna ports. Each antenna port
phase. The maximum entropy 3D channel model consistent is mapped to a group of physical antenna elements which
with the state of knowledge of channel statistics pertaining to carry the same signal and constitute an antenna. The spatial
AoDs and AoAs, turns out to have a systematic structure with multiplexing is performed across the ports.
3
N
1 X p p
[H]su = αn gt (φn , θn , θtilt ) exp (ik(s − 1)dt sin φn sin θn ) gr (ϕn , ϑn ) exp (ik(u − 1)dr sin ϕn sin ϑn ) , (1)
N n=1
In the configuration shown in Fig. 1, there are NBS antenna not a true depiction of the environment. Also, assuming the
ports placed in the horizontal plane, along the êy direction. elevation angle of the antenna boresight to be fixed implies
Each antenna port is mapped to M vertically stacked antenna that the channel’s degrees of freedom in the elevation are
elements that determine the effective port pattern. The same not being exploited. The dynamic adaptation of the downtilt
Tx signal is fed to all elements in a port with corresponding angles can open up several possibilities for 3D beamforming
weights wk ’s, k = 1, . . . , M , in order to achieve the desired that can bring about significant performance gains. Therefore
directivity. In the current activity around 3D channel modeling the extension of the existing 2D models to the 3D case needs
by the 3GPP, the proposed method for 3D beamforming is to take into account the elevation angles of the propagation
to adjust the weights applied to the elements in a port to paths and introduce the parameter θtilt into the expressions of
obtain the desired downtilt angle for that port [8]. The same the antenna pattern rather than assuming it to be fixed. These
is done for the other ports. The MS sees each antenna port extensions have started to emerge recently [24], [32], [30].
as a single antenna because all the elements carry the same Based on the aforementioned standards and for the antenna
signal. Therefore, we are interested in the channel between the configuration in Fig. 1, the effective 3D channel between the
transmitting antenna port and the receiver side. BS antenna port s and the MS antenna port u is given by
(1), where φn and θn are the azimuth and elevation AoD of
B. 3D MIMO Channel Model the nth path respectively, ϕn and ϑn are the azimuth and
elevation AoA of the nth path respectively and θtilt is the
The mobile communication standards like 3GPP SCM [29],
ITU [30] and WINNER [31] follow a system level, stochastic
elevation angle of the antenna boresight.
th
p αn is the complex
random amplitude of the n path. Also gt (φn , θn , θtilt ) and
channel modeling approach wherein, the propagation paths are p
gr (ϕn , ϑn ) are the global patterns of Tx and Rx antennas
described through statistical parameters, like delay, amplitude, respectively. dt and dr are the separations between Tx antenna
AoA and AoD. These models generate channel realizations ports and Rx antenna ports respectively and k is the wave
by summing the contributions of N multipaths (plane waves) number that equals 2π λ , where λ is the wavelength of the carrier
with specific parameters. In the existing 2D channel models frequency. The entries [33],
like the 3GPP SCM and WINNER, the propagation paths are
described using azimuth angles alone. Moreover, the value of
the elevation angle of the antenna boresight (θtilt ) is assumed [at (φ, θ)]s = exp (ik(s − 1)dt sin φ sin θ) , (2)
to be fixed at π2 in these 2D models. However, as shown by [ar (ϕ, ϑ)]u = exp (ik(u − 1)dr sin ϕ sin ϑ) (3)
several measurement campaigns [4], [5], there is a significant
component of energy that is radiated in the elevation so are the array responses of sth Tx and uth Rx antenna respec-
characterizing the propagation paths in the azimuth alone is tively. Fig. 2 shows the 3D channel model being considered.
To enable an abstraction of the role played by antenna
elements in performing the downtilt, ITU approximates the
global pattern of each port by a narrow beam as [30], [33],
Antenna port Antenna port
1 NBS p
[17dBi − min{−(AH (φ) + AV (θ, θtilt )), 20dB}]lin , (4)
w M( tilt)
k Azimuth plane
Multipath n
w 2( tilt) tilt
tilt y
w 1( tilt)
Antenna boresight
Antenna boresight
expression for the statistical distribution of the MI in the where C is given by (14).
general SINR regime for a single Rx antenna and any arbitrary This quadratic form in Normal RVs will play a crucial role
finite number of Tx antennas. The result is then extended to in the subsequent analysis in this work. The CF of a quadratic
an arbitrary finite number of Rx antennas in the low SINR form in Normal RVs is given by [35], [36],
regime. H 1
We consider a point-to-point communication link where H E[ej(α Cα) ] = . (16)
det[IN − j N ((B ΩB) ◦ (AH A)T )]
1 H
is a NM S × NBS MIMO channel matrix from (10). The
channel is linear and time-invariant. The received complex It is easy to see that Q the denominator of (16) can be
N
baseband signal y ∈ CNM S ×1 at the MS is given by, equivalently expressed as i=1 (1 − jλi ), where λi ’s are the
eigenvalues of C. (16) has a form similar to the CF of the
y = Hs + n, (11) sum of i.i.dPunit parameter exponential RVs scaled by λi ’s, i.e.
N
where s ∈ CNBS ×1 is the Tx signal from the BS and n ∈ αH Cα ∼ i=1 λi Exp(1). Consequently, the closed-form ex-
CNM S ×1 is the complex additive noise such that E[nnH ] = pression of the statistical distribution of (R+σ 2 INM S )−1 HHH
R + σ 2 I, where R is the covariance matrix of interference is given by [37],
experienced by the MS and σ 2 is the noise variance at the N
λN
H
X
i 1 x
MS. P[α Cα ≤ x] = 1 − exp − ,
N λi λi
In this context, the MI of the 3D MIMO system is given i=1
Q
(λi − λl )
by, l6=i
2 2
I(σ ) = log det(INM S + (R + σ INM S ) −1 H
HH ). (12) The rest of the analysis can be completed through the
transformation log(1 + x).
