IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO.

5, SEPTEMBER 2008

parameters in Table I and using three internal LDPC iterations as well

as an LDPC output block length of Kldpc = 12 000 bits.
VII. C ONCLUSION
In this paper, we have proposed a novel symbol-based iterative
scheme that exploits the advantages of nonbinary LDPC codes [7],
the rate-1 inner codes in [8], and the STBC-SP scheme in [4]. Our
investigations have demonstrated that attractive performance improvements may be achieved by the proposed scheme over the equivalentthroughput bit-based schemes. Subsequently, novel nonbinary EXIT
charts were used to study the convergence of the proposed symbolbased scheme. By contrast, binary EXIT charts were used to explore the convergence of the bit-based binary LDPC-coded STBC-SP
schemes. Again, it was demonstrated both by EXIT chart analysis and
by the corresponding BER performance curves that the symbol-based
scheme is capable of outperforming its bit-based counterpart, and
both designs had an edge over Alamoutis now classic STBC scheme,
dispensing with the SP-based joint design of the QPSK ST symbols.
Our future research will consider similar differentially encoded lowcomplexity SP designs for the sake of requiring no channel estimation
as well as ST-equalized systems.

3205

[15] L. Bahl, J. Cocke, F. Jelinek, and J. Raviv, Optimal decoding of

linear codes for minimizing symbol error rate, IEEE Trans. Inf. Theory,
vol. IT-20, no. 2, pp. 284287, Mar. 1974.
[16] S. ten Brink, Designing iterative decoding schemes with the extrinsic
information transfer chart, AE Int. J. Electron. Commun., vol. 54, no. 6,
pp. 389398, Nov. 2000.
[17] I. Land, P. Hoeher, and S. Gligorevic, Computation of symbol-wise
mutual information in transmission systems with logAPP decoders and
application to EXIT charts, in Proc. Int. ITG Conf. SCC, Erlangen,
Germany, Jan. 2004, pp. 195202.

Recovering Signal Energy From the

Cyclic Prefix in OFDM
Gregory E. Bottomley, Fellow, IEEE, and
Leif R. Wilhelmsson, Senior Member, IEEE

The authors would like to thank the anonymous reviewers for their
insightful comments.

AbstractIn orthogonal frequency-division multiplexing (OFDM) systems, a cyclic prefix (CP) is often added at the transmitter and discarded
at the receiver. When the length of the CP exceeds the delay spread of
the channel, a portion of the CP can be used to recover additional signal
energy. In the past, Nyquist windowing techniques have been proposed
to recover signal energy, thereby improving performance. In this paper,
linear maximum-likelihood (ML) and minimum-mean-square-error
(MMSE) approaches are developed, which further improve performance,
particularly when the signal-to-noise ratio (SNR) is low. Simpler twovalued windowing (TVW) solutions are also provided, which generally
employ one or two non-Nyquist windows.

R EFERENCES

Index TermsCyclic prefix (CP), demodulation, maximum-likelihood

(ML) detection, multicarrier transmission, orthogonal frequency-division
multiplexing (OFDM).

ACKNOWLEDGMENT

[1] L. Hanzo, T. H. Liew, and B. L. Yeap, Turbo Coding, Turbo Equalisation and SpaceTime Coding: For Transmission Over Fading Channels.
Chichester, U.K.: Wiley, 2002.
[2] S. Alamouti, A simple transmit diversity technique for wireless communications, IEEE J. Sel. Areas Commun., vol. 16, no. 8, pp. 14511458,
Oct. 1998.
[3] V. Tarokh, H. Jafarkhani, and A. Calderbank, Spacetime block codes
from orthogonal designs, IEEE Trans. Inf. Theory, vol. 45, no. 5,
pp. 14561467, Jul. 1999.
[4] W. Su, Z. Safar, and K. J. R. Liu, Spacetime signal design for timecorrelated Rayleigh fading channels, in Proc. IEEE ICC, Anchorage,
AK, 2003, vol. 5, pp. 31753179.
[5] R. Gallager, Low-density parity check codes, IEEE Trans. Inf. Theory,
vol. IT-8, no. 1, pp. 2128, Jan. 1962.
[6] C. Berrou, A. Glavieux, and P. Thitimajshima, Near Shannon limit
error-correcting coding and decoding: Turbo-codes, in Proc. IEEE ICC,
Geneva, Switzerland, May 1993, pp. 10641070.
[7] M. C. Davey and D. J. C. MacKay, Low density parity check codes over
GF (q), IEEE Commun. Lett., vol. 2, no. 6, pp. 165167, Jun. 1998.
[8] D. Divsalar, S. Dolinar, and F. Pollara, Serial concatenated trellis coded
modulation with rate-1 inner code, in Proc. IEEE GLOBECOM, San
Francisco, CA, Nov. 27Dec. 1, 2000, vol. 2, pp. 777782.
[9] B. Scanavino, G. Montorsi, and S. Benedetto, Convergence properties
of iterative decoders working at bit and symbol level, in Proc. IEEE
GLOBECOM, San Antonio, TX, 2001, vol. 2, pp. 10371041.
[10] A. Grant, Convergence of non-binary iterative decoding, in Proc. IEEE
GLOBECOM, San Antonio, TX, Nov. 2001, vol. 2, pp. 10581062.
[11] J. Kliewer, S.-X. Ng, and L. Hanzo, Efficient computation of EXIT functions for nonbinary iterative decoding, IEEE Trans. Commun., vol. 54,
no. 12, pp. 21332136, Dec. 2006.
[12] T. J. Richardson and R. Urbanke, The capacity of low-density paritycheck codes under message-passing decoding, IEEE Trans. Inf. Theory,
vol. 47, no. 2, pp. 599618, Feb. 2001.
[13] J. H. Conway and N. J. Sloane, Sphere Packings, Lattices and Groups.
New York: Springer-Verlag, 1999.
[14] O. Alamri, B. L. Yeap, and L. Hanzo, A turbo detection and spherepacking-modulation-aided spacetime coding scheme, IEEE Trans. Veh.
Technol., vol. 56, no. 2, pp. 575582, Mar. 2007.

