MIMO Channel Hardening: A Physical Model based

Analysis
Matthieu Roy, Stéphane Paquelet, Luc Le Magoarou, Matthieu Crussière

To cite this version:

Matthieu Roy, Stéphane Paquelet, Luc Le Magoarou, Matthieu Crussière. MIMO Channel Hardening:
A Physical Model based Analysis. 2018. �hal-01770014�

HAL Id: hal-01770014

https://hal.archives-ouvertes.fr/hal-01770014
Preprint submitted on 18 Apr 2018

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 MIMO Channel Hardening:
A Physical Model based Analysis
Matthieu Roy† , Stéphane Paquelet† , Luc Le Magoarou† , Matthieu Crussière‡
† <>
b com Rennes, France
‡
Univ Rennes, INSA Rennes, IETR - UMR 6164 F-35000 Rennes, France

Abstract—In a multiple-input-multiple-output (MIMO) based on the asymptotic behavior of the channel gain for large
communication system, the multipath fading is averaged over antenna arrays. This definition was applied to pinhole channels,
radio links. This well-known channel hardening phenomenon i.i.d. correlated and uncorrelated Rayleigh fading models [7].
plays a central role in the design of massive MIMO systems. The
aim of this paper is to study channel hardening using a physical Contributions. Complementary to this pioneer work, we
channel model in which the influences of propagation rays and propose a non-asymptotic analysis of channel hardening, as
antenna array topologies are highlighted. A measure of channel well as new derivations of the coefficient of variation of the
hardening is derived through the coefficient of variation of the channel not limited to classically assumed Rayleigh fading
channel gain. Our analyses and closed form results based on the
used physical model are consistent with those of the literature models. Indeed, channel hardening is analyzed herein using
relying on more abstract Rayleigh fading models, but offer a physically motivated ray-based channel model widely used
further insights on the relationship with channel characteristics. in wave propagation. Our approach is consistent with previous
Index Terms—channel hardening, physical model, MIMO studies [4], [6], but gives deeper insights on channel hardening.
In particular, we managed to provide an expression of the
I. I NTRODUCTION channel hardening measure in which the contributions of
Over the last decades, multi-antenna techniques have been the transmit and receive antenna arrays, and the propagation
identified as key technologies to improve the throughput and conditions can easily be identified, and thus interpreted.
reliability of future communication systems. They offer a
Notations. Upper case and lower case bold symbols are used
potential massive improvement of spectral efficiency over
for matrices and vectors. z ∗ denotes the conjugate of z. ~u
classical SISO (single-input-single-output) systems proportionally
stands for a three-dimensional (3D) vector. h.,.i and ~a ·~u denote
to the number of involved antennas. This promising gain has
the inner product between two vectors of CN and 3D vectors,
been quantified in terms of capacity in the seminal work of
respectively. [H]p,q is the element of matrix H at row p and
Telatar [1] and has recently been even more emphasized with
