2234
5, MAY 2011
Synthesizing Low Autocorrelation and Low PAPR
OFDM Sequences Under Spectral Constraints
Through Convex Optimization and GS Algorithm
Lung-Sheng Tsai, Student Member, IEEE, Wei-Ho Chung, Member, IEEE, and Da-shan Shiu, Member, IEEE
Abstract—Sequences with low autocorrelation (AC) and low
peak to average power ratio (PAPR) are desired in many commu-
nication or signal processing applications. This work investigates
the synthesis of OFDM sequences with the desired low AC and low
PAPR under spectral constraints. The spectral constraints limit
the maximum allowable power on each subcarrier to avoid inter-
ference on particular reserved bands and undesirable DC-offset.
The first part of this work discusses the design of sequences with
periodic AC property under spectral constraints. Specifically, we
present a convex optimization method for synthesizing OFDM
sequences with minimized peak sidelobe level (PSL) or weighted
sum of sidelobe levels (WSSL) of the periodic AC function within
specified time lags. The design objective to be minimized can
also be a certain convex function of the sequence AC sidelobes.
Furthermore, we present methods based on Gerchberg-Saxton
(GS) algorithm to decrease the PAPR of the sequences, while
maintaining the optimized AC characteristic. The second part of
this work investigates the aperiodic AC under spectral constraints.
In this case, the optimal sequence design problem is nonconvex.
By relaxing the nonconvex problem to a convex problem, we
provide lower bounds for PSL and WSSL of the aperiodic AC
function. Based on the optimal solution for the relaxed convex
problem, we present an efficient algorithm to find sequences with
low PAPR and near-optimal aperiodic AC property while the
spectral constraints are satisfied.
Index Terms—Aperiodic, auto-correlation, low correlation zone,
peak to average power ratio, periodic, training sequence, zero cor-
relation zone.
I. INTRODUCTION
S EQUENCES with low autocorrelation (AC) are crucial
components in communications and signal processing ap-
plications. The sequences with low AC can be used as probing
Manuscript received September 06, 2010; revised January 09, 2011; accepted
January 12, 2011. Date of publication January 28, 2011; date of current version
April 13, 2011. The associate editor coordinating the review of this manuscript
and approving it for publication was Prof. Philippe Ciblat. This work was sup-
ported by the National Science Council of Taiwan, by Grant 97-2221-E-002-013
and 99-2218-E-001-004.
L.-S. Tsai was with the Graduate Institute of Communication Engineering,
National Taiwan University, Taiwan. He is currently with the Research Center
for Information Technology Innovation, Academia Sinica, Taiwan (e-mail:
longson@ieee.org).
W.-H. Chung is with the Research Center for Information Technology Inno-
vation, Academia Sinica, Taiwan (e-mail: whc@citi.sinica.edu.tw).
D. Shiu is with the Department of Electrical Engineering and the Graduate
Institute of Communication Engineering, National Taiwan University, Taiwan
(e-mail: dsshiu@cc.ee.ntu.edu.tw).
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/TSP.2011.2108652
1053-587X/$26.00 © 2011 IEEE
IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO.
signals to estimate the channel impulse response [1]–[3], or
as training sequences for correlation-based synchronization
[4]–[6] and frequency offset estimation [7]. For example, the
Zadoff-Chu sequences [8], [9] have been adopted as the pre-
amble sequences in 3GPP long-term evolution (LTE) [10]. In
these applications, the low AC property reduces or eliminates
self-interference caused by multi-paths. Such sequences can
also be used as the signature codes, e.g., the well-known -se-
quences adopted in code division multiple access (CDMA)
systems [11]–[13]. The design of sequences with good AC
properties has been the focus of research in the past decades,
due to its crucial roles in various systems. Readers are re-
ferred to the books [14], [15] and the references therein for
more details. Recently, families of sequences referred to as
zero-correlation zone (ZCZ) sequences and low-correlation
zone (LCZ) sequences have attracted numerous research in-
terests. The research on ZCZ and LCZ sequences concerns
the design of sequences achieving zero correlation or low
correlation within a desired time-delay interval. The sequences
have been investigated and used as the spreading sequences for
quasi-synchronous CDMA systems [16], [17], or as the training
sequences for the systems with limited channel delay spread
[18]–[20].
In addition to the good AC property, it is also desired to ob-
tain sequences with low peak to average power ratio (PAPR) due
to the limited amplification range of the RF amplifier. There
exists literature investigating sequences with low AC and low
PAPR jointly. For instance, the aforementioned Zadoff-Chu se-
quences exhibit the optimal periodic AC property, as well as the
appealing characteristic of PAPR to be 1. Methods for synthe-
sizing unimodular sequences with low periodic AC and aperi-
odic AC have been proposed in [21] and [22].
While the existing works have substantially contributed to
the design and theoretical understanding of the sequences, the
sequence design with spectral constraints had been rarely dis-
cussed. The spectral constraints on transmitted signals are usu-
ally required in communication systems to avoid causing inter-
ference. For example, in OFDM systems, certain subcarriers are
reserved and are prohibited to transmit signals [23]. The DC
subcarrier is reserved, i.e., nulled, to avoid offsets in the D/A
and A/D converter in RF systems. In addition, the guardbands
at spectrum edges are reserved to prevent interferences on adja-
cent frequency channels. As another example, in radar applica-
tions, signals must avoid occupying the bands reserved for nav-
igation or military operations [24]. A possible solution for gen-
erating sequences satisfying the spectrum constraints is to filter
TSAI et al.: SYNTHESIZING LOW AC AND LOW PAPR OFDM SEQUENCES 2235

