Iterative Channel Estimation in LTE Systems

Florent Kadrija, Michal

Simko and Markus Rupp
Institute of Telecommunications, Vienna University of Technology
Gusshausstrasse 25/389, A-1040 Vienna, Austria
Email: [email protected]
Web: http://www.nt.tuwien.ac.at/ltesimulator
AbstractIn this paper, we discuss iterative channel estimation
for the Long Term Evolution (LTE) downlink. The LTE standard
utilizes coherent detection and therefore to achieve a high data
rate over mobile radio channels, it is essential to obtain high
accurate channel state information at the receiver side. For
channel estimation purposes, the LTE standard provides training
data known as pilot symbols. The open question is whether
the accuracy of channel estimation based on pilot symbols is
sufcient to achieve such high data rate transmission. The quality
of the channel estimates can be further enhanced if soft or hard
information bits, which are available after the decoding stage,
are utilized by the channel estimator to re-estimate the channel.
In order to quantify the gain provided by the iterative channel
estimation, several scenarios have been considered in this work.
Index TermsLTE, OFDM, Channel Estimation, Iterative
Receiver.
I. INTRODUCTION
Channel estimation plays a crucial role in the performance
of wireless communication systems, since its knowledge is
utilized to detect the data symbols [1]. With enhancing the
channel estimation performance, also the performance of the
entire system is improved [2]. Many investigations have been
conducted on channel estimation in LTE [3], [4], [5]. The
majority of them investigate non-iterative channel estimation
algorithms based only on the known pilot symbols provided by
the LTE standard. In recent years, receivers with iterative chan-
nel estimation have become more and more attractive because
of their superb performance [6]. Hard or soft information bits
from the decoding stage are utilized to estimate the data sym-
bols and then applied to further improve the channel estimation
[7], [8]. In this paper we discuss and analyse the performance
of different iterative channel estimation algorithms for the
block fading case.
The main contributions of the paper are:
We calculate the soft estimated data symbols for the LTE
standard (according to [9]) for 4-Quadrature Amplitude
Modulation (QAM), 16-QAM and 64-QAM constella-
tions symbols.
We derive various iterative channel estimation algorithms
with different complexities and performance behaviours.
We investigate the impact of processing the soft (a-
posteriori or extrinsic) and hard feedback information for
channel estimation.
The remainder of the paper is organized as follows. In
Section II, the system model is described. Various iterative
channel estimation algorithms are presented in Section III.
In Section IV, we evaluate the performance of the channel
estimators in terms of data throughput and Mean Square Error
(MSE). Finally, Section V concludes the paper with a summary
of most important points obtained in this work.
II. SYSTEM MODEL
In this section, we briey give an overview of the pilot
symbol structure dened in the LTE standard and discuss the
structure of an iterative receiver. Afterwards, we introduce a
signal model and we show how the soft symbols are calculated.
A. Overview
In LTE systems, the initial channel estimates can be ob-
tained based on pilot symbols, which are placed in the
Orthogonal Frequency Division Multiplexing (OFDM) time-
frequency grid. In Figure 1, the pilot symbol structure for one
subframe corresponding to normal cyclic prex conguration
for up to four transmit antenna ports is shown (for differ-
ent antenna ports different color boxes) [9]. Pilot symbols
transmitted from multiple antenna ports are orthogonal to
each other. That is, when one antenna port transmits a pilot
symbol, the other antenna ports do not transmit. For the
iterative channel estimation we take advantage of hard or soft
estimation of data symbols (white boxes), available after the
decoding stage, which in addition to the pilot symbols are
utilized to re-estimate the channel after the initial estimation.
R0
R0
R0
R0
R0
R0
R0
R0 R1
R1
R1
R1
R1
R1
R1
R1 R2
R2
R2
R2
R3
R3
R3
R3
R0
R1
R2
R3
Pilot symbols from
Antenna 0
Pilot symbols from
Antenna 1
Pilot symbols from
Antenna 2
Pilot symbols from
Antenna 3
Data symbols
S
u
b
c
a
r
r
i
e
r
s