column q. kHkF , khk and khkp stand for the Frobenius norm,
the newly introduced massive MIMO paradigm [2].
the euclidean norm and the p-norm, respectively. HH and HT
Moving from SISO to MIMO, the reliability of communication denotes the conjugate transpose and the transpose matrices.
systems improves tremendously. On the one hand in SISO, H̄ denote the normalized matrix H/||H||F . E{.} and Var{.}
the signal is emitted from one single antenna and captured denote the expectation and variance.
at the receive antenna as a sum of constructive or destructive
echoes. This results in fading effects leading to a potentially very
unstable signal to noise ratio (SNR) depending on the richness II. C HANNEL M ODEL
of the scattering environment. On the other hand in a MIMO We consider a narrowband MIMO system (interpretable as
system, with appropriate precoding, small-scale multipath fading an OFDM subcarrier) with Nt antennas at the transmitter and
is averaged over the multiple transmit and receive antennas. This Nr antennas at the receiver, such that
yields a strong reduction of the received power fluctuations, hence y = Hx+n,
the channel gain becomes locally deterministic essentially driven with x ∈ CNt ×1 , y ∈ CNr ×1 and n ∈ CNr ×1 the vectors of
by its large-scale properties. This effect, sometimes referred to as transmit, receive and noise samples, respectively. H ∈ CNr ×Nt
channel hardening [3] has recently been given a formal definition is the MIMO channel matrix, whose entries [H]i,j are the
based on the channel power fluctuations [4]. Indeed, studies on complex gains of the SISO links between transmit antenna j
the stability of the SNR are essential to the practical design and receive antenna i. The capacity of the MIMO channel can
of MIMO systems, in particular on scheduling, rate feedback, be expressed as [1]
channel coding and modulation dimensioning [2], [3], [5]. From
the definition in [4], we propose in this paper a comprehensive
study on channel hardening through a statistical analysis C = log2 (det(INt +ρQ̄H̄H H̄)) bps/Hz, (1)
of received power variations derived from the propagation
Pt
characteristics of a generic ray-based spatial channel model. where ρ = N 0
||H||2F with Q̄ ∈ CNt ×Nt , Pt and N0 the input
Related work. Channel hardening, measured as the channel gain correlation matrix (precoding), emitted power and noise power.
variance, has recently been studied from several points of view. C is a monotonic function of the optimal received SNR ρ [8],
The authors in [6] used data from measurement campaigns and hence kHk2F directly influences the capacity of the MIMO
extracted the variance of the received power. A rigorous definition channel. It is then of high interest studying the spatial channel
of channel hardening was then given in the seminal work [4] gain variations to predict the stability of the capacity.
 In the sequel, we will consider that the channel matrix H is
obtained from the following generic multi-path 3D ray-based 0.6
model considering planar wavefronts [9], [10], [7, p. 485]
P
p X 0.5
H(f ) = Nt Nr cp er (~urx,p )et (~utx,p )H . (2)
p=1 P=2
0.4 P=4