I. I NTRODUCTION
Orthogonal frequency-division multiplexing (OFDM), which was
originally developed in the late 1950s and 1960s (see references in [1]),
is being used or considered in a variety of wireless communication
systems. With OFDM, a block of symbols is sent in parallel on
different frequency subcarriers. Wireless communication channels are
often dispersive, introducing interference between and within blocks
of symbols. Ideally, all symbols should be jointly detected [2], which
requires significant receiver complexity. To simplify receiver processing, part of the block of symbols is copied and preappended at the
transmitter, forming a cyclic prefix (CP) [3]. At the receiver, this
portion of the signal is typically discarded, thus avoiding interblock
interference and allowing for simple subcarrier separation using a fast
Fourier transform (FFT).
The CP may be longer than is needed for a particular users receiver.
For broadcast, the CP should be designed for the worst-case delay
spread, which typically is not experienced by all users. For multipleaccess, where different subcarriers are allocated to different users, the
CP should be designed for the user with the largest delay spread. In
Manuscript received January 30, 2007; revised August 22, 2007 and
October 29, 2007. This work was presented in part at the IEEE Vehicular
Technology Conference, Montral, QC, Canada, September 2528, 2006. The
review of this paper was coordinated by Prof. H. Nguyen.
G. E. Bottomley is with Ericsson Inc., Research Triangle Park, NC 27709
USA.
L. R. Wilhelmsson is with Ericsson AB, Lund, Sweden.
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TVT.2007.914057

0018-9545/$25.00 2008 IEEE

3206

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 5, SEPTEMBER 2008

both cases, the delay spread experienced by a particular user can be

less than the length of the CP.
In this situation, the clear portion of the CP can be used to
recover additional signal energy [4][7].1 In [4], a Nyquist window
is applied to the data prior to demodulation. Such a window preserves
orthogonality between subcarriers and provides robustness to residual frequency offset. In [5], assuming a nondispersive channel, the
shape of the Nyquist window is optimized to minimize the sum of
noise and interference powers for a given residual frequency offset.
When the residual frequency offset is zero, the solution is zero forcing in that it preserves orthogonality between subcarriers. Also, the
design of the window becomes independent of the signal quality. A
dispersive channel and a more general window shape are considered
in [6]. However, when the residual frequency offset is zero, it can
be shown that this approach is equivalent to the approach in [5].
In [7], a Nyquist window based on a second-order polynomial is
proposed.
In this paper, a maximum-likelihood (ML) approach for using the
clear portion of the CP with a dispersive channel is developed. In
general, ML approaches provide good performance and suitable soft
information for forward error correction (FEC) decoding. The ML formulation treats the subcarriers as code-division multiplexing (CDM)
chip sequences, a technique inspired by Scaglione and Giannakis [8],
which showed that chip sequences with robust cross-correlation properties have OFDM properties. As in similar formulations for CDM systems [9], [10], interference is approximated as Gaussian. This results
in a linear solution that can be expressed in terms of a linear minimummean-square-error (MMSE) solution. Based on this solution, two nonNyquist windowing approaches are developed for the case of white
noise.
Gains in performance over discarding the CP are bounded by the
fraction of transmitted power given to the CP. For example, when the
CP is 26.6% overhead, as assumed in [5], the performance gain is
bounded by 1 dB. Such gains may be important in extending coverage,
particularly for broadcast systems.
The remainder of this paper is organized as follows. A system model
is given in Section II. In Section III, an ML receiver is developed,
which includes a matrix prefilter. Two-valued windowing (TVW)
solutions that avoid the matrix prefilter are developed in Sections IV
and V. Performance results are given in Section VI, and Section VII
concludes this paper.
II. S YSTEM M ODEL
As in [4] and [5], a discrete-time system model is used. Vector
notation is used, which is similar to [11]. Each of K subcarriers is
represented as a K-chip sequence of unity amplitude chips (samples)
using K 1 column vector tk . These column vectors are collected
together in a K K matrix T = [t1 tK ]. Preappending an
L-chip CP gives the (K + L)-length extended sequence