(
f
r
e
q
u
e
n
c
y
)
OFDM symbols (time)
Fig. 1. Pilot symbol structure.
B. Iterative Receiver
Figure 2 depicts the iterative receiver structure. The channel
state information knowledge, derived from the pilot symbols,
is utilized for demodulation and soft de-mapping of the OFDM
symbols. Afterwards, the detector provides to the decoding
unit soft information in terms of Log Likelihood Ratios (LLRs)
Copyright 2013 IEEE. Published in the proceedings of WSA 2013, Stuttgart, Germany, 2013.
for each transmitted information coded bit. In order to take
advantage of the feedback information, the decoding unit has
to be modied and new components such as the soft symbol
mapper has to be added to the receiver (the iterative receiver
part in Figure 2). The modied decoding unit, besides hard
decoded bits, outputs also soft estimated coded bits. This
feedback information is utilized by the (soft) symbol mapper
to calculate the estimated data symbols, which in turn, together
with the pilot symbols are applied by the channel estimator to
re-estimate the channel. The process is performed again and
again and new channel estimates are obtained. Through the
successive iterations, the overall system performance can be
improved.
OFDM
Dem.
OFDM
Dem.
Detector
&
Demapper
Decoding Unit
Decoding Unit
Iterative
Channel
Estimator
Coding
Unit
Coding
Unit
Soft Symbol Mapper
Soft Symbol Mapper
Symbol
Mapper
Symbol
Mapper
Pilot
Symbols
Decoded
bits
Hard
bits
Soft
bits
Hard
bits
Decoded
bits
Non-iterative receiver part
Iterative receiver part
1
Nr
.
.
.
Fig. 2. Iterative receiver structure.
C. Signal Model
The received frequency domain signal r
n
r
,n
s
(after OFDM
demodulation) at the receive antenna port n
r
and OFDM
symbol n
s
is obtained as
r
n
r
,n
s
=
N
T

n
t
=1
S
n
t
,n
s
h
n
t
,n
r
+n
n
r
,n
s
, (1)
where S
n
t
,n
s
= diag(s
n
t
,n
s
) is a diagonal matrix and s
n
t
,n
s
is the transmit signal vector at the transmit antenna port n
t
(n
t
= 1 . . . N
T
) and OFDM symbol n
s
, which comprises the
pilots s
p,n
t
,n
s
,n
k
and data subcarriers s
d,n
t
,n
s
,n
k
, where n
k
is
the subcarrier index. The vector h
n
t
,n
r
is the channel vector
in the frequency domain and n
n
r
,n
s
is the additive zero mean
white Gaussian noise vector with the variance
2
n
. The length
of vectors s
n
t
,n
s
, h
n
t
,n
r
and n
n
r
,n
s
is N
K
, which corresponds
to the number of subcarriers within an OFDM symbol n
s
. If we
concatenate Equation (1) for N
S
OFDM symbols, we obtain
r
n
r
=
N
T

n
t
=1
S
n
t
h
n
t
,n
r
+n
n
r
, (2)
where S
n
t
is of size (N
K
N
S
) N
K
, and r
n
r
and n
n
r
are
of size (N
K
N
S
) 1. On the one hand, for initial channel
estimation the sum in Equation (2) disappears because if
there is a pilot symbol located at the n
t
-th transmit antenna
port within the n
s
-th OFDM symbol on the n
k
-th subcarrier,
symbols at the remaining transmit antenna ports at the same
positions are zero. Therefore, Equation (2) extracted at pilot
symbols p is simply r
p,n
r
= S
p,n
t
h
p,n
t
,n
r
+ n
p,n
r
. On the
other hand, for iterative channel estimation, the matrix S
n
t
,n
s
from Equation (1) is replaced by the matrix

S
n
t
,n
s
, in which
the data symbols are replaced by the estimated soft or hard
data symbols. Thus, Equation (2) for the iterative case can be
written as
r
n
r
=

