Channel Measurements for Large Antenna Arrays

Jakob Hoydis, Cornelis Hoek, Thorsten Wild, and Stephan ten Brink
Bell Labs, Alcatel-Lucent, Lorenzstraße 10, 70435 Stuttgart, Germany
{firstname.lastname}@alcatel-lucent.com

Abstract—Equipping base stations (BSs) with very large an-

tenna arrays is a promising way to increase the spectral and 5m
energy efficiency of mobile communication systems without the 15
need for new cell sites. However, the prominently theoretical 10
works on this topic are based on several crucial assumptions 5
about the wireless channel which have not been sufficiently 1
validated by measurements. In this paper, we report on an
outdoor measurement campaign with a scalable virtual antenna
array consisting of up to 112 elements. The large amount of
acquired data allows us to study several important aspects of
large-scale MIMO systems. For example, we partially confirm the
theoretical results based on uncorrelated channels which predict 30
that the channels at different positions become more and more 25
orthogonal as the number of antennas grows. However, for the 20
measured channels, the marginal gain of an additional antenna 16
quickly diminishes. Nevertheless, our results indicate that most
of the theoretical benefits of large-scale MIMO could be realized Fig. 1. Overview of the measurement area on the Alcatel-Lucent campus.
also over the measured channels. The transmit antenna array is loctated on top of the large building on the left.
Measurements were taken at 30 different positions indicated by the red dots.
I. I NTRODUCTION
The use of very large antenna arrays at the base stations
(BSs) has been recently advocated as an effective means to large-scale fading and a varying angular power spectrum along
significantly increase the capacity of cellular communication the antenna elements and argue that the wide-sense stationary
systems while possibly reducing their energy consumption [1], assumption for such channels does not hold. Also first large-
[2], [3]. The beauty of this approach is its simplicity: no new scale MIMO prototypes appear. In [7], a scalable hardware
cell sites are required, no technical changes need to be made at architecture based on an FPGA platform is described and some
the terminals, and simple linear precoders are close to optimal. measurement results as well as implementation aspects are
However, these advantages of large-scale or “massive” MIMO discussed. The last references indicate that the research on
systems come only into play under favorable propagation large-scale MIMO has left the purely theoretical stage. It is
characteristics of the radio channel [4]. In short, “‘favorable” now important to validate if the theoretical performance gains
means here that the channel vectors between different user can be realized in practice.
terminals (UTs) become orthogonal as number of BS-antennas In this paper, we describe a recently developed test-bed
grows. It is unclear though, if this is the case for real channels for outdoor channel measurements with very large antenna
in urban environments. arrays and discuss some of our first measurement results. In
The literature on channel measurements with large antenna particular, we study to which extent the performance predic-
arrays is still very limited. The authors of [5] report on some tions based on independent and identically distributed (i.i.d.)
experiments with a cylindrical indoor antenna array with 128 channel vectors hold for real channels. We then analyze several
elements. They consider outdoor-to-indoor transmissions and other aspects such as the correlation coefficient of the channel
observe that the correlation between the channels cannot be vectors at different positions and the condition number of the
arbitrarily reduced by increasing the number of antennas. channel matrix. In summary, our results indicate that despite
Already from 20 antennas on, there is hardly any advantage fundamental differences between the i.i.d. and the measured
from the use of additional antennas. A possible explanation for channels, a large fraction of the theoretical performance could
this is that the physical channel offers only a limited number be achieved also in practice.
of physical directions or degrees of freedom which cannot be
increased by adding transmit antennas. This gives rise to the II. M EASUREMENT SETUP
conclusion that the effectiveness of large-scale MIMO strongly The channel measurements were conducted outdoors on the
depends on the radio environment. Another measurement Alcatel-Lucent campus in Stuttgart, Germany. For a graphical
campaign of outdoor-to-outdoor channels with a very large illustration, see Fig. 1. The setup consisted of a transmit
virtual linear antenna array (128 elements, λ/2-spacing, 7.3 m antenna array mounted on the top of a large building at a height
width) is described in [6]. Interestingly, the authors observe of 20 m (shown on the left) and two mobile single-antenna

978-1-4673-0762-8/12/$31.00 ©2012 IEEE 811

 Fixed 6.1cm
1m 3.5°

Rotating 16 positions

Fig. 2. Schematic of the transmit antenna array. One antenna element remains Fig. 3. Top view of the rotating antenna array. Channel measurements were
fixed while 7 antenna elements with λ/2 vertical spacing can be rotated to taken successively for 16 fixed angular positions to emulate a cylindrical array
different angular positions. of 112 antennas.