(KL+1:K) T

ek = (tk

(1:K) T

(tk

T

(1)

Fig. 1. OFDM example (K = 4 subcarriers, L = 2 chips CP, and

D = 2 paths in a channel).

energy per chip (per subcarrier), and s = [s1 sK ]T is a vector

of complex modem symbols. The symbol energy Es is equally divided
among all the chips (Ec = Es /(K + L)). We assume that each sk is
randomly drawn from the set {Sm ; m = 1, . . . , M }.
The channel is modeled as a sample-spaced finite-impulse response
filter with D taps and impulse response, i.e.,

1 Note that this requires accurate estimation of the delay spread as well as
accurate synchronization.

cd ( dTs )

(2)

d=0

where cd is a complex channel coefficient, and Ts is the sample period.

The channel response is assumed to be known, although it would be
estimated in practice. In addition, noise is added.
An example of a channel-filtered transmitted signal is shown in
Fig. 1 for a two-tap channel (D = 2), K = 4, and L = 2. Consider
block 1. With traditional reception, the entire CP (two chips) of path 0
would be discarded. However, when the clear portion of the CP is kept
or recycled, the last chip of the CP is kept.
In general, L = L D + 1 chips of the CP are kept. We denote
the resulting sequence for subcarrier k, after removing a portion of the
CP, as (K + L ) 1 vector fk , which is given by


fk =

(KL +1:K)
tk

T 

(1:KL )
tk

T 

(KL +1:K)
tk

T T
.
(3)

These sequences are collected into a (K + L ) K matrix F.

Over the portion of the signal that is kept, the delayed image of the
transmitted signal appears as a cyclic shift of the transmitted sequence,
which can be expressed as a phase shift applied to the undelayed
image. Thus, the image, delayed d chips, can be expressed as (pk )d fk ,
where pk = exp(j2(k 1)/K). As a result, the (K + L ) 1
received vector over the portion of the signal that is kept can be
expressed as
r = FHs + n
where H = diag(h1

(m:n)

denotes the
where superscript T denotes transpose, and tk
subsequence consisting of elements m to n. The extended sequences
are collected into (K + L) K matrix E.
Assuming equal-average-energy symbols on the different subcarriers, the transmitted chip sequence for one
block of symbols can be
expressed as the (K + L) 1 vector y = Ec Es, where Ec is the

D1

c( ) =

hk =

(4)

hK ), and

D1

Ec

(pk )d cd .

(5)

d=0

Thus, hk is the effective channel coefficient for subcarrier k.2 The

noise vector n is assumed to be a vector of complex Gaussian random
variables with zero mean and covariance CN = E{nnH }.

2 Equation

(5) is equivalent to that given in [12].

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 5, SEPTEMBER 2008

3207

III. ML R ECEIVER
With an ML formulation, the likelihood of the received signal r
given that symbol s equals Sm is maximized over m. From (4), r
conditioned on s = Sm can be expressed as
r = h f Sm +

hk fk sk + n.

(6)

k=

The first term on the right-hand side is the signal term; the remaining
terms are interference and noise (I + N or impairment). The interference is approximated as Gaussian, which is reasonable if the number of
summation terms is large (K large) and there are no dominant terms.3
Thus, the overall impairment is complex Gaussian with zero mean
and covariance, i.e.,
CI+N () =

Fig. 2. Receiver block diagram (K = 4 subcarriers, L = 2 chips CP, and

D = 2 paths in a channel).

Applying (13) to (7) and substituting the result into (12) gives
w = a v
where

|hk |2 fk fkH + CN = Cr h f fH

(7)

a = 1/ 1 h fH v

k=

where

(14)

(15)

v = h C1
r f .

(16)

Cr =

|hk |2 fk fkH + CN

(8)

k=1

is the (K + L ) (K + L ) covariance matrix of r, and superscript

H denotes conjugate transpose.
With this Gaussian model, the likelihood of r given that s = Sm
is [13]
Pr{r|s = Sm }
=

exp (r h f Sm )H C1
I+N ()(r h f Sm )
K+L |CI+N ()|

r, 1)
z = DH TH y(
(9)

where | | denotes determinant. A log-likelihood metric (LLM) is

obtained by taking the log and ignoring terms independent of m, giving

z } |Sm |2 h wH f
LLM{s = Sm } = 2Re {Sm

(10)

z = wH r
w = h C1
I+N ()f

(11)
(12)

where

and superscript denotes complex conjugate. See [9] and [10] for a
similar result for CDM.
The LLM in (10) consists of two terms. The first term involves
combining the received samples to form decision variable z . The
second term is a signal bias term that depends on the amplitude of
the hypothesized symbol. For phase-only modulations, this term can
be ignored. Also, for the special case of binary phase-shift keying
(BPSK), we are ultimately interested in the log-likelihood ratio, which
is proportional to the real part of z in (11). For quaternary phase-shift
keying, the real and imaginary parts are used.

