4784 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO.

9, SEPTEMBER 2009

Decision-Directed Least-Squares Phase

Perturbation Compensation in OFDM Systems
Ioannis Dagres, Student Member, IEEE, and Andreas Polydoros, Fellow, IEEE

Abstract—A low-complexity decision-directed iterative scheme the common-phase rotation (CPR) are not adequate. More
is proposed for the estimation and mitigation of strong phase elaborate phase-process estimation and compensation methods
noise plus frequency offset in OFDM-modulated signals, trans- are needed, namely those that also target the ICI effect. Several
mitted over frequency-selective or frequency-flat channels. It is
based on a time-domain, windowed least-squares estimation algo- such methods have already been proposed in the literature
rithm of the total phase perturbation vector, developed through for PHN and FO estimation and compensation in both the
proper system modeling and optimized so as to minimize the data-detection stage (DDS) and the channel-estimation stage
residual inter-carrier interference. An extension of the diagonal- (CES). In [13], two decision-directed approaches are proposed
loading technique is also introduced in order to enhance the employing the maximum likelihood (ML) as well as the
convergence of the algorithm in high-SNR regimes. The resulting
scheme outperforms previously proposed approaches in terms linear minimum mean square error (LMMSE) criteria for
of achievable error rate in the data-detection stage while, in the estimation of the ICI terms at the DDS. The LMMSE
the channel-estimation stage, it attains performance near the approach has shown to provide good performance at the cost
Cramer-Rao bound with comparatively smaller computational of computational complexity. A lower-complexity LMMSE
complexity. estimator was proposed in [14] targeting the higher-energy
Index Terms—Orthogonal frequency division multiplexing ICI terms in the DDS. In [15], a scheme for estimating the
(OFDM), phase noise (PHN), channel estimation, least squares channel and initial frequency offset (IFO) in the presence of
(LS), frequency offset. PHN was proposed using training symbols. It was shown that
channel estimation close to the respective Cramer-Rao bound
I. I NTRODUCTION (CRB) is feasible at all SNR’s of interest. In [16], a systematic
probabilistic framework was presented that allows for joint
O RTHOGONAL frequency division multiplexing
(OFDM) systems operating in strong levels of phase-
noise (PHN) and frequency offset (FO) are known to suffer
estimation of data symbols plus PHN. A key difference of
this last work with respect to those based on hard decisions
is the use of probability distributions for the data (i.e., soft
from significant performance degradation due to the high
decisions).
levels of total-phase perturbation (TPP) induced inter-carrier
In this paper, a parameterized windowed least-squares
interference (ICI), a fact which limits their achievable
(WLS) estimator is proposed via proper system modeling
performance [1]–[6] and thus makes the use of higher-
applied to both estimation stages. The window is optimized
order constellations for increased transmission throughput
so as to minimize a newly introduced metric, namely the
infeasible. Scenarios of strong phase errors are becoming
average post-compensation error variance (A-PCEV) of the
of importance due to recent interest in millimeter-wave
residual phase. It will be shown that A-PCEV is the domi-
low-cost radios [7], [8]. In high frequencies, phase-related
nant term of the residual (e.g., post-compensation) ICI and
errors are more pronounced because the effect of a noisy
thus determines system performance. It will be analytically
oscillator grows with the square of the carrier frequency.
computed for arbitrary phase-error models and closed-form
In the world of ”Dirty RF” as described in [9], new digital
expressions for near-optimal windows will be derived for zero-
signal processing techniques are needed to deal with higher
mean FO, Wiener and first-order autoregressive PHN mod-
levels of impairments, thus relaxing the requirements on
els, respectively. Furthermore, the diagonal-loading approach
future RF sub-systems.
(proposed in [17] for providing robustness to a general class
In such strongly phase-impaired scenarios, typical solu-
of estimators in the presence of model mismatch) will be
tions (e.g., [10]–[12]) that only address the correction of
employed to enhance convergence of the iterative estimation
Manuscript received October 24, 2008; revised April 10, 2008; accepted scheme, in those high-SNR regions where the effect of data
May 29, 2009. The associate editor coordinating the review of this paper and decision errors dominates performance. In addition, channel,
approving it for publication was J. Coon. IFO estimation and data equalization will also be based on
This work has been performed in the framework of the project “EPEAEK II
- PYTHAGORAS II - Support to Research Groups in Universities,” co-funded the same types of LS estimators already proposed for non-
by the European Social Fund and Greek National Resources. The material of PHN systems, thus keeping the overall system complexity
this paper was presented in part at the 10th IEEE International Symposium low. A high-SNR approximation of the variance of the LS
on Spread Spectrum Techniques and Applications (ISSSTA), Bologna, Italy,
August 2008. estimator for the initial FO is also provided, which variance
I. Dagres is with the Department of Electrical and Computer Engineering, characterizes the level of the residual frequency offset (RFO).
University of Patras, Patra, Greece (e-mail: [email protected]). The time-domain LMMSE estimator (proposed in [13] in the
A. Polydoros is with the Department of Physics, National and Kapodistrian
University of Athens, Athens, Greece. frequency domain), is generalized for the case of PHN plus
Digital Object Identifier 10.1109/TWC.2009.081420 FO by computing and exploiting their combined statistics.
1536-1276/09$25.00 ⃝
c 2009 IEEE
DAGRES and POLYDOROS: DECISION-DIRECTED LEAST-SQUARES PHASE PERTURBATION COMPENSATION IN OFDM SYSTEMS 4785