the existing sequences with good AC property through a spec- use to denote its periodic autocorrelation and to de-
trum mask [18]. However, because the filtered sequences are not note its aperiodic correlation at time-delay . The sequences and
explicitly optimized toward the desired AC property under the their elements discussed in this paper are complex-valued. Each
constraints specified by the spectrum mask, the AC properties sequence in our discussion is normalized with unit Euclidean
of the filtered sequences are not as good as those of the original norm. The Hermitian and transpose operators over matrices or
sequences. A recent work [25] considered the waveform design vectors are denoted by and , respectively.
for cognitive radar system. The method presented in [25] is used Definition 1: (Periodic AC): Let denote (mod ). The
to obtain unimodular sequences with suppressed power in arbi- periodic AC function of a length- sequence is defined by
trarily chosen bands and with low aperiodic AC sidelobe level.
In contrast to that work, we focus on designing the sequences (1)
strictly satisfying the given spectral constraints.
Our main goal in this work is to design sequences with good
AC property and low PAPR under the condition where the given Definition 2: (Aperiodic AC): The aperiodic AC function of
spectral constraints must be met. Specifically, our design ob- a length- sequences is defined by
jectives are the metrics of the goodness in the AC property in-
cluding the peak sidelobe level (PSL) and the weighted sum of
sidelobe levels (WSSL) of the AC function. The design objec- (2)
tive can also be a certain convex function of the sequence AC
sidelobes. The two metrics PSL and WSSL had been adopted
as design criteria in [21], [22], [25]–[27]. To demonstrate an
example of practical application, we will present the connec-
Definition 3: [Discrete Fourier Transform (DFT) and In-
tion between the AC metrics and the channel estimation error
verse-DFT (IDFT) Matrices]: For a length- sequence , its
when the proposed sequence is applied as the training sequence
-point DFT, denoted by ,
for channel estimation. The scope of this work is not limited to
is given by . We define as the
the optimal training sequence design for a particular estimator,
transformation matrix such that , and define .
and instead, we look for the sequences with good AC properties
The matrices and are referred to as a DFT matrix and an
which facilitate various applications. The cases for both of the
IDFT matrix. The sequence is called a time-domain sequence
periodic and the aperiodic AC properties will be examined.
while the sequence is called a transform-domain sequence.
In this paper, we first consider the periodic AC and provide
Definition 4: (Power Spectrum of a Sequence): For a
a convex optimization approach to the synthesis of AC-opti-
length- sequence, using as an example, the vector
mized sequences under the given spectral constraints. As will be
shown, the AC-optimized sequence is not unique, and our pro- is the power spectrum
posed approach provides a sufficient condition to guarantee the vector of .
periodic AC optimality for the synthesized sequences. Based on The following property was introduced in [14] and will be
this condition, we present an algorithm to further decrease the used in later sections of this paper.
sequence PAPR while maintaining the optimal AC property. Property 1: The DFT of the periodic AC function
Subsequently, we investigate the aperiodic case. Our studies is equal to , where denotes
show the optimal sequence design problem in this case is non- the element-wise multiplication operator.
convex. We relax the nonconvex constraints and obtain lower The property below follows immediately.
bounds for the PSL and WSSL of the aperiodic AC function. We Property 2: Let denote the DFT of
also provide algorithms to synthesize sequences with near-op- . Then we have
timal aperiodic AC properties and together with low PAPR.
The rest of this paper is organized as follows. In Section II, (3)
we define notations and review properties that are used in the
subsequent discourse. Section III presents methods to synthe- Property 3: A sequence is with impulse-like AC
size spectrum-constrained sequences with optimal periodic AC property if and only if is equal to a constant for all .
property and low PAPR. In Section IV, we investigate the se- Property 2 shows that, for a sequence, its power spectrum
quence synthesis when considering the optimality on the ape- determines its periodic AC function. This property serves as
riodic AC function. Finally, concluding remarks are given in our basis to design AC-optimized sequences under spectral
Section V. constraints.

