SEVENTH FRAMEWORK PROGRAMME

THEME 3
Information and Communication Technologies

D4.2 - Channel estimation techniques

Full Project Title:

High capacity network Architecture with Remote radio heads

& Parasitic antenna arrays

Grant Agreement Number: 318489

Version no

V1.0

D4.2 Channel estimation techniques

Objectives
The main objective of deliverable D4.2 is to address the problems in channel estimation in
MIMO systems equipped with single-RF transmitters. The first issue is channel modeling in
such a system. In fact, due to different implementation technique used in MIMO transmitters
with single-RF, channel can be defined in different ways. Therefore, the first part of deliverable
D4.2 addresses channel modeling in MIMO systems with single-RF transmitters. Next we discuss
about channel estimation in ESPAR (electronically steerable parasitic array radiator) as one of
the candidates of single-RF MIMO transmitters. Finally, the problem of pilot contamination is
discussed.

Task participation
NTNU: Mohammad Ali Sedaghat, Ralf R. Muller
UEDIN: Lin Zhou, Tharm Ratnarajah
AIT: Vlasis Barousis, Constantinos Papadias

Key results
I. C HANNEL M ODELING IN MIMO SYSTEM WITH SINGLE -RF TRANSMITTERS
In Multiple-Input Multiple-Output (MIMO) systems, channel transfer matrix is defined based
on relation between sources and signals detected by receive antenna elements. This means that
channel includes mutual coupling effect of antenna elements at transmitter and receiver as well
as effects of scatters and all objects placed between receiver and transmitter. In other words, it
is preferred to consider all non-deterministic effects as channel effect so that one can estimate
them by some data-aided or non-data-aided methods.
From circuit point of view, one can assume that channel is relation between the voltage fed
to transmit antenna elements (or currents on transmit antennas) and the voltage detected by
receives antenna elements (or currents on receive antennas) [1]. In classical MIMO systems,
July 15, 2014

DRAFT

HARP D4.2 Channel estimation techniques

transmitters consist of some voltage sources and some fixed matching networks hence the
channel definition can be easily done [1]. However, in single-RF MIMO transmitters there is
only one voltage source. Tunable impedances modulate signals in single-RF MIMO transmitters.
Therefore, channel cannot be defined based on relation between the voltage source on transmitter
and the signal detected by receive antennas. The problem is that due to uncertainty in load
modulator circuit [2] the relation between control signals and currents on the transmit antennas
is not totally deterministic. This makes the channel modeling complicated. One may include the
process of converting control signals to transmit antennas currents in channel but this makes the
channel non-linear.
The uncertainty in load modulators can be neglected if a good EVR (Error Vector Magnitude)
is achieved on the transmit antennas. This needs a very high resolution in current on transmit
antennas. Therefore, if we assume that the currents on transmit antennas represent the input
signal of channel with an acceptable EVR, one can define the relation between the currents on
transmit antennas and the signal at receiver as the channel transfer matrix. To define such a
channel, we analyze the circuit model of ESPAR in the next part.
A. Channel model in ESPAR
So far, ESPAR has been modeled in beam-space domain [3]. In beam-space modeling, instead
of assigning the diverse signals for transmission to different antenna elements, symbols are
mapped onto orthogonal radiating modes as
Pm () =

T 1
X

s[m]
n n (),

s[m] F

(1)

n=0

In (1), F =

s[1] s[2] . . . s[M ]

is a set of all possible symbol vectors for transmission and T

is the number of transmit antenna elements. M different loading sets are pre-defined and thus, the
ESPAR array produces the appropriate radiation pattern that corresponds to any possible symbol
vector for transmission according to (1). Although this scheme is efficient as the orthogonal
patterns allow for uncorrelated transmissions, it suffers from the following major drawback: One
1
needs to compute analytically the orthogonal modes {n ()}Tn=0
and eventually to determine

all possible radiation patterns in (1), as well as the corresponding pre-defined set of loading
1
values. The analytical computation of {n ()}Tn=0
can be found in the relevant literature in the

HARP D4.2 Channel estimation techniques

Fig. 1: Network model of a MIMO system with ESPAR configuration.