H is fixed during the communication interval, so we do not
need to time average the MI. P[I(σ 2 ) ≤ y] = P[log(1 + x) ≤ y] = Fx (ey − 1). (17)
The theoretical result coincides very well with the simulated
A. General SINR Regime CDF for the general SINR regime, as will be illustrated later
The exact MI given in (12) has been mostly characterized in the simulations section.
in literature for the cases when HHH is a Wishart matrix [9]–
[11], [13]. Wishart matrices are random matrices with inde- B. Low SINR Regime
pendent entries and have been frequently studied in the context It is well known that the performance of the channel can
of wireless communications. However the special structure of severely deteriorate in the low SINR regime, unless intelligent
our channel model reduces the degree of randomness as now precoding and signaling schemes are employed [38]. It is
H is a diagonal matrix of random Gaussian RVs as opposed important to develop analytical expressions for the MI of
to the whole matrix having random entries, which implies that MIMO channels in this regime to allow for the analysis of
we can not employ the results developed for Wishart matrices. various signaling and precoding schemes. We develop such
However fortunately for the single-bounce scattering model in an analytical expression for the 3D channel model in (10) by
(10), we can obtain results by exploiting the quadratic nature exploiting its structure. The closed-form expression for the
of the entries of HHH . For the case, when MS is equipped CDF of the MI in the low SINR regime is now presented in
with a single antenna, an exact closed-form solution can be the following theorem.
obtained and is provided in the following theorem. Theorem 2: In the low SINR regime, the approximate
Theorem 1: Let the MS be equipped with a single Rx statistical distribution of the MI can be expressed in a closed-
antenna and the BS with any arbitrary number of Tx antennas, form as,
then the statistical distribution of MI in the general SINR
N
λN
regime can be expressed in a closed-form as, 2
X
i 1 x
P[I(σ ) ≤ x] = 1 − exp − ,
N N λi λi (18)
λN (ey − 1)
1
Q
2
X
i
i=1 (λi − λl )
P[I(σ ) ≤ y] = 1 − N
exp − , l6=i
Q λi λi
i=1 (λi − λl ) where λi ’s are the eigenvalues of C given in (14).
l6=i
(13) Proof: The proof of Theorem 2 follows from making use
of the relation that for δ with ||δ||2 ≈ 0,
where λi ’s are the eigenvalues of C given that,
1 log det(X + δX) = log det(X) + T r(X−1 δX). (19)
C = (BH ΩB) ◦ (AH A)T , (14)
N In the low SINR regime, ||(R + σ 2 INM S )−1 ||2 ≈ 0 so taking
where Ω = (R + σ 2 IM S )−1 . δ as (R + σ 2 INM S )−1 HHH and X as INM S yields,
Proof: The proof follows from realizing that when the
I(σ 2 ) = T r((R + σ 2 INM S )−1 HHH ). (20)
number of antennas at the MS is one, (R + σ 2 INM S )−1 HHH
is a scalar and can be expressed as a quadratic form in α. The Observe from (10) that the trace of HHH can be expressed
MI is given as a function of this quadratic form as, as a quadratic form given by,
I(σ 2 ) = log (1 + αH Cα). (15) T r(ΩHHH ) = αH Cα, (21)
6
<(Ck,l ) −=(Ck,l )
<(α) k,l <(α)
<(αH Ck,l α) = [<(α)T =(α)T ] = [<(α)T
=(α) T
](C ) . (29)
=(Ck,l ) <(Ck,l ) =(α) < =(α)
=(Ck,l ) <(Ck,l )
<(α) k,l <(α)
=(αH Ck,l α) = [<(α)T =(α)T ] = [<(α)T
=(α)T
](C ) . (30)
−<(Ck,l ) =(Ck,l ) =(α) = =(α)
1,1 1,1
1
+ T r(C1,1 1,1 T 1 1,1 M,M
+ T r(C1,1 M,M T
4N [T r(C< (C< )) < (C< ) )] ... 4N [T r(C< (C= )) < (C= ) )]
1,2 1,1
1
4N [T r(C< (C< )) + T r(C< (C1,1
1,2 T
< ) )] ... 1
4N [T r(C 1,2
< (C M,M
= )) + T r(C< (CM,M
1,2
= )T )]
.. .. ..
. . .
M,M
1
4N [T r(C< (C1,1
< )) + T r(CM,M (C1,1 T
< ) )] ... 1
[T r(C M,M
(C M,M
)) + T r(C M,M
(C M,M T
) )]
Θ= < 4N < = < =
. (35)
1 1,1 1,1 1,1 1,1 T 1 1,1 M,M 1,1 M,M T
4N [T r(C = (C < )) + T r(C= (C< ) )] ... 4N [T r(C = (C = )) + T r(C = (C = ) )]
1,2 1,1
1
+ T r(C1,2 1,1 T 1 1,2 M,M 1,2 M,M T
4N [T r(C= (C< )) = (C< ) )] ... 4N [T r(C= (C= )) + T r(C= (C= ) )]
.. ..
..
. . .
