Performance Assessment of MIMO Precoding on

Realistic mmWave Channels

∗ Mattia Rebato, † Luca Rose, ∗ Michele Zorzi
∗
Department of Information Engineering, University of Padova, 35131 Padova, Italy
† Nokia Bell Labs, Paris – Saclay, France
{rebatoma, zorzi}@dei.unipd.it – [email protected]

Abstract—In this paper, the performance of multi-user gNB in order to perform beamforming of information toward
Multiple-Input Multiple-Output (MIMO) systems is evaluated in the served UEs.
terms of SINR and capacity. We focus on the case of a downlink
arXiv:1903.11330v1 [cs.IT] 27 Mar 2019

Many works in the literature focused on the evaluation of

single-cell scenario where different precoders have been studied.
Among the considered precoders, we range from different Grid precoding techniques for MU mmWave systems with massive
of Beams (GoB) optimization approaches to linear precoders MIMO. The closest works to ours are [4]–[7]. In [4], massive
(e.g., matched filtering and zero forcing). This performance MIMO was proposed and studied under the ideal condition
evaluation includes imperfect channel estimation, and is carried of almost infinite antennas. In [5], precoding techniques such
out over two realistic mmWave 5G propagation channels, which as Minimum Mean Square Error (MMSE), Matched Filtering
are simulated following either the measurement campaign done
by New York University (NYU) or the 3GPP channel model. (MF) and Zero Forcing (ZF) were studied under the assump-
Our evaluation allows grasping knowledge on the precoding tion of a Rayleigh channel model and under the condition of
performance in mmWave realistic scenarios. The results highlight perfect Channel State Information (CSI) acquisition. In [6],
the good performance of GoB optimization approaches when a channel estimation errors were introduced to estimate the
realistic channel model with directionality is adopted. implementation loss in terms of precoding gain, whereas in
Index Terms—Millimeter-wave, multi-user MIMO, 5G, inter-
ference optimization, linear precoder, grid of beams.
[7] the authors link the precoding performance with channel
correlation. Finally, a recent piece of work [8] uses a realistic
I. I NTRODUCTION channel model to perform an evaluation of a MU system
in terms of bit error rate as a function of the number of
The volume of mobile data is continuously increasing, antenna elements used at the transmitter side, while however
especially with high capacity applications that are emerging overlooking the effect of different precoding strategies and
together with the next generation (i.e., 5G) of cellular commu- channel estimation errors.
nications [1]. As an enabler for these capacity-intensive appli- From the literature, it emerges that linear precoding schemes
cations, the millimeter wave (mmWave) band (approximately can be used to reach high performance under ideal assump-
between 10 and 300 GHz) has been identified as a promising tions. Less known is however their performance when re-
candidate for future mobile communications [2]. In addition alistic channel models are considered. To be precise, under
to the use of mmWave frequencies, another major aspect of a Rayleigh fading model, it is known that MMSE performs
the new mobile generation is the densification of the network appreciably better in terms of balancing the resources among
applying small cells in large numbers. Furthermore, Multi- the UEs acting as a trade-off between MF and ZF approaches.
User (MU) massive Multiple-Input Multiple-Output (MIMO) However, the Rayleigh fading model oversimplifies the chan-
systems became of high interest as they contribute to reaching nel characterization, resulting in a channel model that does not
the 5G high demands (e.g., in terms of rates and densities), due reflect the real mmWave propagation specifics.
to their ability to greatly increase network capacity [3]. For In 3GPP NR systems, the exploitation of mmWave fre-
this reason, it is important to study and evaluate MU massive quency bands (both at 28 GHz and at 60 GHz) for the next
MIMO systems over 5G mmWave propagation channels. By generation of mobile communications is currently defined [9].
exploiting such technologies, data transmission rates are ex- Within the standard, different types of CSI feedback mecha-
pected to increase in the Radio Access Network (RAN), and nisms have been included to support MIMO transmissions. In
a more efficient use of the radio spectrum can be achieved. particular, release 15 includes Type-I and Type-II codebook
The purpose of MU MIMO systems is to account for CSI feedback, enabling different trade-offs between CSI res-
channel scattering and reflections, thus exploiting the spatial olutions and feedback overhead [10]. More precisely, when a
dimension and creating multiple beams of the signal in the Type-I CSI feedback scheme is adopted, the UE feeds back the
direction of the User Equipments (UEs), so that each user index of a vector taken from a suitable oversampled DFT code-
can benefit from the whole allowed bandwidth at any time book that best approximates the dominant eigenvector of the
instant. This can be achieved by precoding the information at channel matrix; conversely, when Type-II CSI is adopted, the
the Next Generation Node Base (gNB) side. Using a precoder, feedback is composed of a linear combination of two or more
data is distributed on the different antenna elements of the (up to 4 per polarization) vectors taken from the oversampled
 angles of departure by C, with CN ×1 and CN ×M being the generalizations to
Ω("
Ω""
vectors and matrices, respectively. The M ×M identity matrix
gNB , -th
- clusters
is written as IM and the zero matrix of size NT ×M is denoted
array +( subpaths
NLoS cluster