Sh
n
r
+n
n
r
, (3)
with

S, of size (N
K
N
S
) (N
K
N
T
), comprising the pilot
symbols and the soft or hard estimated data symbols.
D. Soft Symbol Calculation
The a-posteriori or extrinsic LLRs of coded bits from the
output of the decoding unit are considered as a-priori informa-
tion and utilized by the soft symbol mapper to calculate the
soft estimated data symbols as [10]
s
d
= E[s
d
] =
M

i=1
s
d,i

N
C

n
b
=1
P
_
f
n
b
(s
d,i
)
_
, (4)
where s
d,i
is one of the constellation points that is formed
by the information bits (f
1
, f
2
, . . . , f
N
C
), M is the number
of symbols in the symbol alphabet and N
C
= log
2
M is the
number of bits per symbol. P(f
n
b
(s
d,i
)) is the probability of
the bit f
n
b
that corresponds to the given symbol constellation
point s
d,i
. For notational simplicity we have omitted the
transmit antenna port n
t
, OFDM symbol n
s
and subcarrier
n
k
indexes from s
d,n
t
,n
s
,n
k
. The LLR of the coded bit f
n
b
is
computed by the log-ratio of the probability of the information
bit being 1 over the probability of the information bit being 0
L(f
n
b
) = ln
P(f
n
b
= 1)
P(f
n
b
= 0)
. (5)
We know that P(f
n
b
= 1) + P(f
n
b
= 0) = 1. Using the
denitions of LLRs above, we nd the probability of the
information bit being 1(0) as
P
_
f
n
b
= 1(0)
_
=
e
1(0)L(f
n
b
)
1 + e
L(f
n
b
)
. (6)
Furthermore, Equation (6) can be written in terms of tangent
hyperbolic functions
P
_
f
n
b
= 1(0)
_
=
1
2
_
1 + () tanh (L(f
n
b
)/2)
_
. (7)
Any Gray-mapped QAM constellation symbol can be divided
into real and imaginary parts s
d
= Re(s
d
)+jIm(s
d
) [10]. This
implies that the real and imaginary components of each soft
symbol are independent. For the real Re( s
d
) and imaginary
Im( s
d
) part of the soft estimated symbol s
d
, by applying
Equation (7) in Equation (4), for 4-QAM, 16-QAM and 64-
QAM (for the LTE standard dened modulation alphabet
according to [9]) we nd
4-QAM
Re( s
d
) = a
1
and Im( s
d
) = a
2
. (8)
16-QAM
Re( s
d
) = a
1
a
3
2a
1
and Im( s
d
) = a
2
a
4
2a
2
. (9)
64-QAM
Re( s
d
) = 4a
1
2a
1
a
3
a
1
a
3
a
5
and (10)
Im( s
d
) = 4a
2
2a
2
a
4
a
2
a
4
a
6
, (11)
where a
n
b
= tanh
_
L(f
n
b
)/2
_
is the tangent hyperbolic func-
tion as a function of the bit LLR L(f
n
b
). The soft estimated
symbols are then obtained as s
d
= Re( s
d
) + jIm( s
d
).
III. ITERATIVE CHANNEL ESTIMATION
In this section, different iterative channel estimation algo-
rithms considered in this paper are derived.
A. Iterative Least Square (LS) channel estimator
Iterative LS channel estimate tries to nd a solution

h
(i)
ILS,n
r
by minimizing the Euclidean norm squared J(

h
(i)
ILS,n
r
) =
_
_
_

S

h
(i)
ILS,n
r
r
n
r
_
_
_
2
2
. The minimum is found by computing the
gradient of the cost function J(

h
(i)
ILS,n
r
) and equating it to
the zero vector. Thus, the iterative LS channel estimator is
derived as [11]

h
(i)
ILS,n
r
= (

S
H

S)
1

S
H
r
n
r
, (12)
under the condition that rank(

S) = N
K
N
T
[12], where i is
the iteration index number. For notational simplicity, we have
omitted the iteration index i in

S
(i)
.
B. Iterative Linear Minimum Mean Square Error (LMMSE)
channel estimator
The iterative LMMSE channel estimator provides estimated
channel coefcients by minimizing the MSE
= E
_
_
_
_h
n
r

h
(i)
ILMMSE,n
r
_
_
_
2
2
_
. (13)
Utilizing the second order statistics of the channel the iterative
LMMSE estimator is derived, which improves the performance
of the iterative LS estimator by ltering its estimate as [11]

