You are on page 1of 11

3568 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO.

9, SEPTEMBER 2008

MLSD Bounds and Receiver Designs for


Clipped OFDM Channels
Fei Peng, Student Member, IEEE, and William E. Ryan, Senior Member, IEEE

Abstract—We derive tight closed-form approximations of sym- modulated sequence considered as a super symbol, there exists
bol and bit error probability on maximum-likelihood sequence a one-to-one mapping between sequences at the OFDM mod-
detection (MLSD) of Nyquist-rate and oversampled OFDM ulator input and sequences at the clipper output. As a sequel
systems over the AWGN, flat and frequency-selective quasi-static
Rayleigh fading channels. Contrary to common knowledge, we to [7], in this paper we derive the closed-form approximations
show that on a frequency-selective channel, clipping at the trans- of the symbol and bit error rate performance for maximum-
mitter leads to achievable frequency-diversity and performance likelihood sequence detection (MLSD) on AWGN, flat and
gain over unclipped systems with MLSD. We derive closed-form frequency-selective quasi-static Rayleigh fading channels. The
expressions of the diversity and the performance gain over the term “MLSD bound” used in this paper emphasizes that the
no-clipping case, as well as the optimal clipping ratio for a given
system. We show that such diversity and performance gain are MLSD approximations can be used as a performance limit for
a result of both clipping and channel frequency-selectivity, and clipped OFDM channels.
they do not occur on the AWGN and flat fading channels. We also It is generally believed that clipping necessarily leads to
present a new near-optimum detection algorithm that achieves degradation in performance [1]-[6]. We show in this paper
the MLSD performance bound with a few iterations for the
clipped OFDM systems with both high and low clipping levels on that for MLSD, this degradation exists for the AWGN and flat
the AWGN and flat Rayleigh fading channels. For the frequency- fading channels, but for frequency-selective channels, clipping
selective channel, we present a receiver design that utilizes the leads to diversity and a performance gain over unclipped
advantage of clipping to some extent. We consider rectangular systems. The effect of clipping is very similar to OFDM
QAM signaling in this paper. systems on fast fading channels [8], [9]. We show that the
Index Terms—Orthogonal frequency-division multiplexing maximum diversity of a clipped OFDM channel is the number
(OFDM), clipping noise, maximum-likelihood sequence detection of channel taps, and we derive the optimal clipping ratio for a
(MLSD), iterative method, performance bounds. given system and channel. The diversity and performance gain
are a result of both clipping and channel frequency-selectivity.
I. I NTRODUCTION Our results indicate that clipping with MLSD provides a way
to achieve diversity over frequency-selective fading channels
I T IS COMMONLY understood that one of the drawbacks
of the orthogonal frequency division multiplexing (OFDM)
transmission technique is the presence of high peaks at the
without resorting to Rake receivers [10] whose construction
relies on channel knowledge.
OFDM modulator output. To alleviate the effects of non- The clipping-induced frequency-diversity was first noticed
linear amplification, these high peaks are often intentionally in [11] for frequency-selective channels. However, our results
clipped prior to amplification and transmission [1]. Receiver- sharpen and extend the results in [11]. We show that diversity
side signal processing is necessary to reduce the loss due to and performance gain over unclipped systems do not come
clipping [2]-[6]. In most of these papers, although slightly from clipping alone. It was also noticed in [12] that appro-
different models were used, the clipper output was generally priate manipulations of the clipper can also take advantage of
treated as a random variable, with the so-called clipping clipping. In this paper, we focus on the soft clipper [1] without
noise approximated as zero-mean Gaussian noise. Capacity any modifications to the clipping process.
analyses based on this model predict significant reduction of We also propose in this paper a soft-output algorithm that
channel capacity due to both a shrunken constellation and achieves the MLSD bounds for both high and low clipping
large clipping noise for low clipping ratios [1], [3]. ratios over the AWGN and flat Rayleigh fading channels.
Our prior work [7] takes an alternative approach to analyze The algorithm is iterative and can be implemented in a fully-
the clipped OFDM channel, considering both AWGN-free and parallel fashion. For frequency-selective channels, we present
AWGN scenarios. We proved that, with the entire OFDM a vector-based modeling approach that represents the clipping
process similar to that for a MIMO channel. By doing so,
Manuscript received April 5, 2007; revised August 18, 2007 and February more sophisticated receiver designs such as the MMSE-SD
4, 2007; accepted February 17, 2008. The associate editor coordinating the technique [9] can be applied. We show that the performance
review of this paper and approving it for publication was D. Dardari. This
work was supported by Intel and by Mitsubishi Electric Research Laborato- of the MMSE-SD with the proposed iterative demapper can
ries. This work was presented in part at Globecom’05 and Globecom’07. closely approach the MLSD bound in the low-to-medium
F. Peng was with the Department of Electrical and Computer Engineering, SNR region over frequency-selective channels. We focus our
the University of Arizona, Tucson, AZ, USA, 85721. He is now with LitePoint
Corporation, Sunnyvale, CA, USA, 94085 (e-mail: fei.peng@ieee.org). discussion on rectangular QAM. The same approach can be
W. E. Ryan is with the Department of Electrical and Computer En- easily extended to other modulation schemes.
gineering, the University of Arizona, Tucson, AZ, USA, 85721 (e-mail:
ryan@ece.arizona.edu). In Section II, we describe the system model. We derive
Digital Object Identifier 10.1109/TWC.2008.070368. the tight MLSD bound approximations in Section III and
1536-1276/08$25.00 
c 2008 IEEE

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:05 from IEEE Xplore. Restrictions apply.
PENG and RYAN: MLSD BOUNDS AND RECEIVER DESIGNS FOR CLIPPED OFDM CHANNELS 3569

