PARAMETRIC VERSUS NONPARAMETRIC

DATA-AIDED CHANNEL ESTIMATION IN A

MULTIPATH FADING ENVIRONMENT
D. Van Welden, M. Moeneclaey and H. Steendam
DIGCOM research group, TELIN Department, Ghent University
Sint-Pietersnieuwstraat 41, 9000 Gent, Belgium
E-mail: {dmvwelde,mm,hs}@telin.ugent.be

Abstract— This paper compares a parametric and a non- with L distinct paths. The received symbol pulse h(t) is given
parametric channel estimation method in a multipath fading by
environment. In the parametric method the gains and delays are
estimated and these estimates are used to compute the samples L−1
of the received symbol pulse. In the nonparametric method the
samples of the received pulse are estimated directly, ignoring the
h(t) = ∑ αl p(t − τl ) (1)
l=0
underlying multipath channel structure. We compare these two
channel estimation methods in terms of accuracy (characterized where p(t) is the transmitted symbol pulse and αl and τl
by the MSE on the samples of the received pulse) and compu- denote the complex gain and the path delay of the l-th
tational complexity. We also investigate the influence of channel propagation path. In order to facilitate channel estimation,
estimation errors on the BER of a decision feedback equalizer. a pilot sequence {b(k), k = 0, . . . , K − 1} is transmitted at
I. I NTRODUCTION regular intervals. We assume that the channel is time-invariant
over the duration between successive pilot sequences. The
When transmitting data over multipath fading channels,
received signal is sampled at a sampling frequency of 1/Ts .
the receiver needs channel state information (CSI) in order
The samples r (mTs ) of the received signal are given by
to detect the transmitted symbols. CSI can be acquired by
means of a parametric or a nonparametric channel estimation r (mTs ) = ∑ b(k)h(mTs − kT ) + w(mTs ) (2)
method. In the parametric method the gains and delays of k
the different paths are estimated, and these estimates are where w(mTs ) is complex-valued white Gaussian noise with
used to compute an estimate of the samples of the received variance N0 /Ts , added by the channel.
symbol pulse. In the nonparametric method the samples of In order to compare the estimator performance of both
the received symbol pulse are estimated directly, ignoring the channel estimation methods, we consider the following mean-
underlying structure imposed by the propagation paths. square error on the samples of the received symbol pulse:
In this paper we compare these two channel estima-

tion methods in a data-aided scenario. For the parametric  2
MSE = E ∑ ĥ (mTs ) − h (mTs ) (3)
method, we consider the iterative space-alternating general- m
ized expectation-maximization (SAGE) algorithm as proposed
in [1], because the maximum likelihood (ML) estimator is where ĥ (mTs ) are the estimated samples of the received
too complex. For the nonparametric method, ML estimation symbol pulse.
is used [2]. For both estimation methods, we simulate the
III. PARAMETRIC CHANNEL ESTIMATION
mean-square error (MSE) on the estimates of the sampled
received pulse, and analytically derive the respective Cramer The parametric method involves the estimation of the path
Rao Bound (CRB) on this MSE. Further, the influence of esti- parameters {αl , τl , l = 0, . . . , L − 1}. It can be verified that
mation errors on the bit error rate (BER) is also investigated by ML estimation yields an L-dimensional search to estimate
considering the signal-to-interference-plus-noise ratio (SINR) the L path delays, after which the L path gain estimates
at the output of a fractionally spaced decision feedback are obtained analytically [5]. In order to avoid the high
equalizer (DFE) [3], [4] that is computed from the estimated computational complexity of the L-dimensional search, we
received pulse. Finally the computational complexity of both resort to the iterative SAGE algorithm [1]. Each SAGE
methods is discussed. iteration involves L steps: in the l-th step of the i-th iteration,
we perform a 1-dimensional search to estimate τl and an
II. S YSTEM M ODEL analytic computation to estimate αl , by using the most recent
Consider the transmission of a linearly modulated signal estimates of the gains and the delays of the other paths for
with symbol period T over a slowly fading multipath channel subtracting their estimated path interference from the received