h
(i)
ILMMSE,n
r
= R
h
n
r
,h
n
r
(R
h
n
r
,h
n
r
+
2
n
(

S
H

S)
1
)
1

h
(i)
ILS,n
r
,
(14)
where R
h
n
r
,h
n
r
is the autocorrelation matrix of the channel
vector h
n
r
, which is assumed to be perfectly known.
C. Iterative Approximate LMMSE (ALMMSE) channel estima-
tor
The computational complexity that comes with iterative
LMMSE is very high because of the matrix inversion in-
volved in Equation (14). Therefore, its use in a real-time
implementation is limited and restricted to a low complexity
implementation. In this subsection, we investigate an approx-
imate algorithm which reduces the complexity of the iterative
LMMSE and preserves its performance. This estimator is
presented in [3] for LTE and [13] for WiMAX but derived
only for initial channel estimation. Therefore, we will modify
this estimator for the iterative case. As presented in [3] and
[13], the main idea of ALMMSE estimator is to approximate
the correlation matrix R
h,h
by utilizing only the correlation of
the L closest subcarriers instead of using the full correlation
between all subcarriers as in LMMSE case. Assuming that
the correlation is frequency independent, the approximated
L L correlation matrix

R
(L)
h,h
is obtained by averaging over
the correlation matrices of sizes L L [14]. As in [3], rst
the number of L closest subcarriers is dened, over which
the autocorrelation matrix

R
(L)
h,h
is approximated. Next, the
channel vector

h
(i)
n
t
,n
r
is dened over the interval of these L
consecutive subcarriers as
_

_
_

h
(i)
n
t
,n
r
,1
. . .

h
(i)
n
t
,n
r
,L
_
T
; k
L+1
2
_

h
(i)
n
t
,n
r
,k
L1
2

. . .

h
(i)
n
t
,n
r
,k+
L1
2

_
T
; otherwise
_

h
(i)
n
t
,n
r
,N
K
L+1
. . .

h
(i)
n
t
,n
r
,N
K
_
T
; k N
K

L1
2
(15)
where k is the subcarrier index. The output of the estimator
for the given interval of L subcarriers becomes

h
(i),(L)
n
r
=
_

h
(i),(L),T
1,n
r
. . .

h
(i),(L),T
N
T
,n
r
_
T
. Furthermore, we derive iterative LS

h
(i),(L)
ILS,n
r
over all subcarriers and dene it over the interval of
L consecutive subcarriers as above
_

_
_
[

h
(i),ILS
1,n
r
]
T
1:L
. . . [

h
(i),ILS
N
T
,n
r
]
T
1:L
_
T
_
[

h
(i),ILS
1,n
r
]
T
k
L1
2
:k+
L1
2

. . . [

h
(i),ILS
N
T
,n
r
]
T
k
L1
2
:k+
L1
2

_
T
_
[

h
(i),ILS
1,n
r
]
T
N
K
L+1:N
K
. . . [

h
(i),ILS
N
T
,n
r
]
T
N
K
L+1:N
K
_
T
(16)
Finally, we calculate the iterative ALMMSE channel estimate
by ltering the iterative LS estimate

h
(i),(L)
ILS,n
r
with the ltering
matrix A
(L)
IALMMSE
as follows

h
(i),(L)
n
r
= A
(L)
IALMMSE

h
(i),(L)
ILS,n
r
,
where the ltering matrix is given by
A
(L)
IALMMSE
=

R
(L)
h
n
r
,h
n
r
_

R
(L)
h
n
r
,h
n
r
+
2
n
_
(

S
H

S)
1
_
(L)
_
1
(17)
Then, we select the k-th subcarrier for the estimated channel
vector