H = diag(Fh) is the channel L−1 gain,2 with a normalized


power, i.e., E[|Hi,i |2 ] = l=0 E[|hl | ] = 1. We assume
the receiver has perfect channel knowledge. In simulations,
the channel is generated using the well known Jakes method.
When oversampling is employed, we replace U by S, as in
Fig. 1b. For the AWGN channel, H = I is the identity matrix.
Fig. 1. The clipped OFDM channel models with (a) Nyquist-rate sampling
For a flat block fading channel H = HI, where H is the
and (b) oversampling of factor J. N is the number of carriers. channel gain for the entire OFDM symbol. For transmission
over the frequency-selective channel, we assume the guard
interval is sufficiently long to avoid inter-OFDM-symbol in-
demonstrate their tightness via simulations. In Section IV, we terference, and the channel stays essentially unchanged within
discuss the diversity and performance gain over the unclipped the duration of one OFDM symbol, so that there is no inter-
system attained on frequency-selective channels. In Section carrier interference.
V, we discuss receiver designs that can closely approach For receiver designs on the frequency-selective channel, we
the MLSD bound on different channels. In Section VI, we rely on an alternative representation of the clipping process
conclude this paper. that is similar to a MIMO channel. We will present this
modeling approach in Section V where it is most relevant.
II. C HANNEL M ODEL The existing clipper models and capacity analyses in the
literature are typically derived from a sample-based statis-
Let an M -ary rectangular QAM sequence be denoted by tical model [1], [9]. Though accurate for the sample-based
the N -vector X. For Nyquist-rate sampling (Fig. 1a), x is statistical behavior of the soft clipper input and output, they
the N -point IFFT of X. For oversampled OFDM ([1], [2]) fail to describe the inter-dependence between OFDM carriers
with a oversampling factor J (Fig. 1b), x is the JN - induced by clipping.
point IFFT of X with zero padding. Note that Fig. 1b is In [7], we studied the capacity of the clipped OFDM
the simplified equivalent model [1] where R is the receiver channel by exploiting its deterministic nature. By considering
OFDM demodulator output. If we assume symmetric (unity- the IFFT unit and the soft clipper together as a new transform,
gain) FFT/IFFT operations, theaverage  signal power
 √ at the we showed that a one-to-one mapping between sequences X
2 2
input of the clipper is Es = E |x| = E |X| / J. The at the IFFT input and the clipper output u is guaranteed with
soft clipper output is u = g(x) where, for the i-th sample of probability one for large N , provided sequences of identical
x,  elements are avoided. The analysis and algorithm presented in
Aejφ(xi ) , |xi | > A this paper are motivated by these results.
g(xi ) = , (1)
xi , |xi | ≤ A
III. T HE MLSD B OUNDS
where A > 0 is the clipping amplitude and φ(xi ) is the In this section, we derive tight MLSD approximations on
phase of xi . For simplicity, we abuse the notation slightly Nyquist-rate and oversampled OFDM systems on the AWGN,
by allowing the argument of g(·) to either be a vector or a
flat and frequency-selective Rayleigh fading channels. We note
scalar. The clipping ratio γ is defined as that the term “MLSD bound” does not refer to bounding the

γ = A/ Es . (2) MLSD performance. Instead, it emphasizes the use of the
MLSD performance as benchmarks. In the derivations below,
Throughout the literature, the clipper output is modeled as a we assume Gray labeling so that the symbol-error probability
random variable. In [1], the clipper output is expressed as Ps is related to the bit-error probability Pb by [10]
ui = g(xi ) = αxi + di (3) Ps ≈ Pb log2 M . (6)
where Thus, we focus on deriving Ps . We also assume that the

√ source symbols are i.i.d. uniform, with sequences of identical
E [x u] 2 π
α= = 1 − e−γ + γerfc (γ) (4) elements avoided via scrambling [7]. Further, the clipping
Es 2
ratio for the M -QAM modulation is assumed to be above
and where x and u are generic clipper inputs and outputs, the threshold defined in [7], so that the one-to-one mapping
respectively. Note that (4) is accurate for a soft clipper (1) between IFFT input sequences and clipper output sequences
only when the clipper input is complex Gaussian. The clipping is maintained. For simplicity, we first focus on Nyquist-rate
noise di is approximated as white Gaussian noise [1], [2]. We sampling and treat the oversampling case later. In addition,
call this model a statistical model. Instead of using this model, the MLSD bound will be derived assuming the MLSD is
we shall adopt a more deterministic approach. performed per OFDM symbol (a length-N symbol sequence).
Let F be the symmetric FFT matrix, and h be the channel However, it can be easily generalized to multiple OFDM
impulse response of length L. For quasi-static fading channels, symbols.
the OFDM demodulator output is
A. The MLSD Bound Approximation for Nyquist-Rate-
R = HU + W , (5) Sampled OFDM
where U = Fu, W is circularly symmetric complex white Consider two rectangular-QAM symbol sequences X and
Gaussian noise with variance N0 , and the diagonal matrix Y , the corresponding IFFT outputs x and y, and respective

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:05 from IEEE Xplore. Restrictions apply.
3570 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 9, SEPTEMBER 2008


N
k  
N
k
Ps (H) = Pblk,k (H) = P (u) P [A(u, k)|u, H] (7)
N u
N
k=1 k=1

clipper outputs u and v. Let U = Fu, V = Fv, and u be we have equation (13) on the top  of the next page, where
the transmitted sequence. The received vector at the OFDM we used the assumption that σζ2 = E[|hi |2 ] = 1 and that
demodulator output is R = HU + W in accordance with the Hn are identically distributed (but correlated). (13) is the
(5). With Gray labeling, the conditional symbol error rate well-known formula for uncoded fading channels. As we have
Ps (H) given knowledge of the channel can be expressed seen above, the proposed approximation technique works well
as (7) where Pblk,k (H) denotes the probability that a block for unclipped systems.
(length-N symbol sequence) contains k symbol errors at the If there is clipping, (10) is no longer true, and
2
MLSD output, P (u) is the probability that sequence u (and H (U − V )n becomes random for different sequences. In
equivalently X) is transmitted, and A(u, k) denotes the set this case, we have
of all sequence v different from u producing k symbol errors 1 
in the block. Let dmin be the minimum Euclidean distance Ps (H) ≈ Ec [P (u → v|H)]
N
of the signal constellation. To evaluate P [A(u, k)|u, H], two v∈S1 (u)
⎡ ⎛ ⎞⎤
approximations can be made. The first one is to restrict N −1 √
1  4 M − 4 ⎣ ⎝ H (U − V )n ⎠⎦
2
A(u, k) to the set of sequences v whose corresponding Y ≈ √ Ec Q
N n=0 M 2N0
differs from X by only a single element indexed by k, and
X − Y 2 = d2min , leading to a lower bound. The second one where the expectation Ec is taken
is to expand P [A(u, k)|u, H] by the pairwise union upper √ over the inherent random-
ness due to clipping. Since Q ( x) is a convex function, using
bound. As we shall see later, the compound effect of applying Jensen’s inequality, we have
both approaches leads to a close approximation. ⎛ 
 ⎞
We first apply this bounding technique for unclipped sys- √  2
tems and examine its effectiveness. As described above, we
−1
1  4 M −4 ⎜
N
 Ec H (U − V )n ⎟
Ps (H)  √ Q⎜⎝
⎟.

approximate the MLSD performance by N n=0 M 2N0
1  
(14)
Ps (H) ≈ P (u) P (u → v|H) (8)
N u It can be proved and readily verified via simulations that the
v∈S1 (u)
2

larger N is, the smaller the variance of H (U − V )n is. As
H(U−V )2 a result, the above lower approximation is
where P (u → v|H) = Q 2N0 [10] is the  tight for large
 N.
2
To proceed, we need to analyze Ec U − V n . Since
pairwise error probability that the MLSD considers v more    
2 2
probable given u is transmitted, and S1 (u) is the set of v Ec U − V n = Ec u − vn with symmetric FFT, we
whose corresponding set of Y differ from X by a single  
2
2 examine Ec u − vn at the clipper output. Since we only
element with X − Y 2 = d2min . Denote by H (U − V )n =
2
|Hn,n (Un − Vn )| the Euclidean distance corresponding to deal with the case where X and Y differ by the n-th
such an error event. We then element, the difference between the √ i-th element of x and y
√ have equation (9) on the top
of the next page, where 4 √MM−4 is the average number of is xi − yi = (Xn − Yn ) ej2πin/N / N . For large N , |xi − yi |
QAM neighbors at distance dmin . is thus a very small number compared with the clipping
If there is no clipping, then amplitude A when |Xn − Yn | = dmin . Hence, for our purpose,
it is reasonable to assume that xi is clipped whenever yi is
2
6Es |Hn | clipped. Letting Nc be the number of clipped elements in u
H (U − V )2n = H (X − Y )2n = |Hn |2 d2min =
M −1 and v, we have
(10)   N −N  
Ec u − v2n
dmin +Nc Ec |uik − vik |2 . (15)
c 2
where Hn is the n-th diagonal element of the diagonal matrix
H. Hence, N
⎛ ⎞ where 0 ≤ ik < Nc is the index for clipped elements.
 