very special case of ESPARs with dipoles [4], but cannot be achieved in the more general case
of arbitrarily-shaped elements.
In fact, it is more convenient to define the currents on transmit antennas as input signal instead
of defining the signals in beam-space domain. A circuit model of ESPAR is shown in Fig. 1 [5].
We consider one active antenna and TP parasitic antennas at the transmitter side and R receiver
antennas at the receiver side. Each parasitic antenna is connected to one varactor diode. These
T = TP + 1 antennas at the transmitter side and R receive antennas at the receiver side can
be modeled by a multiport network as shown in Fig. 1. We use impedance model Z for this
network. Based on the unilateral approximation [6] we define Z as

Z AT
0
,
Z=
Z ART Z AR

(2)

where Z AT and Z AR are the impedance matrices of the transmitter and receiver antennas,
respectively and Z ART models the mutual impedance between transmitter and receiver antennas.
The source is also modeled by a voltage source v0 and R0 as the internal resistance. In Fig.
1, the varactor diode connected to the ith parasitic antenna is modeled by an impedance zi .
Furthermore, we consider additive white Gaussian noise in the receiver.
We start with the network equations in the transmitter side as follows
v0 = R0 I0 + vt0 vt0 = v0 R0 I0 .

(3)

HARP D4.2 Channel estimation techniques

Moreover, in the ith input port, we have

vti = zi Ii .
From the network model we have
h
vT = vt0

vtT

iT

(4)

= Z AT

I0

IT

iT

(5)

= Z AT iT .
Therefore, from (3)-(5) it can be written that
v0 Z T iT = Z AT iT iT = (Z T + Z AT )1 v0 ,

(6)

where
v0 =

v0 0

iT

(7)

and
Z T = diag{R0 , z1 , , zT }.

(8)

Now, we obtain the load voltage, vL , as follows

vL = Q Z ART (Z T + Z AT )1 v0 + n ,

(9)

where Q = zL (zL I + Z AR )1 and

n=

n1 n2

nR

iT

(10)

is the vector of receive antennas noise. The impedance of each varactor diode can be modeled
by a purely imaginary impedance as
zi = jxi ,

(11)

where i {1, , TP }. Let us define Z AT = jX AT + RAT . Then, (9) can be written as

vL = Q Z ART (jX T + RAT )1 v0 + n ,

(12)

where
X T = diag{0, x1 , , xTP } + X AT ,
RAT = diag{R0 , 0, , 0} + RAT .

(13)
(14)

HARP D4.2 Channel estimation techniques

The eq. (12) shows the relation between the output signal and the control signals is non-linear.
However, by using switch diodes and digital control signals, except at the short switching time,
the system is linear. This fact reveals a huge advantage of switching load modulators compared
to loads modulators with varactor diodes. However, switching diodes create out of band radiation
which can be alleviated by analog spectral shaping filter. In case of switching diodes one can
assume approximately a linear channel model.
If we define channel as the relation between the signal at receiver and the currents on transmit
antennas, one can describe the system as
y = Hi + n

(15)

where H is the (R T ) channel transfer matrix and n denotes the additive white Gaussian
noise (AWGN). From (12) and (15), the channel matrix H can be expressed as
H = QZ ART .

(16)

In the channel transfer matrix Q represents the effect of mutual coupling between antenna
elements at receiver and Z ART relates to the physical channel. Note that since we define the
currents as transmit signals, mutual coupling effect of transmit antenna elements does not appear
in the channel transfer matrix. This is actually an advantage for single-RF MIMO transmitters
since then mutual coupling of transmit antenna elements can be alleviated in base-band domain.
Correlation among channel coefficients appears in the factor Z ART . Any correlation effects
are captured indirectly by setting the desired inter-element spacing. In particular, the channel
can be mathematically expressed as [7]
Z ART = AR H p AT
T

(17)

where H p is a (LL) diagonal matrix that contains the complex path gains between the transmitter and the receiver and for a rich propagation environment they can be modeled as independent
and identically distributed (i.i.d.) random variables. Furthermore, AR : (R L) , AT : (T L)
are the steering matrices of the receiving and the transmitting arrays respectively, with the m-th
column representing the steering vector of the array towards the direction of arrival/departure
m .

