1390

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 56, NO. 9, SEPTEMBER 2008

The Continuous-Time Peak-to-Average Power Ratio of

OFDM Signals Using Complex Modulation Schemes
K. Daniel Wong, Senior Member, IEEE, Man-On Pun, Member, IEEE, and H. Vincent Poor, Fellow, IEEE
AbstractComputing the continuous-time peak-to-averagepower ratio (PAPR) of OFDM signals is computationally challenging. The pioneering work by Tellambura applies only to
OFDM signals using real-valued modulation schemes. In this
paper, a practical technique for evaluating the continuous-time
PAPR of OFDM signals using complex modulation is presented.
Using the proposed scheme, it is confirmed that the four-time
oversampled discrete-time PAPR is a good approximation of
the continuous-time PAPR even for complex OFDM signals.
Furthermore, the proposed scheme is employed to verify some
existing analytical bounds on continuous PAPR in the literature.
Index TermsOFDM, peak-to-average-power ratio, PAPR,
multicarrier modulation.

I. I NTRODUCTION

NE of the major challenges of Orthogonal Frequency

Division Multiplexing (OFDM) is that the output signal
may have a potentially very large peak-to-average power
ratio (PAPR, also known as PAR). The resulting technical
challenges, as well as PAPR-reduction techniques and related
issues, have been widely studied and reported in the research
literature [1].
Since the actual signal that enters the power amplifiers is
a continuous-time signal, we ultimately want to reduce the
PAPR of the continuous-time OFDM signal (we call this
the continuous-time PAPR for convenience). However, the
evaluation of the continuous-time PAPR is analytically nontrivial and computationally expensive. Therefore, most PAPRreduction techniques focus on discrete-time approximations of
the continuous-time PAPR. The discrete-time approximations
result in what we call the discrete-time PAPR.
In a ground-breaking paper [2], Tellambura investigated the
differences between the continuous-time PAPR and discretetime PAPR. To do this, Tellambura introduced a practical
scheme to compute the continuous-time PAPR, using Chebyshev polynomials of the first kind. The scheme was then
used to obtain numerical results. Based on these results, a
common rule-of-thumb that has since emerged in the OFDM
research community is that the discrete-time PAPR with fourtime oversampling is a sufficiently accurate approximation of
the continuous-time PAPR [1].
Paper approved by N. C. Beaulieu, the Editor for Wireless Communication
Theory of the IEEE Communications Society. Manuscript received December
2, 2006; revised July 2. 2007. This research was supported in part by the U.
S. National Science Foundation under Grant No. ANI-03-38807, and in part
by the Croucher Foundation under a post-doctoral fellowship.
K. D. Wong is with the department of Information Technology, Malaysia
University of Science and Technology, Petaling Jaya, Selangor, Malaysia (email: dwong@must.edu.my).
M.-O. Pun and H. V. Poor are with the Department of Electrical Engineering, Princeton University, Princeton, NJ 08544 USA.
Digital Object Identifier 10.1109/TCOMM.2008.060438.