II. PRELIMINARIES B. Problem Formulation

A. Definitions and Basic Properties
The notations are defined in this section, mostly following
the conventions in [14] and [26]. We denote a sequence con-
sisting of elements by the form of either a set or
an -by-1 vector . Given a sequence , we
We aim to design an OFDM sequence under spectral con-
straints so that the AC property of is optimized. To formu-
late this design problem, first the goodness of the AC property
needs to be quantified. There are two commonly practiced met-
rics for this quantification. The first metric is the peak sidelobe
level (PSL) of the AC function [26] at specified time-lags. We
2236
denote the specified time-lags of interest by a set . The second
metric is the weighted sum of the sidelobe levels (WSSL) of the
AC function [27]. We denote the weighting coefficients by an
-by-1 vector with . These
two metrics are formally formulated as
(4)
(5)
The above metrics are used when the periodic AC function
is considered. For the aperiodic case, we replace with
. The set is called a low-correlation zone.
We describe the spectral constraints by a vector
, where denotes the maximum allowable
power on the th subcarrier. This vector is referred to as
a spectrum mask. The sequence is said to satisfy the spec-
trum-mask constraint if for all . Our main goal
is to obtain sequences satisfying the spectrum-mask constraint
and with minimized PSL or WSSL.
III. SPECTRUM-CONSTRAINED SEQUENCES WITH OPTIMIZED
PERIODIC AC PROPERTY
This section presents our framework for designing spectrum-
constrained sequences with optimal periodic AC property. We
reformulate the design problem for the optimal sequences as
a convex optimization problem. The optimal solution for this
convex problem serves as a sufficient condition to guarantee the
periodic AC-optimality. Based on this condition, the designing
strategies for sequences with both desired periodic AC property
and low PAPR will be discussed. At the end of this section,
we present a possible application of using the AC-optimized
sequences as the training sequence for channel estimation. The
connection between the channel estimation error and the PSL/
WSSL of sequence’s periodic AC function will be presented.
A. Sufficient Condition to Minimize PSL or WSSL
To formulate the sequence design problem as a convex opti-
mization problem, we start with reformulating the expressions
of the constraints and the objective function. We define as
the power spectrum vector of the desired sequence , i.e.,
with . By Parseval’s Theorem, we
have , so the unit Euclidean-
norm constraint is equivalent to
(6)
The spectrum-mask constraint is given by
(7)
Following Corollary 1 and representing the IDFT matrix as
, we have , and the two metrics,
PSL and WSSL, can be written as
(8)
IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 5, MAY 2011
(9)
To convert the problem of minimizing PSL or WSSL to a
convex optimization problem, we express the absolute values
of the complex inner-products in (8) and (9) as functions of real
variables. Given , we define a 2-by- matrix as
(10)
where and are -by-1 vectors consisting of the
real part and the imaginary part of the vector , so the two
elements of correspond to the real part and imaginary part
of . We then have
(11)
(12)
As a result, the PSL/WSSL minimization problem subject to
(6) and (7) is formulated as a second-order cone programming
(SOCP) problem with the variable , which belongs to the cat-
egory of convex optimization problems [28].
In general, the AC metric can be formulated as a more sophis-
ticated function of the AC sidelobe levels to serve as a design
criterion for a particular estimator. Supposing a nonnegative real
vector is a power spectrum vector of a sequence, we consider
an objective function given by
(13)
where , and with
. It is noted that the previously defined
is convex and nonnegative. It can be shown
that is convex if is convex and the extended-value extension
of , denoted by , is nondecreasing in each argument (referred
to [28, pp. 86–87]). Here assigns to points not in the do-
main of . We name such a convex a convex generalized AC
metric. Under the spectrum-constraint and the unit-norm con-
straint, the problem of minimizing the convex generalized AC
metric is an SOCP problem. Such SOCP formulations on the
generalized metrics facilitate the designs with more generic cri-
teria for future applications. In subsequent discussions we focus
on the two metrics PSL and WSSL.
We use packages CVX, SDPT3, and SeDuMi [29]–[31] to
solve the SOCP problems. The optimized values of the objective
functions represent the achievable lower bounds of the PSL and
WSSL under the spectral constraints; the optimal solution of
the sequence power spectrum is a sufficient condition for the
AC-optimality of sequences.
The sufficient condition indicates that the optimality can be
achieved by properly arranging the power spectrum of a se-
quence, independent of the phase at each subcarrier. This prop-
erty also implies that the phases of the elements of the AC-opti-
mized sequence in the transform domain are free variables that
can be used to further design sequences exhibiting other desired
characteristics, e.g., low PAPR, while the optimality of the AC
TSAI et al.: SYNTHESIZING LOW AC AND LOW PAPR OFDM SEQUENCES
property is maintained. In the remainder of this section, we will
present methods to design spectrally constrained sequences with
both low correlation and low PAPR.
There are various approaches for adjusting the phases for
the transform-domain sequence with the optimal power spec-
trum. For instance, in the random search approach, multiple
sets of random phases are generated and the sequence with the
lowest PAPR is selected. However, this approach results in ac-
ceptable outcomes only for the case of short sequence length.
In the following, we present an efficient algorithm, called Ger-
chberg-Saxton algorithm, to obtain low-PAPR sequences of op-
timal AC property.
B. Gerchberg-Saxton Algorithm
Gerchberg-Saxton (GS) algorithm [32] deals with a nearest
vector problem in the following form:
(14)
(15)
(16)
where , and are given.
The GS algorithm is an iterative algorithm that updates the
variables and alternatively in each iteration, and it eventually
achieves a local or global minimum solution. In this algorithm,
the initial values for and are randomly assigned. In the th
iteration, this algorithm performs two main steps:
1) fix ; find the optimal that minimizes .
2) fix ; find the optimal that minimizes .
For the termination condition, the iteration process is stopped if
either the number of iterations exceeds a prescribed threshold
or the difference of the cost function obtained in two consec-
utive iterations is less than a given threshold. It was proved in
[33] (and also outlined in [34]) that the cost function is a
monotonically decreasing function of the number of iterations
. Thus, this algorithm is guaranteed to converge, either to a
local or global minimum solution.
C. Spectrum-Constrained Sequences With Low Periodic AC
and Low PAPR
In this section, we investigate the design for sequences with
both low AC and low PAPR. The PAPR of a sequence is de-
fined by
(17)
Unimodular sequences achieve the lowest PAPR, i.e.,
. Stoica et al. [21] had proposed applying the GS algorithm
to synthesize unimodular sequences with impulse-like periodic
correlation. Their method is based on the property that the DFT
of a sequence with impulse-like periodic correlation has con-
stant magnitude, as elaborated in Property 3. The GS algorithm
was used to find a unimodular sequence with an almost flat
power spectrum by minimizing the following criterion [21]:
(18)
2237
where , , and the phases of and
are free real variables.
The spectrum constraints are further considered in the fol-
lowing. We set as a transform-domain sequence with the
power spectrum which minimizes the PSL (or WSSL) and
satisfies the spectrum-mask requirement. To improve Stoica’s
method to synthesize sequences under the spectrum-mask
constraint, we modify the criterion in (18) as
(19)
where , , is the th ele-
ment of the optimal for the PSL/WSSL minimization problem,
and the phases are free real variables. After
performing the GS algorithm to minimize , we denote the re-
sulting solution pair as . If , it can be shown
that . In this case the unimodular sequence
has the optimal AC property and also satisfies the spec-
tral requirements. For , the sequence exhibits
optimal AC property and satisfies the spectral constraints, while
its PAPR is near-optimal. On the other hand, has optimal
PAPR, while its AC behavior is suboptimal and there may be
slight power leakage out of the spectrum mask. One may choose
or by considering the requirements in applica-
tions.
Finally, we remark that the sequence design problem is an
off-line problem, so the GS algorithm can be performed repeat-
edly with different initial values and the sequence with the best
desired property can be selected as the final solution.
D. Numerical Examples
The settings in IEEE 802.11a standard are used in the fol-
lowing examples to demonstrate the proposed design. The se-
quence length is 64. There are 64 subcarriers, and 12 of them
are nulled at the DC-subcarrier and guard bands. The set of in-
dexes of these null subcarriers is . The
spectrum mask is set as
(20)
The length of cyclic-prefix (CP) is 16. Let the parameter
denote the size of the low correlation zone. We set
and because the length of
the channel response is usually less than the CP length.
We obtain the sequence pair by performing
the GS algorithm to minimize the PSL or WSSL. When the
WSSL metric is used, the weighting vector is set as
(21)
Fig. 1 illustrates the sequence which achieves the min-
imum WSSL. The corresponding PAPR is only slightly greater
than 1. Fig. 2 shows the unimodular sequence which has
WSSL slightly greater than the minimum. However, as pre-
viously mentioned in Section III-C, the obtained unimodular
sequence has slight spectral leakage into the reserved bands.
2238 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 5, MAY 2011