…
as 0NT ×M . Finally, we generally indicate with X
b the Frobenius
LoS ray normalized matrix of X.
)*
antenna
elements II. S YSTEM M ODEL
UE1
We consider a narrowband single-cell downlink multi-user
MIMO mmWave system where a single gNB sector with
UE1 NT transmit antennas is serving M single-antenna UEs.2
ℎ"# The channel to the m-th user is assumed narrowband and is
ℎ$# described by the vector of coefficients hm ∈ CNT ×1 , and its
UE2
gNB j-th element describes the channel response between the j-
ℎ%# th transmitting antenna element and the receive antenna. This
UE3 input-output relationship can be described as
ℎ&#
UE4 ym = hH
m x + nm , m ∈ {1, 2, . . . , M } (1)
where x is the NT ×1 transmitted vector signal, ym ∈ C is the
Figure 1: Illustration of the mmWave system model considered (bottom) and received signal, and nm is the noise term. Assuming PM to use
representation of the channel model used for each link in the framework (top). a precoder, the transmitted vector signal is x = i=1 wi si ,
where si is the data symbol and wi is the NT × 1 linear
DFT codebook. In this latter case, both the indices of the
precoding vector.
chosen vectors and the linear combination coefficients are fed
Aggregating together the precoding vectors of all the
back to gNB. Finally, it is worth observing that the accuracy of
M UEs we can define the precoding matrix W =
a Type-II CSI feedback scheme is larger, and so is the resulting T ×M
[w1 , . . . , wM ] ∈ CN . We
 note that, in order to respect the
overhead [11]. The reason behind such mechanisms is to be 
power constraint E kWsk2 = 1, we normalize the precoding
found in the attempt to reduce the amount of CSI acquisition
matrix with the Frobenius norm as follows W c = W . Using
overhead while exploiting MIMO advantages, such as spatial kWkF
multiplexing and beamforming. Although at the moment full this notation, it is possible to write the system input-output
CSI1 is not included in the standard, ongoing discussions are equation as
attempting to assess the trade-off between precoding gain and y = HH Ws
c +n (2)
overhead cost. where y, s and n are vectors with dimension M × 1, while
Differently from the prior art, the objective of this study channel matrix H is defined in CNT ×M .
is twofold. First, we aim at evaluating the performance of Finally, we define H̄(p) as the M × M equivalent matrix
diverse precoders when a realistic channel is considered, obtained with the product
where “realistic” denotes both the adoption of a channel
model supported by experimental evidence and the inclusion H̄(p) = HH W
c (p) (3)
of CSI imperfections. Second, we compare the aforementioned where superscript p is used to identify the different precoding
linear precoders against Grid of Beams (GoB) optimization approaches evaluated as described in the following.
approaches, with the goal of assessing the gain of linear
precoders overs simpler (and less demanding in term of CSI) A. Channel Models
GoB approaches. In our evaluation, MIMO channel vectors h are generated
As reported in Figure 1, we consider a scenario with both a according to three distinct statistical channel models. The first
realistic sectorization and an antenna array radiation pattern, as model under analysis is a standard Rayleigh fading channel
suggested by the 3GPP specifications in [12]. Moreover, two model; the second is derived from a set of extensive measure-
measurement-based realistic channel models are considered, ment campaigns in New York City by NYU–Wireless [13]; the
one from New York University (NYU) [13] and one from last model considered is the one provided by the 3GPP [14],
3GPP as suggested in [14], both used to evaluate and compare which was obtained from multiple measurement campaigns
the performance of different precoders. from different research groups all around the world. For this
Notation: In this paper, column vectors and matrices study, we adopt the channel model with the settings of the
are respectively denoted by boldface lowercase and uppercase Urban Macro (UMa) scenario.
letters. We identify with XH the conjugate transpose of X, and Both the realistic models (i.e., NYU and 3GPP) are based
the Frobenius norm is denoted k · kF while the Euclidean norm on the WINNER II channel characterization [15], and consider
is denoted as k · k. The set of all complex numbers is denoted
2 We note that the number of UEs that can be simultaneously supported by
1 According to the 3GPP terminology, the term full CSI is known as explicit the gNB sector is less than or equal to the number of antenna elements, i.e.,
CSI. M ≤ NT .
macro-level scattering paths and sub-paths. Some minor differ- pointing towards a direction. According to this principle, the
ences are present in the settings of the models, nevertheless, entire codebook spans the whole effective area.3
a major difference is identified in the number of paths and Two different GoB metrics and thus optimization criteria
sub-paths considered. The NYU characterization has higher are considered in this study. First, for each active UE, we
(GoB )
directionality obtained by assuming a smaller number of paths. identify the precoder wm P which maximizes the received