HARP D4.2 Channel estimation techniques

For uniform linear arrays (ULAs), this vector is a function of m and the inter-element spacing
d:

am
(d
,

)
=

u
m
u

1
ej2du cos m
..
.
ej2(u1)du cos m

, u = {T, R}

(18)

Note that channel modeling in single-RF MIMO proposed in [2] can be also derived by the
same method as presented in this section.
II. T RAINING SEQUENCE DESIGN IN ESPAR
In this section, we design pilots for channel estimation purpose in ESPAR configuration.
However, since ESPAR generates signals with limited support it is important to consider the
limitation in ESPAR. It has been shown that ESPAR is able to transmit M-PSK signal [8].
Therefore, we propose to use a DFT matrix as pilot

r
P

P=

e(2 / L)1

e(2 / L)2

e(2 / L)(L1)

e(2 / L)2

e(2 / L)4

e(2 / L)2(L1)

1
..
.

e(2 / L)3
..
.

e(2 / L)6
..
.

...

e(2 / L)3(L1)
..
.

1 e(2 / L)(T 1)1 e(2 / L)(T 1)2

e(2 / L)(T 1)(L1)

where P is the total energy. Using the DFT matrix ensures that the power in active antenna
is always nonzero. We use Maximum Likelihood (ML) method to estimate channel using the
proposed pilots. The MMSE (Minimum Mean Square Error) is obtained as
2T 2R
MMSE = R tr{(P P ) } =
LP
2

(19)

It is observed that estimated error is proportional to the number of receive antennas, the square
of the number and transmit antennas, and noise power. In addition, it is inversely proportional
to transmitter signal power and the length of training sequence.
Here the simulation results are presented. We consider a 2-dimensional 3-element ESPAR
antenna to analyze the beam-space MIMO system. Two diverse symbols are transmitted in

HARP D4.2 Channel estimation techniques

BPSK: LS estimation in Beam space MIMO

10

Optimal
training
with L=2/8/16
1

BER: ML detection

10

10

10

10

Length=16, optimal training

Length=16, nonoptimal training
Length=8, optimal training
Length=8, nonoptimal training
Length=2, optimal training
Length=2, nonoptimal training
2

10

12

14

16

18

SNR (dB)

Fig. 2: BER performance of the proposed method for BPSK modulation.

the beam-space with BPSK and QPSK. ML detector is used. We compare the performance of
random training sequence and the proposed training sequence with lengths 2, 8, 16. We plot BER
performance versus SNR for BPSK and QPSK modulations in Fig. 2 and Fig. 3, respectively.
It is observed that the proposed training sequence outperforms the system compared to the
random training sequences. In Fig. 4, the MSE performance of the channel estimation methods
is plotted. It is shown that the proposed training sequence has a better performance comparing
to the random generated training sequences.
III. C HANNEL ESTIMATION IN MULTIUSER AND MULTICELL SCENARIO
In this section, we describe problem of pilot contamination in multiuser MIMO systems.
The problem of pilot contamination has been recently discovered in massive MIMO literature.
However, this problem also exists in the multiuser scenario which has been considered in
HARP. As explained, there are some ways to define channel in MIMO systems with singleRF transmitters and still it is not completely clear which one is the best. To describe the
problem of pilot contamination lets assume that the single-RF transmitter presented in [2] can set
currents on antennas similar to classical MIMO transmitters. Therefore, the pilot contamination

HARP D4.2 Channel estimation techniques

QPSK: LS estimation in Beam space MIMO

BER: ML detection

10

Optimal
training
with L=2/8/16

10

10

10

Length=16, optimal training

Length=16, nonoptimal training
Length=8, optimal training
Length=8, nonoptimal training
Length=2, optimal training
Length=2, nonoptimal training

10
SNR (dB)

15

Fig. 3: BER performance of the proposed method for QPSK modulation.

problem is quite similar to the problem in massive MIMO systems. In the remaining parts of this
section, we present a pilot decontamination method based on subspace projection to improve
channel estimation in massive multi-antenna systems which has been proposed recently in [9]
(for extended version one can see [10]).
In massive MIMO, given a rich scattering environment, the access point is able to concentrate
the radiated energy within an area (or volume) around the designated receiver that is only a
fraction of a squared (or cubed) wavelength [11]. This is in contrast to earlier methods of array
processing in wireless communications based upon angles of arrival and departure which are
limited to concentrate their energy into a certain direction rather than an area limited in both
angular and radial domain.
Since power in the far-field can be amplified (in theory) beyond limits by means of array gains,
the ultimate limit of a single-cell massive MIMO system is only constrained by the coherence
time of the channel [12]. The number of terminals needs to stay below the coherence time to
have sufficient degrees of freedom for channel estimation. In practice, the coherence time needs
to be significantly larger than the number of terminals to also allow for data transmission in
addition to pilot symbols for channel estimation.