The weight solution in (12) involves inverting a matrix that is

different for each symbol index . Complexity can be reduced by
expressing the weights in terms of a matrix inverse that is independent
of . To achieve this, we use the following identity [14]:
(A+BCD)
3 For

=A

B(C

+DA

B)

DA

(17)

where


y(
r, ) =

r(L

+1:K)


(18)

(
r(K+1:K+L ) +
r(1:L ) )

and D is a diagonal matrix with elements d = a h . The K 1

vector function y(
r, ) adds the clear portion of the CP to the end
of the data block and scales the result by , which is set to 1 in (17).
Later, this function will be used with other values.
The receiver structure is illustrated in Fig. 2 for the case of K = 4,
L = 2, and D = 2 (see Fig. 1). The serial-to-parallel converter collects
received samples into a vector. A portion of the CP is discarded, and
the remainder is prefiltered by multiplying by a matrix (C1
r ). Then,
the CP filtered value is added to the end of the filtered received vector,
and an FFT is taken (multiplication by TH ). While the FFT makes the
demodulator efficient, the prefilter matrix multiplication is additional
effort not required by the traditional solution. (The Nyquist windowing
approach has the same structure, except that the matrix prefilter is a
diagonal matrix.)
For this and the other receivers considered, h wH f is purely real
and nonnegative, so that the signal-to-interference-plus-noise ratio
(SINR) is given by [9]

A. Complexity Reduction

Observe that these weights depend on inverting Cr , which is independent of . It can be shown that v corresponds to linear MMSE
estimation of s .
To see how an FFT can be used, we substitute (14) and (16) into (11)
r, where
r = C1
to obtain z = a h fH
r r. Thus, the received signal
vector is preprocessed by multiplying by the inverse of Cr . Collecting
decision variables into a vector and applying (3) gives

2 

SINR = h wH f  / wH CI+N w .

(19)

For the ML solution, substituting (14) and (16) into (19) gives

2 H 1
SINR = |h |2 fH C1
r f / 1 |h | f Cr f .

(20)

. (13)

reasonable CP lengths, adjacent subcarriers do not appear to entirely

dominate the summation, as the Gaussian assumption was found to accurately
predict bit-error-rate (BER) performance.

B. Special Case: Nondispersive Channel and White Noise

If the channel is nondispersive (D = 1) and the noise is stationary
and white (CN = N0 IK+L , where Im is an m m identity matrix),

3208

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 5, SEPTEMBER 2008

then the entire CP can be used (L = L). As shown in the Appendix,
(17) then simplifies to


z=

h
N0


TH y(r, )

(21)

where y(r, ) is defined in (18), and , h, and are given in (48)(50),

as shown in the Appendix, respectively.
Thus, the received CP is added to the end of the block, and the sum
is scaled by before taking the FFT and then scaling by h /N0 .
This can be interpreted as using a TVW, with values 1 in the middle
and on the edges. At high Es /N0 , approaches 0.5, giving a Nyquist
window that is the same as the optimized Nyquist window determined
in [5] for nondispersive channels with zero frequency offset.
The SINR is the same for all subcarriers. For c0 = 1, (19) eventually
gives


SINR =

1 + L/K +

2
E /N0
1+L/K s

1 + L/K + (2 L/K)Es /N0

Es
.
N0

The term I() accounts for intersubcarrier interference. In general, it

is different for each subcarrier. When the channel is nondispersive, it
can be shown that I() = L(K L)Ec |c0 |2 .
The SINR is given by
SINR =

A2
[K + (2 1)L ]2 |h |2
.
=
2

[K + (22 1)L ] N0 + (2 1)2 I()

To determine the best choice for ( ), we take the derivative of SINR

w.r.t. , set the derivative to zero, and solve for  . To remove the
averaged
dependence on , we replace I() with its average value I,
over all subcarriers. Using (27)
I = (1/K)

K


I() = (1/K)

K 


=1

2

|hk |2 fH fk  .

(29)

=1 k=

It is convenient to approximate |hk |2 with its average value given by

(22)

|hk |2 = (1/K)

K


D1

|hk |2 = Ec

k=1

Observe how the SINR depends on the ratio L/K, as well as the
symbol signal-to-noise ratio (SNR) Es /N0 . At low Es /N0 , the SINR
approaches Es /N0 , allowing full recovery of the signal energy. With
traditional discarding of the CP, the SINR is (Es /N0 )/(1 + L/K).