M,M
1
4N [T r(C= (C1,1
< )) + T r(CM,M
= (C1,1 T
< ) )] ... 1 M,M
4N [T r(C= (CM,M
= )) + T r(CM,M
= (CM,M
= )T )]
containing the real and imaginary parts of the quadratic terms. B. Distribution of Mutual Information
After studying the asymptotic behaviour of <(HHH ),
<(αH C1,1 α) <(αH C1,M α)
...
H 2,1 =(HHH ) for the proposed entropy maximizing channel, the
<(α C α)
1 ... <(αH C2,M α)
HHH = .. ..
CDF of the MI is worked out. To this end, note that (12) can
N
..
. . . be equivalently expressed as,
<(αH CM,1 α) ... <(αH CM,M α)
1,1
I(σ 2 ) = log det(R + σ 2 IM + HHH ) − log det(R + σ 2 IM ).
H
=(αH C1,M α)
=(α C α) ... (36)
j =(αH C2,1 α) ... =(αH C2,M α)
+ . (31) Decomposing ((R+σ 2 IM )+HHH ) into real and imaginary
.. .. ..
N
. . .
H M,1
parts as,
=(α C α) ... =(αH CM,M α)
<((R + σ 2 IM ) + HHH ) + j =((R + σ 2 IM ) + HHH ),
Note that E[<(α)2 ] = E[=(α)2 ] = 12 . With this decom- = Y + j Z, (37)
position at hand, we now state in the following theorem the
asymptotic behavior of the joint distribution of the entries of and denoting the entries of R + σ 2 IM by ζij , we can extend
<(HHH ) and =(HHH ) stacked in vector x given by, the result of Theorem 3 to the entries of <((R + σ 2 IM ) +
HHH ) and =((R + σ 2 IM ) + HHH ), which will also show the
x = [vec(<(HHH ))T vec(=(HHH ))T ]T , (32) convergence behavior in (33) under assumptions A-1 and A-2
with mean vector m given by (38) and the same covariance.
where the operator vec(.) maps an M ×M matrix to an M 2 ×1 The mean matrix M for ((R + σ 2 IM ) + HHH ) is expressed
vector by stacking the rows of the matrix. as,
Theorem 3: Let H be a NM S × NBS MIMO channel matrix
from (10), then under the setting of assumptions A-1 and A-2,
x behaves as multivariate Gaussian such that the MGF of x M = M1 + j M2 , (39)
has the following convergence behavior,
where,
" √ T !#
N s (x − m) 1
<(ζ11 ) + 2N1 1,1
T r(C< ) ... <(ζ1M ) + 1 1,M
E exp √ − exp →
− 0, (33) 2N T r(C< )
sT Θs 2 <(ζ21 ) + 1 T r(C 2,1 ) . . . <(ζ2M ) + 1 2,M
2N < 2N T r(C< )
M1 = ,
.. .. ..
where, . . .
1 M,1 1 M,M
<(ζM 1 ) + 2N T r(C< ) . . . <(ζM M ) + 2N (C< )
1
m= [T r(C1,1 M,M
< ) . . . T r(C< ), T r(C1,1 M,M T
= ) . . . T r(C= )] , (40)
2N 1,1 1,M
(34)
1 1
=(ζ11 ) + 2N T r(C= ) ... =(ζ1M ) + 2N T r(C= )
=(ζ21 ) + 1 2,1 1 2,M
2N T r(C= ) ... =(ζ2M ) + 2N T r(C= )
and Θ is given by (35). M2 =
.. .. .. .
The proof of Theorem 3 is postponed to Appendix A. Note . . .
1 M,1 1 M,M
that this theorem can be used in applications involving the =(ζM 1 ) + 2N T r(C= ) ... =(ζM M ) + 2N (C= )
statistical distribution of HHH other than the MI analysis. (41)
8
1 1 1
m = [<(ζ11 ) + T r(C1,1
< ), <(ζ12 ) + T r(C1,2
< ), . . . , <(ζM M ) + T r(CM,M
< ),
2N 2N 2N
1 1 1
=(ζ11 ) + T r(C1,1
= ), =(ζ12 ) + T r(C1,2
= ), . . . , =(ζM M ) + T r(CM,M
= )]T (38)
2N 2N 2N
Cell Edge
Before presenting the Gaussian approximation to the statis-
tical distribution of the MI in the asymptotic limit, we define
H1 Hint
a matrix M e and a 2M 2 × 1 vector f(M)e as,
e = M1 −M2
M , (42)
M2 M1
BS 1 MS of interest BS 2
ISD=500 m
e l+(k−1)M = [det(D1 ) + · · · + det(D2M )]|l+(k−1)M ,
f(M)|
Fig. 3. Example scenario.
k = 1, . . . , 2M, l = 1, . . . , M, (43)
CDF
0.5
the following assumption,
0.4
Assumption A-3. In the large NBS , N regime presented in
0.3
A-1,
0.2
e T Θ f(M)
lim inf f(M) e > 0, (44) 0.1
0
where Θ is the covariance matrix from (35). This assumption 0 0.5 1 1.5 2 2.5
Mutual information (bps/Hz)
3 3.5 4
1
1 Monte Carlo − θ tilt=90o
Monte Carlo − SNR=−7 dB 0.9
0.9 Theoretical − SNR=−7 dB Theoretical − θtilt=90o
Monte Carlo − SNR=−3 dB 0.8
0.8 Monte Carlo − θ tilt=96o
Theoretical − SNR=−3 dB
0.7 Monte Carlo − SNR=1 dB
0.7 Theoretical − θtilt=96o
Theoretical − SNR=1 dB 0.6 Monte Carlo − θ tilt=99o
0.6 Monte Carlo − SNR=5 dB
CDF
Theoretical − SNR=5 dB 0.5 Theoretical − θtilt=99o
CDF
0.5
Monte Carlo − θ tilt=102o
0.4
0.4
Theoretical − θtilt=102o
0.3
0.3
0.2
0.2
0.1
0.1
0
0 0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1 1.2 1.4
Mutual information (bps/Hz)
Mutual information (bps/Hz)
Fig. 5. Comparison of Monte Carlo simulated CDF and theoretical CDF Fig. 6. Comparison in a multi-cell environment for different values of θtilt .
derived in (18) for the low SINR regime.