HARP D4.2 Channel estimation techniques

MSE of LS estimation in Beam space MIMO

10

MSE

10

10

10

10

Length=16, optimum training, simulation

Length=16, nonoptimum training, simulation
Length=16, optimum training, analytical
Length=2, optimum training, simulation
Length=2, nonoptimum training, simulation
Length=2, optimum training, analytical

10
SNR (dB)

15

20

Fig. 4: MSE performance of LS estimation in beam-space MIMO.

In [12], however, a pessimistic conclusion about the performance of massive MIMO in multicellular systems was reached. Based on the explicit assumption of no coordination among cells
and on the implicit assumption of linear channel estimation [12, Eq. (5)], it was concluded
that the array gain can be achieved only for data detection, but not for channel estimation. The
author argued that channel state information, though not required to be perfect, must have at
least a certain quality in order to utilize unlimited array gains. As a result, pilot interference from
neighboring cells would limit the ability to obtain sufficiently accurate channel estimates and
be the new bottleneck of the system. This effect, commonly referred to as pilot contamination
[13], is treated as a fundamental effect in many works, e.g. [12][17].
In the next part of this section, we describe the idea behind the proposed blind pilot decontamination.
A. General idea of the proposed blind pilot decontamination algorithm
Lets first introduce the considered channel model. Let the frequency-flat, block-fading propagation channel be described by the matrix equation
Y = HX + Z,

(20)

HARP D4.2 Channel estimation techniques

10

where X CT C is the transmitted data (eventually multiplexed with pilot symbols), H CRT
is the channel matrix of unknown propagation coefficients, Y CRC is the received signal, and
Z CRC is the total impairment. Furthermore, we assume that channel, data, and impairment
have zero mean, i.e. E X = E H = E Z = 0. The impairment includes both thermal noise and
interference from other cells and is, in general, neither white nor Gaussian. We also introduce
the normalized coherence time
C
(21)
R
which is assumed to be a finite, non-zero constant. Note that C in (21) is the coherence time of
=

the channel measured in symbol intervals.

Then consider the channel model (20) for a single active transmit antenna, i.e. T = 1 and look
for the matched filter m such that the signal-to-noise ratio (SNR) at its output is maximum.
In white noise, maximizing the SNR is equivalent to maximizing the total received power
normalized by the power gain of the filter. Thus, the optimum filter is given by
m J m
m = argmax
m m
m
with


J= E
YY .
X,Z|H

(22)

(23)

It is a well-known result of linear algebra that the vector m maximizing the right hand side of

(22), commonly referred to as the Rayleigh quotient, is that eigenvector of J that corresponds
to the largest eigenvalue of J . Since we do not know the matrix J , we have to cope with the
approximate solution
m Y Y m
.
(24)
m m
m
This approximation is tight for large number of antenna elements, i.e. we have the almost sure
m = argmax

convergence of the inner product

|hm ; m i| ||m || ||m ||

(25)

as R , if the largest eigenvalue of the noise is negligible against the largest eigenvalue of
the signal, i.e.
m ZZ m
m
m m
= 0.
lim
R
m HXX H m
max
m
m m
max

(26)

HARP D4.2 Channel estimation techniques

11

Note that the limit R implies C due to (21).

For finite number of antennas, there are better approximations for J than Y Y , e.g. Gestimation [18]. However, such methods exceed the scope of the present work and are left for
future research. G-estimation will further improve the performance of the proposed subspace
projection method, in practice.
The limit condition (26) is not hard to fulfill. In fact, it holds true for independent constant
variance entries in Z, H, and X. To see this note that the largest eigenvalue of ZZ scales
linearly with R, as the number of entries in Z CRC grows quadratically, but the number
of non-zero eigenvalues grows linearly. At the same time the largest eigenvalue of HXX H
grows quadratically with R, as the number of entries in H CRT grows linearly, the number
of entries in X CT C grows linearly, but the number of non-zero eigenvalues is T and thus
constant. The detailed description of algorithm is given in the next part.
B. Detailed algorithm
In this part, the details of proposed algorithm are presented. Consider the singular value
decomposition
Y = U V

(27)

with unitary matrices U CRR and V CCC and the R C diagonal matrix with
diagonal entries 1 2 min{R,C} sorted in non-increasing order. As shown in [19],
the columns of U are highly correlated with the columns of H. Based on this observation, [19]
proposes two algorithms for improved nonlinear estimation of the channel matrix H.
In the sequel, we pursue a strategy different from the one in [19]. We decompose the matrix
of left singular vectors
U = [S|N ]

(28)

into the signal space basis S CRT and the null space basis N CR(RT ) . Now, we project
the received signal onto the signal subspace and get
Y = S Y .