h
(i),(L)
n
r

h
(i)
n
r
,k
=
_

_
_
[

h
(i),(L)
1,n
r
]
T
k
. . . [

h
(i),(L)
N
T
,n
r
]
T
k
_
T
_
[

h
(i),(L)
1,n
r
]
T

L+1
2

. . . [

h
(i),(L)
N
T
,n
r
]
T

L+1
2

_
T
_
[

h
(i),(L)
1,n
r
]
T
L+kN
K
. . . [

h
(i),(L)
N
T
,n
r
]
T
L+kN
K
_
T
(18)
IV. SIMULATION RESULTS
In this section, we discuss the performance of the it-
erative channel estimation algorithms based on simulation
results. We compare the iterative approaches with the pilot
based approaches. Naturally, the question whether to apply
a-posteriori, extrinsic or hard estimated data symbols in the
feedback loop arises. In the following, we investigate the
impact of processing this feedback information for channel
estimation. We distinguish between initial (pilot based), a-
posteriori, extrinsic, hard and perfect cases. In the initial case,
in the plots denoted by It.1, the channel is estimated based only
on pilot symbols. When the system knows the channel state
information perfectly then this is recognized as the perfect case
and denoted as PERFECT. The utilization of the a-posteriori,
extrinsic and hard feedback information in the feedback loop
is denoted (in the front of the abbreviations of the channel
estimation algorithms) by app, ext and hard, respectively.
The channel estimators are compared with each other in terms
of MSE and throughput versus Signal to Noise Ratio (SNR).
All results are obtained using the LTE Link Level Simulator
version r1089 [15], [16]. Table I presents the most important
simulation settings.
Parameter Value
Bandwidth 1.4 MHz
Number of subframes 1000
Number of iterations 5
Number of closest subcarrier L 12 (for ALMMSE)
Number of transmit antennas 2
Number of receive antennas 2
Channel Quality Indicator (CQI) 4,7,10
Modulation order 4-,16-, and 64-QAM
Transmission mode Transmit Diversity (TxD)
Channel type ITU PedB
Detector SSD
TABLE I
SIMULATION SETTINGS
Until otherwise stated, for an iterative ALMMSE estimator
we chose to use L=12 closest subcarrier for the following
simulations. Figure 3 depicts the throughput versus the num-
ber of iterations for the three channel estimation algorithms
when a-posteriori feedback information is utilized. The highest
performance gain is achieved from the rst to the second
iteration. Then, from the second iteration on the throughput is
increased slightly and after the fth iteration the performance
is improved negligibly. Therefore, in the following simulations
only ve iterations between the channel estimator and the de-
coding unit are considered. For further analysis of the iterative
channel estimation algorithms, we consider only the 4-QAM
modulation order. The extension to other modulation orders
is straightforward. For comparison purposes, in Figure 4, the
throughput versus SNR for 4-QAM, 16-QAM and 64-QAM
modulation orders, when a-posteriori feedback information is
utilized, is plotted. As for 4-QAM, approximately the same
performance pattern are also observed for 16-QAM and 64-
QAM.
A. A-posteriori, extrinsic and hard feedback information
In this subsection, we analyse the performance of the system
when the a-posteriori, extrinsic and hard feedback information
are utilized in the feedback loop. Figure 5 depicts the through-
put (analysed at throughput 0.4 Mbit/s) versus SNR, where
the channel estimation algorithms (initial and iterative case)
for the above mentioned feedback information after the fth
iteration are compared to each other. We observe that the best
system performance is achieved when a-posteriori feedback
1 2 3 4 5
0
0.5
1
1.5
2
2.5
Number of iterations
T
h
r
o
u
g
h
p
u
t

[
M
b
i
t
/
s
]


LS
ALMMSE
LMMSE
PERFECT
64-QAM, CQI=10, SNR=10 dB
16-QAM, CQI=7, SNR=4 dB
4-QAM, CQI=4, SNR=-2 dB
Fig. 3. Throughput versus the number of iterations for a-posteriori feedback
information.
-10 -5 0 5 10 15 20
0
0.5
1
1.5
2
2.5
SNR [dB]
T
h
r
o
u
g
h
p
u
t

[
M
b
i
t
/
s
]


LS It.1
app-LS It.5
ALMMSE It.1
app-ALMMSE It.5
LMMSE It.1
app-LMMSE It.5
PERFECT
Fig. 4. Throughput for 4-,16- and 64-QAM modulation orders for a-posteriori
feedback information.
information is utilized in the feedback loop. The improvement
of the a-posteriori case with respect to the extrinsic case is
approximately 0.1 dB for the three iterative channel estimation
algorithms. On the other hand, if we observe the case when
hard feedback information is utilized we see that the iterative
hard LS estimator does not improve with respect to the initial
(pilot based) case, iterative hard ALMMSE estimator obtains
approximately 0.3 dB SNR gain and iterative hard LMMSE
estimator 0.15 dB gain compared to their initial cases. Clearly,
the soft (a-posteriori or extrinsic) case outperforms the hard
case. A-posteriori LS estimator obtains approximately 0.35
dB SNR gain, a-posteriori ALMMSE estimator 0.8 dB and
a-posteriori LMMSE estimator 0.4 dB gain with respect to
the hard case. Obviously, the major drawback utilizing hard
feedback information is the error propagation which is more
severe compared to the soft case. Since the highest system
performance is achieved when a-posteriori information is
utilized, in the following, we compare the channel estimation
algorithms to each other for the a-posteriori feedback infor-
mation case.
-10 -8 -6 -4 -2 0 2 4 6 8 10
0
0.1
0.2
0.3
0.4
0.5
SNR [dB]
T
h
r
o
u
g
h
p
u
t