0-7803-9785-1/06/$20.00 2006
c IEEE 25
 Part Number of complex multiplications
signal. Denoting the resulting path parameter estimates as
1 (KNs + N p ) Nτ
{α̂l , τ̂l , l = 0, . . . , L − 1}, the estimates of the samples of
2 Nτ L (L + 3) + (Nit − 1) Nτ L (L + 1)
1
2
the received symbol pulse h(t) are computed as 3 LN

L−1 TABLE I
ĥ(mTs ) = ∑ α̂l p(mTs − τ̂l ) (4) C OMPUTATIONAL COMPLEXITY FOR THE PARAMETRIC CHANNEL
l=0 ESTIMATION METHOD

A. Cramer Rao bound

We define δ (m) as the estimation error on h(mTs ):
IV. N ONPARAMETRIC CHANNEL ESTIMATION
L−1 L−1
δ (m) = ∑ α̂l p(mTs − τ̂l ) − ∑ αl p(mTs − τl ) (5) The nonparametric method directly estimates the samples
l=0 l=0 {h(mTs )} from the received signal, without exploiting the
For small estimation errors on {αl , τl , l = 0, . . . , L − 1}, relation (1). ML estimation [2] involves the multiplication of
δ (m) can be approximated as a pilot-dependent matrix and the vector of received signal
samples . This yields
L−1 L−1  −1 H
δ (m)
∑ (∆αlℜ + j∆αlℑ ) p(mTs −τl )− ∑ αl ∆τl ṗ(mTs −τl ) ĥ = BH B B r (10)
l=0 l=0
(6) where B is a pilot-dependent matrix and r is the vector of
where ∆αlℜ + j∆αlℑ = α̂l − αl and ∆τl = τ̂l − τl . This expres- received signal samples.
sion can be written in a more compact form by using vector
notation: A. Cramer Rao bound
δ (m) = pTm ε (7)
It can be shown that the Fisher information matrix Jnp of
where ε = [∆α0ℜ , . . . , ∆αL−1ℜ , ∆α0ℑ , . . . , ∆αL−1ℑ , ∆τ0 , . . . , {ℜ (h) , ℑ (h)}is given by [6]
∆τL−1 ]T and p (m) = [p(mTs − τ0 ), . . . , p(mTs − τL−1 ),
    
jp(mTs − τ0 ), . . . , jp(mTs − τL−1 ), −α0 ṗ(mTs − τ0 ), . . . , 2Ts ℜ  BH B −ℑ  BH B
Jnp = (11)
−αL−1 ṗ(mTs − τL−1 )]T . Using (7), the MSE (3) can be writ- N0 ℑ BH B ℜ BH B
ten as:
  which yields the following CRB:
MSE = ∑ pTm E εεH p∗m (8)
m


 2  
This MSE can be lower bounded by the CRB: MSE = E ∑ ĥ (mTs ) − h (mTs )  ≤ tr J−1
np (12)
 
∑ pTm E εεH p∗m ≥ ∑ pTm J−1 ∗ m
p pm (9)
m m where tr (X) denotes the trace of X.
where J p denotes the Fischer information matrix related to
the estimation of {αl , τl , l = 0, . . . , L − 1} [6]. B. Computational complexity
The computational complexity of the nonparametric chan-
B. Computational complexity
nel estimation method is determined by the computational
The parametric channel estimation method can be divided complexity of the multiplication of the pilot-dependent ma-
 −1
in three parts: A first part involves applying the received trix BH B BH with dimensions (Ns (K − 1) + N) × N and
signal to a filter matched to the transmitted pilot signal, and the
computing the matched filter output at Nτ instants with spacing  Hvector
−1 of received signal samples y because the matrix
B B BH can be stored at the receiver. This results in a
∆, such that Nτ ∆ covers the uncertainty region of the path total number
delays. A second parts consists of the different iterations of
the SAGE algorithm for estimating the path gains and delays. N2 − 1
Nmult,np = (K − 1) N + +1 (13)
The third part is the computation of the estimated samples of Ns
the received symbol pulse from the gain and delay estimates.
of complex multiplications for the nonparametric channel
Denoting by K the number of pilot symbols, 1/Ts the estimation method. N can be substituted by Ns DT + 1 in (13)
sampling rate at the receiver, N p Ts the duration of the transmit yielding
pulse p(t), Nit the number of iterations performed in the SAGE
algorithm, and N the number of estimated samples of the Nmult,np = (K − 1) Ns DT + Ns D2T + 2DT + K (14)
received pulse, the number of multiplications (of complex
numbers) required by the different phases is shown in Table where DT denotes the duration of the received pulse expressed
I. in symbol periods.