Unfortunately, Tellamburas method [2] applies only to

real-valued modulation schemes like BPSK (and results were
only presented for N=32 BPSK-OFDM, where N is the
total number of subcarriers), but not complex-valued schemes
like QPSK. To circumvent this shortcoming, [3] extended
Tellamburas method to complex modulation schemes, using
Chebyshev polynomials of both the first and second kinds.
However, neither [2] nor [3] present any analysis of the error
from using the discrete-time PAPR instead of continuoustime PAPR. Thus, even though the empirical distribution of
the continuous-time PAPR and the four-time oversampled
discrete-time PAPR may look close, there is no guarantee that
the error is bounded. Some analytical bounds have been provided in [4][6]. However, due to the lack of computationally
feasible methods to obtain the continuous-time PAPR, [4]
[6] used the discrete-time PAPR to verify their continuoustime PAPR bounds. Furthermore, the distribution of the error
has not been investigated before. To obtain the empirical
distribution of the error in the general case of complex-valued
modulation also requires a feasible scheme to compute the
continuous-time PAPR.
In this paper, we introduce a computational method that
is more general than Tellamburas [2], to find the peaks for
OFDM signals with arbitrary complex-valued modulations.
We express the instantaneous envelope power as a polynomial
of powers of tan(t), the roots of whose derivative can be
solved by using widely available mathematical software such
as Mathematica or Matlab. In contrast with [3], the proposed
method only employs Chebyshev polynomials of the first kind.
Also, because of the one-to-one relationship between tan(t)
and t in 0 t 1, the new method does not require
breaking the problem into two domains (0 t 0.5 and
0.5 t 1) and carefully mapping the roots differently for
each domain. Furthermore, comparisons are made between the
distribution of the continuous-time PAPR obtained through the
proposed method with the discrete-time PAPR obtained from
oversampled signals and some of the analytical upper bounds
derived in [4][6]. This has not been done before, to the best of
the authors knowledge. We also introduce a new probabilistic
bound that generalizes a bound of Sharif [4]. Moreover, we
discuss relationships between some of these bounds.
II. A NALYTICAL M ODEL
The baseband continuous-time OFDM signal with N carriers can be expressed as
N 1
1 
x(t) =
sn ej2nt ,
N n=0

(1)

where {sn } are data symbols and t is normalized with respect

to the OFDM symbol duration.

c 2008 IEEE
0090-6778/08$25.00 

WONG et al.: THE CONTINUOUS-TIME PEAK-TO-AVERAGE POWER RATIO OF OFDM SIGNALS USING COMPLEX MODULATION SCHEMES

With unity average power, the continuous-time PAPR, c ,

is defined as
2
c = max |x(t)| .
(2)
t

c measures the instantaneous envelope peak power of the

baseband signal and represents the maximal PAPR. It is nontrivial to compute. Tellamburas method [2] works only for
the special case of real-valued modulation.
As a computationally feasible alternative, the discrete-time
PAPR, d , is often used instead of c and defined as
d =

max

0kLN 1

|Ck | ,

(3)

1391

III. P ROPOSED METHOD

All trigonometric functions of an angle may be expressed
as rational expressions in terms of t = tan(/2) [7]. Let
x = tan(t). Substituting (1 x2 )/(1 + x2 ) for cos(2t) and
2x/(1 + x2 ) for sin(2t), and letting k = cos(/2k) and
k = sin(/2k), we have


N 1
1 x2
2 
Pa (x) = 1 +
k T k
+
N
1 + x2
k=1


N 1
1 x2
2x
2 
k Tk k
+ k
. (9)
N
1 + x2
1 + x2
k=1

where

N 1
1 
sn ej2nk/LN ,
Ck =
N n=0

(4)

with L being the oversampling rate.

2
Let Pa (t) = |x(t)| . Without loss of generality, no assumptions are made on the modulation scheme used to generate
{sn }. It can be easily shown that
Pa (t) = 1 +

N 1
2 
[k cos(2kt) + k sin(2kt)] ,
N

(5)

k=1

where k and k are defined as follows:

N 1k


sm sm+k ,
k = 1, 2, , N 1, (6)
k = R

We need only to find the roots of Pa (x)/x, since

Pa (t)/t = Pa (x)/x( sec2 (t)). Because Tk (x) is an
order-k polynomial, the highest power of 1/(1 + x2 ) in (9) is
N 1. Hence we can remove the denominator and thus obtain
a polynomial Q(x) by writing
Pa (x)
.
(10)
x
Q(x) is a polynomial of degree at most 2N in x and all
roots of Pa (x)/x are also roots of Q(x). Thus, Pa (x)/x
has at most 2N roots. Pa (x) can be routinely computed from
(9) by expanding the Chebyshev polynomials, factoring out
1/(1 + x2 )N , and collecting terms. We may then evaluate the
values of Pa (x) at the real roots, and the maximum is c .
Q(x) = (1 + x2 )N

m=0

and
k = I

N 1k


IV. N UMERICAL P ROCEDURE S UMMARY


sm sm+k

k = 1, 2, , N 1.