(29)

The null space basis N is not required in the sequel. In fact, there is no need to compute the
full singular value decomposition (27). Only the basis of the signal subspace S is needed and
there are efficient algorithms available to exclusively calculate S.

HARP D4.2 Channel estimation techniques

12

Consider now the massive MIMO case, i.e. R T : The T -dimensional signal subspace is
much smaller than the min{R, C}-dimensional full space, which the noise lives in. White noise
is evenly distributed in all dimensions of the full space. Thus, the influence of white noise onto
the signal subspace becomes negligible as min{R, C} . In other words: The considerations
for the largest eigenvalue in (26) and its corresponding eigenvector in (25) are equally valid for
the T largest eigenvalues and their corresponding eigenvectors, as long as T is finite.
Using the algorithm above, we can achieve an array gain even without the need for estimating
the channel coefficients. In fact, channel estimation can be delayed until the received signal has
been projected onto the signal subspace and the dominant part of the white noise has already
been suppressed.
In order to save complexity it is sensible not to estimate the channel matrix H, at all. Instead,
we directly consider the subspace channel
+Z

Y = HX

(30)

CT T by standard methods of
and estimate the much smaller subspace channel matrix H
linear channel estimation based on pilot symbols. Hereby, we have transformed the problem of
channel estimation for highly asymmetric massive MIMO systems into the well-explored problem
of channel estimation for classical symmetric MIMO systems. Although the data dependent
= S Z CT C is not independent from the data X,
projection (29) implies that the noise Z
neglecting this dependence is an admissible approximation that becomes exact due to (25), as
the number of receive antennas R grows large.
In addition to white noise, there is co-channel interference from L neighboring cells. For sake
of notational convenience, we assume that the number of transmit antennas is identical in all
cells and equal to T . The interference from neighboring cells is anything but white. It is the
more colored, the smaller the ratio
=

T
R

(31)

which will be called load in the following. Any R-dimensional channel vector is orthogonal to
any other channel vector in the limit R [12]. This holds regardless whether the two channel
vectors correspond to transmitters in the same cell or in different cells. In the limit of zero load,
i.e. 0, we have an even stronger result: the subspace spanned by the co-channel interference

HARP D4.2 Channel estimation techniques

13

is orthogonal to the signal subspace.1 That means that in the limit R/T , the (L + 1)T
largest singular values of the received signal matrix Y become identical to the Euclidean norms
of the (L + 1)T channel vectors. We only need to identify which singular values correspond to
channel vectors from transmitters inside the cell as opposed to channel vectors from transmitters
in neighboring cells. Then, we can remove the interference from neighboring cells by subspace
projection.
C. Performance analysis and numerical results
In this part, performance analysis and also a summary of results of the proposed algorithm
are presented. For the complete analysis and results one can see [10]. We have demonstrated
above, that the proposed algorithm works in principle in massive MIMO systems if the number
of receive antennas is much larger than the product of transmit antennas and neighboring cells.
In practical systems, the number of transmit and receive antennas is finite and the load can be
made very small but not arbitrarily small. The standard assumption of massive MIMO systems,
i.e. R , while T staying finite, gives overoptimistic results for finite systems of practical
interest. In order to find the limits of the subspace projection method, we need to consider a
more refined limit in the number of antennas. For that purpose, we utilized the asymptotic scale
invariance of the eigenvalue spectra of large random matrices: For appropriate normalization of
the entries, the spectra of random matrices are hardly affected (asymptotically invariant) if all
matrix dimensions scale proportionally. Thus, a useful and insightful approach to understand
the behavior of a real cellular massive MIMO network consists in assuming that both T and R
grow large with a small, but fixed ratio . This approach is tantamount to studying a system
with T = 5n transmit antennas, R = 300n receive antennas and coherence time C = 100n for
n and assume that the result for a system with T = 5, R = 300 and C = 100 hardly
differs.
We decompose the impairment process
Z = W + H IX I
1

(32)

Note that the pairwise orthogonality of channel vectors holds for R , in general, and does not require 0. However,

the orthogonality of subspaces requires 0 in addition to R , as the accumulation of T = R vanishing pairwise

correlations is not vanishing, in general.