1 1
o
Monte Carlo, SNR=−10 dB Monte Carlo − θ tilt=90
0.9 Theoretical, SNR=−10 dB 0.9
Theoretical − θtilt=90o
Monte Carlo, SNR=−3.75 dB
0.8 Theoretical, SNR=−3.75 dB 0.8 Monte Carlo − θ tilt=96o
Monte Carlo, SNR=2.5 dB
0.7 0.7 Theoretical − θtilt=96o
Theoretical, SNR=2.5 dB
Monte Carlo, SNR=8.75 dB Monte Carlo − θ tilt=99o
0.6 0.6
Theoretical, SNR=8.75 dB
Theoretical − θtilt=99o
CDF
CDF
Monte Carlo, SNR=15 dB
0.5 0.5
Theoretical, SNR=15 dB Monte Carlo − θ tilt=102o
0.4 0.4 Theoretical − θtilt=102o
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15
Fig. 7. Comparison of Monte Carlo simulated CDF and asymptotic Fig. 8. Comparison of Monte Carlo simulated CDF and asymptotic
theoretical CDF in (45). theoretical CDF in a mutli-cell environment.
LoS angle with the MS, i.e. θtilt ≈ 96o . We plot the CDF for currently being devoted to produce accurate 3D channel mod-
the exact expression of MI for a single Rx antenna provided els that will be of fundamental practical use in analyzing
in (13) for SN R = 5dB. The perfect fit of the theoretical these elevation beamforming techniques. In this work, we
result to the simulated result for the CDF of the MI is again proposed a 3D channel model that is consistent with the
evident. state of available knowledge on channel parameters, and is
We now validate the result obtained for the CDF of MI in inspired from system level stochastic channel models that
the asymptotic limit, as NBS and N grow large, for the cases have appeared in standards like 3GPP, ITU and WINNER.
with finite values of NBS and N . The first result deals with The principle of maximum entropy was used to determine
scenario when the BS 2 is operating in a different frequency the distribution of the MIMO channel matrix based on the
band as BS 1. Fig. 7 shows that there is quite a good fit prior information on angles of departure and arrival. The
between the asymptotic theoretical result and the simulated resulting 3D channel model is shown to have a systematic
result for finite sized simulated system with NBS = 60, a structure that can be exploited to effectively derive the closed-
reasonable number for large MIMO systems and NM S = 4. form expressions for the CDF of the MI. Specifically, we
The number of paths is taken to be of the same order as NBS . characterized the CDF in the general and low SINR regimes.
Even at the highest simulated value of SN R, the asymptotic An exact closed-form expression for the general SINR regime
mean and asymptotic variance show only a 1.2 and 3.2 percent was derived for systems with a single receive antenna. An
relative error respectively. This close match illustrates the asymptotic analysis, in the number of paths and BS antennas,
usefulness of the asymptotic approach in characterizing the of the achievable MI is also provided and validated for finite-
MI for 3D maximum entropy channels for any number of sized systems via numerical results. The derived analytical
Rx antennas. Hence, as far as performance analysis based expressions were shown to closely coincide with the simulated
on the achievable MI is considered, the asymptotic statistical results for the proposed 3D channel model. An important
distribution obtained can be used for the evaluation of future observation made was that the meticulous selection of the
3D beamforming strategies. Finally the multi-cell scenario downtilt angles at the transmitting BS can have a significant
is considered in Fig. 8, where we plot the CDF of the MI impact on the performance gains, confirming the potential of
achieved by the proposed entropy maximizing channel in the elevation beamforming. We believe that the results presented
asymptotic limit at different values of antenna boresight angles will enable a fair evaluation of the 3D channels being outlined
of the serving BS. We again deal with the worst case scenario in the future generation of mobile communication standards.
of the MS being in the direct boresight of the interfering
cell. Comparison with the simulated CDF of MI obtained at
A PPENDIX A
SN R = 5dB validates the accuracy of the derived theoretical
P ROOF OF T HEOREM 3
asymptotic distribution and illustrates the impact 3D elevation
beamforming can have on the system performance through the The mean and variance of every quadratic term that consti-
meticulous selection of downtilt angles. For the case on hand, tutes the real and imaginary parts of HHH , i.e. <(T k,l ) and
the MI of the system is again maximized when the serving =(T k,l ) (note that these terms are formed as a consequence
BS also sets its antenna boresight angles equal to the elevation of the decomposition of the original matrix in (27) into real
LoS angle with the MS. and imaginary parts and hence are real quadratic terms), can
be computed as given in (52) and (53) respectively [44].