(7)

m=0

with () denoting complex conjugation and R {} and I {}

being the real and imaginary part of the enclosed quantity,
respectively.
its
Clearly, a necessary condition for Pa (t) to achieve

Pa (t) 

maximum at t , i.e. maxt Pa (t) = Pa (t ), is t  = 0.

t=t
Thus, a practical approach to computing Pa (t ) is to first find
the roots of Pa (t)/t followed by comparing the values of
Pa (t) at only the real roots.
Following the approach of [2], we denote by Tk (t) =
cos(k cos1 t) the kth-order Chebyshev polynomial. For each
k, Tk (x) can be expressed as a kth-degree polynomial in
terms of x, where T0 (x) = 1, T1 (x) = x and Tk+1 (x) =
2xTk (x) Tk1 (x) for k > 1. Exploiting the equalities
Tk (cos ) = cos k and sin() = cos( 2 ), we can rewrite
(5) in terms of Chebyshev polynomials as

Pa (t) =

1+

N 1
2 
k Tk (cos(2t)) +
N
k=1

2
N

N
1

k=1



) .
k Tk cos(2t
2k

(8)

Being different from the BPSK-OFDM systems considered

in [2], the complex OFDM signal introduces the second term
on the right hand side (R.H.S.) of (8), which presents a major
challenge in obtaining exact c values.

The proposed method for computing the continuous-time

PAPR for a given symbol set {sn } and number of subcarriers
N is summarized as follows.
1) Compute k and k for k = 1, 2, , N 1 according
to (6) and (7);
2) Compute Pa (x) according to (9), expanding and collecting the coefficients of the different powers of x;
3) Find the derivative of Pa (x);
4) Find the roots of Q(x), and hence of Pa (x)/x using
standard polynomial root finding algorithms;
5) Keep only the real roots of Q(x);
6) Evaluate and compare the values of Pa (x) at the real
roots, and obtain c .
Each step is straightforwardly handled by common mathematical software like Mathematica or Matlab. In our experiments, we have found that step 2 (expanding and simplifying
Pa (x)), while conceptually easy, may dominate the computation time, especially
for large N . In particular, expandingand

simplifying Tk k (1 x2 )/(1 + x2 ) + k (2x)/(1 + x2 ) is

a time consuming operation for large k. For a given N ,
pre-computing these terms helps to significantly reduce the
computation time.
V. S IMULATION R ESULTS
In this section, we evaluate the proposed scheme using a
QPSK-OFDM system with N = 32.
Fig. 1 shows the complementary cumulative distribution
function (CCDF) of d with different oversampling rates,
L = 1, 2, 4, 8. The CCDF of c labeled as continuous-time

1392

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 56, NO. 9, SEPTEMBER 2008

10

10

Empirical mean
Empirical median
SGHUB
SGHMUB
WHUB

10

10

10

Normalized error (dB)

Prob (PAR >)

L increases

Continuoustime
L=1
L=2
L=4
L=8

10

15

20

25

Fig. 1.

(dB)

10

11

is also plotted in Fig. 1. As indicated in Fig. 1, d obtained

from oversampled signals approaches c as L increases, and
d obtained with with a oversampling rate greater than or equal
to L = 4 is an accurate approximation of c . These results
agree with those reported in [2] where real-valued OFDM
signals (and BPSK-OFDM in particular) were considered.
Next, we define the absolute (unnormalized) error between
d and c as a function of the oversampling rate, L, as
(11)

Following [4], we also define the normalized error between

d and c as
a (L)
.
(12)
(L) =
d (L)

It has been shown [4] that, for L > / 2, is upper

bounded by
2 /2
.
(13)
0 < (L) < 2
L 2 /2
2

If we may assume that arg maxt |x(t)| is uniformly distributed between sampling points (used to obtain d ), then
a probabilistic
bound can be derived (see Appendix) for

