Joint Channel Estimation and Decoding Using

Gaussian Approximation in a Factor Graph

over Multipath Channel
Yang Liu∗† , Loı̈c Brunel∗ , and Joseph J. Boutros‡
∗ Mitsubishi Electric ITE-TCL, 1 allée de Beaulieu, 35700 Rennes, France - l.brunel@fr.merce.mee.com
† ENST, 46 rue Barrault, 75013 Paris, France - yliu@enst.fr
‡ Texas A&M University at Qatar, PO Box 23874, Doha, Qatar - boutros@tamu.edu

Abstract—Joint channel estimation and decoding using be-
lief propagation on factor graphs requires the quantization of
probability densities since continuous parameters are involved.
We propose to replace these densities by standard messages
where the channel estimate is accurately modeled as a Gaussian
mixture over multipah channel. Upward messages include symbol
extrinsic information and downward messages carry mean values
and variances for the Gaussian modeled channel estimate. Such
unquantized message propagation leads to a complexity reduction
and a performance improvement. Over multipath channel, the
proposed belief propagation almost achieves the performance of
iterative APP equalizer and outperforms MMSE equalizer.
I. I NTRODUCTION
Propagating messages in a suitable factor graph [1] is
a systematic tool for deriving iterative algorithms. Among
various receiver issues solved using the belief propagation
algorithm (BP), also called sum-product algorithm [2], we
can cite decoding, channel estimation, synchronization and
detection [3]. [4] presents a BP handling continuous variables,
in which canonical distributions are used for quantizing prob-
ability distributions, in order to propagate discrete probability
distributions. However, the degree of quantization has a strong
impact on estimation accuracy and performance. Even adapt-
ing the quantization in each iteration of BP, as proposed in [5]
and [6], does not fill the complexity gap between BP and other
However, the quantization method will make BP unfeasible
over multipath channel. Thus, the proposed Gaussian approx-
imation is required.
The paper is structured as follows. Section II explains how
the transmission system is modeled using a factor graph and
how BP is applied. Section III presents the approximation of
the distribution of channel estimates over multipath channel in
BP by a mixture of Gaussian distributions. In Section IV, APPs
are computed from the approximated distribution. Continuous
upward messages in the factor graph are presented in Sec-
tion V. The paper ends with simulation results in Section VI.
In the sequel, messages that are not based on quantized
densities will be referred as continuous messages.
II. S YSTEM M ODEL AND FACTOR G RAPH
We consider a coded system with transmission over a
single-input single-output (SISO) multi-path channel as shown
in Fig. 1. An information binary sequence bi is encoded,
h
s Encoder c x y Channel Estimation ŝ
& Mapper ISI &
Interleaver Channel Decoding


2
n ∼ CN 0, 2σ

algorithms. Instead of relying on quantization, we propose

Fig. 1. System model.
here to model probability distributions as mixtures of Gaus-
sian distributions. It allows for estimation improvement and
complexity reduction simultaneously. The BP with Gaussian interleaved and modulated into N BPSK symbols xk . After
approximation over single path channel has been presented convolution with an impulse response made of L taps, i.e.,
in [7]. However, in many practical communication systems, L complex Gaussian channel coefficients hl ∼ CN (0, 1/L),
symbols are transmitted over a channel with intersymbol


and addition of a complex Gaussian noise nk ∼ CN 0, 2σn2 ,
interference (ISI). In this paper, we focus on BP with Gaussian the channel outputs yk are processed by a receiver performing
approximation over multipath channel. joint channel estimation and decoding. Finally, the receiver
Over ISI channel, received symbols are usually processed by outputs the estimated information sequence bˆi . The system
an equalizer in the receiver. A number of important equalizers model is described by
have been presented in former works, including iterative
L−1
a posteriori probability (APP) equalizer [8] and minimum X
mean square error (MMSE) equalizer [9]; however, all these yk = hl xk−l + nk , 0 ≤ k ≤ N − 1. (1)
l=0
equalizers have to work together with a channel estimator
to obtain the channel coefficients. A factor graph with BP We re-write (1) in matrix form:
can help defining the iterative receiver in a systematic way
and implementing joint channel estimation and decoding. yk = xTk h + nk , (2)

