EM-BASED SEMI-BLIND MIMO-OFDM CHANNEL ESTIMATION

Abdelhamid Ladaycia1 , Adel Belouchrani2 , Karim Abed-Meraim3 , Anissa Mokraoui1

1
L2TI, Institut Galilée, Université Paris 13, Sorbonne Paris Cité, France
2
Electrical Engineering Department/LDCCP, Ecole Nationale Polytechnique, Algeria
3
PRISME, Université d’Orléans, France
{abdelhamid.ladaycia, anissa.mokraoui}@univ-paris13.fr, adel.belouchrani@enp.edu.dz, karim.abed-meraim@univ-orleans.fr

ABSTRACT in term of the channel parameters to be estimated. Instead of

This paper deals with semi-blind (SB) channel estimation
of Multiple-Input Multiple-Output Orthogonal Frequency-
Division Multiplexing (MIMO-OFDM) wireless communi-
cations system in the uplink transmission. Herein, we propose
a new channel estimation approach using the well known EM
technique. More precisely, we derive first the SB EM algo-
rithm in the MIMO case. Then, a parallelizable version of
this algorithm is introduced relying on the decomposition of
the MIMO system into several MISO systems. Finally, we
propose a reduced cost EM version where only the lattice
points in the neighboring of the pilot-based detected symbols
are considered.
Index Terms— MIMO-OFDM, MISO-OFDM, EM algo-
rithm, Channel estimation, Cramèr Rao bound.
1. INTRODUCTION
With advances in technology, demand for high-speed Inter-
net and mobile communications is becoming increasingly
important. A Multiple-Input Multiple-Output Orthogonal
Frequency-Division Multiplexing (MIMO-OFDM) wireless
communications system provides many advantages as the
channel capacity enhancement and the improvement of the
communication reliability. However to achieve good perfor-
mance, the receiver should have an accurate channel state
information (CSI).
Many channel estimation approaches have been devel-
oped and can be divided into three main classes: blind [1],[2],
pilot-based [3] and semi-blind methods [4]. For the latter
class, it is shown in [5] that the SB approach improves the
throughput by reducing the training sequences up to 95%.
Furthermore, in [6], semi-blind approaches are used to re-
duce the transmitted power (’green communications’).
The objective of this paper is to propose an efficient semi-
blind channel estimation for MIMO-OFDM system based on
the EM algorithm. Indeed, the latter is a widely used tech-
nique that has been already considered in the literature for
the MIMO blind channel estimation. In particular, Our EM-
based algorithm is distinct from the previous ones ([7], [8])
estimating the subcarrier channel coefficients, we estimate di-
rectly the channel taps so one can obtain a significant gain as
analyzed in [9].
Furthermore, we propose two EM-based algorithms. The
first one considers the MIMO-OFDM block as one system
to estimate one channel vector through the iterative process.
The second one, important for the case of parallel processing
machine, decomposes the MIMO-OFDM system into parallel
MISO-OFDM systems to estimate the different vector chan-
nel taps independently for each receiver. In the case of un-
derdetermined system (i.e. number of transmitters is greater
than the receivers one), where the traditional methods could
not estimate the transmit data, we succeed through the two
proposed EM-based algorithms to estimate the channel taps
and data properly.
Another originality of this work, consists of a simplified
EM-based, denoted S-EM, method that allows to reduce the
computational heaviness based on an initial estimation of the
channel and the data using the pilots.
2. DATA MODEL
We consider a MIMO system composed of Nt transmit an-
tennas and Nr receive antennas. The transmitted signal is
assumed to be an OFDM one, composed of K samples (sub-
carriers) and L Cyclic Prefix (CP) samples. The CP length
is assumed to be greater or equal to the maximum multipath
channel delay denoted N (i.e. N ≤ L). After removing the
CP and taking the K-point DFT, the received signal at the k-
th sub-carrier by the r-th receive antenna, denoted yr (k), is
given by:
Nt N
X X −1
nk
yr (k) = hri (n)wK di (k)+vr (k) 0 ≤ k ≤ K−1,
i=1 n=0
(1)
where di (k) is the transmitted data at subcarrier k by the
i-th transmitter. The noise v is assumed to be additive
independent
h white
i Circular Complex Gaussian satisfying
H
E v(k)v(i) = σv2 IK δki ; (.)H being the Hermitian op-

978-1-5386-4658-8/18/$31.00 ©2018 IEEE 3899 ICASSP 2018

erator; σv2 the noise variance; IK the identity matrix of size 3.2. MIMO channel estimation
K × K and δki the Kronecker symbol. hri (n) represents the
Under the data model assumption, the likelihood function is
n-th channel coefficient between the i-th transmitter and the
nk expressed by:
r-th receiver. wK represents the (n, k)-th coefficient of the
K-DFT point matrix. Kp −1 K−1
p (y; θ) = Π p (y (k) ; θ) Π p (y (k) ; θ) , (7)
Equation (1) can be rewritten in the following matrix form: k=0 k=Kp

yr (k) = wT (k)Hr d(k) + vr (k), (2) where p (y (k); θ) ∼ N W(k)Hdp (k), σv2 I , for