N −1 √ 2 2
4 M − 4 ⎝ |Hn | dmin ⎠ Since |uik − vik | is merely the amplitude of the difference
P1 (u → v|H) ≥ √ Q . of two vectors (a complex number is viewed
M 2N0  as a vector), we
v=u n=0 2
(11) take a geometric approach to derive Ec |uik − vik | . Let
2 2
If the channel is the AWGN channel, Hn = 1, ∀n. From (8) δ 2 = |xik − yik | , 2 = |uik − vik | , ρ1 = |xik |, and ρ2 =
and (11) we have |yik |. As can be easily seen from Fig. 2, using the law of
√   ρ2 +ρ2 −δ 2
cosines for angle ζ, we have 2 = 2A2 1 − 2 2ρ11ρ2 . Thus
no clipping 4 M −4 3Es
Ps,AWGN ≈ √ Q (12)    
M (M − 1)N0 1 1 1 d2 1
E[ 2 ] ≈ A2 δ 2 E ≈ A2 min E 2 (16)
which is the well-known formula in [10]. For a flat or 2 ρ1 ρ2 2 N ρ
2
frequency-selective Rayleigh fading channel, |Hn | is expo- where 1/2 is included because E[ 2 ] is independent of
2
nentially distributed. Defining a random variable ζ ≡ |Hn | , the vector’s direction, and we used the approximation that

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:05 from IEEE Xplore. Restrictions apply.
PENG and RYAN: MLSD BOUNDS AND RECEIVER DESIGNS FOR CLIPPED OFDM CHANNELS 3571

√ ⎛ ⎞
 
N −1 2
4 M − 4 ⎝ H (U − V )n ⎠
P (u → v|H) ≥ √ Q , (9)
M 2N0
v∈S1 (u) n=0

no clipping 1  
Ps,fading = EH [Ps (H)] ≈ P (u) EH [P1 (u → v|H)]
N u
v=u
√  
2 M −2 Es /N0
= √ 1−
M Es /N0 + 2 (M − 1) /3
(13)

given in (4), we have


 
Ec |Un − Vn |2 = |mean (Un − Vn )|2 + var (Un − Vn )
= α2 d2min + σΔd
2
(21)
where applying (17), it can be shown for i = n,
 
  Ec U − V 2n − α2 d2min
2 2
σΔd ≡ Ec |Ui − Vi | =
 N
−γ 2 2
" 2# 2
1−e + γ E1 γ /2 − α
Fig. 2. The geometry used to compute E[2 ]. = d2min , (22)
N
 
which is independent of i. Define η ≡ Ec H (U − V )2n ,
δ is very small compared with ρ. The conditional pdf and let H be the vector containing diagonal entries H. From
for the amplitude of clipped samples is f (x|x > A) = 2 N of−1 2
f (x)
(20)-(22) we have η = α2 d2min |Hn | + σΔd 2
i=0 |Hi | =
2x −x2 /Es A2 /Es
P (x>A) = E e e . Thus, given x > A, H H Δ(n) H where Δ(n) is a diagonal matrix with Δnn =
(n)
1 ! ∞s 1 2x −x2 /E A2 /E 2 " #
E x2 = A x 2 Es e s
e s
dx = E1s eγ E1 γ 2 2 2 2 (n) 2
α dmin + σΔd , and Δi,i = σΔd for i = n. Note that as
! ∞ −t
where E1 (x) ≡ x e /t dt denotes the well-known expo- γ → ∞, α2 → 1 and σΔd 2
→ 0, which reduces to the case
nential integral function with parameter x. Hence, E[ 2 ] ≈ without clipping.
A2 d2min 1 γ 2 " 2 # d2min 2 γ 2 " 2 #
2N E s e E1 γ = 2N γ e E1 γ . Thus, using Nc ≈ Therefore, by averaging over the channel statistics, we have
−γ 2
Ne for large N , (15) becomes clipped
Ps,freq-sel = EH [Ps (H)]
  $ % √ ⎡ ⎛ ⎞⎤
2 2 1 " # 
N −1
Ec u − vn
1 − e−γ d2min + γ 2 E1 γ 2 d2min . (17) 4 M −4 Δ (n) H
H
H
2 ≈ √ EH ⎣Q ⎝ ⎠⎦ .
n=0 N M 2N0
We note that the above equation holds only for the Nyquist-
rate sampling. (23)
For the AWGN channel, H = I. From (14) and (17), The above expression holds for any channel statistics. Below
the symbol error probability over the AWGN channel with we shall focus on Rayleigh fading channels. Let Φη (s)
(n)

MLSD is approximated by (18) and Pbclipped can then be be the moment generating function (MGF) of the random
easily computed. Denote" the# SNR loss due to clipping by variable η having the quadratic form H H Δ(n) H. Since Δ(n)
2
Δ(γ) ≡ 1 − e−γ + γ 2 E1 γ 2 /2. For the flat Rayleigh fading is a Hermitian matrix, and H is a zero-mean circularly
channel, H = HI. From (14) and (17), by averaging over the symmetric  complex Gaussian vector with covariance matrix
exponential random variable |H|2 , we have R ≡ E HH H , the MGF can be shown as in (24), where
√   (n)
δi are the eigenvalues of RΔ(n) . Using partial fraction
clipped 2 M −2 Δ(γ)Es /N0 (n) L−1 (n) (n) (n)
Ps,flat ≈ √ 1− . expansion, we have Ki = k=0,k=i δi /(δi − δk ).
M Δ(γ)Es /N0 + 2 (M − 1) /3 (n)
Clearly, Φη (s) is identical for all n. Thus, (23) reduces to
(19) ⎡ ⎛ ⎞⎤
For frequency-selective quasi-static channels, since H is √
4 M − 4 H H
Δ (n) H
diagonal,
clipped
Ps,freq-sel ≈ √ EH ⎣Q ⎝ ⎠⎦(25)
M 2N0
  N
−1  
2 2 2
Ec H (U − V )n = |Hi | Ec |Ui − Vi | . (20) for any choice of n from 0 to N −1. The inverse Laplace trans-
(n) (n) L−1 Ki(n) −η/λ(n)
i=0 form of Φη (s) results in pη (η) = i=0 (n) e i . Let
λi
Since it is easily seen that |Ec [Un − Vn ]| = αdmin with α Λ be a diagonal matrix with its entries Λk,l = Δk,l /d2min .