978-1-4244-5213-4/09/ $26.00 ©2009 IEEE 3164

 T
where xk = (xk , xk−1 , · · · , xk−L+1 ) represents the symbol µx →f
k k+1
T
vector at time instant k and h = (h0 , · · · , hL−1 ) represents µf →x
ξk,1
µf
the ISI channel. k k k+2 →xk
g
p (g)
xk
g
p (g) µf →g ··· ··· µf
0 µf →g N −1 →g exk
k

(a) (b)
µf →g
y0 f0 y1 f1 y2 f2 y3 f3 y4 f4 k
···

yk fk
x0 x1 x2 x3 x4
··· µx µx →f g
k−2 →fk k k p (g)
= ξk−2,2 = ξk,0
µf →g
µx 0 ··· ···
µf
N −1 →g
k−1 →fk µf
k−1 →g
µf
k+1 →g
= ξk−1,1 µg→f
k
CODE
(c) (d)

Fig. 2. Factor graph for multipath channel. µg→f

k
µf
k+1 →xk
yk fk
The corresponding factor graph for 3 taps is built in Fig. 2 µf →x
k k
µf
k+2 →xk
following [4], where g is a quantized estimate of h and p (g) µx
k−2 →fk
µx →f
k k xk
represents the quantized distribution of the known a priori of = ξk−2,2 = ξk,0 ` ´
µf →x P xk
g. In this paper, we take the estimate from pilots as the a priori k k−1
of g: (e) (f)
Lp −1
Y Fig. 3. Message propagation in factor graph.
p (g) , pp (g) = µfp,k →g , (3)
k=0
as shown in Fig. 3(a). For upward messages, in node CODE, a
forward-backward algorithm computes the extrinsic informa- III. D ISTRIBUTION OF C HANNEL E STIMATE
tion for each deinterleaved coded bit. Taking interleaving into In the iterative receiver, initial estimate is obtained from
account, the extrinsic information exk is propagated to nodes known pilots and subsequent estimates from data symbols.
xk . In node xk , the message µxk →fk+l = ξk,l is obtained by Thus, the distribution of channel estimate will differ depending
multiplying all messages into xk : on the iteration.
L−1
Y
µxk →fk+l = ξk,l = exk µfk+i →xk , (4) A. Estimation based on pilots
i=0 Lp pilots xp,k (0 6 k 6 Lp − 1) are included in the trans-
i6=l
mitted sequence. From the Lp messages µfp,k →g and (3), we
as shown in Fig. 3(b). From each node fk to node g, a discrete get the discrete distribution of g [2]:
distribution µfk →g of the quantized estimate of h is computed
q Lp −1
−1
LX
!
and propagated based on a marginalization of the likelihood Y |yp,k − xTp,k gc |2
p (yk |xk , g) with respect to the transmitted symbol xk , as pp (g) ∝ δ (g − gc ) exp − ,
c=0
2σn2
shown in Fig. 3(c). k=0
For downward messages, the message µg→fk is calculated (6)
as shown in Fig. 3(d): where
L−1
N −1 X
yp,k = hl xp,k−l + np,k ; (7)
Y
µg→fk = pp (g) µfi →g . (5)
i=0 l=0
i6=k
T
T
xp,k = (xp,k , · · · , xp,k−L+1 ) ; gc = g0c , · · · , gL−1
c
is
By multiplying message µg→fk and all messages from
a quantization codebook of size Lq for channel estimate’s
xk,l = (xk , xk−1 , · · · , xk−l+1 , xk−l−1 , · · · , xk−L+1 ) into fk
probability density function (pdf) and δ (·) denotes the Dirac
(Fig. 3(e)), the APP of each transmitted symbol xk−l is com-
delta function. Using a constant amplitude zero autocorrelation
puted, marginalizing the likelihood p (yk |xk , g) with respect to
(CAZAC) sequence as pilots [10], (6) can be approximated as
g and xk,l . The final APP of each coded bit P (xk ) is obtained
q
by multiplying all messages from node fk to xk (Fig. 3(f)) LX −1 
Lp Ep