To summarize, the contributions of this paper are: (a) the channel. At the receiver, after removing the cyclic prefix, the
introduction of a particular time-domain model for the TPP, received sampled signal can be represented in vector notation
which is then applied to the sequence of unknown phases as
in the observed total record; (b) the development of an LS 𝒚 = diag{𝒑𝑒 }𝒈 + 𝒏 (1)
estimator (applicable to different models of the phase pertur-
bations), based on the modeling approach of (a) above; (c) an where 𝒈 ≜ 𝑭 𝐻 𝑯𝒔, 𝑭 is the unitary DFT matrix of length
elaborate but eventually simple procedure for optimizing the 𝑁 , 𝑯 = diag{𝒉𝑓 (} where) 𝒉𝑓 is the channel’s frequency
associated window upon which the estimator of (b) is based; response, 𝒏 ∼ 𝒩ℂ 0, 𝜎𝑛2 𝑰 is the additive white Gaussian
and (d), an improvement (iterations combined with diagonal noise (AWGN) vector and 𝒑𝑒 is the exponentiated TPP vector.
loading) that enhance further the performance of the estimator The latter can be written as
of (b). The main value of this collective procedure is uniformly 𝒑 = 𝜽 + 𝝐 + 𝜗CP 1 (2)
near-optimal performance in CES (as assessed by respective
bounds) and better performance than competitive schemes in with 𝜽 the PHN angle vector, 𝝐 = 𝜀𝒗 the FO vector describing
DDS, both achieved with comparatively low computational the initial or residual, depending on whether the channel or
complexity (as assessed by the complexity of other referenced the data estimation stage is considered, 𝜀 is the normalized
methods). FO with ∣𝜀∣ ⩽ 0.5, 𝒗 = (2𝜋/𝑁 )[1, 2, ..., 𝑁 ]𝑇 , and finally 𝜗CP
This paper is organized as follows: the standard OFDM is the initial uncompensated total phase rotation. After a DFT
system model is described in Section II, while the proposed on the received sampled sequence, (1) becomes
alternative in Section III. The conventional and the proposed
estimators are introduced in Section IV. In Section V, the A- 𝒚 𝑓 = 𝑐(0)𝒈 𝑓 + 𝒏ICI + 𝒏𝑓 (3)
PCEV is derived analytically as a function of the window where 𝒈 𝑓 = 𝑯𝒔, 𝒏𝑓 is the frequency representation of 𝒏
parameters. Additionally, closed-form expressions for near- (also white), 𝒏ICI is the inter-carrier noise vector whose 𝑖th
optimal window parameters are derived and its accuracy are element is [14]
quantified via performance assessment. Section VI describes
𝑁
∑
the application of the proposed estimator on an iterative LS-
𝑛ICI
𝑖 = 𝑔𝑘𝑓 𝑐(𝑖 − 𝑘) (4)
based OFDM scheme. Section VII outlines the diagonal-
𝑘=1,𝑘∕=𝑖
loading technique and its incorporation in the iterative phase-
estimation process, along with the algorithmic steps of the pro- with
𝑁
posed scheme. Some complexity considerations are provided 1 ∑ 𝑗2𝜋(𝑘−1)𝑚/𝑁 𝑒
𝑐(𝑚) = 𝑒 𝑝𝑘 (5)
in Section VIII. Simulations results are presented in Section 𝑁
𝑘=1
IX, followed by conclusions in Section X.
Notation: ⊙ denotes element-wise multiplication; (⋅)∗ , (⋅)𝑇 We note/that, as per (5), 𝑐(0) is the CPR. The system SNR is
and (⋅)𝐻 denote conjugation, transpose and Hermitian trans- 𝛾s ≜ 𝜎𝒈2 𝜎𝒏 2
.
pose, respectively; E𝑥 [⋅] and Var𝑥 [⋅] denotes for the expected Two important PHN types exist [5]: one results from a
value and the variance operators with respect to the proba- system that is only frequency-locked and the other from its
bility destiny function (pdf) of the random variable (R.V.) 𝑥, phased-locked counterpart. For frequency-locked systems, the
respectively (when 𝑥(:) is used it corresponds to the joint pdf Wiener model is employed [2], where the PHN process is a
of all 𝑥𝑖 R.V.’s involved in the operand); Im{⋅} and Re{⋅} are sampled version of a continuous-time Wiener process. The
the imaginary and the real part operators respectively; 1 and 0 𝑖th sample value of the discrete-time
( ) PHN vector 𝜽 is 𝜃𝑖 =
are the all-ones and the all zero vectors respectively; 𝒩 (0, 𝑹) 𝜃𝑖−1 + 𝜑𝑖 , where 𝜑𝑖 = 𝑁 0, 𝜎𝜑2 . The (𝑖, 𝑘)th element of the
and 𝒩ℂ (0, 𝑹) are real and circularly symmetric complex correlation matrix of 𝜽 and its exponentiated( PHN vector 𝜽)𝑒
are 𝜎𝜑2 min{𝑖, 𝑘} (assuming 𝜃0 = 0) and exp −𝜎𝜑2 ∣𝑖 − 𝑘∣ /2
Gaussian random vectors
( )with mean 0 and correlation matrix
𝑹, respectively; 𝜒2 𝜈, 𝜎 2 is a chi-square R.V. with variance respectively. For phase-locked systems, the Gaussian PHN
𝜎 2 and 𝜈 degrees of freedom; upper-case bold letters represent model is typically used, where 𝜃𝑖 is modeled as a stationary
matrices while lower-case bold letters represent vectors (sizes random process. The (𝑖, 𝑘)th element of the discrete-time
are specifically defined if not obvious); (𝑿)𝑖,𝑗 is the (𝑖, 𝑗)th correlation matrix is equal to 𝑅 (∣𝑖 − 𝑘∣ 𝑇𝑠 ), where 𝑇𝑠 is
element of a matrix 𝑿, and 𝑥𝑖 is the 𝑖th element of vector 𝒙; the sampling period of the continuous-time autocorrelation
the exponentiated form of the real-valued vector 𝒙 is defined function 𝑅(𝑡). A first-order autoregressive model will be used,
𝑇 described by the recursive equation 𝜃𝑖 = 𝑎𝜃𝑖−1 + 𝜑𝑖 , where
as 𝒙𝑒 ≜ [exp (𝑗𝑥1 ) , ..., exp (𝑗𝑥𝑁 )] ; diag{𝒙} is the diagonal
matrix with the elements of vector 𝒙 on its diagonal. 𝑹𝒙 𝜑𝑖 defined as in the Wiener model. The (𝑖, 𝑘)th element of
denotes the correlation matrix of a vector 𝒙. the corresponding correlation matrices of 𝜽 and 𝜽 𝑒 , assuming
2 ∣𝑖−𝑘∣
∣𝑎∣ <( 1 and steady state conditions,
) are 𝜎𝜑 𝑎 /1 − 𝑎2 and
2 ∣𝑖−𝑘∣ 2
exp −𝜎𝜑 (1 − 𝑎 )/(1 − 𝑎 ) , respectively. Note that the
II. S TANDARD OFDM S YSTEM M ODEL
first-order model is the same as the one used in [15] and [16]
The symbol vector 𝒔 of length 𝑁 , normalized in power and after appropriate choice of system parameters. The normalized
assumed uncorrelated so that E{𝒔𝒔𝐻 } = 𝑰, is transmitted over FO is assumed to be a uniform R.V. 𝜀 ∼ 𝑈 (−𝛿, 𝛿), 𝛿 > 0 with
the sub-carriers of the OFDM block. After an 𝑁 -point unitary 𝜎𝜀2 = 𝛿 2 /3. In the case of IFO, 𝛿 can be set to the worst case
IDFT at the transmitter the output is serialized, a cyclic prefix expected while, for RFO, 𝛿 depends on the performance of the
extension of length 𝑁CP is added and the signal is sent to the employed IFO estimation algorithm (since RFO is the residual
4786 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 9, SEPTEMBER 2009