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:05 from IEEE Xplore. Restrictions apply.
3572 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 9, SEPTEMBER 2008

√ ⎛  ⎞
2
clipped 4 M − 4 ⎝ 3 1 − e−γ + γ 2 E1 (γ 2 ) /2 Es ⎠
Ps,AWGN ≈ √ Q (18)
M M −1 N0

1 &
L−1
1  K (n)
L−1
Φ(n)
η (s) = $ % = (n)
= i
(n)
(24)
det I − sRΔ(n) i=0 1 − sδ i i=0 1 − sδ i

√  '

clipped 4 M −4 η
Ps,freq-sel ≈ √ Eη Q
M 2N0
√  
2 M −2 
L−1
3λi Es /N0
= √ Ki 1 −
M i=0
2(M − 1) + 3λi Es /N0
(26)

−1 −2 −1
10 10 10
no clipping MLSD bnd, γ=1.2, Nyquist
MLSD bnd simulation, exhaustive search, γ=1.2, Nyquist
MLSD sim, Nyquist MLSD approx, γ=1.2, oversampled
MLSD sim, J=2 simulation, exhaustive search, γ=1.2, J=2
−2 MLSD sim, J=16 −2
10 simulation, exhaustive search, γ=1.2, J=16
10
MLSD bnd, no clipping
simulation, exhaustive search, no clipping
BER
BER

−3
10

−3 −3
10 10

(a) AWGN (b) flat Rayleigh −4


−4
10 −4 10
10 10 15 20 25 30 35
4 6 8 10 15 20 25 30 35 Eb/N0
Eb/N0, dB

Fig. 4. MLSD bounds and ML exhaustive search simulations for QPSK and
Fig. 3. Comparison of MLSD bounds and exhaustive ML search simulations 8-OFDM over frequency-selective quasi-static Rayleigh fading channels with
of QPSK and 8-OFDM over (a) the AWGN channel and (b) the flat Rayleigh σh2 = [0.6, 0.4]. The approximated MLSD bound for oversampled systems
fading channel. γ = 1.2. J is the oversampling factor. is identical for all J ≥ 2.

Omitting the index n, it is clear from (21) and (22) that channel. For a system with a clipping ratio of γ = 1.2 and one
Δ = Λd2min where δi = λi d2min . Thus, by averaging over without clipping, the exhaustive search simulations agree with
exponential random variables and omitting the index n, we the MLSD bounds very well in all cases. We comment that
have the symbol error probability expressed in (26), where without clipping, the OFDM subcarriers become independent,
λi = eig (RΛ) is a function of the clipping and MLSD is reduced to maximum-likelihood symbol-by-
( ratio γ and the symbol detection, which is also depicted in the figures.
channel covariance matrix R, and Ki = L−1 k=0,k=i λi /(λi −
λk ).
To demonstrate the tightness of the approximations (18), B. The MLSD Bound Approximation for Over-Sampled
(19), and (26) for Nyquist-rate sampling, we simulated QPSK OFDM
with 8-OFDM over the AWGN channel, the flat Rayleigh When oversampling and filtering is employed as shown
fading channel and a two-tap frequency-selective quasi-static in Fig. 1b, the above analysis still holds except that filter-
fading channel with a channel delay profile σh2 = [0.6, 0.4]. ing (out-of-band signal removal) may destroy the one-to-one
The MLSD is performed via exhaustive search, i.e., the best mapping between X and S, leading to degradation in MLSD
sequence from all 48 = 65536 possibilities is selected as the performance at high SNR. However, for large N , following
final decision given a received sequence. the analysis in [7], it can be proved that X is still one-
The comparisons over the AWGN and flat Rayleigh fading to-one mapped to S with probability one. Thus the MLSD
channels are demonstrated in Fig. 3. “MLSD sim” denotes ex- approximations (18), (19) and (26) for Nyquist-rate sampling
haustive search simulations explained above. Fig. 4 compares are lower bounds for the oversampled case in corresponding
the MLSD bounds and simulations on the frequency-selective channels as well. In fact, they serve as very tight lower bounds

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:05 from IEEE Xplore. Restrictions apply.
PENG and RYAN: MLSD BOUNDS AND RECEIVER DESIGNS FOR CLIPPED OFDM CHANNELS 3573

for the oversampled cases. we have


The direct derivation of MLSD approximations for the √ )π/2   
oversampled case follows the same procedure as that for the clipped 4 M −4 H H Λ(n) H
Ps,freq-sel ≈ √ EH exp − dθ
Nyquist-rate sampling,
 except  that for the AWGN and flat fad- π M 2N0 sin2 θ
2 0
ing channels Ec U − V n must be modified, and Λ needs √ )π/2 

to be updated for the frequency-selective channel. It can be 4 M −4 1


= √ Φη − dθ
readily verified via simulations that for an oversampling factor π M 2N0 sin2 θ
of J ≥ 2,the total power of out-of-band “clipping noise” 0
" # √
&
)π/2L−1 −1
2
is roughly 1 − e−γ + γ 2 E1 γ 2 /2 − α2 d2min /2, which is 4 M −4 λi d2min
  = √ 1+ dθ.
2
approximately the amount of reduction of Ec U − V n π M i=0
2N0 sin2 θ
0
after filtering. The above expression comes from J = 2. It
becomes a reasonable approximation for J > 2 because a Let L* ≤ L be the number of eigenvalues satisfying
larger J does not lead to a substantially increased power of
2
λi dmin /(2N0 ) 1. By only accounting for the L* multiplica-
out-of-band clipping noise. This agrees with the observaion tive terms in the integrand above, we can further approximate
in [1] that J > 2 yields very similar PAPR reduction results the above expression as
compared with J = 2. Therefore, we can modify (17) and (22)

accordingly. More importantly, this suggests that the MLSD
clipped 4 M −4 &  (M − 1) sin2 θ

)π/2L−1
*
limit for oversampled cases is approximately independent of Ps,freq-sel ≈ √ dθ
J ≥ 2. π M i=0
3λi Es /N0
0
Indeed, as we can observe from Fig. 3 and Fig. 4 that the √ ⎛ ⎞
*
&
L−1 
−L*
difference between J = 2 and J = 16 is almost negligible 4 M −4⎝ M − 1 ⎠ Es *
= √ I(L)
over all three channel types. Note that to avoid an over- π M i=0
3λi N0
crowded figure, we did not add the MLSD approximation 
−L*
for oversampled OFDM in Fig. 3. In Fig. 4, the MLSD Es
∝ . (27)
approximation for oversampled OFDM is identical for all N0
J ≥ 2. As suggested above, its difference from the MLSD !
* ≡ π/2 sin2L* θdθ was used on the second line.
where I(L)
for Nyquist-rate sampling is negligible. In both figures, to 0
reduce figure-crowdness, we omitted the case J = 8, but the Thus, the diversity d of the system is the number of “not very
performances of J = 8 and J = 16 are essentially identical. small” eigenvalues, more precisely,