III. E VALUATION
receivers which were placed at 2 m distance from each other on
top of a car. Measurements were taken at 30 different positions The acquired data allows for the study of various aspects
throughout the campus at a carrier frequency of 2.6 GHz with of large MIMO systems. First, we will compare the measured
a bandwidth of 20 MHz, corresponding to 1200 active sub- channels against randomly created channels with i.i.d. entries
carriers (15 kHz spacing). Only every 3rd sub-carrier was in terms of their capacity and achievable sum-rates with linear
considered for the evaluation, leading to a total of 400 chan- precoders. Second, we will study the orthogonality between
nel coefficients per antenna and measurement position. The the channel vectors at different measurement positions by
measurement positions were selected to provide a good mix various metrics, namely the correlation coefficient and the
of different channel conditions (line-of-sight (LOS), non-LOS inverse condition number.
(NLOS), shadowing from buildings) which can be considered
A. Capacity and achievable rates with linear precoding
as representative for a residential urban area. The transmit
N N
antenna consists of a vertical array of 8 custom-built, dual We denote by hi,f,N ∈ C and hj,f,N ∈ C the channel
polarized elements with λ/2-spacing (only one polarization vectors on sub-carrier f at positions i and j, respectively,
direction was used). Seven of these elements can be rotated on when only the first N of the available 112 antennas are used.
a circle of 1 m radius to simulate a large cylindrical antenna These N antennas are selected at random and we average
array (see Fig. 2). The remaining antenna element is fixed over different choices. This is done to avoid correlation effects
(at sufficient distance from the array to avoid reflections) and between neighboring antennas1 which is an important aspect
only used to estimate the possible carrier phase offset between to study in its own regard but which is not considered in this
different positions. This setup was used to obtain channel paper. The channel vectors at each position are normalized
measurements for a “virtual” antenna array with 112 elements, such that the average received power over all antennas and
corresponding to 16 different angular positions (see Fig. 3). sub-carriers is equal to one. We define the channel matrix
N ×K
The angular separation between two adjacent positions was H = [hi1 ,f,N · · · hiK ,f,N ] ∈ C which is constructed
chosen to be 3.5◦ , which corresponds to a horizontal spacing from the channel vectors hik ,f,N at K ≤ N randomly
of 6.1 cm, slightly more than half of the wavelength. chosen measurement positions i1 , i2 , . . . , iK . For notational
At each measurements point on the campus, the channel simplicity, we ignore the dependence of H on N , K, f , and
vectors corresponding to each of the 16 angular positions were the explicit positions ik , k = 1, . . . , K. If we assume that the
successively recorded. Since the campus is rather quiet without BS transmits data simultaneously to K user terminals (UTs),
K
significant traffic or other moving objects, the channel between the vector y ∈ C of the received signals is given as
the different measurements can be assumed to be quasi-static. √
y = SNRHH s + n (1)
The result of this measurement campaign is essentially a radio
channel map of the campus (2 × 30 = 60 different locations) K
where s ∈ C is the transmit vector satisfying E sH s =1,
 
with an antenna array of 112 elements, taken over 400 sub- SNR is the transmit signal-to-noise-ratio (SNR), and n ∼
carriers. Thus, the available data is sufficient to calculate
averages over different positions as well as averages over 1 Especially, the correlation between horizontal and vertical neighboring
different sub-carriers. antennas it not the same due to a different angular spread in both dimensions.