0-7803-9785-1/06/$20.00 2006
c IEEE 26
 3
10 25

20
2
10
15
1
10 10

0 5
10

SINR (dB)
MSE

0
−1
10
−5

10
−2 −10

MSE parametric method (1st iteration)

−15
−3 MSE parametric method (7th iteration) perfect channnel knowledge
10 MSE nonparametric method parametric method (1st iteration)
CRB parametric method −20 parametric method (7th iteration)
CRB nonparametric method nonparametric method
−4
10 −25
−10 −5 0 5 10 15 20 25 30 −10 −5 0 5 10 15 20 25 30
ES/N0 (dB) ES/N0 (dB)

Fig. 1. Comparison between the two estimation methods in terms of the Fig. 2. Comparison between the two estimation methods in terms of the
MSE SINR at the DFE output

V. SINR AT THE OUTPUT OF A FRACTIONALLY SPACED

DFE complex-valued Gaussian random variables with zero mean
and variance 1/2 (Rayleigh fading). The sample period Ts is
In this contribution we consider a fractionally spaced set to T /2 so that no aliasing occurs. Taking into account the
MMSE DFE [4] as receiver. The DFE consists of a forward duration (20T ) of the transmit pulse, the uncertainty (T ) on τ0
filter heq which has KFF1 anti-causal coefficients and 1+KFF2 and the delay spread (2T ) of the channel, the received pulse
causal coefficients, and a feedback filter hFB which has KFB h (t) has N = 47 samples taken at rate 1/Ts = 2/T . Hence, the
coefficients. We denote the transmitted data symbols also as parametric method involves the estimation of 2 delays and 2
b (k). The input u (k) to the symbol-by-symbol detector is complex gains, while the nonparametric method consists of
KFF2 KFB estimating 47 samples h (mTs ). The parametric method uses
u (k) = ∑ heq (i) r (kNs − i) − ∑ hFB (m) b̂ (k − m) (15) Nτ = 31 and ∆ = T /10.
i=−KFF1 m=1
For the two estimation methods, Figure 1 shows, as a
where b̂ (k − m) are the past decisions and r (kTs ) is the function of Es /N0 , the MSE (3) averaged over different real-
received signal (2) but with {b(k)} now denoting the unknown izations of the channel and of the pilot sequence along with
data symbols instead of pilot symbols. The equalizer coeffi- their respective CRB. We observe that the parametric method
cients heq and hFB are selected such that E |u (k) − b (k)|2 yields the smaller CRB; this is because the parametric method
is minimized (under the assumption that past decisions are exploits the underlying multipath channel model. The MSE
correct). In order to compute the optimum equalizer coeffi- resulting from the nonparametric method essentially coincides
cients, the samples {h (mTs )} of the received symbol pulse with the corresponding CRB. After a sufficient number of
are needed. As these samples are iterations, the MSE resulting from the parametric method