of this initial estimation step). The discrete-time correlation IV. P ROPOSED E STIMATOR
matrices of 𝝐 and 𝝐𝑒 are Standard OFDM systems consider only the CPR part for
estimation and correction. The LS-CPR estimator proposed in
(𝑹𝝐 )𝑖,𝑘 = (2𝜋/𝑁 )2 𝑖𝑘𝜎𝜀2 (6) [10] and [19] is a robust solution with respect to the achievable
BER (for moderate phase perturbations) and is given by
and { ( )
sin((𝑖−𝑘) 2𝜋
𝑁 𝛿)
𝑐ˆ(0) = ℒ𝒮 𝒚 𝑓 ∣𝒈 𝑓 , 𝐷𝑓
, 𝑖 ∕= 𝑘 /
(𝑹 )𝑖,𝑘
𝝐𝑒 (𝑖−𝑘) 2𝜋
𝑁 𝛿 (7) ∑ ( 𝑓 )∗ 𝑓 ∑ 𝑓 2
1 ,𝑖 = 𝑘 ≜ 𝑔 𝑘 𝑦𝑘 𝑔𝑘 (13)
𝑘∈𝐷𝑓 𝑘∈𝐷𝑓
respectively. Eq. (6) applies to all zero-mean FO models with
variance 𝜎𝜀2 . The ℒ𝒮 () operator defined in (13) is a function of an obser-
vation vector (in this case, 𝒚 𝑓 ), further conditioned on another
parameter vector (in this case, 𝒈 𝑓 ) plus a set of indices (in this
III. A LTERNATIVE S YSTEM M ODEL R EPRESENTATION case, 𝐷𝑓 ). As it is well known, this LS solution also satisfies
a number of other criteria (e.g., maximum likelihood, etc.) for
The purpose of this Section is the proposition of a proper classic linear parameter estimation under white noise (which
description model for the received sample sequence, different is the case in [10] under appropriate, valid approximations).
from (1), which allows for low-complexity LS-based estima- The set 𝐷𝑓 = {𝑖1 , 𝑖2 , ..., 𝑖𝑠 } represents either pilot-bearing
tion of the TPP vector. Precisely, 𝑁 different instantiations sub-carrier indices (in a pilot-based estimation mode) or data-
of a time-domain model are applied to the data record. Each bearing sub-carrier indices (in a decision-directed mode) or a
instantiation targets a specific (say, the 𝑙th ) element of this TPP combination of both. As mentioned, however, in environments
vector, which is viewed as a CPR to all received samples (but with strong phase perturbations, an estimation and compensa-
dependent on 𝑙). Thus, targeting 𝑝𝑙 , 𝒑 can be rewritten as tion procedure addressing the TPP vector is necessary. Using
the statistics derived in Section II, the time-domain LMMSE
𝒑 = 𝑝𝑙 1 + 𝜽(𝑙) + 𝝐(𝑙) (8) estimator can be shown to be [20],
( )−1
where 𝑝𝑙 = 𝜃𝑙 + 𝜗𝑐𝑝 + 𝜖𝑙 as defined in (2), 𝝐(𝑙) ≜ 𝝐 − 𝜖𝑙 1 𝒑ˆ𝑒 = 𝑹𝒑𝑒 𝑮𝐻 𝑮𝑹𝒑𝑒 𝑮𝐻 + 𝜎𝑛2 𝑰 𝒚 (14)
and 𝜽(𝑙) ≜ 𝜽 − 𝜃𝑙 1. Using (8), together with the small-angle with 𝑮 = diag{𝒈} and 𝑹𝒑𝑒 = 𝑹𝜽 𝑒 ⊙ 𝑹𝝐𝑒 since PHN and
approximation 𝑒𝑗𝜙 ≈ 1 + 𝑗𝜙 for 𝜽(𝑙) + 𝝐(𝑙), (1) becomes FO can be safely assumed independent. The computational
burden of this solution is related to solving an order-𝑁
𝒚 = 𝑝𝑒𝑙 𝒈 + 𝜽 n (𝑙) + 𝝐n (𝑙) + 𝒏 (9) system of equations, which can be prohibitive. Decision-
directed schemes employing the LMMSE criterion for PHN
where 𝜽 n (𝑙) ≜ 𝑗𝑝𝑒𝑙 𝒈 ⊙ 𝜽(𝑙) and 𝝐n (𝑙) ≜ 𝑗𝑝𝑒𝑙 𝒈 ⊙ 𝝐(𝑙). The only estimation, were proposed in [13] and [14]. In [13], the
superscript ’n’ denotes ”noise”, since these terms are treated frequency-domain equivalent of (14) has been proposed, where
as such in the estimation process. The proposed model is the PHN Fourier coefficients (i.e., the Fourier transform of 𝜽 𝑒 )
parameterized by 𝑙 and is linear in 𝑝𝑒𝑙 . Naturally, the accuracy were estimated.
of the model decreases for samples far away from 𝑙 due to The proposed LS estimator below (see (15)) attempts to
the adoption of the small-angle approximation but, as will capture the best elements of the two previous estimators,
be shown later, only samples close to 𝑙 will be used by the i.e. the low-complexity aspect of the CPR estimator of (13)
proposed estimator at SNR’s of interest. The statistics of 𝝐(𝑙) yet applied to the richer problem of TPP estimation as per
and 𝜽(𝑙) will now be derived, since they are necessary for the (14). To proceed, we first note the duality existing between
optimization of the proposed estimator. Using (6), it can be the frequency-domain model in (3) and the time-domain
shown that 𝝐(𝑙) is zero-mean with alternative model we propose in (9). We emphasize that this
( ) 2
duality pertains to each component of the TPP vector, and that
𝑹𝝐(𝑙) 𝑖,𝑘 = (2𝜋/𝑁 ) (𝑙 − 𝑖)(𝑙 − 𝑘)𝜎𝜀2 (10) the exact same modeling approximating arguments as in the
derivation of (13) can be employed. Consequently, a window-
Furthermore, the statistics of 𝜽(𝑙) are determined by the parameterized LS estimator is adopted and optimized. Let
employed PHN model: for a Wiener-type PHN, it contains 𝐷(𝑙) ≜ {𝑙−𝐿- (𝑙), ..., 𝑙+𝐿+ (𝑙)} be the set of (𝐿- (𝑙)+𝐿+ (𝑙)+1)
samples
( ) different Wiener processes, resulting in 𝜽(𝑙) ∼
of two time-contiguous sample indices around 𝑙, namely a window
𝒩 0, 𝑹𝜽(𝑙) with within which the estimation of the 𝑙th phase element will
⎧ 2 be confined. In accordance with the above arguments, the
( ) ⎨ 𝜎𝜑 min{𝑙 − 𝑖, 𝑙 − 𝑘}, 𝑖, 𝑘 < 𝑙 proposed time-domain WLS estimator for 𝑝𝑒𝑙 is chosen as
𝑹𝜽(𝑙) 𝑖,𝑘 = 𝜎 2 min{𝑖 − 𝑙, 𝑘 − 𝑙}, 𝑖, 𝑘 > 𝑙 (11)
⎩ 𝜑 𝑝ˆ𝑒𝑙 ≜ ℒ𝒮 (𝒚∣𝒈, 𝐷(𝑙)) (15)
0 otherwise
( ) It is parameterized by 𝐷(𝑙), to be optimized later. It can be
For the Gaussian PHN model, again 𝜽(𝑙) ∼ 𝒩 0, 𝑹𝜽(𝑙) with proved, using the statistics derived in Appendices B-C, that
( ) 𝑝𝑒𝑙 ] = 𝑝𝑒𝑙 , i.e., it is an unbiased
its conditional mean E𝑝ˆ𝑒𝑙 ∣𝑝𝑒𝑙 [ˆ
R𝜃(𝑙) 𝑖,𝑘 = 𝑅 (0) + 𝑅 (∣𝑖 − 𝑘∣ 𝑇𝑠 ) estimator with a corresponding estimation noise R.V. given by
− 𝑅 (∣𝑙 − 𝑖∣ 𝑇𝑠 ) − 𝑅 (∣𝑙 − 𝑘∣ 𝑇𝑠 ) (12) 𝑛total = ℒ𝒮 (𝜽 n (𝑙) + 𝝐n (𝑙) + 𝒏∣𝒈, 𝐷(𝑙)) (16)
DAGRES and POLYDOROS: DECISION-DIRECTED LEAST-SQUARES PHASE PERTURBATION COMPENSATION IN OFDM SYSTEMS 4787