(28)

|cd |2 .

(30)

d=0

This is expected to slightly alter  (less so with mild dispersion).

With this approximation, (29) becomes

I |hk |2 (K + L )2 + (1/K)

K


fH (FFH )fH

(31)

=1

IV. TVW R ECEIVER

The ML solution requires matrix prefiltering of the received data
before applying the FFT. However, if the channel is nondispersive and
the noise is stationary and white, then the solution simplifies to a TVW
approach (values 1 and ). The scaling factor depends on the SNR,
which is the same for all subcarriers.
When the channel is dispersive, a low-complexity solution is obtained by imposing the simpler structure of the nondispersive case.
Thus, the matrix prefilter is replaced with a TVW operation (a diagonal
matrix prefilter), and the scaling factor is optimized.4 It is assumed
that the noise is stationary and white.
Specifically, the proposed solution determines the K 1 decision
variable vector z using
z = DH TH y(r, )

(23)

To provide further simplification, we use the fact that F has the same
structure as E, except that L is replaced with L . Thus, a relation
similar to (45) applies so that (31) simplifies to

D1

I L (K L )Ec

|cd |2 .

(32)

d=0

Observe that when the channel is nondispersive, the approximation is

exact.
Substituting (32) in (28) and maximizing SINR w.r.t. eventually
gives
 =

1 + K
1 + 2K

(33)

where

D1

where D is diagonal with elements d =  h /N0 . The scaling factor

 will be determined later.
From (6) and (23), z can be expressed as

(34)

d=0

(24)

Observe that (33) is similar to (50) for the nondispersive case, except
that has been replaced by . The SINR for the TVW solution
can be found by substituting (33) into (28). The SINR for the other
approaches can be obtained using (19).

A =  [K + (2 1)L ] |h |2 /N0

(25)

A. Symbol Metric and Scaling Factor Determination

Using the model in (24) and ignoring terms independent of m, the
LLM is given by

and the impairment i has power

2 =

|cd |2 .

z = A Sm + i
where

= (Ec /N0 )

2 |h |
N02


2

K + (22 1)L N0 + (2 1)2 I()

with
I() =

2

|hk |2 fH fk  .

(26)


z } |Sm |2 q
LLM{s = Sm } = 2Re {Sm

(35)

where z = (A /2 ) z , and q = A2 /2 = SINR .

The scaling factor  is determined by setting z = z , giving
(27)

k=

4 Use of different values for different subcarriers is explored in the next

section.

 =

[K +

(22

K + (2 1)L
1)L ] + (2 1)2 I()/N0

(36)

where is given by (33). We can approximate  and q by replacing

I() with (32).

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 5, SEPTEMBER 2008

Fig. 3. SINR gain (relative to discarding the CP) versus SNR (Es /N0 ),
K = 64 subcarriers, L = 17 chips CP, and a two-path (D = 2) channel.
Results independent of the subcarrier are provided for the RCOS and ONW
approaches. Best- and worst-case gains (over subcarriers) are provided for the
ML, TVW, and TVW2 approaches.

V. TVW2 R ECEIVER
A second approximate TVW approach (TVW2) can be obtained by
using different values (different windows) for each subcarrier. This
is obtained by not replacing I() in (28) with an average value. It can
be shown that SINR is then maximized using
 =
where


 =

1 + K
1 + 2K

1
L (K L )



(37)

I0 ()
N0


.

(38)

Note the similarity between (33) and (37).

Different windows for different subcarriers can efficiently be
achieved by employing two windows, corresponding to the extremes
of = 0.5 and = 1, and differently interpolating the results for each
subcarrier. Specifically

z = DH BTH y(r, 0.5) + (I B)TH y(r, 1)

(39)

where
B = diag {2 (1 1 )

2 (1 K )} .

(40)

From (39), we can see that two FFTs are needed: one for each window.
VI. P ERFORMANCE E XAMPLES
In Fig. 3, the SINR gain (relative to discarding the CP) is shown
as a function
of the SNR
(Es /N0 ) for a two-path channel (D = 2,

c0 = 0.6661, and c1 = 0.3339). The system in [5] is assumed,

in which there are 64 subcarriers (K = 64), and the CP consists of
17 chips (L = 17, L = 16). Gains were analytically determined using (19) and (28) for each subcarrier.
For the ML, TVW, and TVW2 approaches, the gains vary with the
subcarrier. Both minimum- and maximum-gain results are provided.
The minimum- and maximum-gain results correspond to the lowestand highest-frequency subcarriers, respectively. This can be explained
as follows: More gain occurs when performance is noise limited,
as these approaches trade interference to reduce noise. The channel
is low pass, providing more gain at low-frequency subcarriers. As
interference is larger from adjacent subcarriers, this maximizes I0 ()
at the lowest subcarrier and minimizes I0 () at the highest subcarrier.