2
and then propagated to node CODE. The whole process of pp (g) ∝ δ (g − gc ) exp − 2 |gc − h| , (8)
2σn
propagating upward and downward messages is then iterated. c=0

3165
 j
where Ep represents the pilot energy. Hence, pp (g) can be 2) With low SNR, all probabilities ξi−l,l are close to 1/2 for
2σn2 BPSK modulation. In (10), there are 2(N −1)L items. However,
approximated as one Gaussian distribution CN (h, ).
Lp Ep when considering the first exponential item in (10), only the
2
2
B. Estimation based on data sequences with Ulj − Vlj ≥ (N − 1) , i.e., with Ulj =
T
Let si,m = (si,0,m , si,1,m , · · · , si,L−1,m ) , where 0 ≤ m ≤ N − 1 or Vlj = N − 1, are not close to zero. Therefore, there
L
2 − 1, represent the mth possible symbol vector and ξi−1,l m are only 2L dominant terms:
represent the probability that xi−l = si,l,m . The product q
N −1 LX −1 L−1
Y  
Y (N − 1) c 2
µfi →g can be expressed as pd,k (g) [2]: pd,k (g) ∝ δ (g − gc ) βl exp − |gl − h l |
c=0
2σn2
i=0 l=0
i6=k
 
(N − 1) c 2
+ (1 − βl ) exp − |gl + hl | ,
N
Y −1 q
LX −1 2σn2
pd,k (g) ∝ µfi →g ∝ δ (g − gc ) × (12)
i=0 c=0
i6=k j
L  where βl denotes the normalized product of ξi−l,l ; for a single
−1 2X−1
N
!
Y |yi − sTi,m gc |2 L−1
Y
m
 channel tap, the distribution pd,k (gl ) is:
exp − ξi−l,l
i=0

m=0
2σn2  q−1
l=0   
X (N − 1) c 2
i6=k
pd,k (gl ) ∝ δ (gl − glc )
βl exp − |gl − hl |
q
LX −1
c=0
2σn2
∝ δ (g − gc ) × 
(N − 1) c

2
c=0 + (1 − βl ) exp − |gl + h l | . (13)
  2σn2
(N −1)L
2 X −1  1 NX −1
 NY−1 L−1
From (11), (12) and (13), for each channel tap, the pdf
jT
Y j
2
exp  −
 2σ 2 |y i − si g c | ξi−l,l ,
j=0 n i=0

i=0 l=0
pd,k (gl ) can be approximated as a mixture of two Gaussian
i6=k i6=k distributions; for the whole ISI channel, the pdf pd,k (g) can be
(9) approximated as a mixture of multiple Gaussian distributions
 T which are the product of all pdfs of each tap with variance
where sji = sji,0 , sji,1 , · · · , sji,L−1 is the value of symbol 2σn2
.
j
xi in sequence j and ξi−l,l is the probability that xi−l equals N −1
sji,l . After some calculations and approximations, (9) can be
IV. APP E VALUATION FROM D OWNWARD M ESSAGES
approximated as
q
LX −1 With the conclusions in Section III, the known a priori
pd,k (g) ∝ δ (g − gc ) × discrete channel distribution pp (g) can be approximated as
c=0 one Gaussian distribution, and the discrete distribution of the
    2  N −1
Ulj − Vlj
(N −1)L
X −1 L−1
2   Y
Y  (N − 1)   
  product µfi →g = pd,k (g) can be approximated as a
c
exp − 2
 gl − h l  ×

i=0
j=0
  2σ n
 (N − 1) i6=k
 l=0    
   2 
 mixture of multiple Gaussian distributions, where pp (g) uses
 (N − 1) |h |2
 j
Ul − Vl j
 N −1
 

 pilots and pd,k (g) uses the messages in the current iteration.
l  Y j Furthermore, together with the conclusion in [7], for each tap,
exp − 1 − ξi−l,l  ,