To be precise, the NYU model considered a maximum of power among all possible precoder vectors z in the codebook
4 main paths (defined as clusters), while the 3GPP channel Z, thus
(GoBP )
model can reach a maximum of 20 clusters in Non Line of wm = arg max |hH 2
m wz | . (5)
z∈Z
Sight (NLoS) conditions. Here, by directionality, we mean
the ability of the channel and beamforming, to focus the We identify it with the acronym GoBP , and the respective
power in a specific direction. Together with the channel, precoding matrix is derived as
also a measurement-based distance-dependent path loss model h
(GoB ) (GoB )
i
is considered with Line of Sight (LoS), NLoS and outage W(GoBP ) = w1 P , . . . , wM P . (6)
conditions. b) Grid of beams (SLNR optimization): Similarly, we
At the transmitter side, we model the antenna of the consid- study an alternative in which the precoder is chosen by
ered sector as a Uniform Planar Array (UPA). In this manner, maximizing the Signal to Leakage plus Noise Ratio (SLNR)
the beamforming can be performed in both the azimuth and for each single UE m. We define it as GoBSLNR and the
elevation dimensions. Furthermore, we precisely model each optimization expression becomes
element radiation pattern following the 3GPP specifications !
in [12] and [14], as done in our previous work [16]. We (GoBSLNR )
|hH
m,m wz |
2
wm = arg max (7)
σ 2 + i6=m |hH
P 2
consider the superposition of element radiation pattern and z∈Z m,i wz |
array factor in order to gather a precise knowledge of the
array radiation effects due to beamforming. This permits a then, the precoder matrix W(GoBSLNR ) is derived as in (6).
careful characterization of the steering beams, and therefore The rationale behind this choice is that the sum of SLNRs
a precise knowledge of the amount of power irradiated by is a close approximation of the sum of Signal to Interference
the antenna array in all directions. Thus, we are realistically plus Noise Ratios (SINRs), with the advantage of being
computing both the desired and the interfering signals. A computationally much easier to perform. This stems mainly
complete explanation of the relationship between array and from the fact that whereas the sum SINR maximization would
element patterns can be found in [12] and [16]. require an exhaustive search for all possible beams and all
The channel of each link is computed with a set of K users in the cell, the sum SLNR can be maximized by simply
clusters and Lk sub-paths per cluster (as shown in the top maximizing the SLNR of each UE.
part of Figure 1), and is represented as c) Matched filter precoder [17]: The MF, also known as
conjugate beamforming, maximizes the power of the received
Lk
K X
X signal, without any interference consideration. It is optimum
h= gkl FT (Ωkl ) uT (Ωkl ) (4) when the noise power received by the UE is much stronger
k=1 l=1
than the interference that would result from the transmitted
where gkl is the small-scale fading gain of sub-path l in signals intended to be received by the co-scheduled UEs. For
cluster k, and uTX is the 3D spatial signature vector of this reason, it is optimum for noise-limited scenarios.4 Its
the transmitter. Furthermore, for brevity, we use subscript or precoding matrix is expressed as
superscript T, referring to a transmitter related term. Moreover,
Ωkl = (θkl , φkl ) is the angular spread of vertical and horizon- W(MF) = H.
b (8)
tal Angles of Departure (AoD) for sub-path l in cluster k [13]. The gNB computes the precoding matrix after estimating the
Finally, FT is the field factor term of the transmitting array. channel so as to direct the useful energy in the direction of
Detailed explanation on how to precisely compute all these each UE. In our evaluation, we assume complete knowledge
channel terms can be found in [13] and [16]. of the channel and we use this assumption for the calculation
B. Precoders Considered of this and the next precoders.
d) Zero-forcing precoder: An evolution of MF linear
With the intent to perform a study of the different precoding processing can be used to limit the detrimental effects of
techniques while realistically modeling the channel, we discuss multi-user interference. The ZF precoder tries to cancel the
in the following paragraphs all the approaches evaluated and power of the interference, and therefore is an optimal solution
provide details on how they are computed. for interference-limited scenarios. This interference canceling
a) Grid of beams (power optimization): This approach
3 This principle is an assumption adopted for this evaluation. Different
consists in the use of a codebook Z of precomputed precoders
that will be tested with the aim to choose the one that codebook designs can also be applied in our optimization.
4 Noise-limited and interference-limited scenario refer, respectively, to the
maximizes a specific metric. Each precoder vector in the case in which the noise power is greater than the interference power and
codebook represents a Discrete Fourier Transform (DFT) beam vice-versa.
 1 1 1
GoB P GoB P GoB P
GoB SLNR GoB SLNR GoB SLNR
0.8 MF 0.8 MF 0.8 MF
ZF ZF ZF
MMSE MMSE MMSE
0.6 0.6 0.6
CDF