3209

Fig. 4. Modem BER versus SNR (Es /N0 ), K = 64 subcarriers,

L = 17 chips CP, and a two-path (D = 2) channel. Results independent of the
subcarrier are provided for the RCOS and ONW approaches. Best-case results
(over subcarriers) are provided for the ML, TVW, and TVW2 approaches.

As a result, performance is most noise limited at the highest subcarrier

and least noise limited at the lowest subcarrier.
Results, which are independent of the subcarrier, are also
provided for the raised-cosine (RCOS) approach of Muschalik
[4] and the optimum Nyquist windowing (ONW) approach of
MllerWeinfurtner [5]. While the latter was not intended for dispersive channels, it is, in this case, equivalent to the approach in [6],
which considers dispersive channels. Results are not provided for the
approach in [7], although its performance should be no better than
ONW. For the RCOS approach, the roll-off is given by L /K [4]. For
the ONW approach, a TVW results, with values 0 and 0.5 [5].
At a low SNR, the gains over the best existing approach (ONW)
are highest. For the TVW and TVW2 approaches, approaches 1,
effectively giving equal weight to all received samples, giving up
orthogonality between subcarriers. This makes sense, as signal energy
recovery (matching to signal) is most important when noise dominates.
At a high SNR, the ML, TVW, and TVW2 approaches perform
similarly to the ONW approach. For the TVW and TVW2 approaches,
approaches 1/2, giving the same Nyquist window used by ONW.
Here, interference dominates, so that maintaining orthogonality (zero
forcing) makes sense. Observe that the TVW approach shows a slight
loss relative to the ML solution for the worst-case gain results, but it
shows a more significant loss for the best-case gain results. In contrast,
the TVW2 approach provides gains close to the ML solution in both
cases.
Theoretical modem
BER results for BPSK, which are obtained
using BER = 0.5 erfc{ SINR} [15], are shown in Figs. 4 and 5. For
comparison purposes, results for the conventional (CONV) approach
of removing the CP are also given. Observe from Fig. 5 that the gains
in terms of the required SNR to achieve a certain BER are comparable
to the SNR gains shown in Fig. 3.
A second example,
corresponding

to a four-path
channel (D = 4,

=
14,
c
=
0.5324,
c
=
0.2668,
c
=
0.1337, and c3 =
L
0
1
2

0.0670), is given in Fig. 6. Most trends are similar. However, note

that with more dispersion, the TVW approach worst-case gain can be
less than the gain of the ONW approach. With more dispersion, the
SINR of the different carriers varies more, so that the approximation of
I() by I becomes more severe, degrading performance. By contrast,
the TVW2 approach still performs close to ML, providing gains over
existing approaches.
Thus, for both channels considered, the gains are large at low
to moderate SNRs and small at a high SNR. Such gains would be
important to system coverage (providing service at large distances
from the transmitter) with low to moderate data rates and significant
FEC coding. Such a situation leads to low modem SNR levels.

3210

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 5, SEPTEMBER 2008

and V = [v1
gives

vK ]. Solving for V and applying (13) eventually

H
V = IK+L + C1
N FH(FH)

C1
N FH.

(43)

Assuming white stationary noise (CN = N0 IK+L ) and a nondispersive channel (FH = hE), (43) becomes

V = IK+L + |h|2 /N0 EEH

(1/N0 )E.

(44)

It can be shown that


Fig. 5. Modem BER versus SNR (Es /N0 ), K = 64 subcarriers,
L = 17 chips CP, and a two-path (D = 2) channel. Results independent of the
subcarrier are provided for the RCOS and ONW approaches. Best-case results
(over subcarriers) are provided for the ML, TVW, and TVW2 approaches.
Narrow ranges of BER and SNR values are shown.

EE

=K

IL
0
IL

0
IKL
0

IL
0
IL


.

(45)

Substituting (45) in (44) eventually gives

V = (h/N0 )

1
T(KL+1:K)
1+2K
1
T(1:KL)
1+K
1
T(KL+1:K)
1+2K

(46)

where = |h|2 /N0 , and T(m:n) denotes the submatrix consisting of

rows m to n of T. Substituting (46) in (14) and (11) eventually gives


z=

h
N0


TH y(r, )

(47)

2
(|c0 |2 Es /N0 )
1+L/K
2L/K
(|c0 |2 Es /N0 )
1+L/K

(48)

where
Fig. 6. SINR gain (relative to discarding the CP) versus SNR (Es /N0 ),
K = 64 subcarriers, L = 17 chips CP, and a four-path (D = 4) channel.
Results independent of the subcarrier are provided for the RCOS and ONW
approaches. Best- and worst-case gains (over subcarriers) are provided for the
ML, TVW, and TVW2 approaches.