VI. C ONCLUSION 1
E[<(T k,l )] = E[|<(α)|2 ]T r(Ck,l
< ), (52)
The prospect of enhancing system performance through N
elevation beamforming has stirred a growing interest among 1
E[=(T k,l )] = E[|=(α)|2 ]T r(Ck,l
= ).
researchers in wireless communications. A large effort is N
11
1
V ar[<(T k,l )] = E[|<(α)|2 ]2 [T r(Ck,l k,l k,l k,l T
< C< ) + T r(C< (C< ) )], (53)
N2
1
V ar[=(T k,l )] = 2 E[|=(α)|2 ]2 [T r(Ck,l k,l k,l k,l T
= C= ) + T r(C= (C= ) )].
N
0 0 1 0 0 0 0
Cov[</=(T k,l )</=(T k ,l )] = 2
E[|</=(α)|2 ]2 [T r(Ck,l
</= Ck</=
,l
) + T r(Ck,l
</= (Ck</=
,l T
) )]. (54)
N
1
x= [<(αH C1,1 α) . . . <(αH CM,M α), =(αH C1,1 α) . . . =(αH CM,M α)]T ,
N
1
m= [T r(C1,1 M,M
< ) . . . T r(C< ), T r(C1,1
= ) . . . T r(C=
M,M T
)] . (56)
2N
T
1 T T
<(α) T
E exp g [x − m] = E exp <(α) =(α) Ξ −g m . (57)
N =(α)
1,1 1,1
1
+ T r(C1,1 1,1 T 1 1,1 M,M 1,1
(CM,M )T )]
4N [T r(C< (C< )) < (C< ) )] ... 4N [T r(C< (C= )) + T r(C< =
1,2 1,1
1
4N [T r(C< (C< )) + T r(C1,2 1,1 T
< (C< ) )] ... 1
4N [T r(C 1,2
< (C M,M
= )) 1,2 M,M T
+ T r(C< (C= ) )]
.. .. ..
. . .
M,M
1
4N [T r(C< (C1,1
< )) + T r(CM,M (C<1,1 T
) )] . . . 1
[T r(C M,M
(C M,M
)) + T r(C M,M
(C M,M T
) )]
Θ=
< 4N < = < =
(60)
1 1,1 1,1 1,1 1,1 T 1 1,1 M,M 1,1 M,M T
4N [T r(C = (C < )) + T r(C= (C< ) )] ... 4N [T r(C = (C = )) + T r(C = (C = ) )]
1,2 1,1
1
+ T r(C1,2 1,1 T 1 1,2 M,M 1,2 M,M T
4N [T r(C = (C < )) = (C < ) )] ... 4N [T r(C = (C = )) + T r(C = (C = ) )]
.. ..
..
. . .
M,M 1,1 M,M 1,1 T M,M
1
4N [T r(C = (C < )) + T r(C = (C < ) )] . . . 1
4N [T r(C= (CM,M
= )) + T r(CM,M
= (CM,M
= )T )]
Similarly, the covariance between two entries of <(HHH ) in x is asymptotically Normal. To characterize the joint distri-
and =(HHH ) is given by (54). The proof for covariance is bution of the entries of <(HHH ), =(HHH ), the behaviour of
straightforward and can be done by realizing that if X = their linear combination, i.e. gT [x−m], where g is any arbitrary
αH Aα and Y = αH Bα where α and A and B are real then, vector, is studied. The MGF of this linear combination is given
N
X by (57), where Ξ = [g1 C1,1 M,M
< + ... + gM 2 C< + gM2 +1 C1,1
= +
E[XY ] = E[ αi αj αk αl Aij Bkl ] M,M <(α)
... + g2M 2 C= ]. Note that <(α)T =(α)T Ξ
is
i,j,k,l=1 =(α)
N
X N
X also a quadratic form in Normal RVs (<(α), =(α)). Extending
= E[|αi |4 ]Aii Bii + E[|αi |2 ]2 Aii Bkk the convergence result in (55) for a quadratic form in Normal
i=1 i=j,k=l,i6=k RVs with s = 1, E[|<(α)|2 ] = E[|=(α)|2 ] = 21 , we have
N
X N
X
√ T
+ E[|αi |2 ]2 Aij Bij + E[|αi |2 ]2 Aij Bji N g [x − m] − exp 1 →
E exp q − 0, (58)
i=k,j=l,i6=j i=l,j=k,i6=j 1
[T r(ΞΞ) + T r(Ξ(Ξ)H )] 2
4N
and the cumulant (µ4 − 3µ22 )
for Gaussian RV’s is 0.