 not known to the receiver, is close to the corresponding CRB at small and moderate
the estimated samples ĥ (mTs ) are used instead, yielding
suboptimum equalizer coefficients. Es /N0 , but exhibits an error floor at high Es /N0 . This floor
The equalizer output u(k) from (15) consists of a useful is caused by the discretization of the unknown delays when
term (proportional to b(k)), a residual ISI term (linear com- performing the 1-dimensional searches associated with the
bination of symbols b(m) with m = k) and an additive noise SAGE algorithm (we have verified that a finer discretization
term. As a performance measure we consider the SINR at the indeed lowers the floor).
equalizer output, which is defined as the ratio of the power In Figure 2 the SINR at the output of a T /2-spaced
of the useful term to the power of the residual ISI plus noise. DFE, with KFF1 = 16 anti-causal forward coefficients, 1 +
KFF2 = 17 causal forward coefficients and KFB = 1 feedback
VI. S IMULATION R ESULTS coefficient, is shown in order to illustrate the effect of channel
The transmit pulse is a square root raised cosine pulse with estimation errors on the reliability of the communication link.
25% roll-off, truncated to a length of 20 symbol intervals. The The equalizer coefficients are computed from the estimated
pilot sequence consists of 10 BPSK symbols. We consider a samples of the received symbol pulse. We observe that
multipath fading channel with 2 paths. We assume that τ0 is the SINR associated with the parametric channel estimation
uniformly distributed in (0, T ), the delay difference τ1 − τ0 method is approximately 1 dB below the SINR corresponding
is uniformly distributed in (0, 2T ). The channel gains αl are to perfect channel estimates for moderate ES /N0 , whereas the

0-7803-9785-1/06/$20.00 2006
c IEEE 27
SINR resulting from the nonparametric method is about 5 dB lower computational complexity than the parametric chan-
less than the SINR corresponding to the parametric method. nel estimation method.
Hence for the example considered, the nonparametric method Our numerical example involves two Rayleigh fading paths
yields an SINR degradation of 4 dB as compared to the of equal average power with uniformly distributed path delays,
parametric method. but the parametric and nonparametric estimation methods
Let us now discuss the computational complexity. For the can equally be applied for another number of paths, other
parametric method the example mentioned here results in fading statistics, other power delay profiles and other path
3225 complex multiplications when 7 SAGE iterations are delay distributions. Also in these cases, the parametric method
carried out. 1860 complex multiplications or 60% of the total outperforms the nonparametric method, because the former
amount is needed for the first part of the algorithm. For the takes advantage of the multipath structure of the channel.
nonparametric method the required number of complex mul-
R EFERENCES
tiplications for the considered situation equals 1528. Hence
the parametric method needs 2 times as much multiplications [1] B. Fleury, M. Tschudin, R. Heddergott, D. Dahlhaus, and K. Pedersen.
Channel Parameter Estimation in Mobile Radio Environments Using the
as the nonparametric method. SAGE Algorithm. IEEE Journal on Selected Areas in Communications,
17(3):434–450, March 1999.
VII. C ONCLUSION AND REMARKS [2] H. Meyr, M. Moeneclaey, S.A. Fechtel. Synchronization, Channel
In this contribution we have compared parametric and Estimation, and Signal Processing, volume 2 of Digital Communication
Receivers. John Wiley & Sons, 1997.
nonparametric channel estimation from pilot symbols. We [3] R.D. Gitlin, J.F. Hayes and S.B. Weinstein. Data Communications
derived analytically the Cramer-Rao lowerbound on the MSE Principles. Plenum Press, 1992.
of the samples of the received symbol pulse for both methods. [4] S.U.H. Qureshi. ”Adaptive Equalization”. Proceedings of the IEEE,
73(9):1349–1387, September 1985.
For Es /N0 values of practical interest, we have shown that the [5] E.G. Strom and F. Malmsten. A maximum likelihood approach for
former method outperforms the latter method, both in terms of estimating DS-CDMA multipath fading channels. IEEE Journal on
estimation accuracy and detector performance. However, the Selected Areas in Communications, 18(1):132–140, January 2000.
[6] H.L. Van Trees. Detection, Estimation,and Modulation Theory, Part I.
nonparametric channel estimation method has a considerably Wiley and Sons, October 2001.

0-7803-9785-1/06/$20.00 2006
c IEEE 28