h=

1+
1 + 2K
=
1 + (2K L)
1+

Es /(K + L) c0 = (

Ec )c0

(49)

VII. C ONCLUSION
In this paper, ML and MMSE receivers are developed for an
OFDM system when the delay spread is less than the length of
the CP. These solutions can be expressed as a matrix prefilter followed by an FFT and individual subcarrier scaling. If the channel
is nondispersive, the matrix prefilter simplifies to a TVW operation.
Lower complexity solutions are obtained by using one or two TVW
operation(s) (TVW and TVW2), even when the channel is dispersive.
Performance results show that at a low SNR, the ML and TVW2
solutions gain over existing Nyquist windowing approaches. The TVW
solution also gains over existing approaches when the dispersion
is mild.
A PPENDIX
In this Appendix, the ML solution is simplified for the special case
in which the channel is nondispersive and the noise is white and
stationary. Substituting (8) in (16), applying (13), and collecting results
into a matrix gives
V = C1
N FH VX

(41)

where
X = [x1

xK ] = (FH)H C1
N FH

(42)

1+
1 + K
=
1 + 2K
1+

1
(|c0 |2 Es /N0 )
1+L/K
.
2
(|c0 |2 Es /N0 )
1+L/K

(50)

ACKNOWLEDGMENT
The authors would like to thank J.-C. Guey, A. S. Khayrallah,
K. Zangi, and the anonymous reviewers for their helpful comments.
R EFERENCES
[1] R. W. Chang and R. A. Gibby, A theoretical study of performance
of an orthogonal multiplexing data transmission scheme, IEEE Trans.
Commun. Technol., vol. COM-16, no. 4, pp. 529540, Aug. 1968.
[2] A. M. Tonello, Performance limits for filtered multitone modulation in
fading channels, IEEE Trans. Wireless Commun., vol. 4, no. 5, pp. 2121
2135, Sep. 2005.
[3] S. B. Weinstein and P. M. Ebert, Data transmission by frequency
division-multiplexing using the discrete Fourier transform, IEEE Trans.
Commun., vol. COM-19, no. 5, pp. 628634, Oct. 1971.
[4] C. Muschalik, Improving an OFDM reception using an adaptive Nyquist
windowing, IEEE Trans. Consum. Electron., vol. 42, no. 3, pp. 259269,
Aug. 1996.
[5] S. H. Mller-Weinfurtner, Optimum Nyquist windowing in OFDM receivers, IEEE Trans. Commun., vol. 49, no. 3, pp. 417420, Mar. 2001.
[6] A. J. Redfern, Receiver window design for multicarrier communication
systems, IEEE J. Sel. Areas Commun., vol. 20, no. 5, pp. 10291036,
Jun. 2002.

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 5, SEPTEMBER 2008

[7] R. Song and S.-H. Leung, A novel OFDM receiver with second order
polynomial Nyquist window function, IEEE Commun. Lett., vol. 9, no. 5,
pp. 391393, May 2005.
[8] A. Scaglione and G. B. Giannakis, Design of user codes in QS-CDMA
systems for MUI elimination in unknown multipath, IEEE Commun.
Lett., vol. 3, no. 2, pp. 2527, Feb. 1999.
[9] G. E. Bottomley, T. Ottosson, and Y.-P. E. Wang, A generalized RAKE
receiver for interference suppression, IEEE J. Sel. Areas Commun.,
vol. 18, no. 8, pp. 15361545, Aug. 2000.
[10] S. Tantikovit, A. U. H. Sheikh, and M. Z. Wang, Combining schemes
in rake receiver for low spreading factor long-code W-CDMA systems,
Electron. Lett., vol. 36, no. 22, pp. 18721874, Oct. 26, 2000.
[11] B. Muquet, Z. Wang, G. B. Giannakis, M. de Courville, and
P. Duhamel, Cyclic prefixing or zero padding for wireless multicarrier
transmissions? IEEE Trans. Commun., vol. 50, no. 12, pp. 21362148,
Dec. 2002.
[12] X. Wang and K. J. R. Liu, Performance analysis for adaptive channel
estimation exploiting cyclic prefix in multicarrier modulation systems,
IEEE Trans. Commun., vol. 51, no. 1, pp. 94105, Jan. 2003.
[13] A. D. Whalen, Detection of Signals in Noise. Orlando, FL: Academic,
1971.
[14] T. Kailath, Linear Systems. Singapore: Prentice-Hall, 1998.
[15] J. G. Proakis, Digital Communications, 2nd ed. New York:
McGraw-Hill, 1989.