T
where d(k) = [d1 (k), · · · , dNt (k)] , k = 0, · · · , Kp − 1, dp (k) being the pilot vector at the k-th
h
k (N −1)k
iT sub-carrier, and
and w(k) = 1 wK , · · · , wK . Hr is given by
|D|
  X
hr1 (0) ··· hrNt (0) p (y (k) ; θ) = p (y (k)|dξ ; θ) p (dξ ), (8)
Hr =  .. .. .. ξ=1
. (3)
 
. . .

hr1 (N − 1) · · · hrNt (N − 1) with p (y (k)|dξ ; θ) ∼ N W(k)Hdξ , σv2 I
When considering all the outputs in a single vector y(k) =
T 3.2.1. E-step
[y1 (k), · · · , yNr (k)] , the previous model can be written in  
the following compact form: After some straightforward derivations, Q θ, θ [i] can be
y (k) = W(k)Hd(k) + v(k), (4) written as:
T  KP p −1
where W(k) = INr ⊗ w (k) (⊗ being the Kronecker prod- 
uct) and H = [HT1 , · · · , HTNr ]T Q θ, θ [i] = log p (y(k)|dp (k); θ)
k=0
In the sequel the received OFDM symbols are assumed (9)
P |D|
K−1 P  
[i]
to be i.i.d and arranged according to the comb-type scheme + αk,ξ θ log p (y(k)|dξ ; θ),
k=Kp ξ=1
with Kp sub-carriers dedicated to pilots corresponding (upon
appropriate permutation) to k = 0, · · · , Kp − 1 and Kd data where
sub-carriers. Also, we denote by D (respectively |D|) the  
finite set of all possible realizations of the data vector d (re-
  p y (k) |dξ ; θ [i] p (dξ )
αk,ξ θ [i] = . (10)
spectively its cardinal). |D|
P  [i]

p y (k) |dξ0 ; θ p (dξ0 )
ξ 0 =1
3. ML CHANNEL ESTIMATION
In this paper, all the realizations dξ are equi-probable and
In the sequel, the unknown parameters are grouped in θ con- hence the term p (dξ ) can be ignored in equation (10).
taining the channel taps (vec (H) or vec (Hr )) and the noise
power σv 2 (for simplicity, the signal power is assumed to be 3.2.2. M-step
known).
The goal of the M-step is to find θ, i.e. the channel matrix H
and the noise power σv2 that maximizes the auxiliary function:
3.1. EM algorithm  
θ [i+1] = arg max Q θ, θ [i] . (11)
The EM algorithm looks for finding the ML estimate of the θ
unknown parameters using the marginal likelihood of the ob-  
served data y by an iterative optimization process. By setting to zero the derivative of Q θ, θ [i] in (9) w.r.t.
Considering y as the incomplete data and d as the missing H, one obtains:
data, the EM-algorithm is based on the two following steps:  
"
KPp −1  
[i+1] ∗ H
• Expectation step (E-step): vec H = dp (k) dp (k)T ⊗ W(k) W(k)
  k=0
Q θ, θ [i] = Ed|y,θ[i] [log p (y|d, θ)]  −1
#
(5) K−1 |D|  
H
P P [i] ∗ T
+ αk,ξ θ dξ dξ ⊗ W(k) W(k)
• Maximization step (M-step): "k=Kp ξ=1
KPp −1  
  H H
θ [i+1] = arg max Q θ, θ [i] (6) × vec W(k) yp (k) dp (k)
θ k=0
#
This process is proven in [10] to increase the ML value, i.e P |D|
K−1   
H

αk,ξ θ [i] vec W(k) y (k) dξ H .
P
+
p (y|d, θ), and hence leads to the algorithm’s convergence to k=Kp ξ=1
a local maximum point. (12)

3900
 Parameters Specifications
y1 (k ) EM H1 ,  12

Equalization/
Number of pilot sub-carriers Kp = 8

Decision
. . . .
. . . .

....
dˆ (k ) Number of data OFDM symbols Nd = 16
Number of data sub-carriers Kd = 56
y Nr (k ) EM H Nr ,  N2 r Pilot signal power σp2 = 13 dBm
Data signal power σ 2d = 10 dBm
Fig. 1: MIMO-OFDM system model using Nr parallel Number of sub-carriers K = 64
MISO-OFDM systems.
Table 1: Simulation parameters.
 