We repeat that the estimator in (15) is of the same operator place on a frame-by-frame basis. For the ‘per-symbol’ case,
form as in (13), but the estimator now pertains to the new time- (21) can be computed as (see Appendix B)
domain model of (9) for each time sample 𝑙. The parameters ∑ ∑ ( ) 2 2
𝐿- (𝑙), 𝐿+ (𝑙) will define a feasible window 𝐷(𝑙) if they satisfy 𝜎𝑛4 /4 + R𝜀(𝑙)+𝜃(𝑙) 𝑘,𝑖 ∣𝑔𝑘 ∣ ∣𝑔𝑖 ∣
𝑘∈𝐷(𝑙) 𝑖∈𝐷(𝑙)
the following boundary conditions: Var [𝑟𝑙 ] = ( )2
∑ 2
0 ⩽ 𝐿- (𝑙) < 𝑙 and 0 ⩽ 𝐿+ (𝑙) ⩽ 𝑁 − 𝑙 (17) ∣𝑔𝑘 ∣
𝑘∈𝐷(𝑙)
It is interesting to note that, for 𝐿- (𝑙) = 𝑙 − 1 and 𝐿+ (𝑙) = ( ) ( ) ( ) (23)
𝑁 − 𝑙, 𝑝ˆ𝑒𝑙 = 𝑐ˆ(0) for every 𝑙 (𝑙 = 1...𝑁 ), namely the TPP where 𝑹𝝐(𝑙)+𝜽(𝑙) 𝑘,𝑖 = 𝑹𝝐(𝑙) 𝑘,𝑖 + 𝑹𝜽(𝑙) 𝑘,𝑖 . To solve (22)
estimator reduces to the classic CPR estimator. For 𝐿- (𝑙) = using (23) for a given 𝒈 (and for every 𝑙) requires integer-
𝐿+ (𝑙) = 0, this TPP estimator reduces to the ML estimator programming techniques (applied for every OFDM symbol
proposed in [13]. Furthermore, the above expressions assume separately). It can be expected to result in better performance
a known 𝒈 (or estimated sufficiently accurately) either in comparing to the ‘per-frame’ scenario since the estimated
CES or DDS; details about the practical estimation of 𝒈 in a value of 𝒈 is exploited precisely, but it suffers from prohibitive
decision-directed manner will be given in Section VI. Because computational complexity. We thus focus instead on the ‘per-
of the exponentiated form of the TPP, only the real angle frame’ scenario where the computation of (21) is based on
𝑝ˆ𝑙 = arg (ˆ𝑝𝑒𝑙 ) of the complex R.V. 𝑝ˆ𝑒𝑙 is of interest in the approximate statistics of 𝒈. In the following sub-Section A,
phase-compensation process. Each time-domain sample is de- closed-form expressions for all the three terms of (21) are
rotated via a multiplication by exp (−𝑗 𝑝ˆ𝑙 ). Thus, the residual provided based on the approximate statistics of 𝒈. In order to
phase error for sample 𝑙 after this phase compensation is avoid a numerical procedure for (22), a closed-form solution
for a near-optimal window is derived in sub-Section B. In sub-
𝑟𝑙 = arg (𝑝𝑒𝑙 exp (−𝑗 𝑝ˆ𝑙 )) = arg (𝑝𝑒𝑙 (ˆ
𝑝𝑒𝑙 )∗ ) (18)
Section C, the accuracy of the analysis, the quality of the ICI
In Appendix A, the post-compensation (residual) total ICI approximation and of the proposed solutions are assessed by
power level is approximated by the power of the average simulations in frequency flat and frequency-selective channels.
residual phase error. Because the estimator is unbiased this
power level equals the A-PCEV, i.e., A. Analytic derivation of PCEV
𝑁
2 1 ∑ Initially, closed-form expressions for the three terms of (21)
𝜎
˜ICI ≈ Var [𝑟𝑙 ] (19)
𝑁 are derived. The following results are expressed as functions
𝑙=1
of 𝐿- , 𝐿+ (we drop the index 𝑙 for convenience), due to
We will refer to (19) as the A-PCEV criterion. It will be the symmetry of the statistics around 𝑙; however, 𝑙 must be
employed to optimize (15) in order to minimize the residual accounted for since it defines the boundary conditions (17) for
ICI, since minimal ICI (when modeled as noise) means 𝐿- (𝑙), 𝐿+ (𝑙). We will refer to the unconstrained PCEV
optimal performance in both CES and DDS regimes. ( when )
the (17) is not taken into account. Modeling 𝒈 as 𝒩ℂ 0, 𝜎𝑔2 𝑰 ,
accurate for frequency flat channel, then (see Appendix C)
V. O PTIMUM W INDOW
[ ] 1
We now address the optimal window under this A-PCEV Var 𝑛LS
𝑙 = (24)
2𝛾𝑠 𝐿1,1
-,+
criterion. A small-angle approximation for 𝑟𝑙 in (18) under the
proposed estimator, when 𝑝ˆ𝑒𝑙 ≈ 𝑝𝑒𝑙 (∣ˆ
𝑝𝑒𝑙 ∣ ≈ 1 ), is and
{ } [ ]
𝑟𝑙 ≈ Im 𝑝𝑒𝑙 (ˆ𝑝𝑒𝑙 )∗ (20) Var 𝜖LS
𝑙 =
( )2
Using (15),(16) and (20), the post-compensation error variance ( )2 4𝐿3,3 + 6𝐿 2,2
+ 2𝐿 1,1
+ 3 𝐿 2,1
− 𝐿 2,1
2𝜋 𝜎𝜀2 -,+ -,+ -,+ +,+ -,- (25)
(PCEV) of 𝑟𝑙 is, ( )( )
𝑁 12 1,1 1,1
𝐿-,+ + 1 𝐿-,+ + 2
[ ] [ ] [ ]
Var [𝑟𝑙 ] ≈ Var 𝑛LS
𝑙 + Var 𝜖LS
𝑙 + Var 𝜃𝑙LS (21)
with 𝐿𝑝-,+1 ,𝑝2 ≜ (𝐿- )𝑝1 + (𝐿+ )𝑝2 . Then 𝐿𝑝+,+
1 ,𝑝2
and 𝐿𝑝-,-1 ,𝑝2 are
where 𝑛LS LS LS
𝑙 , 𝜖𝑙 and 𝜃𝑙 are R.V’s resulting from the imaginary LS
defined similarly. As for 𝜃𝑙 , the answer depends on the
part of the ℒ𝒮 (⋅∣𝒈, 𝐷(𝑙)) operator on 𝒏, 𝑗𝒈 ⊙ 𝝐(𝑙) and 𝑗𝒈 ⊙
model,
𝜽(𝑙), respectively. To find the optimal window (for each 𝑙) we
seek to solve [ ] 𝜎𝜑2 𝐿3,3 + 3𝐿2,2 + 2𝐿1,1
{ } Var 𝜃𝑙LS = ( -,+ )-,+
( -,+
) (26)
𝑁
1 ∑ 3 𝐿1,1 + 1 𝐿1,1 + 2
-,+ -,+
(𝐿- (𝑙), 𝐿+ (𝑙)) = arg min Var [𝑟𝑙 ] (22)
𝐿- (𝑙)∈{0,...𝑙−1} 𝑁
𝐿+ (𝑙)∈{0,...,𝑁 −𝑙}
𝑙=1 for Wiener, and
To solve (22), closed-form expressions for (21) must be [ ]
derived. The computation of (21) depends on the prevailing Var 𝜃𝑙LS = 𝜌0 +
scenario, since 𝒈 can either be assumed known (estimated ∑
𝐿+ ∑
𝐿+ ∑
𝐿+
sufficiently) when the computation takes place on a symbol- 𝜌0 − 2 𝜌∣𝑖∣ 𝜌∣𝑘−𝑖∣ (27)
𝑖=−𝐿- 𝑖=−𝐿- ,𝑖∕=0, 𝑘=−𝐿- ,𝑘∕=0
by-symbol basis (that is, 𝒈 is channel- and data-dependent), or ( ) + ( )( )
𝒈 can be modeled as stochastic when the computation takes 𝐿1,1
-,+ + 1 𝐿1,1
-,+ + 1 𝐿1,1
-,+ + 2
4788 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 9, SEPTEMBER 2009