CV 2
Such channel consists of a sum of P physical paths where cp is P=6
the complex gain of path p and ~utx,p (resp. ~urx,p ) its direction P=8
of departure - DoD - (resp. of arrival - DoA -). In (2) et and er 0.3
are the so-called steering vectors associated to the transmit and
receive arrays. They contain the path differences of the plane 0.2
wave from one antenna to another and are defined as [10]
0.1
 T
1 ~
a ·~
u
2jπ tx,1λ tx,p 2jπ
atx,N ·~
~
t
utx,p
2 4 6 8 10 12 14 16
et (~utx,p ) = √ e ,···,e λ , (3) Number of antennas N (Rx and Tx)
Nt
and similarly for er (~urx,p ). The steering vectors depend not Fig. 1. Simulated CV 2 for growing number of rays. Asymptotes are the black
only on the DoD/DoA of the impinging rays, but also on the dashed lines.
topology of the antenna arrays. The latter are defined by the
physical channel model that leads to much more interpretable
sets of vectors Atx = {~atx,j } and Arx = {~arx,j } representing
results.
the positions of the antenna elements in each array given an
Assumptions on the channel model. The multipath channel
arbitrary reference.
model described in Section II relies on several parameters
Such channel model has already been widely used (especially governed by some statistical laws. Our aim is to provide an
in its 2D version) [9], [10], verified through measurements analytical analysis of CV while relying on the weakest possible
[11] for millimeter waves and studied in the context of channel set of assumptions on the channel model. Hence, we will consider
estimation [12]. In contrast to Rayleigh channels, it explicitly that:
takes into account the propagation conditions and the topology
• For each ray, gain, DoD and DoA are independent.
of the antenna arrays.
• arg(cp ) ∼ U[0,2π] i.i.d.
In the perspective of the following sections, let c = • ~utx,p and ~urx,p are i.i.d. with distributions Dtx and Drx .
T
[|c1 |,···,|cP |] denote the vector consisting of the amplitudes
of the rays. kck2 is the aggregated power from all rays, corre- The first hypothesis is widely used and simply says that no
sponding to large-scale fading due to path-loss and shadowing. formal relation exists between the gain and the DoD/DoA
of each ray. The second one raisonnably indicates that each
propagated path experiences independent phase rotation without
III. C HANNEL H ARDENING any predominant angle. The last one assumes that all the rays
come from independent directions, with the same distribution
Definition. Due to the multipath behavior of propagation (distributions Dtx at the emitter, Drx at the receiver).
channels, classical SISO systems suffer from a strong fast fading It has indeed been observed through several measurement
phenomenon at the scale of the wavelength resulting in strong campaigns that rays can be grouped into clusters [13], [14].
capacity fluctuations (1). MIMO systems average the fading Considering the limited angular resolution of finite-size antenna
phenomenon over the antennas so that the channel gain varies arrays, it is possible to approximate all rays of the same cluster
much more slowly. This effect is called channel hardening. In as a unique ray without harming a lot the channel description
this paper, the relative variation of the channel gain kHk2F , accuracy [12]. It then makes sense to assume that this last
called coefficient of variation (CV ) is evaluated to quantify the hypothesis is valid for the main DoDs and DoAs of the clusters.
channel hardening effect as previously introduced in [4], [7]: Simulations. A preliminary assessment of the coefficient of
   2
2 Var kHk2F E kHk4F −E kHk2F variation is computed through Monte-Carlo simulations of (4)
CV = 2 = 2 (4) using uniform linear arrays (ULA) with inter-antenna spacing
E{kHkF }2 E{kHkF } 2
λ
In (4) the statistical means are obtained upon the model which of 2 at both the transmitter and receiver and taking a growing
govern the entries of kHk2 given random positions of the number of antennas. A total of P ∈ {2, 4, 5, 6} paths were
transmitter and the receiver. This measure was previously applied randomly generated with Complex Gaussian gains cp ∼ CN (0,1),
to a Nt ×1 correlated Rayleigh channel model h ∼ CN (0,R) uniform DoDs ~utx,p ∼ US2 and DoAs ~urx,p ∼ US2 .
[4], [7, p. 231]. In that Simulation results of CV are reported in Fig. 1 as a function
 Hparticular
2
case, (4) becomes
2 of the number of antennas. It is observed that all curves seem
|h h| −Tr(R) 2
E Tr(R )
CV 2 = = , (5) to reach an asymptote around 1/P for large Nt and Nr . Hence,
Tr(R)2 Tr(R)2
the higher the number of physical paths, the harder the channel.
where the rightmost equality comes from the properties of
The goal of the next sections is to provide further interpretation
Gaussian vectors [7, Lemma B.14]. This result only depends on
of such phenomenon by means of analytical derivations.
the covariance matrix R, from which the influences of antenna
array topology and propagation conditions are not explicitly
IV. D ERIVATION OF CV 2
identified. Moreover, small-scale and large-scale phenomena are
not easily separated either. In this paper, (4) is studied using a In this section, CV 2 is analytically analyzed from (4).
Expectation of the channel gain. From (2) and (3) the channel
9 ULA, Nt=16
gain kHk2F = Tr(HH H) can beX written as UCA, Nt=16
8 UPA, Nt=16
2
kHkF = Nt Nr c∗p cp0 γp,p0 ,
7 asymptote
p,p0
where the term γp,p0 is given by 6