Fig. 1. A sequence (~ x F ~u ) with minimized WSSL and its AC func-

Fig. 3. A sequence (~ 
x F ~u ) with minimized PSL and its AC function,
given L = 16. The minimized PSL is 0.1093. The sequence PAPR is 1.1626.
tion. The minimized WSSL is 0.0851. The sequence PAPR is 1.8401. The solid
The solid line in the plot of the AC function stands for the AC of the 64-QAM
line in the plot of the AC function stands for the AC of the 64-QAM quantized
quantized sequence, with PSL = 0:1541 and PAPR = 1:3439.
sequence, with WSSL = 0:0851 and PAPR = 1:9095.

Fig. 2. A unimodular sequence (~ 

x ~x ) with near-minimum WSSL and its
Fig. 4. A unimodular sequence (~ 
x ~x ) with near-minimum PSL and its
AC function, given L = 16. The corresponding PSL is 0.125.
AC function. The corresponding WSSL is 0.0914.

AC properties of the quantized sequences are slightly degraded

Figs. 3 and 4 illustrate the synthesized sequences and
in these examples.
, respectively, when the PSL metric is used.
F. Training Sequences for a Least-Square Channel Estimator
E. Sequences Belonging to Finite Constellations
Here we consider the training sequence design for channel
In the above discussions, we have presented complex-valued
estimation. This is to demonstrate an application with sequence-
sequences with good AC properties. In many applications, it
design criteria in the form of the convex generalized AC metric
may be desired to synthesize OFDM sequences whose trans-
where the proposed algorithms can be applied.
form-domain elements belong to particular constellations, e.g.,
Suppose the training sequence is transmitted through a
4-QAM in IEEE 802.11a standard [23]. With the additional con-
time-invariant channel . The received
straint of the constellation on sequence constellations, the AC
signal, denoted by , can be expressed as
optimization problem becomes a mixed-integer programming
(MIP) problem, which is generally NP-hard. (22)
The optimal solution for such an MIP problem can be ap- where is the additive noise vector, and the matrix formed
proximated through quantization of the sequence obtained from by the training sequence is given by
the convex and GS techniques presented in this work. We im-
plement examples by quantizing the nonzero elements of the
AC-optimized sequences in Figs. 1 and 3 to 64-QAM constel- .. .. .. (23)
lations in transform domain. The AC functions of the quantized . . .
sequences are also shown in the two figures. We observe that the
TSAI et al.: SYNTHESIZING LOW AC AND LOW PAPR OFDM SEQUENCES 2239