+ + ,+
+ + +
As predicted, the MLSD performance with oversampling is
d=L * Es ≡ + λi +λi M − 1 1 + ≤ L, (28)
slightly worse than that with Nyquist-rate sampling when SNR N0 + + 3 Es /N0 +
becomes large, and there is little difference between J = 2 where | · | here denotes the size of a set. For practical purpose,
and J = 16. As a result, below we focus our discussion may be considered as 10 times larger. Since λi depends on
on the Nyquist-rate sampling, with an understanding that the the clipping ratio γ, L* is also a function of γ.
oversampled case follows the same trend. The diversity is a result of both clipping and channel
frequency-selectivity. For flat Rayleigh fading channels, the
IV. I MPLICATIONS AND D ISCUSSIONS OF THE MLSD equivalent tapped-delay line model has a single tap, i.e.
B OUNDS L = 1 [10]. As a result, there is no additional achievable
diversity on the flat Rayleigh fading channel or the AWGN
Observe the surprising result in Fig. 4 that the OFDM channel with clipping. The MLSD approximations (18) and
system with a quite low clipping ratio of 1.2 is superior to (19) confirm this analysis. Clipping on these two channels only
the unclipped OFDM system by almost 10 dB at a BER of results in performance degradation over unclipped systems,
10−4 . Further, the clipped-OFDM curve has a much steeper even with MLSD. On the other hand, if there is no clipping
slope, an indication of the higher diversity due to clipping but the channel is frequency-selective, OFDM subcarriers are
attained by the MLSD receiver. On the other hand, the same independent. In this case, as mentioned earlier, MLSD is
phenomenon is not observed for the AWGN and flat fading identical to ML-symbol-by-symbol detection, which means
channels, as depicted in Fig. 3. Below, we will take a more again d = 1. Mathematically, without clipping, Λ is an all-
thorough look at the diversity and performance gain attained zero matrix except the n-th diagonal entry being 1. As such,
by the clipped system over frequency-selective channels. λn = 1 and λi = 0 for i = n, which indicates d = 1. Hence,
the diversity does not come from either clipping or frequency-
selectivity alone. Thus, we call the achievable diversity for a
A. The Diversity of Clipped OFDM Systems over Frequency- clipped OFDM system over frequency-selective channels the
Selective Channels clipping-induced frequency-diversity.
The diversity d is defined as d ≡ − log(Ps )/ log(Es /N0 )
[10], which describes the rate of decay of error probability as a B. The Optimal Clipping Ratio of Clipped OFDM Systems
function of Es /N0 . We start by taking an alternative approach over Frequency-Selective Channels
in the derivation of the MLSD bound approximation. Using  define ρ ≡ Es /N0 . If there is no
To simplify notation,
Craig’s formula for the Q-function [13], from (24) and (25), clipping, using 1 − 1/(1 + x) ≈ x/2 for small x, the

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:05 from IEEE Xplore. Restrictions apply.
3574 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 9, SEPTEMBER 2008

−1 1.02
10
MLSD bnd, N=128
d=1 MLSD bnd, N=8

normalized performance gain Δ(γ)


simulation, exhaustive search 1
−2
10

0.98
−3
10 no clipping
BER

0.96
σ2h = [0.4 0.3 0.3]
−4
10 γ = 1.1 σ2h=[0.6 0.4]
0.94
d=3 γ=1.8 σ2h=[0.4 0.3 0.3]
−5
10 0.92 σ2=[0.6 0.4]
h
σ2h = [0.8 0.2]
σ2=[0.8 0.2]
γ = 0.9 d=2 h
−6 0.9
10 0.6 0.7 0.8 0.9 1 1.1
10 15 20 25 30 35
E /N , dB clipping ratio, γ
b 0

Fig. 6. Normalized performance gain for 16-QAM and 128-OFDM over


Fig. 5. MLSD bounds and simulations for QPSK and 8-OFDM over frequency-selective quasi-static Rayleigh fading channels with different clip-
frequency-selective quasi-static Rayleigh fading channels with different clip- ping ratios and channel delay profiles.
ping ratios and channel delay profiles. d indicates the diversity of the system.
−2
10
MLSD approximation (13) for√ the no clipping case can be
no clipping M −1 2(M−1)
approximated as Ps,fading > √ M 3ρ . Using (27) and
the above equation, for large SNR, the performance gain in dB −3
10
over the unclipped system that operates at ρdB = 10 log10 (ρ)
is then
1 
L̃−1
BER

L̃ − 1 1 −4
10
Δρ (γ) = 10 log10 λi + ρdB − C(L̄), (29)
L̃ i=0 L̃ L̃
 
" M−1 #L̃−1 γ=0.6
where C(L̄) = 10 log10 2I(L̃) 3 /π . This perfor- −5
10 γ=0.85
mance difference increases linearly with the operating SNR γ=1.1
γ=2.5
ρdB of the unclipped system. Again, the performance gain is
no clipping
a result of both clipping and channel frequency-selectivity. −6
10
Intuitively, as the clipping ratio decreases, the amount of 15 20 25 30
E /N , dB
power spread from a subcarrier to others relative to the b 0

remaining power of its own becomes larger, which means that


Fig. 7. MLSD bounds for 16-QAM and 128-OFDM over a frequency-
subcarriers are more correlated and the MLSD performance selective quasi-static Rayleigh fading channel with various clipping ratios
becomes better. However, due to clipping, the total transmitted and a delay profile σh2 = [0.4, 0.3, 0.3].

power also decreases at the same time, leading to lower


average received SNR. As a result, there exists a point where same slope (or diversity) as that without clipping, agreeing
the correlation effect over the reduced SNR becomes the with conventional expectations. The lower the clipping ratio
greatest. The clipping ratio corresponding to this point is the is, the larger the performance degradation due to clipping
optimal clipping ratio for the frequency-selective channel. becomes. However, for frequency-selective channels, interest-
At a fixed ρdB , for a given system and channel, Δρ (γ) only ingly and surprisingly, as we predicted, clipping results in
depends on the eigenvalues λi = λi (γ). Thus, the optimal diversity and performance gain over unclipped systems.
clipping ratio for a given frequency-selective channel and
Fig. 5 illustrates the MLSD approximations together with
OFDM system can be found as
- L−1 . -L−1 . simulations for different channel delay profiles and different
1  & 1/L clipping ratios. Again, exhaustive search over 65536 possible
γopt (L) = max log λi (γ) = max λi (γ) . sequences for each received sequence was performed in the
γ L i=0 γ
i=0 simulations. For a two-tap channel, at high SNR, we observe
(30)
that the diversity (or the slope of the curve times 10) is
As a result, the optimal clipping ratio is the one that maximizes
indeed d = 2 regardless of the clipping ratio and the channel
the geometric mean of the eigenvalues of RΛ. Note that at
* instead. delay spread, as predicted by (28). At low SNR, the diversity
low SNR, it is more relevant to consider γopt (L)
is reduced to 1, which also agrees with (28). The same
phenomenon is observed for a three-tap channel for which
C. Discussions on the MLSD Bound d = 3. In addition, for N > 8, there is little difference in
As we discussed above, for the AWGN and flat Rayleigh the MLSD bound approximations, since the eigenvalues stay
fading channels, the MLSD performance with clipping has the roughly unchanged for large N .