CDF

CDF
0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
-40 -20 0 20 40 -40 -20 0 20 40 -40 -20 0 20 40
SINR [dB] SINR [dB] SINR [dB]

(a) Rayleigh fading channel. (b) NYU channel model. (c) 3GPP channel model.
Figure 2: Empirical CDF of the SINR for the different precoders evaluated. In these figures, we used a fixed number of UEs M = 4 and PT = 30 dBm.
Table I: List of parameters used in our evaluation. Unless specified otherwise,
property is obtained at the price of a slightly complex precoder these settings are considered as default in all the studies carried out during
computation and of a reduced received power. The precoding our evaluation.
matrix is designed according to the ZF criterion [18] and is
Value Meaning and (Notation)
given by
W(ZF) = H( b H b −1
b H H) (9) 28 GHz carrier frequency (f )
7 dB noise figure (NF)
which simply denotes the right pseudo-inverse of the matrix 100 m transmitter receiver distance
b H. 4 # of served UEs (M )
H 64 total # of antennas per gNB (NT )
e) MMSE precoder: Differently from the last two pre- [8 × 8] vertical and horizontal UPA configuration
coders considered, the MMSE precoding strategy (also known λ[0.7, 0.5] vertical and horizontal UPA element spacing
as Kalman filter precoder) maximizes the sum of the SINR. 30 dBm transmitted power (PT )
6 # bits phase shifters
Therefore, it optimizes the received power while minimizing 0.99 imperfect channel metric (τ )5
the interference signal. It can be considered as a solution in
between the MF and the ZF precoders. The precoding matrix c e(p) ,
b HW
consideration, the equivalent matrix becomes H̄e = H
is expressed as (p)
where the precoder W c e has been calculated considering the
 −1 imperfect channel He .
W(MMSE) = H b H b HHb + 1 IM (10)
SNR III. C OMPARISON RESULTS
and it is possible to prove that it can be expressed as a linear In this section we provide some simulation results to com-
combination of MF and ZF precoders [19]. pare the performance of the different precoders considered,
which will be assessed in terms of SINR and achievable
C. Imperfect channel estimate system capacity. Before examining in detail all the results, we
Focusing on a realistic system, achieving a complete and briefly report here the SINR expression used in our evaluation.
correct knowledge of the CSI is not feasible in a practical Furthermore, Table I details all the parameters and respective
framework. To be precise, typical mmWave implementation values adopted.
does not have direct access to the signals received on each The first metric considered in our evaluation is the SINR,
gNB antenna, so learning the channel on each antenna element we calculate it for each UE m as follows
is currently extremely difficult and almost infeasible. For this (p)
(p) |h̄m,m |2
reason, we consider the performance in case the transmitter has SINRm = (p)
(12)
1
P
an imperfect channel estimate. The channel estimation error SNR + i6=m |h̄m,i |2
is modeled following a Gauss-Markov formulation, where the where SNR is computed using the transmitted power, the path
imperfect channel He is obtained using the true channel H as −1
loss `, and the thermal noise σ 2 as PTσ`2 . We note that
follows p each UE’s SINR is affected by the accurate antenna array
He = τ H + 1 − τ 2 E (11) radiation pattern that is computed considering the field factor
where each term of the matrix E follows a circularly sym- term into the channel gains, as previously described in (4).
metric Normal distribution CN (0, 1). Moreover, the scalar Finally, superscript (p) is used to identify the M × 1 vector
parameter τ ∈ [0, 1] is used to indicate the quality of the of the equivalent matrix H̄(p) obtained with the corresponding
precoder Wc (p) . We recall that the precoding matrix is included
channel estimation, where τ = 1 corresponds to perfect
estimation of the channel whereas τ = 0 corresponds to having into the equivalent matrix as done in (3).
only the random channel E [20]. This parameter depends on 5 The value τ = 0.99 identifies an optimistic channel imperfection. As
factors such as the time/power spent on pilot-based channel discussed later in the results, even with a small error in the CSI the degradation
estimation. As done in (3), and with the imperfect channel is notable.
 8 8 1 1
GoB P
7 GoB SLNR 7
MF 0.8 0.8
6 ZF 6
MMSE
5 5 0.6 0.6