) 2
γp,p0 = her (~urx,p ),er (~urx,p0 )ihet (~utx,p ),et (~utx,p0 )i∗ .

Nt 2( tx,
5
Using the hypothesis arg(cp ) ∼ U[0,2π] i.i.d. introduced in the
channel model and γp,p = 1, the expectation of the channel gain 4
can further be expressed as 3
E kHk2F = Nt Nr E kck2 .
 
(6) 2
Thus the average channel gain increases linearly with Nr and
1
Nt , which is consistent with the expected beamforming gain Nt
and the fact that the received power linearly depends on Nr . 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00
d/
Coefficient of variation. The coefficient of variation CV is
derived Fig. 2. Numerical evaluation of E(Atx , US2 ) for various array types and
( using the previous hypotheses and (6). We introduce:
increasing antenna spacing ∆d. The values are normalized so the asymptote is 1.
E 2 (Atx ,Dtx ) = E |het (~utx,p ),et ∗ (~utx,p0 )i|2

 (7) B. Small-scale fading
E 2 (Arx ,Drx ) = E |her (~urx,p ),er (~urx,p0 )i|2 .
These quantities are the second moments of the inner products The coefficient of variation particularized with the statistical
of the transmit/receive steering vectors associated to two distinct conditional model can easily be proven to be:
2 2 2 2
CVkck 2 = E (Atx ,Dtx )E (Arx ,Drx )α (c)
rays. They represent the correlation between two rays as observed 
by the system. They can also be interpreted as the average Ec|kck2 kck44 (9)
inability of the antenna arrays to discriminate two rays given where α2 (c) = 1− 4
.
kck
a specific topology and ray distribution. From such definitions, The small-scale fading contribution to CV 2 thus consists of
and based on the derivations given in Appendix A, CV 2 can be a product of the quantities defined in (7) that depend only on
expressed as a sum of two terms,  the antenna array topologies (Atx /Arx ) and ray distributions
E kck4 −kck44 (Dtx /Drx ) multiplied by a propagation conditions factor α2 (c)
CV 2 =E 2 (Atx ,Dtx )E 2 (Arx ,Drx ) 2
E{kck2 } that depends only on the statistics of the ray powers c.
 (8) Ray correlations. This paragraph focuses on the quantity
Var kck2
+ 2 . E 2 (Atx , Dtx ) (the study is done only at the emitter, the obtained
E{kck2 } results being equally validat the receiver). Eq. (7) yields
Note that this result only relies on the assumptions introduced N
t
2 
in section II. The second term can be identified as the 1 X atx,i ·(~
~ utx,p −~
u ) 
tx,p0 
E 2 (Atx , Dtx ) = 2 E  e2jπ  .
 λ
contribution of the spatial large-scale phenomena since it simply Nt   
i=1
consists in the coefficient of variation of the previously defined
A well-known situation is when the inner sum involves exponen-
large-scale fading parameter kck2 of the channel. To allow
tials of independent uniformly distributed phases and hence corre-
local channel behavior interpretation, conditioning the statistical
sponds to a random walk with Nt steps of unit length. The above
model by kck2 is required. It results in the cancellation of the
expectation then consists in the second moment of a Rayleigh dis-
large-scale variations contribution of CV 2 which reduces to
tribution and E 2 (Atx , Dtx ) = N1t . A necessary condition to such
what is called hereafter small-scale fading.
a case is to have (at least) a half wavelength antenna spacing ∆d
V. I NTERPRETATIONS to ensure that phases are spread over [0,2π]. On the other hand,
A. Large-scale fading phase independences are expected to occur for asymptotically
2
The contribution of large-scale fading in CV is basically large ∆d. It is however shown hereafter that such assumption
the coefficient of variation of the total aggregated power kck2 turns out to be valid for much more raisonnable value of ∆d.
of the rays. To better emphasize its behavior, let us consider a Numerical evaluations of E 2 are performed versus ∆d (Fig. 2),
simple example with independent |cp |2 of mean µ and variance and versus N (Fig. 3). Uniformly distributed rays over the 3D
t
σ 2 . The resulting largescale fading term is then
2 unit sphere (D tx = Drx = US2 ) and Uniform Linear, Circular and
Var kck2
 
1 σ Planar Arrays (ULA, UCA and UPA) are considered. As a re-
2 = .
E{kck2 } P µ minder, the smaller E(Atx ,Dtx ) the better the channel hardening.
In Fig. 2, E 2 reaches the asymptote 1/Nt for all array types with
It clearly appears that more rays lead to reduced large-scale
∆d = λ2 and remains almost constant for larger ∆d. Fig. 3 shows
variations. This stems from the fact that any shadowing 2
phenomenon is well averaged over P independent rays, hence that E merely follows the 1/Nt law whatever the array type.
becoming almost deterministic in rich scattering environments. We thus conclude that the independent uniform phases situation
This result explains the floor levels obtained for various P in discussed above λis a sufficient model for any array topology
our previous simulations in Section II and is consistent with given that ∆d ≥ 2 . It is therefore assumed in the sequel that,
the literature on correlated Rayleigh fading channels where high
rank correlation matrices provide a stronger channel hardening E 2 (Atx ,US2 )≈1/Nt , E 2 (Arx ,US2 )≈1/Nr .
effect than low rank ones [7].
 assumes a rich scattering environment. Using (5) with R = I:
0.5 1/Nt 1
ULA 2
CViid = .
UCA Nt Nr
0.4 UPA Using the realistic model in a rich scattering environment, the
large-scale part of (8) vanishes leading to a deterministic kck2
0.3
)