L > / 2 as:


2 2 /2
P 0 < (L) < 2
,
(14)
L 2 2 /2
where 0 < < 1. In particular, setting = 1/2 gives the
following upper bound for the median.
median (L) <

2 /8
.
L2 2 /8

(15)

Another upper bound has been independently derived for

L > 1 in [5], [6].
1
0 < (L)
1.
(16)

cos( 2L
)
Interestingly, it is shown in the Appendix that the upper bound
shown in (16) approximates that for the median given in (15)
and is strictly tighter than that in (13). In the sequel, the upper
bounds shown in (13), (15) and (16) are referred to as SGHUB, SGH-MUB and WH-UB, respectively.

4
5
Oversampling rate (L)

0.1

0.12

0.14

Normalized errors as a function of L.

Fig. 2.

PAPR distribution for QPSK OFDM systems with N = 32.

a (L) = c d (L).

30
1

Prob (a (8) > )

10

10

10

10

CCDF for all c

10

CCDF for the smallest 33% c

CCDF for the middle 33% c
CCDF for the largest 33% c

10

Fig. 3.

0.02

0.04

0.06
0.08
Estimation error,

CCDF of the absolute (unnormalized) error a (8).

Fig. 2 shows the average and median of (L) as a function

of L. The upper bounds for (L) shown in (13), (15) and
(16) are also plotted in Fig. 2. Inspection of Fig. 2 confirms
that WH-UB can be approximated by SGH-MUB but is much
tighter than SGH-UB.
Next, we investigate the relationship between the empirical
probability density of the error a (L) and the value of c . In
the next experiment, we fix L = 8 and investigate the CCDF of
a (8) obtained with different values of c . More specifically,
we rank all c obtained by the proposed scheme and group
them into three equal value ranges, namely the smallest 33%,
the middle 33% and the largest 33%.
In Fig. 3, the curve labeled CCDF for all c takes into
account all the a (8) from the experiment while the curve
labeled CCDF for the smaller 33% c is the CCDF for
a (8) corresponding to the c values in the bottom 33% range.
Similarly, the middle 33% and the largest 33% refer to
the ranges from the 33rd to 66th percentile, and 66th to 100th
percentile, respectively. These last three curves can be viewed
as distributions conditioned on the range of c . From Fig. 3,
it can be seen that the error tends to increase as c increases
(so d would be more likely to under-estimate c in the worst

WONG et al.: THE CONTINUOUS-TIME PEAK-TO-AVERAGE POWER RATIO OF OFDM SIGNALS USING COMPLEX MODULATION SCHEMES

Prob (PAR >)

10

1393

A PPENDIX

10

10

In this Appendix, we first prove the bound (14) and then we

derive the relationships among the three upper bounds shown
in (13), (15) and (16) for L > 1.
Proof of Probabilistic Bound. For L times oversampling,
the distance between samples is 2/LN . The location of the
peak must be within /LN of the closest sampling point.
Let x0 be the location of the closest sampling point and xp
be the location of the peak. The bound (14) is based on the
assumption that xp x0 is uniformly distributed in the region
[/LN, /LN ]. Then

Simulated c
10

10

Analytical upper bound (L=4)

Analytical upper bound (L=8)
Analytical upper bound (L=16)

P (|xp x0 | x) = ,
0

6
(dB)

10

12

Fig. 4. CCDF of c and the corresponding upper bounds given in [4] with
L = 4, 8, 16.

cases). Similar results were obtained when the data was broken
up into nine segments rather than three, and also for other
values of L. The concentration of the densities on the left
(smaller error values) is noteworthy.
Finally, we compare the CCDF of c against its upper bound
given in [4] as
P r (c > ) < LN e

(1 2L
2)

(17)