[
M
b
i
t
/
s
]


-2.5 -2 -1.5 -1 -0.5 0
0.4


LS It.1
ALMMSE It.1
LMMSE It.1
app-LS It.5
app-ALMMSE It.5
app-LMMSE It.5
ext-LS It.5
ext-ALMMSE It.5
ext-LMMSE It.5
hard-LS It.5
hard-ALMMSE It.5
hard-LMMSE It.5
PERFECT
Fig. 5. Comparison for a-posteriori, extrinsic and hard feedback information for the channel estimators in terms of throughput.
B. Comparison of the channel estimation algorithms for a-
posteriori feedback information
First, we investigate the ALMMSE estimator in dependence
of the closest subcarrier index L. In Figure 6 and Figure 7, the
performance of the iterative (for the fth iteration) ALMMSE
estimator for different values of L, and in comparison to LS
and LMMSE estimators, is plotted. It can be observed that
with increasing L the MSE decreases while the throughput
increases. For L = N
K
, the iterative ALMMSE estimator
performance is equal to the iterative LMMSE estimator. It
is obvious that with increasing L also the complexity of the
estimator increases. This fact allows adjusting the performance
and complexity of the estimator to achieve a good trade-off.
We see that even for L=2 consecutive subcarriers the iterative
ALMMSE estimator performs better than the iterative LS esti-
mator. Increasing the number L more than 12 the performance
of the system is increased negligibly as compared from L=2
to L=12. Therefore, we conclude that the choice L=12 is a
good trade-off between complexity and performance.
Figure 8 depicts the MSE versus SNR for the three iterative
algorithms (after the fth iteration) in comparison to their
initial cases. For the three iterative channel estimation algo-
rithms, the MSE considerably decreases after ve iterations
compared to the initial case. The MSE decreases most rapidly
for the LMMSE, followed by the ALMMSE, and then the
LS. Figure 9 depicts the throughput versus SNR. For iterative
LS we observe approximately 0.3 dB SNR gain, for iterative
ALMMSE 1.1 dB and for iterative LMMSE 0.6 dB with
respect to their initial cases. Iterative ALMMSE outperforms
the initial LMMSE for about 0.35 dB and loses 0.25 dB
compared to iterative LMMSE. On the other hand, iterative
LMMSE channel estimator performs approximately equal to
the system with perfect channel knowledge, with an SNR loss
of 0.05 dB. Table II and III summarize in more detail the
above SNR gain and loss values, measured at throughput of
0.4 Mbit/s. Each iterative estimator algorithm is compared with
its initial estimator as well as with the other iterative estimators
and also with the system with perfect channel knowledge.
-10 -5 0 5 10 15 20
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
10
1
SNR [dB]
M
S
E


app-LS It.5
app-ALMMSE L=2 It.5
app-ALMMSE L=3 It.5
app-ALMMSE L=4 It.5
app-ALMMSE L=6 It.5
app-ALMMSE L=12 It.5
app-ALMMSE L=24 It.5
app-ALMMSE L=36 It.5
app-ALMMSE L=72 It.5
app-LMMSE It.5
Fig. 6. Iterative ALMMSE estimator for different values of L compared to
iterative LS and LMMSE estimators in terms of MSE.
-10 -8 -6 -4 -2 0 2 4 6 8 10
0
0.1
0.2
0.3
0.4
0.5
SNR [dB]
T
h
r
o
u
g
h
p
u
t

[
M
b
i
t
/
s
]


app-LS It.5
app-ALMMSE L=2 It.5
app-ALMMSE L=3 It.5
app-ALMMSE L=4 It.5
app-ALMMSE L=6 It.5
app-ALMMSE L=12 It.5
app-ALMMSE L=24 It.5
app-ALMMSE L=36 It.5
app-ALMMSE L=72 It.5
app-LMMSE It.5
PERFECT
-2.5 -2 -1.5 -1 -0.5
0.4