It it noted that for under the assump-

tion that the CP is added, and for if
the zero-padding (ZP) guard period is added. As will be shown
later, the channel estimation error for the CP-based scenario is
related to the periodic AC of the training sequence. On the other
hand, for the ZP-based scenario, the channel estimation error is
related to the aperiodic AC. In the following, we focus on the
CP-based case since the ZP-based case has similar formulations
for the channel estimation error.
From (22), the LS channel estimator becomes

(24)

if the -by- matrix has full rank. If is a zero-

mean complex white Gaussian vector with covariance matrix
, the corresponding mean-squared estimation error (MSEE) Fig. 5. A sequence ~
x with minimized PSL given L = 5 and its AC function.
is [2], [35] The minimized PSL is 0.

(25) IV. SPECTRUM-CONSTRAINED SEQUENCES WITH GOOD

APERIODIC AC PROPERTY
trace (26) In this section, we extend our convex-optimization based
framework for sequence design to the case considering
the aperiodic AC function. We show the design problem for
where is the th smallest eigenvalue of .
AC-optimized sequence is nonconvex. By relaxing a nonconvex
To design the training sequence, instead of directly mini-
constraint, the problem is converted to a convex problem, and
mizing the MSEE in (26), the authors in [35] suggested to min-
imize a upper bound of the MSEE. The MSEE can be bounded we provide a lower bound for the PSL/WSSL of the aperiodic
AC function. Furthermore, based on the optimal solution of the
by
relaxed problem, we propose a sequence design method that
considers both the sequence aperiodic AC property and PAPR.
(27)
Examples will be presented to demonstrate the sequences
synthesized by our method.
Gerschgorin disc theorem [35], [36] provides the following
lower bound on : A. Lower Bound of the Aperiodic AC Property of
Spectrum-Constrained Sequences
Similar to Section III-A, here we aim at representing the PSL
(28) and WSSL of the aperiodic AC function as a function of
real variables so that the convex optimization tools can be uti-
where is the diagonal element of the matrix , and lized. We define a -by-1 vector associated with the length-
is the absolute sum of the off- sequence as , where zeros are padded.
diagonal elements in the th row of for . The -IDFT of is denoted by . The aperiodic AC function of
From (27) and (28), it was proposed to design training sequences then can be represented by the periodic AC function of , i.e.
by minimizing in [35]. Similarly, we can treat the mini- (29)
mization of the PSL as another approximate design criterion for
We note that . The similar technique of
the optimal training sequence.
zero-padding had been used in [26]. With the relationship
It is noted that both and PSL belong to the category
in (29) and the technique developed in Section III-A for the
of the convex generalized AC metric defined in Section III-A. periodic AC function, the WSSL/PSL metric for can be
Therefore, for the spectrum-constrained case, the problem of represented as a linear functional on the power spectrum of ,
minimizing or PSL is a convex optimization problem which is a real vector.
and can be solved by the algorithms proposed in this work. We define the power spectrum vector of as
Moreover, from (28), a training sequence is guaranteed to be , where is the th element of
optimal if the PSL of its AC function is zero for all delay within the -DFT of . The PSL of the aperiodic AC function
the maximal channel delay spread, . The minimum MSEE then can be written as
is a constant independent of the channel response. Under
the settings used in Section III-D, we find that sequences with
exist given . Fig. 5 shows this optimal training
sequence in both the time domain and the transform domain. (30)
2240 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 5, MAY 2011

where is the th row vector of the -by- The spectral mask constraint in (40) is due to . The
IDFT matrix , and problem of minimizing PSL/WSSL under the above constraints
. The WSSL of the aperiodic AC is not convex with the variable due to the equality constraint
function can be written as in (38). Here we replace this constraint by two relaxed convex
constraints of (39), i.e., the upper bound and lower bound
on , so that the resulting optimization problem is
(31) convex. To see this, we rewrite the expression of as

Next, we convert the constraints on the sequence to the con-

straints on so that the sequence design problem can be for- (41)
mulated as an optimization problem with the variable . Before
formulating the constraints, we consider the following neces-
sary and sufficient condition on the transform-domain sequence
to ensure that is the zero-padded version of .
Property 4: A length- sequence is the zero-padded ver-
sion of a length- sequence if and only if (42)

(32)
Suppose the set of indexes of the null subcarriers is denoted by
(33) . By applying the arithmetic-geometric means inequality to
(42), we have the following lower and upper bounds, denoted
by and , on :
for . In other words, the even-indexed el-
ements of are exactly those of the original transform-do-
(43)
main sequence , while the odd-indexed elements of are
linear combinations of the elements of . Defining
where
, we have , where the ma-
trix is a unitary matrix with elements given by