−2

for Gaussian with 𝜌𝑖 = 𝑅 (∣𝑖∣ 𝑇𝑠 ). Other cases can be

10
δ=0.1
computed via Appendix C by means of the corresponding
SNR=15dB
statistics. We note that (25) holds not only for uniform but
also for any arbitrary zero-mean FO.
σ2φ=0.001

B. Closed-form approximate solution of (22)

ICI power
Having derived all the required terms, the optimal window −3
10

can, in principle, be computed numerically [21]. In order

to further reduce complexity, a closed-form solution will be
derived for solving (22) following a two step procedure: First, σ2=0.0001
φ

a near-optimum window that minimizes the unconstrained A−PCEV analysis

PCEV will be computed. This window will be used for all A−PCEV simulation (AWGN)
A−PCEV simulation (Rayleigh)
those 𝑙 (centered appropriately around each 𝑙) which satisfy SNR=25dB A−PCEV lower bound analysis
2
σICI simulation (Rayleigh)
the boundary conditions (17), whereas for those ‘edge’ values −4
10
of 𝑙 not satisfying (17) a heuristic approach will be adopted. As 0 5 10 15 20
Symmetric Window Length (L)
25 30

shown in Appendix D, the unconstrained PCEV is minimized

by a symmetric window parameterized by 𝐿 if the search is Fig. 1. ICI and A-PCEV performance of the WLS estimator for Wiener PHN
with 𝜎𝜑2 = 10−3 and 10−4 ; SNR=15dB and 25dB; 𝛿 = 0.1, as a function
restricted to odd-length windows (of total length 2𝐿 + 1). So,
of the symmetric window length.
a near-optimum window can be found (‘near optimum’ since
we are not considering even-sized windows) in the solution
space of symmetric windows. This near-optimum[ window ] can
LS C. A-PCEV simulations
be easily
[ computed
] in closed form since Var 𝜖 𝑙 (𝐿) and
Var 𝜃𝑙LS (𝐿) (see Appendix D for related definitions) can be Simulations regarding the A-PCEV under the assumptions
approximated by linear expressions. Eq. (25) can be written of perfectly known channel and data where conducted and are
as shown in Figs. 1-3. The simulations confirm: a) the accuracy
[ ] 2 2 of the analytic derivations of Section A above for the PCEV;
Var 𝜖LS
𝑙 (𝐿) ≈ (2𝜋/𝑁 ) 𝜎𝜀 (𝑎𝜖 𝐿 + 𝑐𝜖 ) (28)
b) the quality of the approximation of (19) as a function of
while (26) and (27) as the residual ICI insofar as it leads to a good-quality optimal-
[ ] window determination; c) the performance loss due to the
Var 𝜃𝑙LS (𝐿) ≈ 𝜎𝜑2 (𝑎𝜃 𝐿 + 𝑐𝜃 ) (29) adoption of a symmetric window (curtailed at the edges) when
compared to the fully optimal window; d) the accuracy of
In (28) and (29), linear approximations involving 𝐿 have been
the practical solution using (30),(31) when compared to the
employed, with 𝑎𝜖 , 𝑐𝜖 , 𝑎𝜃 , 𝑐𝜃 , appropriately chosen parameters.
optimal symmetric window that would result from solving (22)
For Wiener and zero-mean FO we choose 𝑎𝜃 = 𝑎𝜖 = 1/6
numerically with the aid of (24)-(27).
(constant terms are ignored in the minimization procedure).
The simulation parameters are: 128 sub-carriers, 16-QAM
For the first-order Gaussian model, a linear approximation can
constellation and 𝑁CP = 10 samples. The frequency-selective
be used, depending on the value of 𝑎. For 𝑎 > 0.9 we choose
channel corresponds to a Rayleigh channel impulse response
𝑎𝜃 = 0.17. With the help of the above linear approximations
with exponential decay power delay profile of length 𝑁CP .
and temporarily treating 𝐿 as a continuous-valued parameter, 2
In Fig. 1, the residual ICI power, 𝜎 ˜ICI , and its dominant
the unconstrained PCEV confined to symmetrical windows, is
term, A-PCEV, are plotted for the proposed estimator as a
minimized when ˜ of (31). The A-
function of the symmetric window length 𝐿
1 PCEV is both plotted analytically (using Section A) as well as
𝐿𝑐 = √ ( ) (30)
simulated for the AWGN and Rayleigh channels. The lower
2 𝛾𝑠 𝑎𝜃 𝜎𝜑2 + 𝑎𝜖 (2𝜋/𝑁 )2 𝜎𝜀2
bound on A-PCEV, resulting from the numerical solution
of (22) is also shown. The parameter values used for the
An approximate solution for the near-optimal, symmetric
˜ simulations are shown in the respective legends. The results
window for the unconstrained version of (22) is 𝐿- = 𝐿+ = 𝐿
˜ 𝑐 confirm the validity of the theoretical analysis for the A-PCEV
where 𝐿 is the integer closest to 𝐿 and greater or equal to
in both AWGN and Rayleigh environments. Precisely, the
one. Thus, the proposed near-optimal solution for (21) is
results for AWGN are in agreement with the analysis, while
˜ 𝑙 − 1} and 𝐿+ (𝑙) = min{𝐿,
𝐿- (𝑙) = min{𝐿, ˜ 𝑁 − 𝑙} (31) a very small deviation is noticed for the Rayleigh case, as
expected. The fact that A-PCEV is the dominant term of the
For those 𝑙 where 𝐿 ˜ does not satisfy (17), the heuristic ICI (and thus qualifies the use of (19)) is also confirmed since
approach is to ‘clip’ the window at the boundary values the ICI follows the same shape as A-PCVE, thus justifying
(namely 1 and 𝑁 ). In conclusion, the proposed estimator is its use in the determination of the optimal-window length.
given by (15), using (31) for the parameters of the TPP- Finally, the results also confirm that near-optimal performance
related window, while compensation is performed in the (in comparison to the optimal-window choice) can be achieved
classic way: by multiplying each sample in the time domain via the symmetric-window approach: this symmetric window
with exp (−𝑗 𝑝ˆ𝑙 ). exhibits negligible performance degradation versus the optimal
DAGRES and POLYDOROS: DECISION-DIRECTED LEAST-SQUARES PHASE PERTURBATION COMPENSATION IN OFDM SYSTEMS 4789