812
CN (0, IK ) is a vector of complex Gaussian noise. The sum- 50
measured
capacity Csum of this channel is given as, e.g., [8],
i.i.d.
log2 det IN + HH QH (2)


Csum = max 40 Csum
Q0,tr Q≤SNR RMMSE
which can be computed by a simple water-filling algorithm

bits/channel use
RBF
over the eigenvalues of HH H [9] and is achieved by the 30
non-linear dirty-paper coding strategy. Less optimal, but also
less complex are linear beamforming strategies, where the 20
transmitted signal s is given as
1
s= √ Wx (3) 10
tr WWH
N ×K
where W ∈ C is a precoding matrix and x ∼
CN (0, IK ). The achievable sum-rate Rsum with linear pre- 0
−10 −5 0 5 10 15 20
coding is given as SNR (dB)
K
X
Rsum = log2 (1 + γk ) (4) Fig. 4. Average sum-capacity and sum-rates with linear precoding versus
k=1 SNR for N = 10 antennas and K = 8 UTs.

where
2 measured
hH 80 i.i.d.
ik ,f,N wk
γk = 1 (5) Csum
SNR + hH H
ik ,f,N W[k] W[k] hik ,f,N RMMSE
60
bits/channel use

and W[k] is matrix W with its kth column wk removed. We RBF

consider two different types of precoding, namely minimum-
mean-square-error (MMSE) precoding [10] and conjugate
beamforming (BF). The precoding matrices for both schemes 40
are respectively defined as
 −1
K 20
WMMSE = H H H + H
IK (6)
SNR
WBF = H. (7)
0
From the last equations, it is easy to see that WMMSE is simply −10 −5 0 5 10 15 20
a scaled version of WBF if the columns of H are orthogonal, SNR (dB)
i.e., HH H = IK . Thus, the more orthogonal the columns of
H, the closer is the sum-rate with conjugate BF RBF to the Fig. 5. Average sum-capacity and sum-rates with linear precoding versus
SNR for N = 112 antennas and K = 8 UTs.
sum-rate with MMSE precoding RMMSE .
In Figs. 4 and 5, we show the the capacity and sum-rates
as a function of the SNR for N = 10 and N = 112 antennas, the much simpler conjugate beamforming precoding strategy,
respectively. In both plots, we have set K = 8 and averaged we can see that the performance difference between the
over 400 different permutations of random positions i1 , . . . , iK measured and the i.i.d. channels is much larger than for MMSE
and the 400 sub-carriers. For comparison purposes, we also precoding. The main reason for this is that beamforming does
depict all rates for i.i.d. channels, i.e., the channel vectors not avoid multi-user interference. Consequently, the interfer-
are i.i.d. as hi,f,N ∼ CN (0, IN ). We can make the following ence level depends alone on the correlation between different
observations. For N = 10 antennas and at high SNR, there is a channel vectors. We can hence conclude that the measured
significant gap between Csum and RMMSE for the measured and channels are more correlated than the i.i.d. channels. This
the i.i.d. channels. However, for N = 112, this gap vanishes aspect will be discussed in more detail in the next section.
entirely for the i.i.d channels and becomes significantly smaller
for the measured channels. This shows that linear MMSE B. Correlation coefficient
precoding is close to optimal for a large number of antennas.2 We define the correlation coefficient δi,j (f, N ) between the
The performance loss for RMMSE between the i.i.d. and the two channel vectors hi,f,N and hj,f,N as
measured channels is roughly 4 dB for N = 112 and stays
almost constant for the entire range of SNR values. Concerning hH
i,f,N hj,f,N
δi,j (f, N ) = (8)
2 Note that MMSE precoding is identical to zero forcing at high SNR. khi,f,N k khj,f,N k