and small-scale variations reach the upper bound of (10). This

2

yields the limit

2( tx,

P →∞
0.2 CV 2 −−−−→ CViid 2
(12)
which is coherent with the interpretation of the model.
0.1
VI. C ONCLUSION
In this paper, previous studies on channel hardening have
0.0 been extended using a physics-based model. We have separated
0 20 40 60 80 100 120 140 influences of antenna array topologies and propagation char-
Number of antennas Nt
acteristics on the channel hardening phenomenon. Large-scale
Fig. 3. Numerical evaluation of E(Atx ,US2 ) for various antenna arrays at and small-scale contributions to channel variations have been
the half wavelength. The lower, the better. evidenced. Essentially, this paper provides a general framework
to study channel hardening using accurate propagation models.
small scale To illustrate the overall behavior of channel hardening, this
0.35 large scale
large scale + small scale framework have been used with generic model parameters and
0.30 simulation hypotheses. The scaling laws evidenced for simpler channel
0.25 models are conserved provided the antennas are spaced by at
least half a wavelength. The results are consistent with state
0.20 of the art and provide further insights on the influence of array
CV 2

0.15 topology and propagation on channel hardening. The proposed

expression can easily be exploited with various propagation
0.10 environments and array topologies to provide a more precise
0.05 understanding of the phenomenon compared to classical channel
descriptions based on Rayleigh fading models.
0.00
2 4 6 8 10 12 14 16
ACKNOWLEDGMENT
Number of antennas N (Rx and Tx) This work has been performed in the framework of the Horizon 2020 project
ONE5G (ICT-760809) receiving funds from the European Union. The authors
Fig. 4. Comparison between (8) and simulated CV 2 . Uniform distribution would like to acknowledge the contributions of their colleagues in the project,
of DoDs and DoAs over the unit sphere and complex Gaussian gains.
although the views expressed in this contribution are those of the authors and
Propagation conditions. It is now interesting to point out that do not necessarily represent the project.
the propagation factor α(c) introduced in (9) is bounded by A PPENDIX A
0 ≤ α2 (c) ≤ 1−1/P. (10) C OEFFICIENT OF VARIATION (8)
Those bounds are deduced from the following inequality: For the sake of simplicity, an intermediary matrix A is
introduced. It is defined
 by
kck42 /P ≤ kck44 ≤ kck42 . (11)
2|γp,p0 |cos(φp,p0 ) if p 6= p0
The right inequality comes from the convexity of the square [A]p,p0 =
1 if p = p0
function. Equality is achieved when there is only one contributing ∗
2 with φp,p = arg(cp cp0 γp,p ) the whole channel phase
0 0
ray, i.e. no multipath occurs. In that case CVkck 2 = 0 and the

MIMO channel power is deterministic. The left part in (11) is dependence. kHk2F can be written using a quadratic form with
given by Hölder’s inequality. Equality is achieved when there vector c and matrix A, which can be decomposed into two
are P rays of equal power. Then, taking the expectation on each terms I (identity) and J
member in (11) yields (10). kHk2F
In contrast to the large-scale fading, more rays lead to more = cT Ac = cT c+cT Jc
Nt Nr
small-scale fluctuations. It is indeed well known that a richer
scattering environment increases small-scale fading. where J = A−I. E{J} = 0 so:
Comparison with the simulations. Based on the general E kHk4F
= E kck4 +E (cT Jc)2 .
 
(Nt Nr )2
formula given in (8), on the interpretations and evaluations
of its terms, we can derive the expression of channel hardening The ray independence properties yields the following weighted
for the illustrating simulations of Section II: sum of coupled ray powers
X 
1 E (cT Jc)2 = E |cp |2 |cp0 |2 E [J]2p,p0 .
 