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:05 from IEEE Xplore. Restrictions apply.
PENG and RYAN: MLSD BOUNDS AND RECEIVER DESIGNS FOR CLIPPED OFDM CHANNELS 3575


Fig. 6 plots L−1
i=0 log λi /L normalized by the largest value
for 16-QAM and 128-OFDM. It can be observed that different /
P (blm = 0|X)
channel delay spreads may lead to different optimal clipping Llm ≈ log
/
P (blm = 1|X)
ratios. For the three-tap channel, the optimal clipping ratio is 
 0 $ %0 
− 02
predicted to be around 0.85. This is verified in Fig. 7, where 0 H *
Xl ∈χ− exp − N10 0 r − g F X l 0 P (Xl )
γ = 0.85 results in better performance than γ = 0.6 or γ =
= log  0 $ %02 
1.1. This also confirms our previous discussion that as γ → 0,  0 *+ 0
the performance first becomes better and then becomes worse Xl ∈χ+ exp − N10 0r − g FH X l 0 P (Xl )
relative to the unclipped system.
where P (Xl ) is the a priori probability of the l-th symbol,
We remark here that a necessary condition for the MLSD
X* + denotes the a priori estimated sequence except the l-
approximations to hold is the one-to-one mapping between l
the OFDM modulator input X and the clipper output u. This th element whose corresponding m-th bit blm is 1, and X *−
l
condition is satisfied provided√that the
clipping ratio is above similarly defined for blm = 0. Using
M −3 3π
a certain threshold, i.e., γ > 2M−1 for large N , as
2 +  +2
given in [7]. When this condition is not satisfied, the above 0 $ %02 N −1 + 
N −1 +
0 0
* 0 = + 1 *l ej2πil/N ++ ,
bounds become loose and MLSD performance may exhibit an 0r − g FH X +ri − g √ X
l
+ N l=0 +
error floor. In addition, the above MLSD approximations also i=0
" #
become loose when the number of clipped elements in u is together with P (blm ) = e(1−blm )Llm / 1 + eLlm , and
2
not approximately N e−γ , which happens for large clipping  N −1 * j2πil/N
√1 = √1N (X *l − X̂l )ej2πil/N + x̂i , it can
l=0 Xl e
ratios or small N . This is the condition for which (17) holds. N
be shown that Llm = Llm + Llm , where Lelm is the extrinsic
e a

information expressed in (31) on the top of the next page, and


Lalm represents the a priori estimated LLR.
V. N EAR -O PTIMUM S EQUENCE D ETECTION As shown on the top of the next page, we can further
 expand
Lelm using the well-known approximation log [ i exi ] ≈
Since the exact MLSD is of course not practical, the natural max (xi ), to obtain
question to ask is that if the MLSD performance bounds i
can be achieved or closely approached for algorithms with 
N −1
reasonable complexity. This question is particularly relevant Lelm ≈ Lei→blm (33)
for frequency-selective channels where OFDM is famous for i=0
and where clipping results in diversity and performance gains. where Lei→blm is the extrinsic information delivered from the i-
We believe that the answer is yes, and below we present algo- th received time-domain sample ri to blm . The detector output
rithms that can achieve MLSD bounds on the AWGN and flat can be used to compute E[X] to estimate X / for the next
Rayleigh fading channels. For frequency-selective channels, iteration. The initial LLR estimate can be generated based
we first describe a vector-based model of the clipping process, on the statistical model. Since the extrinsic information from
and then we present a receiver design that can utilize the ri is independently computed from other received samples,
benefit of clipping to some extent. (33) can be implemented via N parallel processors. Obviously
this process can be repeated to improve
" # LLR estimates. The
complexity is on the order of O N 2 .
A. A Soft-Output Iterative Demapper When oversampling and filtering is employed, due to
filtering at the transmitter, we can no longer process the
Let’s first consider Nyquist-rate sampling over the AWGN
signal in the time-domain as given above. Instead, the re-
channel. The algorithms we now present are based on the
ceiver can use the equivalent model in Fig. 1b and pro-
deterministic model of the clipper (1) rather than the statistical
cess
+ the$ signal
$ in the frequency-domain
% %+2by simply replacing
model (3) and were motivated by the results of [7]. + +
+ri − g √ 1 * /
Xl − Xl e j2πli/N
/i + in (31) by (34) for
+x
In order to maintain reasonably low complexity, we assume N
that we are given an a priori detected sequence X. / Since 0 ≤ i < N . As a result, the same parallel structure is retained.
in modern communication systems, error-control coding with In this case, an additional JN -point FFT " must be employed
#
soft-input decoding is often implemented, we design a fully- for each Ri . Thus, the complexity is O JN 2 log(JN ) .
parallel iterative soft-output detector. For both flat and frequency-selective fading channels,
Let r = FH R, where R is given in (5). Let χ denote the the receiver can first preprocess the received signal by
set of symbols in the signal constellation. Then M = |χ|. Let normalizing the channel gain, and then use the algorithm for
the AWGN channel described above with R *i = Ri /Hi for
χ+ be the symbols whose bit under consideration is 1, with
χ− similarly defined for 0. Let blm (0 ≤ m < log2 M and +0 ≤ i < N . Alternatively,
$ (34) in (31) can %be modified to+
+ * l −X
/ l j2πlk/(JN ) +2
/k e−j2πik/(JN ) + ,
JN
0 ≤ l < N ) be the m-th bit corresponding to the l-th symbol +Ri − JN k=0 g
√Hi X√
JN
e +x
(or carrier). Its corresponding log-likelihood ratio (LLR) is where 0 ≤ i < N and H"i = H for the #flat fading channel.
defined as Llm ≡ log [P (blm = 0)/P (blm = 1)]. The complexity is also O JN 2 log(JN ) .
Assuming we have previously computed LLR’s as defined As we shall later present, the above algorithm achieves
above, we can estimate the symbol sequence by the soft values the MLSD bound on the AWGN and flat Rayleigh fading
as X/ = E[X], where E[Xi ] = 
S∈χ SP (Xi = S). Then channels. For frequency-selective channels, it can restore the

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:05 from IEEE Xplore. Restrictions apply.
3576 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 9, SEPTEMBER 2008

 
 1
 ++ $
1
$
* /
%
j2πli/N
%+2 
+ a
X* ∈χ−
l
exp − N0 i +ri − g √
N
X l − X l e + x
/ i + + k=m (1 − b lk )Llk
Lelm = log  + $ $ % %+  (31)
 1
 + 1 * / +2  a
* ∈χ+ exp − N i +ri − g Xl − Xl e +x/i + + k=m (1 − blk )Llk
√ j2πli/N
X l 0 N
⎛ ⎡ ⎤
+  $ %
+2
 + + 
≈ ⎝ max ⎣− 1 +ri − g √1 X*l − X/l ej2πli/N + x /i ++ +
1
(1 − blk )Lalk ⎦
i
* ∈χ−
X l
N0 + N N
k=m
⎡ ⎤⎞
+  $ %
+2
1 + 1 + 1 
− max ⎣− +r i − g √ *l − X
X /l ej2πli/N + x /i ++ + (1 − blk )Lalk ⎦⎠
* ∈χ+
X l
N0 + N N
k=m
 + $ $ % %+2 
 1 + 1 * / j2πli/N + 1
 a
 * ∈χ−
X l
exp − N0 +r i − g √
N
X l − X l e + x
/ i + + N k=m (1 − b lk )L lk
≈ log  + $ $ % %+ 
 1 + 1 * / + 2
1
 a
i * ∈χ+
X l
exp − N0 +r i − g √
N
X l − X l e j2πli/N + x
/ i + + N k=m (1 − b lk )Llk


