Int. J. Communications, Network and System Sciences, 2010, 3, 821-825
doi:10.4236/ijcns.2010.310111 Published Online October 2010 (http://www.SciRP.org/journal/ijcns)
Copyright © 2010 SciRes. IJCNS
Using Least Squares Support Vector Machines for
Frequency Estimation
Xiaoyun Teng, Xiaoyi Zhang, Hongyi Yu
Information Science and Technolog y Institute, Zhengzhou, China
E-mail: ieukey@163.com
Received July 6, 2010; revised August 9, 2010; accepted September 11, 2010
Abstract
Frequency estimation is transformed to a pattern recognition problem, and a least squares support vector
machine (LS-SVM) estimator is derived. The estimator can work efficiently without the need of statistics
knowledge of the observations, and the estimation performance is insensitive to the carrier phase. Simulation
results are presented showing that proposed estimators offer better performance than traditional Maximum
Likelihood (ML) estimator at low SNR, since classification-based method does not have the threshold effect
of nonlinear estimation.
Keywords: Carrier Recovery, LS-SVM, Pattern Recognition
1. Introduction
In digital communication systems with burst transmis-
sion, carrier recovery within each information burst is a
critical issue. The estimation of carrier frequency in ad-
ditive noise is one of the very important problems in the
theory and applications of digital signal processing.
Various techniques have been proposed for carrier fre-
quency recovery [1-5]. The estimators in [1] are based on
a maximum likelihood criterion, which is known to be an
excellent estimator, but it suffers from the threshold ef-
fect of nonlinear estimators. Frequency estimation in
colored noise is addressed in [6] and [7], which model
the colored noise as an AR or MA process. However,
most of above estimators require the statistics knowledg e
of the observations, such as, probability density function
(pdf), autocorrelation, etc.
A Support Vector Machine (SVM) [8] uses training
data as an integral element of the function estimation
model as opposed to simply using training data to esti-
mate parameters of an a priori model using maximum
likelihood [9]. The SVM has the advantage over tradi-
tional learning approaches in terms of performance,
complexity and convergence. SVMs have been widely
used in solving classification and function estimation
problems. Recently, SVM has been introduced to com-
munication systems as a new method for channel equali-
zation [9,10] and multiuser detector in CDMA commu-
nications [11]. The least squares support vector machines
(LS-SVM) involves solving linear equations instead of
solving the quadratic programming problem required in
the original SVM. In this paper, we view frequency es-
timation as a pattern recognition problem, and propose a
different frequency estimator based on LS-SVM.
2. Signal Model
Using complex-envelope notation, the observed signal
samples are expressed by


() ,
0,1, ,1, ,1
jn
n
p
rane vn
nNN



(1)
where
is the unknown carrier frequency normalized
to sampling frequency
s
f
, for the sake of simplicity,
1
s
f
,
is an initial random phase, ()vn are additive
noise samples. n
a is the normalized transmitted BPSK
symbol, i.e., 1
n
a
. We consider scenarios where the
signal n
a is known, i.e., a training sequence is trans-
mitted for carrier recovery.
Define the following vector
011
,, ,,
p
T
NN
rr r

r
011
,, ,,
p
T
NN
aa a

A



11
1, ,,,,
p
jN jN
j
Ndiag eee
Ψ
X. Y. TENG ET AL.
Copyright © 2010 SciRes. IJCNS
822
 
0,1,,1 T
vv vN


v (2)
The signal model can be arranged in matrix form as

jN
e
rΨAv (3)
3. SVM Based Frequency Estimation
3.1. SVM Introduction
SVM developed by Vapnik is a new type of learning
machines which has a high generalization performance
and sparse solution. It replaces empirical risk minimiza-
tion by structural risk minimization (SRM). The goal of
SVM is to find the hyperplane that maximizes the mini-
mum distance between any point and the hyperplane.
The idea of SVM can be expressed as follows.
Consider (x ,),1,2,....,
ii
yi N be a linearly separa-
ble training set, wherexd
Rand {1,1}y , which can
be separated by the hyperplane satisfying0
T
wxb
,
where w is the weight vector and b is the bias. If this
hyperplane maximizes the margin, then we need to solve
the following optim izat ion probl em.

2
1
minimize 2
subject to 1
ii
w
ywx b 
(4)
For the inputs data that is not separable, SVM uses
soft margins that can be expressed as follows, by intro-
ducing the non-negative slack variables , 1,...,
iiN
:
2
1
1
minimize 2
subject to ()1
lk
i
i
T
ii i
C
ywx b

w (5)
Data points are penalized if they are misclassified. The
parameter C controls trad eoff between the complexity of
the model and the classification errors.
To construct nonlinear decision functions, SVM maps
the training data nonlinearly into a higher-dimensional
feature space via a kernel function, and constructs a
separating hyperplane with maximum margin there.
The kernel fun c tion is
(, )()()
T
iji j
K
xxx x