(34) (44)

Proof: Because the DFT operation is invertible, we only

need to show that (32) and (33) are true if . The
th element of the -DFT of can be expressed as (45)

(35)

(36)
(46)

(37) These two bounds and are linear combina-

tion of the elements of , and, therefore, they are convex
constraints.
After manipulating the equations, we obtain the results in (32) It then follows that, with the variable , the problem of mini-
and (33). mizing (30) or (31) subject to
We now return to the formulations of the constraints in
the PSL/WSSL minimization problem. From the Parseval’s
theorem, we have and
. As a result, in the transform
(47)
domain, the constraints for the minimization problem include
(48)
(38) (49)

(39) is a convex optimization problem. The lower bounds of PSL and

(40) WSSL then can be efficiently computed.
TSAI et al.: SYNTHESIZING LOW AC AND LOW PAPR OFDM SEQUENCES 2241

B. Spectrum-Constrained Sequences With Low Aperiodic AC

and Low PAPR
The next goal is to find the sequences with good aperiodic AC.
Obviously, the optimal solution of for the problems subject to
(47)–(49) may not be a realizable power spectrum of due to
the relaxation of the constraint . We let the solution
for the relaxed PSL/WSSL minimization problem be denoted by
and treat as a guide for the optimal
power spectrum of and . To find the desired sequence , we
pursue a transform domain sequence , i.e., , such that
and can be as close as possible. This is accomplished by
minimizing the following function:

(50)

where the elements of the variables , are restricted to

Fig. 6. Trade-off between WSSL and PAPR of sequences obtained by applying
the GS algorithm, given L = 16 . The WSSL lower bound is 0.
(51)

When PAPR is further considered, similar to the method used

in [24] and [25], we try to find a solution set that where is a constant. Then the optimal is given by
minimizes the following cost function:
(56)
(52)
where is the th element of , and is the th
The elements of the variables , , and are restricted to element of .
the forms of 4) Iterate between steps 2 and 3 until the termination condi-
tion is met.

C. Numerical Examples
and (53) In the following examples, we adopt the same setting used in
Section III-D. We present the WSSL and PAPR of the synthe-
where the phases of , and for are sized sequences for the cases with and in
free real variables. In (52), the first term is the penalty for the Figs. 6 and 7, respectively. These figures also illustrate the ef-
loss of the AC optimality, and the second term is the penalty for fects of the controlling factor used in (52). We simulate cases
the sequence PAPR. The constant controls the rela- with for . Each simulation point in
tive weighting of the two penalty terms. Clearly, if there exists these figures corresponds to a sequence obtained by the GS al-
such that for , the sequence gorithm with a given . For each point, the phases of the free
is a unimodular sequence satisfying the spectral requirements variables, , , and , are randomly and indepen-
and also has the optimal AC property. dently initialized. The results corresponding to the PSL metric
The GS algorithm is used again to find that lo- are omitted due to the similarity of their characteristics to those
cally minimizes the cost function . It consists of the following corresponding to the WSSL metric.
steps. As shown in Fig. 7, given and , the synthe-
1) Randomly generate initial values of , , and for all sized sequence closely follows the WSSL lower bound. Fig. 8
. presents this sequence in both the time domain and the transform
2) Fix ; update the values of and by the optimal ones domain. We observe that the sequence has a flat power spectrum
that minimize the cost function . They are given by except at the spectrum nulls.
Finally, we consider the effects of the controlling factor .
(54) The WSSL of the obtained sequence at is denoted as
. The results in Figs. 6 and 7 follow a trend that a
where is the th row of , is the th row of , and larger leads to the lower WSSL and larger PAPR for the syn-
is an operator that returns the argument (also known thesized sequences. Nevertheless, we have to remark that the
as phase) of a complex number. GS algorithm does not promise a global minimizer of and
3) Fix and ; update the value of by the optimal one we perform a relaxation to generate a reference power spectrum
that minimizes the cost function . This optimal value can of , which may introduce suboptimality. Our empirical results
be derived by expanding the cost function as [25] show that the property for
is mostly followed in the trend, without strict point-by-point
(55) guarantee.
2242 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 5, MAY 2011

PSL and WSSL of the aperiodic AC function over all possible

sequences. This was accomplished by extending the convex for-
mulation for the optimal sequence design from the periodic-AC
case to the aperiodic-AC case. Algorithms to synthesize spec-
trally constrained sequences with low PSL/WSSL and together
with low PAPR were presented. The effectiveness of the pro-
posed approaches and the desirable properties of the resulting
sequences were demonstrated through numerical examples.