2σn2
 2 

 (N − 1) 
 i=0 
 there is always one dominant Gaussian distribution (with mean
i6=k value hl ). Hence, when calculating APP, we consider only
(10) the dominant one (βl = 1). Then, the discrete distributions
where Ulj (resp. Vlj ) is the number of items with sji,l x∗i−l = +1 of pp (g) and pd,k (g) can both be reduced to L pairs of
2 2
parameters: (ĥp,l , σ̂hp ) for pp (g) and (ĥd,k,l , σ̂hd ) for pd,k (g).
(resp. sji,l x∗i−l = −1) in sequence j.
Thus, pp (g) times pd,k (g) can also be approximated by a
mixture of Gaussian distributions, i.e., the discrete distribution
1) With high SNR, the decoder almost provides perfect
of message µg→fk can be reduced to L pairs of parameters
extrinsic information. Thus, for a single sequence j with
(ĥk,l , σ̂h2 ), denoted as pk (g). Obviously, (ĥk,l , σ̂h2 ) can be
Ulj = N − 1, all ξi−l,l
j
→ 1 and other terms are null: 2 2
calculated from (ĥp,l , σ̂hp ) and (ĥd,k,l , σ̂hd ) that will be shown
q
LX −1 
(N − 1)
 in the following part. Thus, we can calculate each downward
2
pd,k (g) ∝ δ (g − gc ) exp − |g c − h| ; message µfk →xk−l in a continuous way, instead of computing
2σn2
c=0 it for each codebook value gc , and then marginalizing with
(11) respect to g. It reduces the computation complexity.

3166
 N −1 L−1
With the discrete way, the probability of symbol vector xk Y Y j
where ∆j = ξi−l,l . Thus, we can get the distribution
can be calculated as:
i=0 l=0
q
LX −1   i6=k
1  2 of gl from (18):
P (xk = sk,m ) = exp − 2
yk − sTk,m gc  P (gc ) .
c=0
2σn Z Z
0

(14) pd,k (gl ) ∝ · · · pd,k (g) dg , (19)

| {z }
Equation (14) can be written in a continuous way as: g0
 2  0
Z Z  1 

L−1
X   where g = (g0 , · · · , gl−1 , gl+1 , · · · , gL−1 ). After some com-
P (xk = sk,m ) = · · · exp − 2 yk −

sk,l,m gl 
 plex derivation and approximation, the pdf pd,k (gl ) can be
 2σn    approximately written as
| {z } l=0
L
−1)L
L−1 2(N X −1  
Y 1 2 j
pk (gl ) dg0 · · · dgL−1 . (15) pd,k (gl ) ∝ ∆j exp − 2 |gl | Ωk,l ×
j=0
2σn
l=0   
  
With some calculations, we have 

 1

 L−1 Φ j 


0
 
k,l
X
∗ j ∗ j
1 exp 2<e g Φ
l k,l − gl R 0 , (20)
P (xk = sk,m ) ∝ ×  2σn2  0 Ωjk,l0 k,l ,l 
l =0
  
L−1 
 
 

X 2 l0 6=l
σ̂h2 |sk,l,m | + σn2
l=0 N −1 N −1 
 j 2
X X 
where Φjk,l = yi sj∗ j j
 

  2 
 i,l , Ωk,l = si,l  and Rk,l0 ,l =
i=0 i=0
 
 T
yk − sk,m ĥk 
   

  i6=k i6=k
exp − L−1
! , (16) N
X −1
sji,l0 sj∗
i,l . Using (20) and considering normalization, we
 X 
 2 2 2

 2 σ̂h


 |sk,l,m | + σn  

 i=0
l=0 i6=k
 T get
where ĥk = ĥk,0 , · · · , ĥk,L−1 . According to Fig. 3(e) and Z  
Fig. 3(f), we have gl pd,k (gl ) d gl L−1
XΦ
g 1  e 0
k,l e