813
 1 of the matrix H̄H H̄, i.e.,
measured
Average correlation coefficient δi,j (f, N )
i.i.d. largest eigenvalue of H̄H H̄
κK,N = (10)
0.8 smallest eigenvalue of H̄H H̄
N,K
where H̄ = [h̄i1 ,f,N · · · h̄iK ,f,N ] ∈ C is constructed from
0.6 the normalized channel vectors h̄ik ,f,N = hik ,f,N /khik ,f,N k
at K ≤ N randomly chosen measurement positions
i1 , i2 , . . . , iK . The condition number is a widely used indicator
0.4 for the performance of linear receivers/precoders, see, e.g.,
[11], [12], [13]. A large condition number means that the
columns of H̄ are strongly correlated while κK,N = 1
0.2 implies that all columns are orthogonal. In order to obtain a
meaningful metric taking values in a finite range, we consider
0 the inverse of the condition number κ−1 N,K ∈ [0, 1]. As a
20 40 60 80 100 consequence of (9), it follows for i.i.d. channels that
Number of antennas N a.s.
H̄H H̄ −−−−→ IK (11)
N →∞
Fig. 6. Average correlation coefficient δi,j (f, N ) of two randomly picked
channel vectors versus the number of antennas N . and hence κiid N,K → 1, almost surely, as N → ∞. We
would like to remark that the condition number can also lead
to misleading conclusions regarding the system performance.
where k·k denotes the Euclidean norm. Clearly, Even for rank deficient channels, i.e., the smallest eigenvalue
0 ≤ δi,j (N, f ) ≤ 1. A low correlation between the channels is equal to zero, the capacity can still be high (depending on
at two different positions is desirable as it would allow one the remaining eigenmodes of the channel).
to simultaneously serve different UTs with little cross-talk. If Fig. 7 shows the average inverse condition number κ−1 N,K
the elements of hi,f,N are i.i.d., the correlation can be made as a function of N for different numbers of simultaneously
arbitrarily small by increasing N [1], [2], i.e., considered measurement positions K ∈ {2, 4, 6}. For compar-
ison, we also provide results for i.i.d. channels. When only
iid a.s. two channel vectors are considered and N is small, there is
δi,j (f, N ) −−−−→ 0 (9)
N →∞ little difference between κ−1
N,K for the measured and the i.i.d.
a.s. channels. However, this difference becomes more and more
where “−−→” denotes almost sure convergence. pronounced as K and/or N increase. This implies that there
Our first goal is to verify if this assumption also holds is a significant amount of correlation between the channel
for real channels. Fig. 6 shows the correlation coefficient vectors at different positions. We can also see that the larger
δi,j (f, N ) averaged over 400 random pairs of measurement K, the larger is the number of antennas which still lead to
points (i, j) and 400 sub-carriers as a function of the number considerable improvements of the condition number.
of antennas N . For comparison purposes we also depict the To further elaborate on the impact of N and K on κ−1 N,K ,
expected correlation coefficient for Rayleigh fading channels, we show in Figs. 8 and 9 the cumulative distribution function
i.e., hi,f,N
 iid ∼ CN (0,  IN ). In this case, it is easy to show (CDF) of κ−1 N,K for N = 16 and N = 112 antennas, respec-
that E (δi,j (f, N ))2 = N −1 . While the correlation of the tively. The difference between the curves for the measured
measured channels decreases equally fast as the correlation of and the i.i.d. channels is rather small for N = 16. However,
the i.i.d. channels for small N , it decreases at a much slower for N = 112 antennas, this picture changes. In particular, the
rate from N ≥ 10 on. Note that a similar effect has been CDF is spread out over the entire range [0,1] which indicates
observed in [5]. Thus, additional antennas only help up to a that there are several heavily correlated channel vectors which
certain degree to make the channels between two terminals cannot be rendered orthogonal by the use of more antennas.
more orthogonal. This effect depends of course on the radio Thus, user selection becomes a crucial aspect of large-scale
environment. MIMO systems.

IV. C ONCLUSIONS
C. Condition number
In this paper, we have described a setup for channel mea-
The correlation coefficient evaluates the “orthogonality” of surements with large antenna arrays and presented some of the
the channel vectors at two measurement positions. However, first results of a recently conducted measurement campaign.
the main purpose of large-scale MIMO systems is to simulta- These measurements are necessary to verify some of the
neously serve multiple UTs. Thus, it is even more important fundamental assumptions behind the quickly growing body of
to study the joint orthogonality of multiple channel vectors. theoretical works on large-scale MIMO systems. In particular,
To this end, we consider the condition number κN,K ∈ [1, ∞) we have analyzed by different metrics (correlation coefficient,

814
 1 1
N,K measured
Average inverse condition number κ−1
i.i.d.

Cummulative distribution function

K=2
0.8 0.8
K=2
K=4

0.6 0.6

K=6
0.4 K=4 0.4

0.2 0.2
measured
i.i.d.
K=6
0 0
20 40 60 80 100 0 0.2 0.4 0.6 0.8 1
Number of antennas N κ−1
N,K

Fig. 7. Average inverse condition number κ−1 N,K over N for different Fig. 9. CDF of the inverse condition number κ−1N,K for N = 112 antennas
numbers of jointly considered measurement positions K. and different numbers of jointly considered measurement positions K.