Residual Frequency Offset Compensation-Embedded

Turbo Decoder
Kwonhue Choi, Member, IEEE

AbstractWe propose a modified turbo decoder that internally includes

the residual frequency offset compensation function in an iterative decoding process. Based on the recently proposed phase offset compensation
scheme that embeds phase compensation into iterative turbo decoding,
we extend it to cover frequency offset compensation. At each decoding
iteration, we estimate the residual frequency offset from the extrinsic information and compensate it for the next iteration. As the iteration goes on,
the frequency offset is gradually and drastically reduced. Consequently,
the severe turbo decoding failure due to even a small residual frequency
offset can be significantly recovered, and then, almost ideal decoding
performance is achieved. In particular, for power-limited communications
under a very-low-SNR environment where even a tiny frequency offset is
critical to turbo decoding, the proposed scheme is efficient since it does not
require additional transmission power for compensation.
Index TermsCarrier recovery, iterative decoding, low SNR, synchronization, turbo code.

I. I NTRODUCTION
For a power-limited transmission or low-received-SNR environment, powerful channel codings, such as the turbo code or low-density
parity-check code, have been widely used. To achieve the desired error
correction capability, strict synchronization is a crucial prerequisite
prior to the decoding process. It is observed that even a small residual
frequency offset in the decoder input causes a severe effective SNR
reduction, and thus, the decoding performance significantly degrades.
Manuscript received March 1, 2007; revised September 13, 2007,
November 9, 2007, and December 6, 2007. This work was supported in part
by Yeungnam University under research Grant 205A236007 in 2005 and in
part by the Ministry of Knowledge Economy, Korea, under the Information
Technology Research Center support program supervised by the Institute
of Information Technology Assessment under Grant IITA-2008-C1090-08010045. The review of this paper was coordinated by Associate Prof. X.-G. Xia.
The author is with Yeungnam University, Korea (e-mail: [email protected]).
Digital Object Identifier 10.1109/TVT.2007.914482

3211

In general, we can achieve reliable frequency error compensation by

using additional transmission power. However, this is prohibitive in a
high-power efficient system.
Recently, in [1][4], it has been shown that the iterative phase
compensations embedded into a turbolike iterative decoding process
achieve a bit error rate (BER) curve close to that of the ideal phase
recovery. This excellent synchronization performance comes from the
nature of the powerful capability of iterative decoding, particularly for
low SNRs. Unlike other conventional schemes, this algorithm fully
exploits the intrinsic virtue of an iterative mechanism, i.e., gradual
reliability enhancement of the soft output as the iteration goes on,
and applies it to the iterative phase recovery to get the gradual enhancement of phase-estimation reliability. At each decoding iteration,
phase estimation and compensation is performed by using the soft
output. Consequently, the multiple phase compensations, as many as
the decoding iteration number, are performed at each codeword, and
thus, the separate phase recovery is possible without codeword-bycodeword closed-loop phase recovery. This implies that the successive
codeword transmission is not necessary. Furthermore, the additional
complexity of embedding the phase recovery function into a turbo
decoder is quite modest.
However, the phase offset is estimated by averaging all of the
observations in the codeword, and then, all of the codeword samples
are derotated with the same amount of phase offset, which inhibits
this algorithm to be well suited to the case when the phase offset is
time varying within the codeword, say, a residual frequency offset
exists. When there is a residual frequency offset in the codeword
samples, constellations of the codeword samples rotate throughout the
codeword, and the codeword samples phase error drifts. Therefore,
the constant phase compensation for all codeword samples achieves
zero phase error only in the middle of the codeword. Consequently,
the phase offset at the beginning and the end of the codeword becomes
significant for the case when the frequency offset is non-negligible or
the case when the code length is long. Simulation results reveal that
the decoding performance significantly degrades, even with a small
phase rotation. This shortcoming of the constant phase compensation
makes it necessary to consider some modification for frequency offset
compensation.
In [5], the authors proposed carrier frequency offset compensation
algorithms for a turbo decoder. By using the fact that the mean square
value of the soft decision output (so-called MSSO) is maximized
when the frequency and phase offset is correctly compensated, they
have searched the frequency and phase offset pair that maximizes
MSSO. The critical drawback of this algorithm is that it has to perform
multiple turbo decodings for the same codeword (as many as the
number of frequency and phase pairs to be searched) since it only relies
on exhaustive search. Even though this algorithm may achieve almost
ideal synchronization with a wide range of allowable frequency offset
by increasing the search area and resolution, its computation time and
complexity are unacceptable in the practical application.
In this correspondence, we extend the basic concept of the iterative
phase compensation employed in [1], [2], and [4] to include frequency
compensation. At each decoding iteration, the residual frequency is
calculated by simply measuring the slope in the phase offset within
the decoder input sequence and is compensated for the next decoding
iteration. At each iteration, the proposed frequency estimation is very
simple, and thus, the estimation error is larger compared to a more
complicated and more optimized method. However, as the iteration
goes on, the frequency offset is repeatedly calculated and compensated
by using the gradually enhanced decoder output. Finally, we can

0018-9545/$25.00 2008 IEEE