Fig. 7. Trade-off between WSSL and PAPR of sequences obtained by applying
the GS algorithm, given L = 64 . The dashed line indicates the WSSL lower
bound, 0.057.
Fig. 8. A sequence with near-minimum WSSL and its aperiodic AC function,
given L = 64 . The corresponding WSSL is 0.067. The WSSL lower bound
is 0.057.
V. CONCLUSION
This paper presents approaches to the synthesis of OFDM se-
quences with low AC and low PAPR under spectral constraints
through convex optimization techniques. Utilizing the fact that
the power spectrum and the periodic AC function of a sequence
constitute a Fourier transformation pair, we showed that, given
a sequence, the PSL and WSSL of its periodic AC function can
be represented as convex functions of its power spectrum. Con-
sequently, the sequence design problem for an optimal periodic
AC property under the spectral constraints can be formulated as a
convex optimization problem. The convex optimization problem
can be efficiently solved by the existing approaches and tools.
Through the proposed convex formulation, the optimal solution
of the sequence power spectrum is a sufficient condition for the
AC-optimality of sequences. Since the phases of the sequence
components on each subcarrier are unrelated to the constraints
on the sequence power spectrum, we further applied the GS al-
gorithm to find sequences with low PAPR while maintaining the
optimality of the AC property. Moreover, given the spectral con-
straints and sequence length, we provided lower bounds for the
REFERENCES
[1] A. Milewski, “Periodic sequences with optimal properties for channel
estimation and fast start-up equalization,” IBM J. Res. Develop., vol.
27, no. 5, pp. 426–431, 1983.
[2] S. Crozier, D. Falconer, and S. Mahmoud, “Least sum of squared errors
(LSSE) channel estimation,” IEE Proc.-F, vol. 138, pp. 371–378, Aug.
1991.
[3] J. Ng, K. Letaief, and R. Murch, “Complex optimal sequences with
constant magnitude for fast channel estimation initialization,” IEEE
Trans. Commun., vol. 46, pp. 305–308, Mar. 1998.
[4] T. Schmidl and D. Cox, “Robust frequency and timing synchroniza-
tion for OFDM,” IEEE Trans. Commun., vol. 45, pp. 1613–1621, Dec.
1997.
[5] J. Van de Beek, M. Sandell, and P. Borjesson, “ML estimation of time
and frequency offset in OFDM systems,” IEEE Trans. Signal Process.,
vol. 45, pp. 1800–1805, Jul. 1997.
[6] U. Lambrette, M. Speth, and H. Meyr, “OFDM burst frequency syn-
chronization by single carrier training data,” IEEE Commun. Lett., vol.
1, pp. 46–48, Mar. 1997.
[7] H. Minn, X. Fu, and V. Bhargava, “Optimal periodic training signal
for frequency offset estimation in frequency-selective fading channels,”
IEEE Trans. Commun., vol. 54, pp. 1081–1096, Jun. 2006.
[8] R. Frank, S. Zadoff, and R. Heimiller, “Phase shift pulse codes with
good periodic correlation properties (corresp.),” IRE Trans. Inf. Theory,
vol. 8, pp. 381–382, Oct. 1962.
[9] D. Chu, “Polyphase codes with good periodic correlation properties
(corresp.),” IEEE Trans. Inf. Theory, vol. 18, pp. 531–532, July 1972.
[10] 3GPP TS 36.211 V8.2.0, 3rd Generation Partnership Project; Technical
Specification Group Radio Access Network; Evolved Universal Ter-
restrial Radio Access (E-UTRA); Physical Channels and Modulation
(Release 8) 2008.
[11] S. Golomb and L. Welch, Shift Register Sequences. Laguna Hills,
CA: Aegean Park, 1982.
[12] R. Peterson, R. Ziemer, and D. Borth, Introduction to Spread-Spectrum
Communications. Englewood Cliffs, NJ: Prentice-Hall, 1995.
[13] A. Viterbi, CDMA: Principles of Spread Spectrum Communication.
Redwood City, CA: Addison Wesley Longman, 1995.
[14] P. Fan and M. Darnell, Sequence Design for Communications Applica-
tions. Taunton, Somerset, U.K.: Research Studies, 1996.
[15] S. Golomb and G. Gong, Signal Design for Good Correlation. Cam-
bridge, U.K.: Cambridge Univ. Press, 2005.
[16] X. Deng and P. Fan, “Spreading sequence sets with zero correlation
zone,” Electron. Lett., vol. 36, pp. 993–994, May 2000.
[17] S. Park, S. Park, I. Song, and N. Suehiro, “Multiple-access interference
reduction for QS-CDMA systems with a novel class of polyphase se-
quences,” IEEE Trans. Inf. Theory, vol. 46, pp. 1448–1458, July 2000.
[18] E. Larsson and J. Li, “Preamble design for multiple-antenna OFDM-
based WLANs with null subcarriers,” IEEE Signal Process. Lett., vol.
8, pp. 285–288, Nov. 2001.
[19] A. Van Zelst and T. Schenk, “Implementation of a MIMO OFDM-
based wireless LAN system,” IEEE Trans. Signal Process., vol. 52, pp.
483–494, Feb. 2004.
[20] L. Tsai and Y. Su, “Transform domain approach for sequence design
and its applications,” IEEE J. Sel. Areas Commun., vol. 24, pp. 75–83,
Jan. 2006.