CDF
4 4
0.4 GoB
0.4 P
3 3 GoB SLNR
MF
2 2
0.2 ZF
0.2
MMSE
1 1

0 0 0 0
0 20 40 0 20 40 -60 -40 -20 0 20 40 -20 0 20
PT [dBm] SINR [dB]

Figure 3: System capacity for the different precoders, varying the transmitted Figure 4: Empirical CDF of the SINR for the different precoders evaluated
power PT . In this figure, M = 4, NF = 7 dB, and the channel is modeled when an imperfect channel is considered. In this figure, M = 4, NF = 7 dB
as NYU in the left plot and following the 3GPP characterization in the right and τ = 0.99, and the channel is modeled as NYU in the left plot and
plot. following the 3GPP characterization in the right plot.

The first result is a comparison of the SINR values for the power without a preferred direction, realistic mmWave
all the precoders studied under different channel model as- channels present limited multi-paths and rays with a large
sumptions. In order to have a comprehensive view of the portion of the power concentrated in few directions. To further
overall performance, Figure 2 reports the empirical Cumulative support such interpretation, Figure 2b reports the SINR values
Distribution Function (CDF) of the SINR for all the config- under NYU and 3GPP UMa channel characterizations. We
urations considered. The results have been collected over a note that, since the NYU model has fewer clusters than its
sufficient number of repetitions in order to obtain the desired UMa counterpart, the power is concentrated in fewer directions
accuracy, thus precisely evaluating the different precoders. We and thus the GoB approaches have even higher performance.
compare both the NYU and 3GPP channel characterizations With the use of the SINR expression in (12), we can
with a random Rayleigh channel model computed as HR ∼ compute the channel capacity as follows
√1 CN (0N ×M , 1N ×M ). The figures are obtained with a  
M T T (p)
Monte Carlo approach which generates random samples of Cm = log2 1 + SINR(p)m . (13)
channel and environment for all the UEs in each iteration. This metric can be used to evaluate the spectral efficiency of
As expected, the MMSE precoder outperforms all the other each configuration, and we indicate its average value by C̄.
configurations for most of the UEs. Furthermore, due to the Figure 3 plots the average system capacity in the different
high directivity of the NYU channel, both GoB precoders are configurations as a function of the transmit power PT used
able to reach higher SINR values, with respect to the MF, at the gNB side. The growing transmit power increases at the
for more than forty percent of the UEs. We note that for this same rate the received power and the interference levels for the
plot we have used a 7 dB noise figure6 , which corresponds interference-blind precoders (i.e., MF, GoB) hence resulting
to a mostly interference-limited system [21]. Due to lack of in a saturation of the performance. Conversely, interference-
space, we are not reporting here any results with higher noise aware precoders such as ZF and MMSE can have indefinitely
power. However, larger noise values push the system into a growing performance. The figures display the good perfor-
noise-limited regime, hence reducing the performance of the mance of the GoB precoder when the SLNR is optimized.
ZF precoder. We remark that, even if the average MF SINR is higher with
Figure 2 also highlights the lack of fairness among UEs respect to the other configurations, it presents poor fairness as
when the MF precoder is used. Even if the average value of previously discussed.
MF’s SINR is the highest, really high values of SINR are
Comparing the two realistic models, we can notice also in
obtained only for a small percentage of UEs. As we can see
these figures how the directivity of the NYU model results in
in Figure 2a, the precoders which do not require knowledge of
good outcomes for the capacity in the two GoB approaches
the channel (e.g., GoBP and GoBSLNR ) are unable to operate