HARP D4.2 Channel estimation techniques

14

into white noise W and interference from L neighboring cells where interfering data X I
CLT R is transmitted in neighboring cells and received in the cell of interest through the channel
H I CRLT . Combining (20) and (32), we get
Y = HX + H I X I + W .

(33)

Let the entries of the data signal X be iid with zero mean and variance P . Let the entries
of the channel matrix H be also iid with zero mean, but have unit variance. Let the entries of
the matrix of interfering signals X I be iid with zero mean and variance P and let the entries
of the k th column of the matrix of interfering channels H I be independent with zero mean and
variance Ik /P such that the ratio Ik /P accounts for the relative attenuation between out-of-cell
user k and the intracell users. Let the empirical distribution of Ik converge to a limit distribution
as LT which is denoted by PI (). Furthermore, we assume that the elements of the noise
W are independent and identically distributed (iid) with zero-mean and variance W . Let us
denote the asymptotic eigenvalue distribution of Y Y as PY Y (x). In [10, Appendix A], it has
been shown that this asymptotic eigenvalue distribution obeys
sGY Y (s) + 1 =
P T C (sGY Y (s) + 1 ) GY Y (s)
P T C (sGY Y (s) + 1 ) GY Y (s)
Z
xLT C (sGY Y (s) + 1 ) GY Y (s)dPI (x)

xT C (sGY Y (s) + 1 ) GY Y (s)

W C (sGY Y (s) + 1 ) GY Y (s)

(34)

with
GY Y (s) =

dPY Y (x)
xs

(35)

denoting its Stieltjes transform. By means of the Stieltjes inversion formula

p(x) =

1
lim G(x + jy)
y0+

(36)

the asymptotic eigenvalue density pY Y (x) is obtained.

In Fig. 5, the solid lines in red shows the asymptotic eigenvalue distribution of Y Y /R. The
histograms in blue show the empirical eigenvalue distributions of Y Y /R for T = 10, C = 100,
and various values of R. We observe that for sufficiently large number of receive antennas R,
the distribution decomposes into two disjunct bulks: A noise and interference bulk to the left and

0.15

15

R = 30

probability density

probability density

HARP D4.2 Channel estimation techniques

0.1
0.05
0
0

10
eigenvalue

R = 100

0.04

0.02

0
0

20

10
eigenvalue

20

0.015

R = 300

0.01
0.005
0
0

10
eigenvalue

20

probability density

probability density

x 10
4

R = 1000

0
0

10
eigenvalue

Fig. 5: Asymptotic eigenvalue density of the matrix Y Y /R in solid red line for
P = 10 dB, Ik =

P (1+kmodT )
4T

20

1
,
10

k = 1 . . . LT , L = 2, W = 0 dB. The asymptotic eigenvalue

distribution is compared to the empirical eigenvalue density for T = 10, C = 100, and various
values of R given by the histograms in blue.

a bulk of the signal of interest to the right. If the bulks do not overlap, we can blindly separate
the signals of interest from interference and noise.
The bulks are not disjunct in general, but only for certain values of the involved system
parameters. It is therefore of utmost importance for practical design of blind pilot decontamination to know which system parameters do lead to bulk separation. The extremely good match
between the asymptotic distribution and the empirical distribution for finite matrices corroborate
the usefulness to study the support of the asymptotic eigenvalue distribution of Y Y and the

HARP D4.2 Channel estimation techniques

16

asymptotic conditions of bulk separability. We remark that, in principle, the common interference
and noise bulk could also separate into two separate bulks for noise and interference. However,
this happens only if the weakest interfering signal is sufficiently strong in comparison to the
noise power.
IV. P UBLICATION
The publications are as follows

On channel modeling: [5] and [7]

On pilot decontamination: [9], [10], [20], [21]