[21] P. Stoica, H. He, and J. Li, “On designing sequences with impulse-like
periodic correlation,” IEEE Signal Process. Lett., vol. 16, pp. 703–706,
Aug. 2009.
[22] H. He, P. Stoica, and J. Li, “On aperiodic-correlation bounds,” IEEE
Signal Process. Lett., vol. 17, pp. 253–256, Mar. 2010.
[23] IEEE Standard for Information Technology- Telecommunications and
Information Exchange Between Systems-Local and Metropolitan Area
Networks-Specific Requirements-Part 11: Wireless LAN Medium Ac-
cess Control (MAC) and Physical Layer (PHY) Specifications, IEEE
Std 802.11, 1997.
[24] M. Lindenfeld, “Sparse frequency transmit and receive waveform de-
sign,” IEEE Trans. Aerosp. Electron. Syst., vol. 40, pp. 851–861, July
2004.
TSAI et al.: SYNTHESIZING LOW AC AND LOW PAPR OFDM SEQUENCES
[25] H. He, P. Stoica, and J. Li, “Waveform design with stopband and cor-
relation constraints for cognitive radar,” in Proc. 2nd Int. Workshop on
Cognitive Inf. Process., Jun. 2010.
[26] L. R. Welch, “Lower bounds on the maximum cross correlation of sig-
nals,” IEEE Trans. Inf. Theory, vol. 20, pp. 396–399, May 1974.
[27] P. Stoica, H. He, and J. Li, “New algorithms for designing unimod-
ular sequences with good correlation properties,” IEEE Trans. Signal
Process., vol. 57, pp. 1415–1425, Apr. 2009.
[28] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge,
U.K.: Cambridge Univ. Press, 2004.
[29] M. Grant and S. Boyd, CVX: Matlab Software for Disciplined Convex
Programming, Version 1.21 [Online]. Available: http://cvxr.com/cvx
2010
[30] K. Toh, M. Todd, and R. Tutuncu, “SDPT3—A Matlab software
package for semidefinite programming,” Optimiz. Methods Software,
vol. 11, no. 12, pp. 545–581, 1999.
[31] J. Sturm, “Using SeDuMi 1.02, a MATLAB toolbox for optimization
over symmetric cones,” Optimiz. Methods Software, vol. 11, no. 12, pp.
625–653, 1999.
[32] R. Gerchberg and W. Saxton, “A practical algorithm for the determi-
nation of the phase from image and diffraction plane pictures,” Optik,
vol. 35, pp. 237–246, 1972.
[33] J. R. Fienup, “Phase retrieval algorithms: A comparison,” Appl. Opt.,
vol. 21, pp. 2758–2769, Aug. 1982.
[34] W. Roberts, H. He, J. Li, and P. Stoica, “Probing waveform synthesis
and receiver filter design,” IEEE Signal Process. Mag., Jul. 2010.
[35] P. Spasojevic and C. Georghiades, “Complementary sequences for ISI
channel estimation,” IEEE Trans. Inf. Theory, vol. 47, pp. 1145–1152,
Mar. 2001.
[36] R. Horn and C. Johnson, Matrix Analysis. Cambridge, U.K.: Cam-
bridge Univ. Press, 1990.
Lung-Sheng Tsai (S’05) received the B.S. degree in
electrical engineering from the National Tsing Hua
University, Taiwan, in 2002, the M.S. degree in com-
munication engineering from National Chiao Tung
University, Taiwan, in 2004, and the Ph.D. degree
in Graduate Institute of Communication Engineering
from National Taiwan University, in 2010.
He is currently a Research Assistant with the
Research Center for Information Technology In-
novation, Academia Sinica, Taiwan. His research
interests include multiantenna systems and commu-
nication sequences design.
2243
Wei-Ho Chung (M’11) was born in Kaohsiung,
Taiwan, in 1978. He received the B.Sc. and M.Sc.
degrees in electrical engineering from National
Taiwan University, Taipei City, Taiwan, in 2000 and
2002 respectively, and the Ph.D. degree from the
Electrical Engineering Department, University of
California (UCLA), Los Angeles, in 2009.
From 2002 to 2005, he was a system engineer
with ChungHwa TeleCommunications Company.
From 2007 to 2009, he was a Teaching Assistant at
UCLA. In 2008, he was a research intern working on
CDMA systems in Qualcomm Inc. Since January 2010, he has been a faculty
member holding the position as an Assistant Research Fellow in Research
Center for Information Technology Innovation, Academia Sinica, Taiwan.
His research interests include wireless communications, signal processing,
statistical detection and estimation theory, and networks.
Dr. Chung received the Taiwan Merit Scholarship, sponsored by the National
Science Council of Taiwan, from 2005 to 2009.
Da-shan Shiu (M’08) received the B.S.E.E. degree
from National Taiwan University in 1993, and the
Ph.D. degree in electrical engineering and computer
sciences from the University of California, Berkeley,
in 1999.
From 1999 to 2004, he was with Qualcomm. He
joined the Department of Electrical Engineering
and the Graduate Institute of Communication Engi-
neering as an Assistant Professor in the fall of 2004
where he is currently an Associate Professor. His
current research interests include smart antenna sys-
tems, MIMO space-time signal processing, next generation wireless network,
and mesh wireless network.