2
CVillustration = (1−1/P )+1/P.
Nt Nr p6=p 0

Simulation and approximated formula are compared in Fig. 4


Considering i.i.d. rays, all the weights E [J]2p,p0 are
in which small-scale and large-scale contributions are easily identical. Using the weights notations introduced in (7) and
evidenced, as intuitively expected from simulations of Fig. 1. the
 definition
of the 4-norm yields the second order moment
Comparison with the Gaussian i.i.d. model. This model E kHk4F . With the expectation (6) we derive the result (8).
 R EFERENCES
[1] E. Telatar, “Capacity of Multi-antenna Gaussian Channels,”
European Transactions on Telecommunications, vol. 10, no. 6,
pp. 585–595, Nov. 1999.
[2] T. L. Marzetta, Fundamentals of massive MIMO. Cambridge
University Press, 2016.
[3] B. M. Hochwald, T. L. Marzetta, and V. Tarokh, “Multiple-
antenna channel hardening and its implications for rate feedback
and scheduling,” IEEE transactions on Information Theory,
vol. 50, no. 9, pp. 1893–1909, 2004.
[4] H. Q. Ngo and E. G. Larsson, “No downlink pilots are needed
in TDD massive MIMO,” IEEE Transactions on Wireless
Communications, vol. 16, no. 5, pp. 2921–2935, 2017.
[5] E. Björnson, E. de Carvalho, E. G. Larsson, and P. Popovski,
“Random access protocol for massive MIMO: Strongest-user
collision resolution (SUCR),” in 2016 IEEE International
Conference on Communications (ICC), May 2016, pp. 1–6.
[6] . O. Martı́nez, E. De Carvalho, and J. O. Nielsen, “Massive
MIMO properties based on measured channels: Channel
hardening, user decorrelation and channel sparsity,” in Signals,
Systems and Computers, 2016 50th Asilomar Conference on.
IEEE, 2016, pp. 1804–1808.
[7] E. Björnson, J. Hoydis, and L. Sanguinetti, “Massive MIMO
Networks: Spectral, Energy, and Hardware Efficiency,”
Foundations and Trends in Signal Processing, vol. 11, no. 3-4,
pp. 154–655, 2017.
[8] S. Loyka and G. Levin, “On physically-based normalization
of MIMO channel matrices,” IEEE Transactions on Wireless
Communications, vol. 8, no. 3, pp. 1107–1112, 2009.
[9] V. Raghavan and A. M. Sayeed, “Sublinear Capacity Scaling
Laws for Sparse MIMO Channels,” IEEE Transactions on
Information Theory, vol. 57, no. 1, pp. 345–364, Jan. 2011.
[10] T. Zwick, C. Fischer, and W. Wiesbeck, “A stochastic multipath
channel model including path directions for indoor environments,”
IEEE Journal on Selected Areas in Communications, vol. 20,
no. 6, pp. 1178–1192, Aug. 2002.
[11] M. K. Samimi, G. R. MacCartney, S. Sun, and T. S. Rappaport,
“28 GHz Millimeter-Wave Ultrawideband Small-Scale Fading
Models in Wireless Channels,” in 2016 IEEE 83rd Vehicular
Technology Conference (VTC Spring), May 2016, pp. 1–6.
[12] L. Le Magoarou and S. Paquelet, “Parametric channel estimation
for massive MIMO,” arXiv preprint arXiv:1710.08214, 2017.
[13] A. A. Saleh and R. Valenzuela, “A statistical model for indoor
multipath propagation,” IEEE Journal on selected areas in
communications, vol. 5, no. 2, pp. 128–137, 1987.
[14] X. Wu, C.-X. Wang, J. Sun, J. Huang, R. Feng, Y. Yang, and X. Ge,
“60-GHz Millimeter-Wave Channel Measurements and Modeling
for Indoor Office Environments,” IEEE Transactions on Antennas
and Propagation, vol. 65, no. 4, pp. 1912–1924, Apr. 2017.