Given the expressions for mean and covariances of quadratic where,
forms in Gaussian RVs, the Central Limit Theorem for the 1
[T r(ΞΞ) + T r(Ξ(Ξ)H )] = gT Θg, (59)
(k, l)th entry of <(HHH ) or =(HHH ) can be established from 4N
the result in [45] under assumptions A-1 and A-2. Noting that and Θ is √ the 2M 2 × 2M 2 covariance matrix of the entries
E[|<(α)|2 ] = E[|=(α)|2 ] = 12 , this convergence implies that of vector N x given by (60). The entries of Θ are obtained
the MGF of <(T k,l ) has the following behavior, using (54). Therefore,
√
√
1 k,l
s N [<(T k,l ) − 2N
2 " !#
T r(C< )] s N (gT x − gT m)
1
E exp q
− exp →
− 0. E exp − exp →
− 0. (61)
k,l k,l k,l k,l T
2
p
[T r(C< C< )+T r(C< (C< ) )] gT Θg 2
4N
(55) Since gT x behaves as a Gaussian RV with mean gT m
k,l
The MGF of =(T ) has the same behavior with < replaced and variance given by N1 gT Θg, so it can be deduced that
with =. x behaves as multivariate Gaussian distribution with mean m
We stack the entries of <(HHH ), =(HHH ) and their means and covariance N1 Θ. Therefore under assumptions
√ A-1 and A-
in x and m respectively given by (56). From [45], each term 2, the joint distribution of the entries of N ([vec(<(HHH ))T ,
12
1 1 1 1
xn = [<(ζ11 ) + <(αH C1,1 α) . . . <(ζM M ) + <(αH CM,M α), =(ζ11 ) + =(αH C1,1 α) . . . =(ζM M ) + =(αH CM,M α)]T
N N N N
(62)
M1 −M2
5f1 = [det(D1 ) + · · · + det(D2M )]|l+(k−1)M , (68)
M2 M1
l+(k−1)M
2
.5 T M1 −M2 M1 −M2
× 5 f1 Θ 5 f1 , (70)
M1 −M2 M2 M1 M2 M1
det
M2 M1
vec(=(HHH ))T ]T −m) behaves approximately as multivariate We use the following to determine the expression for 5f1 ,
Gaussian with mean 0 and covariance matrix given by Θ. This
Y -Z
completes the proof of Theorem 3. (det(Y + jZ))2 = det . (67)
Z Y
A PPENDIX B th
Therefore the (l + (k − 1)M ) , k = 1, . . . 2M, l =
P ROOF OF T HEOREM 4
5f1 is given
1, . . . M entry of the vector by (68) [Ch 6, [46]],
To complete the characterization of MI, the log det trans- M1 −M2
formation needs to be applied on the entries of Y and Z, where Di is identical to except that the entries
M2 M1
i.e. <((R + σ 2 IM ) + HHH ) and =((R + σ 2 IM ) + HHH ) in the ith row are replaced
by their derivatives with respect to
respectively. These entries can be stacked in vector xn as th M1
(k, l) entry of , k = 1, . . . , 2M, l = 1, . . . , M . Every
shown in (62). M2
To obtain the statistical distribution of MI from the joint M1 −M2
entry of 5f1 will involve the sum of only
distribution of the entries of xn , we succumb to the Taylor M2 M1
series expansion of a real-valued function f (xn ), two non-zero determinants as every (k, l)th entry occurs in
only two rows and,
f (xn ) = f (m) + 5T f (m)(xn − m) + Rn , (63)
.5
where m is the mean vector in (38) and Rn = o(||xn − m||2 ). f20 (det(M1 + jM2 )2 ) = .
2 p M 1 −M2 (69)
We observe that o(||xn − m|| ) is o(1) because xn − m −
→ 0. det
M2 M1
Any mapping of xn − m will therefore be o(1) which implies
that Rn = o(1). Therefore from (65), in the asymptotic limit, the distribution
√
of N log det((R + σ 2 IM ) +HHH )
h i
E[exp (f (xn ) − f (m))] = E exp 5T f (m)(xn − m) + o(1).
1 M1 −M2
(64) − 2 log det can be approximated
M2 M1
Note that E[exp 5T f (m)(xn − m) ] is analogous to by a Gaussian distribution with mean 0 and
variance given by (70), under the assumption that
E exp gT [xn − m] with g = 5f (m). Now using the T M1 −M2 M1 −M2
convergence result in (58) with (59) and invoking Slutsky’s lim inf 5 f1 M2 M1
Θ 5 f1
M2 M1
>
Theorem, given that 0. Using this result together with (36), we can establish the
lim inf 5T f (m)Θ 5 f (m) > 0 results in, statistical distribution of MI in the asymptotic limit for 3D
√
MIMO channels. This completes the proof of Theorem 4.
N [f (xn ) − f (m)] − exp 1
E exp q →
− 0.
2
5T f (m)Θ 5 f (m) R EFERENCES
(65) [1] E. Telatar, “Capacity of multi-antenna Gaussian channels,” European
This is analogous to applying Delta Theorem on the entries Trans. Telecommun., vol. 10, no. 6, pp. 585–595, Nov./Dec. 1999.
[2] G. J. Foschini, “Layered space-time architecture for wireless commu-
of Y+jZ. √ Therefore for f (Y+jZ) = log det(Y+jZ), we can nication in a fading environment when using multi-element antennas,”
say that N (log det(Y + jZ) − log det(M1 + jM2 )) behaves Bell Labs Technical Journal, vol. 1, no. 2, pp. 41–59, 1996.
as a Gaussian RV under assumptions A-1 and A-2, with mean [3] E. Larsson, O. Edfors, F. Tufvesson, and T. Marzetta, “Massive MIMO
for next generation wireless systems,” IEEE Communications Magazine,
0, and variance given by (5T f (M)Θ 5 f (M)). vol. 52, no. 2, pp. 186–195, Feb. 2014.
Define a function h(xn ) = f2 ◦ f1 (xn ), where f1 (xn ) = [4] A. Kuchar, J.-P. Rossi, and E. Bonek, “Directional macro-cell channel
(det(Y + jZ))2 and f2 (x) = 0.5 log x. For this function the characterization from urban measurements,” IEEE Transactions on An-
tennas and Propagation, vol. 48, no. 2, pp. 137–146, Feb. 2000.
gradient is, [5] J. Fuhl, J.-P. Rossi, and E. Bonek, “High-resolution 3-d direction-of-
arrival determination for urban mobile radio,” IEEE Transactions on
5h|xn =m = f20 (det(M1 + jM2 )2 ) × 5f1 |Y=M1 ,Z=M2 . (66) Antennas and Propagation, vol. 45, no. 4, pp. 672–682, Apr. 1997.