in the range of values around 20 dBm of transmitted power.
efficiently when the channel is Rayleigh. Conversely, they
Contrary to expectation, in this particular range, GoB proce-
show SINR values close to those of MF when the channel is
dures can perform appreciably better than ZF if high spectral
modeled following the NYU characterization. This stems from
efficiency is desired, while, if a more energy-efficient operating
the fact that, while a Rayleigh channel is isotropic and scatters
point is chosen, the performance gap narrows, and eventually
6 The noise figure term quantifies the degradation of the Signal to Noise MF outperforms all the other configurations.
Ratio (SNR) due to the noise present in the system. As the last result, Figure 4 reports the empirical CDF of the
Table II: Evaluation of the gaps in the 50th percentile of the SINR expressed Finally, we have studied how the performance behaves when
in dB for the different precoders considered in this evaluation. Table obtained
with a fixed number of UEs M = 4, and PT = 30 dBm.
considering an error in the CSI acquisition. Results show
that, even with a small CSI imprecision, the performance gap
perfect CSI imperfect CSI between linear precoding and GoB vanishes.
(τ = 1) (τ = 0.99) The study of more refined MMSE approaches which include
NYU 3GPP NYU 3GPP per-user power balancing, as well as a theoretical evaluation
MF – GoBSLNR +0.87 −0.80 −0.46 −1.03 of the performance of the various techniques, also for different
ZF – GoBSLNR +1.40 +6.33 +0.89 +5.67 frequency bands, are left for future study.
MMSE – GoBSLNR +4.05 +6.45 +1.36 +5.67
R EFERENCES
SINR when an imperfect channel He is considered. Compar- [1] Cisco, “Cisco Visual Networking Index: Global Mobile Data Traffic
Forecast Update, 20162021,” White Paper, March 2017.
ing it with Figure 2, is it possible to notice how the imperfec- [2] S. Rangan, T. S. Rappaport, and E. Erkip, “Millimeter-Wave Cellular
tion in the CSI results in a degradation of the performance, Wireless Networks: Potentials and Challenges,” Proceedings of the
especially for the linear precoders. More importantly, when an IEEE, vol. 102, no. 3, pp. 366–385, March 2014.
[3] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive
error in the channel estimation is considered, the gap between MIMO for next generation wireless systems,” IEEE Communications
GoB optimization approaches and linear precoders is strongly Magazine, vol. 52, no. 2, pp. 186–195, February 2014.
reduced and, in most cases, GoB is even able to outperform the [4] H. Yang and T. L. Marzetta, “Performance of Conjugate and Zero-
Forcing Beamforming in Large-Scale Antenna Systems,” IEEE J. Sel.
linear precoders. We recall that gathering the CSI necessary Areas Commun., vol. 31, no. 2, pp. 172–179, February 2013.
to use MF, ZF and the MMSE precoders has a cost for the [5] J. Hoydis, S. ten Brink, and M. Debbah, “Massive MIMO in the UL/DL
system that should be properly considered. Furthermore, given of Cellular Networks: How Many Antennas Do We Need?” IEEE J. Sel.
Areas Commun., vol. 31, no. 2, pp. 160–171, February 2013.
the small implementation loss of GoB precoding with respect [6] D. L. Colon, F. H. Gregorio, and J. Cousseau, “Linear precoding
to more refined systems, and considering the high level of in multi-user massive MIMO systems with imperfect channel state
complexity that gathering the necessary CSI would require, it information,” in XVI Workshop on Information Processing and Control
(RPIC), Oct 2015, pp. 1–6.
seems that the additional complexity may not be justified by [7] X. Gao, O. Edfors, F. Rusek, and F. Tufvesson, “Linear Pre-Coding Per-
the modest (or even vanishing) performance improvement. formance in Measured Very-Large MIMO Channels,” in IEEE Vehicular
Remarks: We report in this subsection the main remarks Technology Conference (VTC Fall), Sept 2011, pp. 1–5.