Similarly, setting to zero the derivative of Q θ, θ [i]
3.4. Simplified EM algorithm
w.r.t. σv 2 leads to:
The computational heaviness in equations (12) and (13) is due
p −1
KP 2
{σv } 2 [i+1]
= 1
yp (k) − W(k)H [i+1]
dp (k) to the summation taking all the possible realizations of the
K
k=0 data vector d (i.e. |D|). In this subsection we propose a sim-
plifying method to reduce the summation set from |D| (which
!
P |D|
K−1   2
αk,ξ θ [i] y (k) − W(k)H[i+1] dξ
P
+ . growth exponentially with the number Nt ) to another reduced
k=Kp ξ=1
summation set of size |D0 | proportional to Nt .
(13) The proposed approach is summarized in Fig. 2, where
The algorithm can be summarized as below: we use the Decision Feedback Equalizer technique (DFE) to
 
[0]

[0]
T
2 [0]
re-estimate the channel using the EM-based algorithm. The
Step 1 : Initialization θ = vec H , {σv } which first step consists of estimating the data d̂d using only pilots
represent the standard pilot-based channel and noise es- to estimate ĥop followed by Zero Forcing Equalizer (ZF) and
timates; hard decision. Using d̂d , the summation in equations (12) and
(13) is done on a reduced size set |D0 | corresponding to the
Step 2 : Estimate H[i+1] using H[i] and {σv2 }[i] according to
neighborhood of d̂d defined here as the points differing from
equation (12);
d̂d by at most one entry.
Step 3 : Estimate {σv2 }[i+1] using H[i+1] , H[i] and {σv2 }[i]
according to equation (13); dp
[i+1] y hˆ op dˆ d
hˆ SSB− EM
Step 4 : repeat from step 1 using θ ; LS Equalization + Simplified EM
Channel Estimation Decision Algorithm

Remark: In the case of blind channel estimation, either in the

MIMO or MISO approaches, one can ignore the pilot’s terms
in equations (12) and (13) and take into account only the data
OFDM symbols.
In the sequel, the MIMO system is sub-divided into Nr
parallel MISO systems, for which the EM is applied in a par-
allel scheme.
3.3. MISO channel estimation
Equations (2) and (3), allow the parallel decomposition of
the MIMO system into Nr MISO-OFDM systems, as illus-
trated in Fig. 1. The estimation of the global parameters of the
MIMO-OFDM system is done by concatenating the parame-
ters of the Nr parallel MISO-OFDM system. The parameters
of the r-th MISO-OFDM system are denoted as:
h i
T
θ r = vec (Hr ) , σv2 r (14)
The computation of Hr and σv2 r , using the EM algorithm,
leads to the same expressions as in the MIMO case given in
the previous subsection where H and W are substituted by
Hr and w, respectively.
Fig. 2: Simplified EM algorithm.
4. SIMULATIONS RESULTS
Herein, we analyze the performance of the EM blind and
semi-blind channel estimators in terms of the normalized
Root Mean Square Error (NRMSE) for the two system
configurations presented in this paper i.e MIMO system
(hEM EM
SB (M IM O) and hB (M IM O)) and parallel MISO
systems (hSB (M ISO//) and hEM
EM
B (M ISO//)).
A (2 × 2) MIMO-OFDM communications system with
BPSK data is considered. The IEEE 802.11n training se-
quence and channel model are used in this paper as pilots [11].
Simulation parameters are summarized in Table 1, where the
used IEEE 802.11n channel model is of type B with path de-
lay [0 10 20 30] µs and an average path gains of [0 -4 -8 -12]
dB.
Fig. 3 provides a comparison between the two algorithms
presented in this paper and the performance limits defined
by the Cramèr Rao bound CRB, detailed in [5] versus the

3901
SN R. The curves show clearly that both blind and semi- −1
Nd=16, (4×2) MIMO
10
blind channel estimation algorithms give better results than CRBSB
the case where only pilots are used to estimate the channel, hLS
OP

i.e the least square estimator (hLS

OP ). Also, the semi-blind so-
10
−2
hEM
SB
(MIMO)

lutions show the best results (close to the CRB) with a slight hEM
B
(MIMO)

advantage in favor of the MIMO version. −3 hEM

SB
(MISO//)
10
Fig. 4 represents the simulation results in the case of 4×2 hEM(MISO//)

NRMSE
B

underdetermined MIMO system, where we can see that, even −4

10
in this configuration the EM-based algorithms perform well.
For a given SN R = 10 dB (around the operating mode
of the IEEE 802.11n), Fig. 5 presents the behavior of EM 10
−5

algorithms when increasing the number of data OFDM sym-

bols (Nd ). The analysis of the curve confirms that when the −6
10
number of data OFDM symbols increases, the performance −5 0 5 10 15 20
SNR (dB)
of the EM algorithm in the blind and semi-blind approaches
improves. One can see that in the MISO case one has taken Fig. 4: NRMSE of the EM algorithms versus SNR in the un-
advantage of the parallel computational architecture but at the derdetermined case (Nt > Nr ).
cost of a reduced channel estimation quality as compared to
the MIMO case. SNR=10 dB, (2×2) MIMO
−3
10
In Fig. 6, we present the effect of the used approximation
on the performance of the S-EM algorithm. We can see that
the degradation is relatively minor for a computational gain
that is significant.
−4
10