N−1
= Lei→blm (32)
i=0

+   +2
+ 1 
JN *l − X
/l +
+ X j2πlk/(JN ) −j2πik/(JN ) +
+Ri − √ g √ e /k e
+x + (34)
+ JN k=0 JN +

$ %
unclipped performance, but it cannot utilize the benefit of diag (ς), W = diag |x10 | , |x11 | , · · · , |xJN
1
| , and X =
clipping simply by itself. In fact, we have not found any diag (x) are diagonal matrices of size JN by JN .
existing practical algorithm in the literature that can utilize After the JN -point FFT and removing the out-of-band
the diversity induced by clipping. Below, we introduce a components (taking only the first N elements), the OFDM
vector-based modeling approach of the clipping process which symbol vector is
enables application of more sophisticated signal-processing
algorithms, which, when combined with our algorithm above, S = FN,JN · u = X − FN,JN Ψ (IJN − AW) FH
N,JN X
can utilize the advantage of clipping. (37)

where FN,JN denotes N by JN partial symmetric FFT


B. A Vector-Based Modeling of the Clipping Process and A matrix. The N -point IFFT of the length-N vector S is then
MMSE-SD Receiver Design transmitted through the channel.
We use an approach similar to the one in [4], where an At the receiver, the system can be described as R = HS +
auxiliary indicator variable is employed to aid the modeling W as given in (5). Substituting (37) into the above equation
process. We can write g(xi ) in (1) as yields

R = HΓX + W = GX + W (38)
A
g(xi ) = xi − ςi 1 − xi (35)
|xi | with the N by N clipping matrix Γ = I −
 FN,JN Ψ (IJN − AW) FH . The entries of the channel
0, |xi | ≤ A N,JN
where ςi = 1, |xi | > A
is a clipping indicator variable transfer matrix G = HΓ can be computed as
(similar to [4]). Note that in the receiver the choice of ςi may 

Hk 
JN
be made according to its log-likelihood ratio, thus making Gkl = Hk δk−l − ςm 1 −
A
e−j2πm(k−l)/(JN ) ,
“clipping” probabilistic at the receiver. The term JN m=0 |xm |

(39)
A
ci = ςi 1 − xi (36) where δk−l is the Kronecker delta.
|xi |
As we can see from (38), the vector-based description of
in (35) is the clipping noise. Note that the semi-deterministic the clipped OFDM channel is analogous to that for a fast
clipping noise in this model is different from that discussed fading OFDM channel or a MIMO channel. Thus we have
in (3) in Section II. The following derivations assumes an referred to G as the channel transfer matrix and, in this
oversampling factor of J. To avoid confusion, we denote the setting, we need to jointly estimate the “channel” as well as
N by N identity matrix by I, and the JN by JN identity detect the signal. It can be easily shown that G is a full-
matrix by IJN . The size of the rest matrices and vectors are rank matrix and that most of the power is contained along the
clear within the context. diagonal. The power of each entry of G may be computed
The signal vector at the clipper output in Fig. 1b can be by first noticing that the second√term in (39) is the JN -
written as u = x−Ψ (IJN − AW) x = x−X (IJN − AW) ς point FFT of ςm (1 − A/ |xm |) /√ JN and then computing
where ς is the clipping index vector of length JN , Ψ = the power of ςm (1 − A/ |xm |) / JN . From (36), we then

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:05 from IEEE Xplore. Restrictions apply.
PENG and RYAN: MLSD BOUNDS AND RECEIVER DESIGNS FOR CLIPPED OFDM CHANNELS 3577

   
need to compute E |ci |2 . From (35) and (36), E |ci |2 =
−1
10
 + 
2+
E (|xi | − A) + |xi | > A . After a simple exercise of first
 + 
+
−2

computing E [ |xi || |xi | > A] and E |xi |2 + |xi | > A similar


10
 
2
to that in (III-A), it can be shown that [6] E |Hk,l | =
 2 " √ #  √ 2 " √ #
−3
10
4 1
δk−l 1 − √JN eγ Q γ 2 + JN 1 − 2 πγeγ Q γ 2 ,

BER
  Iter 5 cancel, γ=1.2
2
which agrees with the fact that as γ → ∞, E |Hk,l | → 1. −4 Iter 5 demap, γ = 1.1
10
MLSD bnd, γ = 1.1
Since (38) is similar to that for a MIMO channel, many
Iter 5 demap, γ = 1.2
existing equalization and multi-user detection techniques may MLSD bnd, γ = 1.2
be applied. We adopt the MMSE-SD technique [9] as the pre- −5
10 Iter 5 demap, γ = 1.5
processing unit before sending the received sequence to the MLSD bnd, γ = 1.5
proposed iterative demapper. The MMSE-SD technique es- −6
no clipping
sentially consists of three steps for detecting each modulation 10
8 9 10 11 12 13 14 15
symbol successively: 1) MMSE filter generation and the next- E /N , dB
b 0
symbol selection, 2) MMSE filtering, and 3) interference can-
cellation and nulling. This technique is designed to minimize Fig. 8. Performance of the proposed algorithm (“demap”) for different
clipping levels with five iterations for 16-QAM and 128-OFDM with Nyquist-
the mean-squared error between the MMSE filter output and rate sampling over the AWGN channel.
the desired symbol at each successive stage. It is identical to
the well-known V-BLAST technique [14] for MIMO channels. −1
10
The complexity of the MMSE-SD technique is O(N 3 ) [9].
As mentioned before, the initial estimate of bit LLR’s
for each received OFDM symbol can be obtained via the 10
−2

statistical model. Then E[X] can be computed and used in


the MMSE-SD process. Since the output of the MMSE filter
is approximately Gaussian [15], bit LLR’s are computed at 10
−3

the MMSE-SD output. A block diagram of this scheme can


BER

be found in [6]. These LLR’s are then passed to the proposed


Nyquist, 1 iter
iterative demapper discussed earlier. We note that it is possible
−4
10
Oversample J=2, 1 iter
to iterate between the MMSE-SD unit and the demapper. Nyquist, 2 iter
−5 Oversample J=2, 2 iter
10 Nyquist, 3 iter
C. Results and Discussion
Oversample J=2, 3 iter
In the simulations, we assume uncoded 16-QAM and 128- −6
MLSD bnd
OFDM (N = 128). For both flat and frequency-selective 10
8 9 10 11 12 13 14 15
quasi-static fading channels, we assume that the channel co- E /N , dB
b 0
efficients are Rayleigh distributed and the receiver has perfect
channel knowledge. The iterative cancellation technique [2] is Fig. 9. Performance of the proposed algorithm for 16-QAM and 128-OFDM
simulated for comparison. over the AWGN channel with and without oversampling. Clipping ratio is 1.1.
Fig. 8 plots the MLSD bounds (“MLSD bnd”) and the
performance of the proposed iterative demapping algorithm The proposed algorithm approaches the MLSD bound with
(“demap”) with five iterations for different clipping levels three iterations. On the other hand, the iterative cancellation
over the AWGN channel. “iter 5 cancel” denotes the iterative technique [2] does not perform as well due to the difficulty
cancellation technique [2] based on the statistical model (3) of reducing the pseudo clipping noise when the entire block
with five iterations. For both high and low clipping ratios, the is in a deep fade.
algorithm achieves the MLSD bound (18). Thus the algorithm Fig. 11 illustrates the performance of the receiver consisting
is near optimum, and it also verifies that the bound is very of the MMSE-SD unit and the iterative demapper over a
tight. For a clipping ratio of 1.2, the proposed algorithm frequency-selective Rayleigh fading channel, with a clipping
is asymptotically about 0.75 dB better than the iterative ratio of 1.2. The channel has an exponentially decaying profile
cancellation technique. with 17 taps. “MMSE-SD” denotes MMSE filtering with
Fig. 9 illustrates the performance of the proposed algorithm successive detection [9]. “iter-demap” denotes our iterative
for a clipping ratio of 1.1 which corresponds to a signal power demapper discussed earlier, which by itself restores the un-
loss of around 30%. The performance with Nyquist-rate sam- clipped performance as shown in the figure. Due to its heavy
pling and twice oversampling (J = 2) approaches the MLSD dependence on the a priori estimate, a sophisticated pre-
bound (18) with only a few iterations. However, oversampling processor such as the MMSE-SD is required in order to utilize
and filtering leads to faster convergence, requiring just two the diversity induced by clipping. In our simulations, other
iterations. existing receiver algorithms may also restore the unclipped
Fig. 10 shows the performance of the system over a flat performance on the frequency-selective channel, such as that
block Rayleigh fading channel with Nyquist-rate sampling. in [2]. But we have found no existing practical algorithm