13
[6] J. Koppenborg, H. Halbauer, S. Saur, and C. Hoek, “3D beamforming [30] Report ITU-R M.2135, “Guidelines for evaluation of radio interface
trials with an active antenna array,” in Proc. ITG Workshop Smart technologies for IMT-advanced,” [Online]. Available: http://www.itu.int/
Antennas, pp. 110-114, 2012. pub/R-REP-M.2135-2008/en, 2008.
[7] Q.-U.-A. Nadeem, A. Kammoun, M. Debbah, and M.-S. Alouini, “A [31] IST-4-027756 WINNER II, “D1.1.2, WINNER II Channel
generalized spatial correlation model for 3D MIMO channels based Models,” [Online]. Available: https://www.ist-winner.org/
on the fourier coefficients of power spectrums,” Accepted in IEEE WINNER2-Deliverables/D1.1.2v1.1.pdf, Sept. 2007.
Transactions on Signal Processing, 2015. [32] R1-122034, “Study on 3D channel Model for elevation beamforming
[8] 3GPP TR 36.873 V12.0.0 , “Study on 3D channel model for LTE,” Sep. and FD-MIMO studies for LTE,” 3GPP TSG RAN no. 58, Dec. 2012.
2014. [33] A. Kammoun, H. Khanfir, Z. Altman, M. Debbah, and M. Kamoun,
[9] Y. Zhu, P. Y. Kam, and Y. Xin, “On the mutual information distribution “Preliminary results on 3D channel modeling: From theory to standard-
of MIMO Rician fading channels,” IEEE Transactions on Communica- ization,” IEEE Journal on Selected Areas in Communications, vol. 32,
tions, vol. 57, no. 5, pp. 1453–1462, May 2009. no. 6, pp. 1219–1229, June. 2014.
[10] G. Alfano, A. Lozano, A. M. Tulino, and S. Verdu, “Mutual information [34] P. Almers, E. Bonek, A. Burr, N. Czink, M. Debbah, V. Degli
and eigenvalue distribution of MIMO Ricean channels,” in International Esposti, H. Hofstetter, P. Kyösti, D. Laurenson, G. Matz, A. F.
Symposium on Information Theory and its Applications (ISITA) 2004, Molisch, C. Oestges, and H. Özcelik, “Survey of channel and radio
Oct 2004. propagation models for wireless MIMO systems,” EURASIP Journal
[11] J. Salo, F. Mikas, and P. Vainikainen, “An upper bound on the ergodic on Wireless Communications and Networking Volume 2007, url =
mutual information in Rician fading MIMO channels,” IEEE Transac- http://www.eurecom.fr/publication/2162, 02 2007.
tions on Wireless Communications, vol. 5, no. 6, pp. 1415–1421, June [35] U. Fernández-Plazaola, E. Martos-Naya, J. F. Paris, and J. T.
2006. Entrambasaguas, “Comments on Proakis analysis of the characteristic
function of complex Gaussian quadratic forms,” CoRR, vol.
[12] P. Smith and L. Garth, “Exact capacity distribution for dual MIMO
abs/1212.0382, 2012. [Online]. Available: http://arxiv.org/abs/1212.0382
systems in Ricean fading,” IEEE Communications Letters, vol. 8, no. 1,
[36] G. L. Turin, “The characteristic function of hermitian quadratic forms
pp. 18–20, Jan 2004.
in complex normal variables,” Biometrika, vol. 47, no. 1/2, pp. pp.
[13] P. Smith, L. Garth, and S. Loyka, “Exact capacity distributions for 199–201, 1960. [Online]. Available: http://www.jstor.org/stable/2332977
MIMO systems with small numbers of antennas,” IEEE Communications [37] M. Bibinger, “Notes on the sum and maximum of independent
Letters, vol. 7, no. 10, pp. 481–483, Oct 2003. exponentially distributed random variables with different scale
[14] M. McKay and I. Collings, “General capacity bounds for spatially parameters,” Humboldt-Universität zu Berlin, Tech. Rep., 2013.
correlated Rician MIMO channels,” IEEE Transactions on Information [Online]. Available: \url{http://arxiv.org/abs/1307.3945}
Theory, vol. 51, no. 9, pp. 3121–3145, Sept 2005. [38] X. Wu and R. Srikant, “MIMO channels in the low-SNR Regime:
[15] P. Smith and M. Shafi, “On a Gaussian approximation to the capacity communication rate, error exponent, and signal peakiness,” IEEE Trans-
of wireless MIMO systems,” in IEEE International Conference on actions on Information Theory, vol. 53, no. 4, pp. 1290–1309, April
Communications, ICC 2002., vol. 1, 2002, pp. 406–410. 2007.
[16] Z. Wang and G. Giannakis, “Outage mutual information of space-time [39] H. Yin, D. Gesbert, M. Filippou, and Y. Liu, “A coordinated approach
MIMO channels,” IEEE Transactions on Information Theory, vol. 50, to channel estimation in large-scale multiple-antenna systems,” Selected
no. 4, pp. 657–662, April 2004. Areas in Communications, IEEE Journal on, vol. 31, no. 2, pp. 264–273,
[17] M. Chiani, M. Win, and A. Zanella, “On the capacity of spatially February 2013.
correlated MIMO Rayleigh-fading channels,” IEEE Transactions on [40] J. Zhang, C. Pan, F. Pei, G. Liu, and X. Cheng, “Three-dimensional
Information Theory, vol. 49, no. 10, pp. 2363–2371, Oct 2003. fading channel models: A survey of elevation angle research,” IEEE
[18] M. Kang and M.-S. Alouini, “Capacity of MIMO Rician channels,” IEEE Communications Magazine, vol. 52, no. 6, pp. 218–226, June. 2014.