Fig. 7. Iterative ALMMSE estimator for different values of L compared to
iterative LS and LMMSE estimators in terms of throughput.
-10 -5 0 5 10 15 20
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
10
1
SNR [dB]
M
S
E


LS It.1
app-LS It.5
ALMMSE It.1
app-ALMMSE It.5
LMMSE It.1
app-LMMSE It.5
Fig. 8. Comparison of initial (pilot based) and iterative LS, ALMMSE and
LMMSE estimators in terms of MSE for a-posteriori feedback information.
-10 -8 -6 -4 -2 0 2 4 6 8 10
0
0.1
0.2
0.3
0.4
0.5
SNR [dB]
T
h
r
o
u
g
h
p
u
t

[
M
b
i
t
/
s
]


LS It.1
app-LS It.5
ALMMSE It.1
app-ALMMSE It.5
LMMSE It.1
app-LMMSE It.5
PERFECT
-2.5 -2 -1.5 -1 -0.5 0
0.4



Fig. 9. Comparison of initial (pilot based) and iterative LS, ALMMSE
and LMMSE estimators in terms of throughput for a-posteriori feedback
information.
SNR gain values
Iter. versus Init. Init. LS Init. ALMMSE Init. LMMSE
Iter. LS 0.3 dB - -
Iter. ALMMSE 2.3 dB 1.1 dB 0.35 dB
Iter. LMMSE 2.5 dB 1.35 dB 0.6 dB
TABLE II
SUMMARY OF SNR GAIN VALUES OF THE ITERATIVE OVER THE INITIAL
(PILOT BASED) ESTIMATORS
SNR loss values
Iter. versus Iter. Iter. LS Iter. ALMMSE Iter. LMMSE Perfect
Iter. LS - 2 dB 2.25 dB 2.3 dB
Iter. ALMMSE - - 0.25 dB 0.3 dB
Iter. LMMSE - - - 0.05 dB
TABLE III
SUMMARY OF SNR LOSS VALUES OF THE ITERATIVE ESTIMATORS OVER
EACH OTHER AND OVER THE PERFECT CASE
V. CONCLUSION
We showed the calculation procedure of the soft estimated
data symbols for the LTE standard. Furthermore, we derived
different iterative channel estimation algorithms with different
performance behaviours and complexities. We investigated
the impact of processing the a-posteriori, extrinsic and hard
feedback information for the channel estimation. The sys-
tem with iterative channel estimation achieves a considerable
performance improvement in terms of MSE and throughput
compared to the system with non-iterative channel estimation.
The best performance is achieved when a-posteriori feedback
information is utilized. Among the estimators presented in
this paper, the iterative LS estimator achieves the lowest per-
formance. The iterative LMMSE estimator achieves approx-
imately the performance of the system with perfect channel
knowledge with an SNR loss of 0.05 dB. Despite the fact that
the performance of the iterative LMMSE estimator is excellent,
its complexity is very high for a real-time implementation.
Therefore, the goal of iterative ALMMSE estimator is to
reduce the complexity of the iterative LMMSE estimator and
in turn to preserve its performance. The iterative ALMMSE
estimator exploits only the correlation between the L closest
subcarriers. Thus, there is a trade-off between complexity and
performance by varying the parameter L. Iterative ALMMSE
estimator with L=12 obtains approximately 1.1 dB SNR gain
with respect to the initial case and loses approximately 0.25
dB compared to the iterative LMMSE estimator.
ACKNOWLEDGMENTS
The authors would like to thank the LTE research group
for continuous support and lively discussions. This work
has been funded by the Christian Doppler Laboratory for
Wireless Technologies for Sustainable Mobility, KATHREIN-
Werke KG, and A1 Telekom Austria AG. The nancial support
by the Federal Ministry of Economy, Family and Youth
and the National Foundation for Research, Technology and
Development is gratefully acknowledged.
REFERENCES
[1] Christoph Studer, Markus Wenk, Andreas Burg, and Helmut B olcskei,
Soft-output sphere decoding: Performance and implementation as-
pects, in Proc. of Asilomar Conf. on Signals, Systems, and Computers,
Nov. 2006, pp. 20712076.
[2] M.