(6)
The typical kernel functions include RBF
,
K
xy

22
exp/ 2xy
 and polynomial

,
K
xy

1d
x
y .
We prefer LS-SVM over other models of SVM, for it
offers a fast method for obtaining classifiers with good
generalization performance in many applications. In
LS-SVM, an equality constraint-based formulation is
involved instead of the convex quadratic programming
(QP) pro blem in (5).
22
1
1
minimize 2
subject to ()1
l
i
i
T
ii i
C
ywx b

w (7)
To solve this problem Lagrange multipliers
(,1,..., ;0)
ii
il
are used:
22
11
1()1
2
lN T
P
iiii i
ii
LC yb



 


wwx (8)
The KKT optimality conditions are give n by



1
1
00
00
00
010
l
pii i
i
l
pii
i
pii
i
pT
ii i
i
i
Lwyx
w
Ly
b
LC
Lywx b




 


 
 

(9)
Optimal decision function (ODF) is then given by:


1
sign n
ii i
i
f
xyKxxb

(10)
3.2. SVM-FEA
To the best of our knowledge, SVM has not been im-
plemented yet for parameter estimation in digital commu-
nication systems. We introduce SVM as a new method for
frequency estimation, by building the following frequency
estimator

1
ˆargmin ;
N
n
n





(11)
where the cost function
()
;() ()
jn
nrnesn


 
(12)
In such a scenario, the frequency estimation problem
can be transferred to a pattern recognition problem. The
optimal estimate of
can be attained by minimizing
the classification error.
In a general way, the carrier phase
existing in the
received signal samples is unknown. Thus, the ideal ML
detector is hard to handle the classification problem in
(12). The powerful LS-SVM technique is applied in this
paper.
To fit the support vector machine model, the outpu t of
the channel can be grouped into vectors
 
x( )Re( ),Im( )
jn jn
nrne rne


 
(14)
X. Y. TENG ET AL.
Copyright © 2010 SciRes. IJCNS
823
Where Re{.} and Im{.} mean the real and image
part of {.}, separately. For training purposes, taking
x( )n as the input sequence of SVM, and the transmitted
symbol ()an to be the desired output sequence.
This model of SVM-based frequency estimator, that
we call SVM-FEA, is illustrated in Figure 1.
The optimal estimator cannot be found in a single step
because the input data has the unknown term
j
n
e
. So an
effective searching process is needed to achieve the fre-
quency estimation. The procedure, which is similar to the
coarse search and fine search routine in ML algorithm, is
particularized as follows.
1) Choose a set of
values according to a appropri-
ate interval 1
, i.e.,
1211321
0, ,,...


2) Construct the input sequence w
x using each
value; for example,

11
1
x()Re (),Im ()
jn jn
wnrne rne



 

solve the QP problem and obtain the decision function.
3) Classify w
x and identify the
that minimizes
the classification error, get a approximate estimate ˆ
.
4) Set a refined interval 2
, get a new set of
val-
ues between 1
ˆ
, do 2) and 3) and get a fine estimate
of
.
4. Simulation Results
Computer simulations have been run to check the ana-
lytical results of the previous sections. We will observe
the average estimate
1
1
ˆˆ
m
N
i
i
m
Ef f
N


and the mean square error (MSE) of the estimate

2
1
1
ˆˆ
m
N
i
i
m
M
SE fff
N



Linear LS-SVMs are used with 15 to 50 data samples.
In such a scenario, C is the only parameter to be chosen
by the user during LS-SVM training, C is the upper
Figure 1. SVM based frequency estimator structure.
sponds to assigning a higher penalty to classification
errors. Simulation results showed that it has been more
robust to set C between 0.1 and 10. Specifically, we fix
5C
.
Figure 2. illustrates the average estimates versus f
when SNR = 7 dB, N = 20. The ideal line ˆ
Eff

 is
also shown for comparison. The curves show that the
range over which the estimates are unbiased is about
(–0.4,0.4).
We compare the performance of the proposed SVM
-FEA and the typical ML algorithm in AWGN channel.
MSE of the estimates are compared with the CRLB as
follows
 