Fig. 4 shows the CCDF of c and the corresponding upper

bounds given in [4] with L = 4, 8, 16. Inspection of Fig. 4
reveals that the upper bound becomes tight as L increases.
However, little observable improvement can be achieved beyond L = 8, which indicates that there may be room for
further tightening of the upper bound shown in (17).
VI. C ONCLUSIONS
In this paper, a computationally feasible scheme to evaluate
the continuous-time peak-to-average-power ratio (PAPR) of
OFDM signals using complex modulation schemes has been
introduced. The scheme expresses the instantaneous power as
a polynomial in terms of tan(t).
Using the proposed scheme, we have generalized Tellamburas results [2] to complex-valued modulations and shown
for complex-valued modulations (like QPSK-OFDM) that the
discrete-time PAPR obtained from four-time oversampled signals may be considered a sufficiently accurate approximation
of the continuous-time PAPR.
We have also used our scheme to examine the empirical
distribution of the error between the continuous-time PAPR
and discrete-time PAPR. We have found that the error tends
to increase proportionately with PAPR. Furthermore, we have
used our scheme to obtain empirical verification of several
analytical upper bounds related to continuous-time PAPR
derived in [4], and for which empirical verification had not
been previously available.
Finally, we have generalized (13) from [4], obtaining a new
bound for the median of the continuous-time PAPR that is
almost the same as the bound in [5], [6], and have proven that
the bound shown in (16) is tighter than that in (13) for L > 1.

(18)

where x = /LN and 0 x /LN

The proof of Theorem 2 in Sharif et al. [4] uses the fact
that |xp x0 | /LN . Using (18) instead, we may perform
similar algebraic manipulations as Sharif et al. to obtain a
generalized version of their Theorem 4. Their Theorem 4 is
expressed in (13), which can be seen as a special case of
our (14), namely when approaches 1.
Relationship between bounds. We first recall
 2n

  
(1)n 2L
cos
=
.
(19)
2L
(2n)!
n=0
Since 0 <

< 1 for L > 1, we have

 
 
1
2
1
cos
+O
.
2
2L
8L
L4

2L

(20)

After dropping the higher-order term of L in (20), it is

straightforward to show that
2 /8
2 /2
1
1 2
<
.

cos( 2L )
L 2 /8
L2 2 /2

(21)

Thus, the upper bounds shown in (15) and (16) are approximately equal but strictly tighter than that in (13) for L > 1.
ACKNOWLEDGEMENT
The authors would like to thank the anonymous reviewers
for their constructive suggestions, and in particular for bringing [3], [5] and [6] to our attention.
R EFERENCES
[1] S. H. Han and J. H. Lee, An overview of peak-to-average power
ratio reduction techniques for multicarrier transmission, IEEE Wireless
Commun., vol. 12, no. 2, pp. 5665, Apr. 2005.
[2] C. Tellambura, Computation of the continuous-time PAR of an OFDM
signal with BPSK subcarriers, IEEE Commun. Lett., vol. 5, no. 5, pp.
185187, May 2001.
[3] H. Yu and G. Wei, Computation of the continuous-time PAR of an
OFDM signal, in Proc. 2003 IEEE Intl Conf. Acoust. Speech Signal
Process., Hong Kong, pp. 529531, Apr. 2003.
[4] M. Sharif, M. Gharavi-Alkhansari, and B. H. Khalaj, On the peak-toaverage power of OFDM signals based on oversampling, IEEE Trans.
Commun., vol. 51, no. 1, pp. 7278, Jan. 2003.
[5] G. Wunder and H. Boche, Peak value estimation of bandlimited signals
from their samples, noise enhancement, and a local characterization in the
neighborhood of an extremum, IEEE Trans. Signal Processing, vol. 51,
no. 3, pp. 771780, Mar. 2003.
[6] , Upper bounds on the statistical distribution of the crest-factor in
OFDM transmission, IEEE Trans. Inform. Theory, vol. 4, no. 2, pp.
488494, Feb. 2003.
[7] J. H. Silverman and J. Tate, Rational Points on Elliptic Curves. New
York: Springer, 1992.