Transactions on Wireless Communications, vol. 5, no. 1, pp. 112–122, [41] K. Kalliola, K. Sulonen, H. Laitinen, O. Kivekas, J. Krogerus, and
Jan 2006. P. Vainikainen, “Angular power distribution and mean effective gain of
[19] W. Hachem, O. Khorunzhiy, P. Loubaton, J. Najim, and L. Pastur, “A mobile antenna in different propagation environments,” IEEE Transac-
new approach for mutual information analysis of large dimensional tions on Vehicular Technology, vol. 51, no. 5, pp. 823–838, Sep. 2002.
multi-antenna channels,” IEEE Transactions on Information Theory, [42] W. J. L. Queiroz, F. Madeiro, W. T. A. Lopes, and M. S. Alencar,
vol. 54, no. 9, pp. 3987–4004, 2008. “Spatial correlation for DoA characterization using Von Mises, Cosine,
[20] E. T. Jaynes, “Information theory and statistical mechanics, part 1,” Phys. and Gaussian distributions,” International Journal of Antennas and
Rev., vol. 106, pp. 620–630, 1957. Propagation, vol. 2011, July. 2011.
[21] ——, “Information theory and statistical mechanics, part 2,” Phys. Rev., [43] J. Ren and R. Vaughan, “Spaced antenna design in directional scenarios
vol. 108, pp. 171–190, 1957. using the von mises distribution,” in Vehicular Technology Conference
[22] J. Shore and R. Johnson, “Axiomatic derivation of the principle of Fall (VTC 2009-Fall), Sept. 2009, pp. 1–5.
maximum entropy and the principle of minimum cross-entropy,” IEEE [44] W. Hachem, M. Kharouf, J. Najim, and J. Silverstein, “A CLT for
Transactions on Information Theory, vol. 26, no. 1, pp. 26–37, Jan 1980. information-theoretic statistics of non-centered gram random matrices,”
Random Matrices: Theory and Applications, 01 (02), April 2012.
[23] M. Debbah and R. R. Müller, “MIMO channel modeling and the
[45] R. Bhansali, L. Giraitis, and P. Kokoszka, “Convergence of quadratic
principle of maximum entropy,” IEEE Transactions on Information
forms with nonvanishing diagonal,” Statistics & Probability Letters,
Theory, vol. 51, no. 5, pp. 1667–1690, May 2005.
vol. 77, no. 7, pp. 726–734, April 2007. [Online]. Available:
[24] J. Meinilä, P. Kyösti et al., “D5.3: WINNER+ Final Channel Models http://ideas.repec.org/a/eee/stapro/v77y2007i7p726-734.html
V1.0,” [Online.] Available: http://projects.celtic-initiative.org/winner+ [46] C. D. Meyer, Matrix analysis and applied linear algebra. Philadelphia,
/WINNER+%20Deliverables/D5.3 v1.0.pdf, June. 2010. PA, USA: Society for Industrial and Applied Mathematics, 2000.
[25] F. Athley and M. Johansson, “Impact of electrical and mechanical
antenna tilt on LTE downlink system performance,” in IEEE 71st
Vehicular Technology Conference (VTC 2010-Spring), May 2010, pp.
1–5.
[26] B. Partov, D. Leith, and R. Razavi, “Utility fair optimization of antenna
tilt angles in LTE networks,” IEEE/ACM Transactions on Networking,,
vol. PP, no. 99, pp. 1–1, Jan 2014. Qurrat-Ul-Ain Nadeem Qurrat-Ul-Ain Nadeem
[27] O. N. C. Yilmaz, S. Hämäläinen, and J. Hämäläinen, “Analysis of was born in Lahore, Pakistan. She received her B.S.
antenna parameter optimization space for 3GPP LTE,” in IEEE 70th degree in electrical engineering from the Lahore
Vehicular Technology Conference Fall (VTC 2009-Fall), Sept 2009, pp. University of Management Sciences (LUMS), Pak-
1–5. istan in 2013. She joined King Abdullah Univer-
[28] A. Adhikary, J. Nam, J.-Y. Ahn, and G. Caire, “Joint spatial division sity of Science and Technology (KAUST), Thuwal,
and multiplexing - the large-scale array regime,” IEEE Transactions on Makkah Province, Saudi Arabia in August 2013,
Information Theory, vol. 59, no. 10, pp. 6441–6463, Oct 2013. where she is currently a M.S./Ph.D. student. Her
[29] “Spatial Channel Model for Multiple Input Multiple Output (MIMO) research interests include channel modeling and
Simulations,” [Online]. Available: http://www.3gpp.org/ftp/Specs/ performance analysis of wireless communications
html-info/25996.htm, Sep. 2003. systems.
14
Abla Kammoun Abla Kammoun was born in Sfax, Tunisia. She received
the engineering degree in signal and systems from the Tunisia Polytechnic
School, La Marsa, and the Master’s degree and the Ph.D. degree in digital
communications from Telecom Paris Tech [then Ecole Nationale Supérieure
des Télécommunications (ENST)]. From June 2010 to April 2012, she
has been a Postdoctoral Researcher in the TSI Department, Telecom Paris
Tech. Then she has been at Supélec at the Large Networks and Systems
Group (LANEAS) until December 2013. Currently, she is research scientist
at KAUST university. Her research interests include performance analysis,
random matrix theory, and semi-blind channel estimation.