ĥd,k,l = Zl ≈ Φk,l − Rk,l0 ,l  ,
e 
P (xk = bi ) ∝ pd,k (gl ) d gl Ω
e k,l  0 Ω
e 0
k,l
l =0

L−1 L−1 gl l0 6=l
Y X Y
P (xk+l = sk+l,m ) µxk+l−l0 →fk+l . (17) (21)
l=0 sk+l,m,l =bi 0
l =0 where
l0 6=l
N
X −1 X
Thanks to the computation in (16), a single APP computation Φ
e k,l = yi s∗i,l,m ξi,l
m
, (22)
instead of Lq computations is performed for each symbol i=0 m
vector xk with the Gaussian approximation. Thus, the global i6=k

complexity is much reduced by the Gaussian approximation

in the downward messages. −1
N
! !
X X X
e 0 =
Rk,l ,l
m
si,l0 ,m ξi,l 0 s∗i,l,m ξi,l
m
(23)
V. E STIMATION FROM U PWARD M ESSAGES i=0 m m
i6=k

In order to improve the performance of the Gaussian ap- e k,l ≈ (N −1)Eav . Here, Eav represents the average power
and Ω
proximation, we propose to increase the accuracy of ĥk using
of transmitted symbols.
a continuous upward message.
Following the same steps, the estimation with a CAZAC
Replacing the discrete distribution in (9) by an integral, we
pilot sequence can be calculated as
get continuous pd,k (g):
Lp −1
  1 X
(N −1)L
2 X −1

 1 NX −1 
 ĥp,l ≈ yp,k xp,k−l . (24)
 2 
jT 
 Lp Ep
k=0
pd,k (g) ∝ exp − 2 yi − si g ∆j ,

j=0

 2σn i=0 
  T
 
i6=k With (21), we can get ĥd,k = ĥd,k,0 , · · · , ĥd,k,L−1 for
(18) pd,k (g) by using the messages in current iteration; with (24),

3167
  T -1
2 10
we can get ĥp = ĥp,0 , · · · , ĥp,L−1 . Together with σ̂hd and
2
σ̂hp , we can get
2 2 2 2 10-2
σ̂hp ĥd,k,l + σ̂hd ĥp,l σ̂hp σ̂hd
ĥk,l = 2 + σ̂ 2 and σ̂h2 = 2 + σ̂ 2 , (25)
σ̂hp hd σ̂hp hd
2 2
where the value of σ̂hp and σ̂hd can be obtained from (8) and

BER
10-3
2
σn σn2
(10): for pilot case and for data case.
Lp Ep N −1 APPEQ+EM iter 1
APPEQ+EM iter 2
VI. N UMERICAL RESULTS 10-4 APPEQ+EM iter 5
BP-DUGA iter 1
Some simulation results are shown in this section for an BP-DUGA iter 2
BP-DUGA iter 5
ISI Rayleigh channel with 3-tap rectangular impulse response. APPEQ (PerCSI) iter 5
BP-DUGA (PerCSI) iter 5
10-5
We use a half rate 64-state (133, 171) convolutional code 0 2 4 6 8 10 12 14 16
and BPSK modulation. The pseudo-random interleaver size Transmit Es/N0(dB)

is 1000. The number of pilots is 18.