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:05 from IEEE Xplore. Restrictions apply.
3578 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 9, SEPTEMBER 2008

−1
10 ventional understanding, for uncoded systems over frequency-
selective channels, soft clipping by itself without manipula-
tions such as those in [12] leads to diversity and performance
−2
10 gain in the frequency-selective channel with ML sequence
detection. We emphasize that this phenomenon is a result of
both clipping and channel frequency-selectivity, and does not
come from either alone. We have also shown that our MLSD
BER

−3
10
approximations are very tight, and an optimal clipping ratio
no mitigation exists for a given system. In addition, soft clipping and MLSD
proposed, 1 iter provide a way to achieve diversity over frequency-selective
proposed, 2 iter
−4
10 proposed, 3 iter
fading channels without resorting to Rake receivers [10]
iter 5 cancel whose construction relies on channel knowledge. We presented
MLSD bnd, clipped algorithms that can achieve MLSD bounds on the AWGN and
−5
flat Rayleigh, no clipping flat Rayleigh fading channels. We also demonstrated that with
10
15 20 25 30 35 40 manageable complexity, the clipping-induced diversity and
E /N , dB
b 0 performance gain can be achieved over frequency-selective
channels. Our future work includes improved receiver designs
Fig. 10. Performance of the proposed algorithm for 16-QAM and 128-OFDM for the frequency-selective channel, and extension of the above
over a flat block Rayleigh fading channel with clipping ratio of 1.2.
−1
results to MIMO systems.
10

R EFERENCES
[1] H. Ochiai and H. Imai, “Performance analysis of deliberately clipped
OFDM signals,” IEEE Trans. Commun., vol. 50, pp. 89-101, Jan. 2002.
−2 [2] H. Chen and A. M. Haimovich, “Iterative estimation and cancellation
10
of clipping noise for OFDM signals,” IEEE Commun. Lett., vol. 7, pp.
305-307, July 2003.
[3] J. Tellado, L. M. C. Hoo, and J. M. Cioffi, “Maximum-likelihood
BER

detection of nonlinearly distorted multicarrier symbols by iterative


decoding,” IEEE Trans. Commun., vol. 51, pp. 218-228, Feb. 2003.
−3 [4] D. Declercq and G. B. Giannakis, “Recovering clipped OFDM symbols
10 with Bayesian inference,” in Proc. IEEE Inter. Conf. Acoustics, Speech
MLSD bnd and Signal Proc. (ICASSP’00), vol. 1, pp. 157-160, June 2000.
MMSE−SD + iter−dmap [5] D. Kim and G. L. Stuber, “Clipping noise mitigation for OFDM by
iter−dmap decision-aided reconstruction,” IEEE Commun. Lett., vol. 3, pp. 4-6,
no clipping Jan. 1999.
no mitigation [6] F. Peng and W. E. Ryan, “New approaches to clipped OFDM channels:
−4
10 modeling and receiver design,” in Proc. 2005 IEEE Global Telecommun.
10 15 20 25 30 Conf. (GlobalCom’05), vol. 3, pp. 1490-1494, St. Louis, MO, Dec. 2005.
E /N , dB [7] F. Peng and W. E. Ryan, “On the capacity of clipped OFDM chan-
b 0
nels,” in Proc. 2006 IEEE International Symp. on Information Theory
Fig. 11. Performance for 16-QAM and 128-OFDM over frequency-selective (ISIT’06), pp. 1866-1870, Seattle, WA, July 2006.
quasi-static Rayleigh fading channels with Nyquist-rate sampling and γ = [8] F. Peng and W. E. Ryan, “A low complexity soft demapper for OFDM
1.2. fading channels with ICI,” in Proc. IEEE WCNC 2006, vol. 3, pp. 1549-
1554, Las Vegas, NV, Apr. 2006.
[9] X. Cai and G. B. Giannakis, “Bounding performance and suppressing
that can utilize the diversity induced by clipping on the intercarrier interference in wireless mobile OFDM,” IEEE Trans. Com-
frequency-selective channel. The receiver presented can utilize mun., vol. 51, no. 12, pp. 2047-2056, Dec. 2003.
the diversity in the low-to-medium SNR region. It confirms [10] J. Proakis, Digital Communications. McGraw Hill, 4th edition, 2000.
[11] H. Ochiai, “Performance of optimal and suboptimal detection for
that the benefit of clipping on frequency-selective channels uncoded OFDM system with deliberate clipping and filtering,” in Proc.
not only exists as analyzed throughout this paper, but can also IEEE GlobeCom 2003, pp. 1618-1622, Dec. 2003.
be achieved with algorithms of manageable complexity for [12] H. Qian, R. Raich, and G. T. Zhou, “On the benefits of deliberately
introduced baseband nonlinearities in communication systems,” in Proc.
large N . ICASSP 2004, vol. 2, pp. 17-21, May 2004.
[13] J. W. Craig, “A new, simple and exact result for calculating the
VI. C ONCLUSION probability of error for two-dimensional signal constellations,” in Proc.
IEEE MILCOM’91, pp. 25.5.1-25.5, Boston, MA, 1991.
We derived tight closed-from MLSD approximations for [14] J. Benesty, Y. Huang, and J. Chen, “A fast recursive algorithm for
clipped OFDM systems over the AWGN, flat and frequency- optimum sequential signal detection in a BLAST system,” IEEE Trans.
Signal Processing, vol. 51, pp. 1722-1730, July 2003.
selective Rayleigh fading channels. We considered rectangular [15] X. Wang and H. V. Poor, “Iterative (turbo) soft interference cancellation
QAM with both Nyquist-rate and oversampled OFDM sys- and decoding for coded CDMA,” IEEE Trans. Commun., vol. 47, pp.
tems. More importantly, we have shown that, contrary to con- 1046-1061, July 1999.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:05 from IEEE Xplore. Restrictions apply.

You might also like