[8] M. E. Hassan, A. E. Falou, and C. Langlais, “Performance assessment
raised in our evaluation study. Due to the directionality of of linear precoding for multi-user massive MIMO systems on a realistic
mmWave channels, our results support GoB approaches as a 5G mmWave channel,” in IEEE Middle East and North Africa Commu-
good trade-off between CSI acquisition complexity and perfor- nications Conference (MENACOMM), April 2018.
[9] 3GPP, “TS 38.104 Base Station radio transmission and reception,” 2018.
mance. Given the limited advantage (about +4 dB with MMSE [10] ——, “TS 38.211, Physical Channels and Modulation,” 2018.
and the NYU channel model), there is no strong motivation [11] ——, “TS 38.214, Physical layer procedures for data,” 2018.
to use linear precoders in multi-user systems at mmWave [12] ——, “TS 37.840 Group Radio Access Network; Study of Radio
Frequency and Electromagnetic Compatibility requirements for Active
frequencies. Table II summarizes our findings, reporting the Antenna Array System base station,” 2013.
SINR gaps for the 50th percentile in the different approaches [13] M. Akdeniz, Y. Liu, M. Samimi, S. Sun, S. Rangan, T. Rappaport, and
considered. Although linear precoders can exploit a larger E. Erkip, “Millimeter Wave Channel Modeling and Cellular Capacity
Evaluation,” IEEE J. Sel. Areas Commun, vol. 32, no. 6, pp. 1164–1179,
amount of information on the channel matrix, requiring full June 2014.
CSI at the transmitter, the gain under imperfect CSI can be [14] 3GPP, “TS 38.900 Group Radio Access Network; Study on channel
assessed as less than +1.36 dB with MMSE and the NYU model for frequency spectrum above 6 GHz,” 2016.
[15] P. Kyosti and et al., “WINNER II channel model,” Technical Report
channel model. If the inaccuracy of the channel estimation IST-WINNER D1.1.2 ver 1.1, Sept. 2007.
grows, it is possible to conjecture that the gap would close [16] M. Rebato, L. Resteghini, C. Mazzucco, and M. Zorzi, “Study of
even more, eventually eliding any advantage. A similar trend, Realistic Antenna Patterns in 5G mmWave Cellular Scenarios,” in IEEE
Int. Conf. Commun., Reliability and Modeling Symposium, Kansas City,
though with slightly higher gains, can be observed when the USA, May 2018.
3GPP channel model is considered. [17] M. Joham, W. Utschick, and J. A. Nossek, “Linear transmit processing in
MIMO communications systems,” IEEE Trans. Signal Process., vol. 53,
IV. C ONCLUSION AND FUTURE WORKS no. 8, pp. 2700–2712, Aug 2005.
[18] C. B. Peel, B. M. Hochwald, and A. L. Swindlehurst, “A
In this study, we have highlighted the impact of realistic vector-perturbation technique for near-capacity multiantenna multiuser
mmWave channel behaviors on the downlink mmWave MU communication-part I: channel inversion and regularization,” IEEE
Trans. Commun., vol. 53, no. 1, pp. 195–202, Jan 2005.
MIMO system when different precoders are considered at [19] L. Rose and M. Maso, “Receiver-Centric Inter-Cell Interference Cancel-
the gNB side. Our study led to the following observations. lation in D2D-Assisted Networks,” in IEEE Globecom Workshops, Dec
Under ideal condition (i.e., Rayleigh channel model, perfect 2016.
[20] A. Müller, A. Kammoun, E. Björnson, and M. Debbah, “Efficient linear
CSI), linear precoders largely outperform GoB due to their precoding for massive MIMO systems using truncated polynomial ex-
ability to perfectly adapt to the channel realization. However, pansion,” in IEEE 8th Sensor Array and Multichannel Signal Processing
the directionality present in realistic channel models reduces Workshop (SAM), June 2014, pp. 273–276.
[21] M. Rebato, M. Mezzavilla, S. Rangan, F. Boccardi, and M. Zorzi,
the gap, sometimes even letting the GoB approaches surpass “Understanding Noise and Interference Regimes in 5G Millimeter-Wave
more complex solutions and in most cases not justifying the Cellular Networks,” in 22th European Wireless Conference, May 2016.
additional complexity.