1
R EFERENCES
K=6
Cummulative distribution function

[1] T. L. Marzetta, “Noncooperative cellular wireless with unlimited num-

0.8 K=4 K=2 bers of base station antennas,” IEEE Trans. Wireless Commun., vol. 9,
no. 11, pp. 3590–3600, Nov. 2010.
[2] J. Hoydis, S. ten Brink, and M. Debbah, “Massive MIMO: How
many antennas do we need?” in Proc. IEEE Allerton Conference on
0.6 Communication, Control, and Computing, Urbana-Champaign, IL, US,
Sep. 2011.
[3] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta,
0.4 O. Edfors, and F. Tufvesson, “Scaling up MIMO: Opportunities and
challenges with very large arrays,” IEEE Signal Process. Mag., 2012,
to appear. [Online]. Available: http://arxiv.org/abs/1201.3210
N = 16 [4] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and
0.2 spectral efficiency of very large multiuser MIMO systems,”
measured IEEE Trans. Commun., 2012, submitted. [Online]. Available:
i.i.d. http://arxiv.org/abs/1112.3810
[5] X. Gao, O. Edfors, F. Rusek, and F. Tufvesson, “Linear pre-coding
0 performance in measured very-large MIMO channels,” in Proc. IEEE
0 0.2 0.4 0.6 0.8 1
Vehicular Technology Conf., San Francisco, CA, US, Sep. 2011.
κ−1
N,K [6] S. Payami and F. Tufvesson, “Channel measurements and analysis for
very large array systems at 2.6 Ghz,” in Proc. 6th European Conference
on Antennas and Propagation, Prague, Czech Republic, Mar. 2012.
Fig. 8. CDF of the inverse condition number κ−1 N,K for N = 16 antennas [7] C. W. Shepard, “Argos: Practical base stations for large-scale
and different numbers of jointly considered measurement positions K. beamforming,” Master’s thesis, RICE University, Houston, Texas, USA,
2012. [Online]. Available: http://clay.rice.edu/pubs/MasterThesis.pdf
[8] S. Vishwanath, N. Jindal, and A. Goldsmith, “Duality, achievable rates,
and sum-rate capacity of gaussian MIMO broadcast channels,” IEEE
condition number) to which extent channel orthogonality be- Trans. Inf. Theory, vol. 49, no. 10, pp. 2658– 2668, Oct. 2003.
[9] T. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed.
tween different terminals can be established by scaling up the New York, NY, USA: John Wiley and Sons, Inc., 2006.
number of transmit antennas. Our results show that despite [10] C. B. Peel, B. M. Hochwald, and A. L. Swindlehurst, “A
significant differences between the i.i.d. and the measured vector-perturbation technique for near-capacity multiantenna multiuser
communication-part I: channel inversion and regularization,” IEEE
channels, a huge fraction of the theoretical performance gains Trans. Commun., vol. 53, no. 1, pp. 195– 202, Jan. 2005.
of large antenna arrays could be achieved in practice. [11] H. Artes, D. Seethaler, and F. Hlawatsch, “Efficient detection algorithms
for MIMO channels: a geometrical approach to approximate ML detec-
tion,” IEEE Trans. Signal Process., vol. 51, no. 11, pp. 2808– 2820,
Nov. 2003.
V. ACKNOWLEDGMENTS [12] J. Maurer, G. Matz, and D. Seethaler, “Low-complexity and full-
diversity MIMO detection based on condition number thresholding,” in
Proc. IEEE International Conference on Acoustics, Speech and Signal
We would like to thank Günther Herzog, Günter Kalt- Processing, vol. 3, Apr. 2007, pp. III–61 –III–64.
beitzel, and Jürgen Otterbach from Bell Labs, Alcatel-Lucent, [13] M. Matthaiou, M. R. Mckay, P. Smith, and J. A. Nossek, “On the
Stuttgart, Germany, for their help with the channel measure- condition number distribution of complex Wishart matrices,” IEEE
Trans. Commun., vol. 58, no. 6, pp. 1705–1717, Jun. 2010.
ments and the configuration of the measurement equipment.

815