5. CONCLUSION
NRMSE

hLS
OP
This paper introduced the EM based blind and semi-blind −5
CRBSB
10
channel identification in MIMO-OFDM wireless communica- hEM
SB
(MIMO)
tions system. Through the simulation results, we have shown hEM
B
(MIMO)
that the EM (blind or semi-blind) performs better than the pi- hEM
SB
(MISO//)

lot based channel estimation method. The decomposition of −6

hEM
B
(MISO//)
10
the MIMO system into Nr parallel MISO systems gives good 0 10 20 30 40 50 60 70
Nd number of OFDM symbols
results and is suitable for parallel machine processing allow-
ing reduction of the EM execution time. The latter, is further Fig. 5: NRMSE versus the number of OFDM symbols (Nd ).
reduced by a simplifying method employing a first estimation
of the channel and transmitted data using the pilots.
Nd=16, (4×4) MIMO
−2
10
Nd=16, (2×2) MIMO
−1
10 CRBSB
CRBSB
−3 hLS
OP
hLS 10
OP
hS−EM
−2
10
hEM
SB
(MIMO) SB

hEM
NRMSE

hEM
B
(MIMO) −4 SB
−3 10
10
hEM (MISO//)
NRMSE

SB

hEM
B
(MISO//)
−4 −5
10 10

−5
10
−6
10
−5 0 5 10 15 20
SNR (dB)
−6
10
−5 0 5 10 15 20
SNR (dB)
Fig. 6: S-EM-algorithm performance
Fig. 3: NRMSE of the EM algorithms versus SNR .

3902
 6. REFERENCES

[1] K. Abed-Meraim, W. Qiu, and Y. Hua, “Blind system

identification,” Proceedings of the IEEE, vol. 85, no. 8,
pp. 1310–1322, Aug 1997.
[2] C. Shin, R. W. Heath, and E. J. Powers, “Blind chan-
nel estimation for MIMO-OFDM systems,” IEEE Trans.
on Vehicular Technology, vol. 56, no. 2, pp. 670–685,
March 2007.
[3] O. Simeone, Y. Bar-Ness, and U. Spagnolini, “Pilot-
based channel estimation for OFDM systems by track-
ing the delay-subspace,” IEEE Transactions on Wireless
Communications, vol. 3, no. 1, pp. 315–325, Jan 2004.
[4] W. Yang, Y. Cai, and Y. Xun, “Semi-blind channel es-
timation for OFDM systems,” in 2006 IEEE 63rd Ve-
hicular Technology Conference, vol. 1, May 2006, pp.
226–230.
[5] A. Ladaycia, A. Mokraoui, K. Abed-Meraim, and
A. Belouchrani, “Performance bounds analysis for
semi-blind channel estimation in MIMO-OFDM com-
munications systems,” IEEE Transactions on Wireless
Communications, vol. 16, no. 9, pp. 5925–5938, Sept
2017.
[6] A. Ladaycia, A. Mokraoui, K. Abed-Meraim, and
A. Belouchrani, “Toward green communications us-
ing semi-blind channel estimation,” in Proc. EUSIPCO
2017.
[7] G. Hu and D. Li, “EM-based channel estimation for
MIMO OFDM system,” in 2009 International Confer-
ence on Networks Security, Wireless Communications
and Trusted Computing, vol. 1, April 2009, pp. 159–
162.
[8] S. Park, J. W. Choi, J. Y. Seol, and B. Shim,
“Expectation-maximization-based channel estimation
for multiuser MIMO systems,” IEEE Transactions on
Communications, vol. 65, no. 6, pp. 2397–2410, June
2017.
[9] A. Ladaycia, A. Mokraoui, K. Abed-Meraim, and
A. Belouchrani, “Further investigations on the perfor-
mance bounds of MIMO-OFDM channel estimation,” in
2017 13th International Wireless Communications and
Mobile Computing Conference (IWCMC), June 2017,
pp. 223–228.
[10] N. M. L. A. P. Dempster and D. B. Rubin, “Maxi-
mum likelihood from incomplete data via the EM al-
gorithm,” Journal of the Royal Statistical Society. Series
B (Methodological), vol. 39, no. 1, pp. 1–38, 1977.
[11] R. S. E. Perahi, Next Generation Wireless LANS. New
York, USA: Cambridge University Press, 2008.

3903