The proposed BP with continuous downward and upward
messages (BP-DUGA) is compared with iterative APP equal-
izer (APPEQ) and MMSE equalizer (MMSEEQ), both using
EM channel estimation [11] [12]. From Fig. 4, we observe that
the proposed BP-DUGA has better BER performance with per-
fect channel state information (PerCSI) than MMSE equalizer
– about 1dB for 10−5 – where K = 11 represents the number
of complex-valued tap weight coefficients of the equalizer.
With 5 iterations, it also outperforms MMSEEQ+EM. From
Fig. 5, we observe that BP-DUGA has BER performance
very close to that of iterative APP equalizer with PerCSI
and EM channel estimation. Using continuous downward and
upward messages brings a complexity reduction compared to
the quantization method.
10-1
10-2
BER
Fig. 5. Bit error rate performance comparison: BP-DUGA vs APP Equalizer
with EM channel estimation.
the performance of the more complex APPEQ and outperforms
MMSE equalizer. This paper is focusing on BPSK modulation.
Nevertheless, the extension of the Gaussian approximation
principle to a higher level modulation scheme is natural.
VIII. ACKNOWLEDGEMENT
This work has been performed in the framework of the
CELTIC project CP5-026 WINNER+.
R EFERENCES
[1] H.-A. Loeliger, “An introduction to factor graphs,” IEEE Signal Process.
Mag., vol. 21, pp. 28–41, Jan. 2004.
[2] F. R. Kschischang, B. J. Frey, and H.-A. Loeliger, “Factor graphs and
the sum-product algorithm,” IEEE Trans. Inf. Theory, vol. 47, no. 2, pp.
498–519, Feb. 2001.
[3] H. Wymeersch, H. Steendam, H. Bruneel, and M. Moeneclaey, “Code-
aided frame synchronization and phase ambiguity resolution,” IEEE
Trans. Signal Process., vol. 54, no. 7, pp. 2747–2757, Jul. 2006.
[4] A. P. Worthen and W. E. Stark, “Unified design of iterative receivers
using factor graphs,” IEEE Trans. Inf. Theory, vol. 47, no. 2, pp. 843–
849, Feb. 2001.

10-3 [5] F. Z. Merli and G. M. Vitetta, “A factor graph approach to the iterative
detection of ofdm signals in the presence of carrier frequency offset and
phase noise,” in Proc. IEEE ICC ’07, Jun. 2007, pp. 2865–2870.
MMSEEQ+EM(K=11) iter 1 [6] J. Dauwels, S. Korl, and H.-A. Loeliger, “Particle methods as message
MMSEEQ+EM(K=11) iter 2
10-4 MMSEEQ+EM(K=11) iter 5
passing,” in Proc. IEEE ISIT ’06, Jul. 2007, pp. 2052–2056.
BP-DUGA iter 1 [7] Y. Liu, L. Brunel, and J. Boutros, “Belief propagation with gaussian
BP-DUGA iter 2
BP-DUGA iter 5 approximation for joint channel estimation and decoding,” in Proc.
MMSEEQ (PerCSI) IEEE International Symposium on Personal, Indoor and Mobile Radio
BP-DUGA (PerCSI) iter 5
10-5 Communications, Sep. 2008.
0 2 4 6 8 10 12 14 16 [8] C. Douillard, M. Jzquel, and C. Berrou, “Iterative correction of inter-
Transmit Es/N0(dB) symbol interference: Turbo-equalization,” Eur. Trans. Telecomm, vol. 6,
pp. 507–511, Sep./Oct. 1995.
Fig. 4. Bit error rate performance comparison: BP-DUGA vs MMSE [9] J. G. Proakis, Digital Communications fourth edition. McGraw-Hill
Equalizer with EM channel estimation. Companies, 2001.
[10] B. S. Ünal, A. Berthet, and R. Visoz, “Iterative channel estimation
and coded symbol detection for dispersive channels,” in Proc. IEEE
International Symposium on Personal, Indoor and Mobile Radio Com-
VII. C ONCLUSION munications, vol. 1, Sep./Oct. 2001, pp. C–100 – C–106.
[11] T. Moon, “The expectation-maximization algorithm,” Signal Processing
Thanks to an approximation of the distribution of the Magazine, IEEE, vol. 13, no. 6, pp. 47–60, Nov 1996.
channel estimate as a mixture of Gaussian distributions, we [12] Y. Liu, L. Brunel, and J. Boutros, “EM channel estimation for coded
improved the performance of BP and reduced its complexity OFDM transmissions over frequency-selective channel,” in Proc. IEEE
10th International Symposium on Spread Spectrum Techniques and
by propagating continuous messages in the factor graph for Applications, 2008 (ISSSTA’08), Aug. 2008, pp. 544–549.
multipath channel. The proposed BP-DUGA almost achieves

3168