SNR=20dB, δ=0 is either exactly equal to the one computed by the A-PCEV
metric or is equal to one sample less, which means that a
σ2=0.01
φ
near-optimal window value which minimizes the ICI power is
A−PCEV analysis
A−PCEV simulation (AWGN) practically guaranteed.
−2 A−PCEV simulation (Rayleigh)
10
A−PCEV lower bound analysis
σ2ICI simulation (Rayleigh)
VI. U NKNOWN C HANNEL AND DATA
ICI power

σ2=0.001
φ
In the derivation of the proposed estimator, the exact value
of 𝒈 was assumed to be known whereas in practice it has to
be estimated. In a decision-directed scheme, 𝒈 is estimated
−3
10
iteratively, where at each step the previously estimated vector
σ2φ=0.0001
is used for phase-compensation and then fresh data/channel
decisions are taken at the subsequent step and used, in turn,
in order to update the estimate. At the 𝑖th iteration, the phase-
5 10 15
Symmetric Window Length (L)
20 25 30
compensated samples in the frequency domain are
( ( ) )
Fig. 2. ICI and A-PCEV performance of WLS estimator for Gaussian PHN 𝒚 𝑓 (𝑖) = 𝑭 diag{exp −𝑗 𝒑ˆ(𝑖) } 𝒚 (𝑖) (32)
with 𝑎 = 0.9, 𝜎𝜑 2 = 10−2 , 10−3 and 10−4 , SNR=20dB, and 𝛿 = 0, as a
function of the symmetric window length.
where 𝒑ˆ(𝑖) is the argument of the estimated exponentiated
TPP vector at the 𝑖th iteration (other parameters are defined
20
19 A−PCEV Analysis
accordingly).
18 A−PCEV Simulation (Rayleigh) In CES, a known OFDM symbol is transmitted (a ‘pream-
A−PCEV Approximation (30)
17
16
ICI Simulation (Rayleigh) ble’). Let 𝒅 be the preamble symbol of unit modulus and 𝒉 the
15 channel impulse-response vector of length 𝑁CP . The LS chan-
14
δ=0.1 nel estimator, assuming perfect phase recovery (equivalent to
13
the ML estimator derived in [15]) is
SNR=15dB
Optimized L

12
11
10
9
SNR=20dB
ˆ (𝑖) = √1 𝑾 𝐻 𝑫 𝐻 𝒚 𝑓 (𝑖)
𝒉 (33)
8 𝑁
7
6
SNR=30dB
where 𝑫 = diag{𝒅} and 𝑾 is an 𝑁 × 𝑁CP matrix containing
5
4
the first 𝑁CP columns of 𝑭 such that 𝑯 = diag {𝑾 𝒉}. With
3 𝒑ˆ(0) = 0, it is clear that the estimated channel incorporates the
2
1
CPR of (3) since it was not compensated at the first step. This
0
−4 −3
irresolvable rotation would still exist even if the phase was
estimated first, as shown in [15] where an analytic discussion
10 10
PHN variance (Wiener)
can be found on this matter. It does not cause any performance
Fig. 3. Optimized symmetric window length as a function of the PHN degradation since it is removed by the equalization step. In
variance via analysis and simulation.
Section IX, it is taken into account when simulating the
channel-estimation MSE performance. After the last iteration
the estimated TPP is used for IFO estimation. Since the CPR
(the latter expressed by the A-PCEV lower bound). The reason
is incorporated into the channel model, the estimated phase is
for this slight degradation, which is mainly observed when
the optimal 𝐿 is large and the number of edge (clipped) 𝒑ˆ = arg (𝒑ˆ𝑒 ) = 𝝐¯ + 𝜽¯ + 𝒑ˆerror (34)
samples is also large, is basically the FO term; the last term
of the nominator in (25) increases due to the asymmetry of where 𝒑ˆ𝑒𝑟𝑟𝑜𝑟 is the estimation error and 𝜽, ¯ 𝝐¯ are the resulting
the clipped windows. In Fig. 2, similar scenarios as in Fig. 1 phase and FO vectors after the removal of their samle mean
are plotted, but now for Gaussian PHN with 𝑎 = 0.9 and a values. We caution here that the implementation of the argu-
fixed SNR of 20dB, with no FO. The simulation results again ment operator of (34) must take into account possible absolute
confirm the analysis, while (31) provides answers very close phase jumps greater than 𝜋. In that case, the phases must be
to the optimal solution even for large 𝐿 (as expected due to the unwrapped to their 2𝜋 complement. Using the LS fit approach,
absence of FO). In Fig. 3, the optimized symmetric window proposed in [18] for systems without PHN, the IFO estimate
length is shown as a function of PHN via simulation plus is given by
analysis. Three SNR values are chosen, with the FO fixed at 𝜀ˆ = ℒ𝒮 (𝒑∣¯
ˆ 𝒗 , {1, ..., 𝑁 }) (35)
0.1. The analysis of Section A is always in agreement with the
A-PCEV simulations. The continuous-valued version of sub- where 𝑣¯𝑖 = 𝑣𝑖 − (𝜋 + 1/𝑁 ). The RFO level is determined by
optimal closed-form solution of (30) is also shown. It is close the estimation error variance of (35). An approximation can
to the analytic numerical solution for small window lengths, be computed when 𝜽¯ >> 𝒑ˆerror ,
while it deviates slightly for larger window lengths (again, due / 𝑇
to the FO). The optimized length based on the simulated ICI RFO ≈ 𝒗 ¯𝑇 𝑹𝜽¯𝒗
¯ 𝒗¯ 𝒗¯ (36)
4790 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 9, SEPTEMBER 2009