R EFERENCES

[1] M. Ivrlac and J. Nossek, A multiport theory of communications, in Source and Channel Coding (SCC), 2010 International
ITG Conference on, Jan. 2010, pp. 1 5.
[2] M. A. Sedaghat, R. R. Muller, and G. Fischer, A novel single-RF transmitter for massive MIMO, in Smart Antennas
(WSA), 2014 18th International ITG Workshop on. VDE, 2014, pp. 18.
[3] A. Kalis, A. Kanatas, and C. Papadias, A novel approach to MIMO transmission using a single RF front end, IEEE
Journal on Selected Areas in Communications, vol. 26, no. 6, pp. 972 980, Aug. 2008.
[4] V. Barousis, A. Kanatas, and A. Kalis, Beamspace-domain analysis of single-RF front-end MIMO systems, IEEE
Transactions on Vehicular Technology, vol. 60, no. 3, pp. 11951199, 2011.
[5] M. A. Sedaghat and R. R. Mueller, Large system analysis of low-cost mimo transmitters, in Smart Antennas (WSA),
2013 17th International ITG Workshop on. VDE, 2013, pp. 14.
[6] M. Ivrlac and J. Nossek, Toward a circuit theory of communication, Circuits and Systems I: Regular Papers, IEEE
Transactions on, vol. 57, no. 7, pp. 16631683, 2010.
[7] V. I. Barousis, C. B. Papadias, and R. R. Muller, A new signal model for mimo communication with compact parasitic
arrays, in International Symposium on Communications, Control and Signal Processing, 2014, pp. 15.
[8] O. Alrabadi, C. Papadias, A. Kalis, and R. Prasad, A universal encoding scheme for MIMO transmission using a single
active element for PSK modulation schemes, IEEE Transactions on Wireless Communications, vol. 8, no. 10, pp. 5133
5142, Oct. 2009.
[9] R. R. Muller, M. Vehkapera, and L. Cottatellucci, Blind pilot decontamination, in Smart Antennas (WSA), 2013 17th
International ITG Workshop on.

VDE, 2013, pp. 16.

[10] R. Muller, L. Cottatellucci, and M. Vehkapera, Blind pilot decontamination, arXiv preprint arXiv:1309.6806, 2013.
[11] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta, O. Edfors, and F. Tufvesson, Scaling up MIMO:
Opportunities and challenges with very large arrays, IEEE Signal Processing Magazine, vol. 30, no. 1, pp. 4060, Jan.
2013.
[12] T. L. Marzetta, Noncooperative cellular wireless with unlimited numbers of base station antennas, IEEE Transactions
on Wireless Communications, vol. 9, no. 11, pp. 35903600, Nov. 2010.

HARP D4.2 Channel estimation techniques

17

[13] J. Jose, A. Ashikhmin, T. L. Marzetta, and S. Vishwanath, Pilot contamination and precoding in multi-cell tdd systems,
Wireless Communications, IEEE Transactions on, vol. 10, no. 8, pp. 26402651, 2011.
[14] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, The multicell multiuser mimo uplink with very large antenna arrays and
a finite-dimensional channel, Communications, IEEE Transactions on, vol. 61, no. 6, pp. 23502361, 2013.
[15] B. Gopalakrishnan and N. Jindal, An analysis of pilot contamination on multi-user MIMO cellular systems with many
antennas, in IEEE International Workshop on Signal Processing Advances in Wireless Communications (SPAWC), San
Francisco, CA, USA, Jun. 2011.
[16] F. Fernandez, A. Ashikhmin, and T. Marzetta, Interference reduction on cellular networks with large antenna arrays, in
Proc. of IEEE International Conference on Communications (ICC), Ottawa, Canada, Jun. 2012.
[17] N. Krishnan, R. D. Yates, and N. B. Mandayam, Cellular systems with many antennas: Large system analysis under pilot
contamination, in Communication, Control, and Computing (Allerton), 2012 50th Annual Allerton Conference on. IEEE,
2012, pp. 12201224.
[18] X. Mestre, Improved estimation of eigenvalues and eigenvectors of covariance matrices using their sample estimates,
Information Theory, IEEE Transactions on, vol. 54, no. 11, pp. 51135129, 2008.
[19] H. Q. Ngo and E. G. Larsson, EVD-based channel estimation in multicell multiuser MIMO systems with very large
antenna arrays, in Proc. of IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Kyoto,
Japan, Mar. 2012.
[20] L. Cottatellucci, R. R. Muller, and M. Vehkapera, Analysis of pilot decontamination based on power control, in Vehicular
Technology Conference (VTC Spring), 2013 IEEE 77th.

IEEE, 2013, pp. 15.

[21] R. R. Muller, M. Vehkapera, and L. Cottatellucci, Analysis of blind pilot decontamination, in Signals, Systems and
Computers, 2013 Asilomar Conference on.

IEEE, 2013, pp. 10161020.