ˆ
2
3,
12 1
6,
1
phase is known
SNR NNN
CRLB phase is unknown
SNR NN
 

We first consider the case that the carrier phase is
known, so ideal ML detector can be directly used to han-
dle the classification problem in (13). Thus a ML detec-
tor based frequency estimator can be derived by (12).
The curves of MSE versus SNR are shown in Figure 3.
The simulation results are attained when f = 0.34, N = 18.
The performance curves of the ML single tone, i.e. DA,
frequency estimation algorithms with different N pro-
posed in [1] are also shown for comparison. The likeli-
hood functions are computed through FFT. For the sake
of simplicity, only the coarse search is made and the FFT
length is 32768. It can be seen in Figure 2 that the MSE
performance of the SVM-FEA is close to ML detector
based frequency estimator in this case. And both classi-
fication based frequency estimation algorithm outper-
form the traditional ML estimator at low SNR (< –1dB).
-0.5 -0.4 -0.3-0.2 -0.1 00.1 0.20.3 0.4 0.5
-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
N or m alized fr equen c y off set
Average est imat e
ideal
SVM base d es t im at o r
Figure 2. Average frequency estimate versus f.
X. Y. TENG ET AL.
Copyright © 2010 SciRes. IJCNS
824
-14 -12 -10-8-6-4-20 24 6
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
SNR( dB)
MS E
ML detector bas ed estim ator
ML est imator
SVM bas ed esti mator
CRLB with N=1 8
Figure 3. MSE versus SNR when phase is known.
Figure 4 illustrates the comparison of the MSE per-
formance of the SVM-FEA and ML estimator when the
phase is unknown. The simulation results are attained
when f = 0.34, N = 20 and N = 30. Since all the data
points in the input vectors have the same constant caused
by the carrier phase, SVM-FEA shows almost the same
performance in both cases. The performance of SVM-FEA
improves with the increasing of the data length, which is
not as remarkable as that of ML estimator (the cross of
two curves change from 0 dB to –4 dB).
It is noticed that although SVM-FEA present a sig-
nificant improvement over the ML estimator at low SNR,
it can not reach the CRLB even at high SNR. The reason
is that classification based frequency estimation algo-
rithm identifies the estimate of
corresponding to the
classification error, which is insensitive to a very small
frequency change to a certain extent.
5. Conclusions
In this paper, we formulated frequency estimation of
digital communication signals as a classification problem,
and applied SVM technique to solve it. A primary
searching routine has been proposed to find the optimal
frequency estimate.
Simulations have shown that the SVM provides a ro-
bust method for frequency estimation with following
attractive features: The estimator can work efficiently
without the need of statistics knowledge of the observa-
tions, and the estimation performance is insensitive to the
carrier phase; it shows a better performance than tradi-
tional ML estimator at low SNR, for SVM-FEA has not
the threshold effect of nonlinear estimation.
A main drawback of the proposed algorithm is the
-14 -12-10-8 -6-4 -20 2 4 6
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
SNR(dB)
MS E
S VM b ased esti mator
ML estimator
CRLB with N=18
(a)
-14 -12-10-8 -6-4 -202 46
10
-7
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
SNR(dB)
MS E
S V M based estim ator
ML estimator
CRLB wi th N=4 0
(b)
Figure 4. MSE versus SNR when phase is unknown.
high computational cost, which can be reduced by intro-
ducing faster optimization techniques and improving our
searching routine. Future work will also be carried out on
the generalization of the proposed procedure to multi-
level modulations and other channel conditions, such as
fading channel and colored noise conditions.
6. References
[1] D. C. Rife and R. R. Boorstyn, “Single-Tone Parameter
Estimation from Discrete-Time Observations,” IEEE
Transactions on Information Theory, Vol. 20, No. 5, 1974,
pp. 591-598.
[2] M. Morelli and U. Mengali, “Feedforward Frequency
Estimation for PSK: A Tutorial Review,” European Tran-
X. Y. TENG ET AL.
Copyright © 2010 SciRes. IJCNS
825
sactions on Telecommunications, Vol. 9, No. 2, 1998, pp.
103-116.
[3] W. Y. Kuo and M. P. Fitz, “Frequency Offset Compensation
of Pilot Symbol Assisted Modulation in Frequency Flat
Fading,” IEEE Transactions on Communications, Vol. 45,
No. 11, 1997, pp. 1412-1416.
[4] U. Mengali and M. Morelli, “Data-Aided Frequency
Estimation for Burst Digital Transmission,” I E E E Transactions
on Communications, Vol. 45, No. 1, 1997, pp. 23-25.
[5] Y. Wang, E. Serpedin and P. Ciblat, “Optimal Blind
Carrier Recovery for MPSK Burst Transmissions,” IEEE
Transactions on Communications, Vol. 51, No. 9, 2003, p p.
1571-1581.
[6] D. N. Swingler, “Approximate Bounds on Frequency
Estimaties for Short Cissoids in Colored Noise,” IEEE
Transactions on Signal Processing, Vol. 46, No. 5, 1998,
pp. 1456-1458.
[7] J. Viterbi and A. M. Viterbi, “Nonlinear Estimation of
Psk-Modulated Carrier Phase with Application to Burst
Digital Transmissions,” IEEE Transactions on Information
Theory, Vol. 29, No. 4, 1983, pp. 543-551.
[8] V. N. Vapnik, “Statistical learning theory,” John Wiley &
Sons, Inc., New York, 1998.
[9] D. J. Sebald, “Support Vector Machine Techniques for
Nonlinear Equalization,” IEEE Transactions on Signal
Processing, Vol. 48, No. 11, 2000, pp. 3217-3226.
[10] S. Chen, S. Gunn and C. Harris, “Decision Feedbach
Equalizer Design Using Support Vector Machines,”
Vision, Image and Signal Processing, Vol. 147, No. 3,
2000, pp. 213-219.
[11] F. Perez-Cruz, A. Navia-Vazquez, P. Alarcon-Dianna and
A. Artes-Rodriguez, “SVC-Based Equalizer for Burst
TDMA Transmission,” Signal Processing, Vol. 81, No. 6,
2001, pp. 1681-1693.
[12] J. G. Proakis, “Digital Communications,” 4th Edition,
McGraw-Hill, New York, 2001.