Simko, Q. Wang, and M. Rupp, Optimal pilot symbol power
allocation under time-variant channels, EURASIP Journal on Wireless
Communications and Networking, vol. 2012, pp. 225, 2012.
[3] M. Simko, D. Wu, C. Mehlf uhrer, J. Eilert, and D. Liu, Implementation
aspects of channel estimation for 3GPP LTE terminals, in Proc.
European Wireless 2011, Vienna, April 2011.
[4] A. Ancora, C. Bona, and D.T.M. Slock, Down-sampled impulse
response least-squares channel estimation for LTE OFDMA, in 2007.
ICASSP 2007. IEEE International Conference on Acoustics, Speech and
Signal Processing, April 2007, vol. 3, pp. III293III296.
[5] S. Omar, A. Ancora, and D.T.M. Slock, Performance Analysis
of General Pilot-Aided Linear Channel Estimation in LTE OFDMA
Systems with Application to Simplied MMSE Schemes, in Proc.
of IEEE PIMRC 2008, Cannes, French Riviera, France, Sept. 2008, pp.
16.
[6] Daejung Yoon and Jaekyun Moon, Low-complexity iterative channel
estimation for turbo receivers, IEEE Transactions on Communications,
vol. 60, no. 5, pp. 11821187, 2012.
[7] Luis

Angel Maestro Ruiz de Temi no, Carles Navarro i Manchon,
Christian Rom, Troels B. Srensen, and Preben E. Mogensen, Iterative
channel estimation with robust wiener ltering in LTE downlink, in
VTC Fall, 21-24 September 2008, Calgary, Alberta, Canada, 2008, pp.
15.
[8] L. Boher, R. Legouable, and R. Rabineau, Performance analysis of
iterative receiver in 3GPP/LTE DL MIMO OFDMA system, in Proc.
IEEE 10th International Symposium on Spread Spectrum Techniques
and Applications 2008 (ISSSTA 2008), Aug 2008, pp. 103108.
[9] 3GPP, Evolved Universal Terrestrial Radio Access (E-UTRA); Physical
channels and modulation, TS 36.211, 3rd Generation Partnership
Project (3GPP), Sept. 2008.
[10] Jianfeng Liu, Hilde Vanhaute, Marc Moonen, Andr e Bourdoux, and
Hugo De Man, Efcient computation of symbol statistics from bit
a priori information in turbo receivers, Trans. Comm., vol. 57, no. 7,
pp. 18891891, July 2009.
[11] J. J. van de Beek, O. Edfors, M. Sandell, S. K. Wilson, and P. O.
Borjesson, On channel estimation in OFDM systems, in Proc. IEEE
45th Vehicular Technology Conference (VTC 1995), 1995, vol. 2, pp.
815819.
[12] Xiaolin Hou, Xueyuan Zhao, Changchuan Yin, and Guangxin Yue,
Unied view of channel estimation in MIMO-OFDM systems, in Proc.
of International Conference on Wireless Communications, Networking
and Mobile Computing, 2005, Sept. 2005, vol. 1, pp. 54 58.
[13] Christian Mehlf uhrer, Sebastian Caban, and Markus Rupp, An accurate
and low complex channel estimator for OFDM WiMAX, in Proc.
Third International Symposium on Communications, Control, and Signal
Processing (ISCCSP 2008), St. Julians, Malta, Mar. 2008, pp. 922926.
[14] M. Noh and Y. Lee, Low complexity LMMSE channel estimation for
OFDM, IEEE Proceedings-communications, vol. 153, 2006.
[15] C. Mehlf uhrer, J. Colom Ikuno, M.

Simko, S. Schwarz, M. Wrulich, and
M. Rupp, The Vienna LTE Simulators - Enabling Reproducibility in
Wireless Communications Research, EURASIP Journal on Advances
in Signal Processing, pp. 113, 2011.
[16] C. Mehlf uhrer, M. Wrulich, J. Colom Ikuno, D. Bosanska, and M. Rupp,
Simulating the Long Term Evolution Physical Layer, in Proc. of
EUSIPCO 2009, Glasgow, Scotland, Aug. 2009.