The statistics of 𝜽¯ can be shown after some algebra to be been studied in [14] and [19]. All such approaches model the
𝑁 ( ) ICI as an additional white noise term with power determined
1 ∑
(𝑹𝜽¯)𝑖,𝑘 = (𝑹𝜽 )𝑖,𝑘 + (𝑹𝜽 )𝑖,𝑡 + (𝑹𝜽 )𝑘,𝑡 by the statistics of the PHN. Generalizing those results, by
𝑁
𝑡=1 including the FO perturbation, it is straightforward to show
𝑁 𝑁
(37) that
1 ∑∑ 2
+ 2 (𝑹𝜽 )𝑠,𝑡 𝜎ICI = 1 − 1𝑇 𝑹𝒑𝑒 1/𝑁 2 (40)
𝑁 𝑠=1 𝑡=1 /( 2 )
2 2 2
with the SINR is given by 𝛾ICI ≈ (1 − 𝜎ICI )𝜎𝒈 𝜎𝒏 + 𝜎ICI .
For Wiener PHN, a closed-form expression can be shown Thus, by knowing the TPP statistics, one can approach the
to be calculation of system performance as pertaining to a system
( )
𝑁 1 1 without phase errors but with a lower operating SNR. In other
(𝑹𝜽¯)𝑖,𝑘 = min {𝑖, 𝑘} + + + 𝜎𝜑2
3 6𝑁 2 words, a white interfering signal can effectively model the
(38) performance degradation due to ICI. Following the diagonal
𝑖(2𝑁 − 𝑖 + 1) + 𝑘(2𝑁 − 𝑘 + 1)
− loading approach, 𝐷(𝑙) can then be chosen based on the
2𝑁
SINR, as opposed to the SNR. In most cases, a conservative
This approximation is very accurate in the high SNR regions,
upper bound on the extra noise level is adequate. In CES,
as will be shown via simulations in Section IX. Therefore, it
only a small portion of the ICI affects the estimates due
will also be used as a proxy for the RFO variance in DDS for
to the small dimensionality of the channel; thus, marginal
determining the optimal window. In cases where the original
performance degradation is expected due to model mismatch.
IFO variance is smaller than (36), it is better to simply use a
In order to improve the convergence of the estimator we may
zero estimate for the IFO.
choose the SNR of each intermediate iterative step so as to
In DDS, the channel is assumed known (estimated suffi-
somehow ‘mirror’ the decrease in model mismatch due to
ciently at CES), while initial data decisions are made assuming
better successive data decisions or estimated CSI. This can be
𝒑ˆ(0) = 𝑐ˆ(0)1, i.e., data detection with CPR removal only as
done by using the actual SINR at the first iteration, whereas at
performed in [10]. IFO estimated at CES is assumed already
the very last step only the white noise is taken into account,
removed. In subsequent steps, hard decisions are made using
namely the system SNR. The intermediate SNR values are
the LS estimates (LS equalization) of the data:
heuristically computed via a linear interpolation between the
(𝑖)
𝑠ˆ𝑙 = ℒ𝒮(ŷ𝑓 (𝑖) ∣ĥf , {𝑙}) (39) two extreme values in logarithmic scale. The algorithmic steps
of the WLS-based OFDM scheme are summarized in Table I,
In both DDS and CES, the major step that facilitates initial- where 𝑁iter denotes the number of iterations. The estimated-
ization and, therefore, convergence, is the effective annulment IFO removal step is not included since it is common to all
of the CPR. In the DDS mode this happens explicitly since schemes.
CPR is pre-compensated, whereas in CES this happens via the
incorporation of the CPR into the composite channel model. VIII. C OMPUTATIONAL C OMPLEXITY
In DDS, the computational difference between the WLS and
VII. I MPROVING ROBUSTNESS
the LMMSE is due to the different expressions in estimating
The estimator developed so far assumed perfect channel the exponentiated TPP vector in (15) and (14), respectively.
state information (CSI) and data decisions. This effectively For the WLS, since closed-form expressions for computing
amounts to a model mismatch because the statistics of the the symmetric window length exist, and the nominator of the
received observation are actually computed based on occa- LS is sufficient for the determination of its argument, what
sionally erroneous CSI or data decisions, something that will needs to be computed is (see Table I, Step B, line 6)
adversely affect performance in the high-SNR region where
(𝑖)
∑ ( (𝑖−1) )∗
decision errors due to ICI tend to dominate over thermal-noise 𝑆𝑙 = 𝑔ˆ𝑘 𝑦𝑘 , 𝑙 = 1..𝑁 (41)
effects. The robustness of general estimators under model mis- 𝑘∈𝐷(𝑖) (𝑙)
match has been studied in [17]. When interference is present where the upper index 𝑖 denotes the iteration number. Let
but is not accounted for in the design process, an estimator can 𝐿o be the computed symmetric window length. If we define
be made more robust when its parameters are selected based ( (𝑖−1) )∗
𝒒 (𝑖) = 𝒈 ˆ ⊙ 𝒚 (requiring 𝑁 complex multiplications)
on a hypothesized lower SNR value than the true prevailing (𝑖) ∑𝐿o 𝑖
and 𝑆1 = 𝑘=1 𝑞𝑘 (requiring 𝐿o complex additions), each
SNR. This is called diagonal loading because it is equivalent (𝑖)
to increasing the diagonal elements of the received sample term 𝑆𝑙 can be computed recursively:
covariance matrix by inserting a higher-level white noise. This 𝑆𝑙
(𝑖) (𝑖) (𝑖) (𝑖)
= 𝑆𝑙−1 − 𝑞𝑙−1 + 𝑞𝑙+1 , 𝑙 = 2...𝑁 (42)
needed lower SNR was chosen heuristically in [17], but is
always proportional to the power of the interfering signal with 𝑞𝑘𝑖 = 0 for 1 > 𝑘 > 𝑁 . Thus, 𝑁 − 𝐿o non-zero
therefore, interpreted as a signal to interference plus noise ratio elements of 𝒒 (𝑖) are added and 𝑁 − 𝐿o non-zero elements
(SINR). In adopting this approach for the proposed scheme, of 𝒔𝑖 are subtracted in the total recursion loop, thus leading
the only additional parameter to be calculated is the extra to total 𝑁 complex multiplications, 𝑁 complex additions and
white noise level that shall be injected in the calculation of the 𝑁 − 𝐿o complex subtractions for each iteration, which mean
optimal window size. The degradation in OFDM performance a total complexity of O(𝑁 ). In comparison, the LMMSE
in DDS under PHN-induced ICI with CPR removal only, has requires the solution of an order-𝑁 system of equations
DAGRES and POLYDOROS: DECISION-DIRECTED LEAST-SQUARES PHASE PERTURBATION COMPENSATION IN OFDM SYSTEMS 4791

TABLE I 4

WLS BASED OFDM SCHEME

CPE LS
3.5 ICM 1it
ICM 2it
Step A) Initialization [CES/DES] WLS 1it
3 WLS 2it
1: Compute 𝛾ICI from PHN and [CFO/RFO Bound (36)] WLS 4it
WLS 8it
2: for 𝑖 = 1 to 𝑁iter

SNR degradation
2.5 LMMSE 1it

𝛾 (𝑖) = 𝛾ICI + (𝑖 − 1)(𝛾sys − 𝛾ICI )/(𝑁iter − 1)

LMMSE 2it
3: LMMSE 4it

Compute 𝐷 (𝑖) (𝑙) from (30),(31) using 𝛾 (𝑖)

LMMSE 8it
4: 2

5: end
1.5

Step B) For every [preamble/symbol] , i.e. [CES/DES]

[ / ] 1
1: Compute 𝒚 𝑓 (0) (32) assuming 𝒑 ˆ(0) = 0 𝒑 ˆ(0) = 𝑐ˆ(0)1
[ / ]
2: Compute 𝒉 ˆ (0) from (33) 𝒔(0) from 𝒔 ˆ(0) , (39) 0.5

3: for 𝑖 = 1 : 𝑁iter
[ / ]
−4 −3 −2
10 10 10

4: Compute 𝒈
ˆ (𝑖−1) ˆ (𝑖−1) and
from 𝒉 𝒅 𝒔(𝑖−1) Wiener PHN variance

5: for 𝑙 = 1 : 𝑁 Fig. 4. SNR degradation for attaining BER of 10−2 for various schemes
⎛ ⎞
∑ ( ) (AWGN)
(𝑖) (𝑖−1) ∗
6: 𝑝ˆ𝑙 = arg ⎝ 𝑔ˆ𝑘 𝑦𝑘 ⎠
𝑘∈𝐷 (𝑖) (𝑙)
7: end is order O(𝑁 ), with total complexity O(𝑁iter (log(N)+𝑁CP )𝑁 ).
8: Compute 𝒚 𝑓 (𝑖)
(32) using 𝒑
ˆ (𝑖) In comparison, the modified JCPCE scheme of [15] for
[ / ]
9: ˆ
Compute 𝒉 from (33) 𝒔(𝑖) from 𝒔
(𝑖)
ˆ(𝑖) , (39)
joint channel/FO/PHN estimation requires the solution of an
order-𝑁/2 system for IFO estimation and an order-𝑁 for
10: end
PHN estimation, respectively, therefore a total complexity of
11: Compute 𝜀ˆ (35) from 𝒑ˆ(𝑁iter ) (34) O(𝑁 3 ). Sub-optimal conjugate-gradient approaches have also
[ / ]
12: Output 𝒉ˆ =𝒉 ˆ iter 𝒔 = 𝒔(𝑁iter )
(𝑁 )
been proposed, leading to O(𝑁 log(N)) complexity.

IX. S IMULATION R ESULTS

In this Section, the performance of the proposed scheme
for the same term, therefore O(𝑁 3 ) (in each iteration). All
will be evaluated and compared with previously proposed
other computations such as 𝑁 argument operations, FFT,
schemes. Both flat and frequency-selective channel models
and hard decisions are identical for both algorithms. Thus, (specified in Section V) will be considered in order to as-
since the main computational burden of WLS is the FFT,
sess the robustness of the algorithms under different channel
the total complexity( is of order
) O (𝑁iter 𝑁 log (𝑁 )), whereas
conditions. The algorithms chosen for comparison purposes
for LMMSE it is O 𝑁iter 𝑁 3 . The iterative conditional mode
for the DDS are the ICM of [16] and the LMMSE of [13].
(ICM) algorithm of [16] has the nice property of avoiding
The latter can also be viewed as a special case (full-order esti-
successive data decisions but its complexity is also dictated mation) of the one in [14]. As mentioned, the ICM algorithm
by the same need to solve an order-𝑁 system (for each
performs joint data and PHN detection, whereas [13] derives
soft iteration). An iterative sub-optimal conjugate-gradient
the LMMSE phase estimates via hard decisions on the data.
approach has been proposed therein, which needs approxi- Since both require matrix inversions, iterative modifications
mately 8 iterations (for solving this system of equations inside
have been proposed that converge with a few iterations. In
each soft iteration) in order to converge (for medium PHN).
the simulations below, full-matrix inversion is used for both
Even this suboptimal approach incurs a complexity higher
methods in each iteration so as to achieve their full potential.
than WLS. The algorithm proposed in [14] can indeed trade
For the CES, the modified JCPCE is used for comparison. For
complexity with performance: its computational burden is
initial CPR estimation, the LS pilot-based approach proposed
the solution of an order (2𝐾 + 1) system, where 𝐾 is the
of [10] is used, i.e., (13) with 𝐷𝑓 indicating the pilot indices.
number of the estimated ICI terms in the frequency domain.
The system simulation parameters are the same as in Section
Schemes with complexity lower than the proposed can be V, using 10 pilot sub-carriers to ensure near-optimal CPR
derived but, as will be shown later, the WLS outperforms
estimation.
even the full-order LMMSE (hence, even more so for reduced-
performance schemes). The ML [13] scheme requires the same
complexity as the proposed one (since can be derived as a A. Performance assessment in DDS with perfect CSI (PHN
special case of WLS), but it is inferior in performance by only)
all methods. The computational complexity involved in the In Fig. 4, the SNR degradation (versus a non-TPP system)
CES is not a critical factor as in DDS, since it consumes corresponding to a BER of 10−2 is shown for AWGN, 16-
a small portion of the total system complexity (due to the QAM, Wiener PHN, zero RFO and perfect CSI. The non-TPP
small number of symbols involved). Regarding WLS, the LS SNR is ≈ 13.9 dB. ICM is shown for 2 iterations because
channel estimation is of order O(𝑁CP 𝑁 ) and the FO estimation its performance does not improve with further iterations. The
4792 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 9, SEPTEMBER 2009

6 CPR LS
ICM 2it
WLS 1it
WLS 2it
5 WLS 4it
WLS 8it −3
LMMSE 1it 10
SNR degradation

Channel MSE
4 LMMSE 4it

CRLB
Modified JCPCE
2 −4
WLS 2 iter
10
WLS 3 iter
WLS 4 iter
LS (PHN only)
1

15 20 25 30 35
−4 −3
10 10 SNR (dB)
Gaussian PHN variance (α =0.99)

Fig. 7. Channel MSE estimation performance with PHN and IFO

Fig. 5. SNR degradation for attaining BER of 10−2 for various schemes
(Rayleigh)

−3
modified Moose IFO estimator of [15] is also shown. It is
10
derived by modifying the estimator of [22] so as to take into
account the PHN. It is better in the low-SNR region, but it
σ2=0.001
φ
is quite close to the proposed LS estimator in the high-SNR
2
σφ=0.0001 region. We note, however, that in addition to the computational
10
−4
complexity of an order-(𝑁/2) system, this scheme requires a
RFO variance

very specific pilot structure since half of the sub-carriers of

the preamble should carry pilots (other half must be zero).
σ2φ=0.00001
Finally, we note that since both estimators are not of Bayesian
−5
nature, their performance is dictated by the PHN level in
10
the high-SNR region (regardless the prior IFO statistics). Fig.
7 examines the performance of the proposed scheme in the
Modified Moose
LS approximation
2
σφ=0 CES for 2-4 iterations. The parameters are: Rayleigh channel,
LS
16QAM, Wiener PHN with 𝜎𝜑2 = 10−3 and IFO of 0.5. The
10
−6

10 15 20 25 30 35
CRB plus the performance of the modified JCPCE scheme
SNR dB are also plotted, both derived in [15], which employs the
modified Moose IFO estimator. This scheme exhibits good
Fig. 6. Performance of the LS IFO estimation at different levels of PHN.
performance without requiring iterations but possesses high
complexity (if sup-optimal approaches are not employed). The
performance of WLS and LMMSE is shown for 1,2,4 and 8 proposed WLS scheme achieves nearly the same performance
iterations. We observe that LMMSE performs a bit worse than without matrix inversions or any special preamble scheme.
WLS for the same number of iterations. For SNR degradations The conventional LS channel estimator is also plotted (for
less than 1dB, WLS extends the robustness to PHN by nearly the case of PHN only), in order to show the performance
one order of magnitude in terms of the PHN variance that degradation when PHN is not treated appropriately. As a final
can be handled by the system. For 𝜎𝜑2 < 10−3 , a choice comment, a large value of 𝛿 has been chosen in order to show
of 2 iterations is a good trade-off between performance gain that satisfactory FO estimation as well as channel estimation
and complexity. Fig. 5 repeats the previous comparisons but can be guaranteed (i.e., very close to the respective bounds)
now for a Rayleigh channel with 64-QAM and Gaussian even for this worst case.
PHN (a = 0.99). The non-TPP SNR is ≈ 26.1 dB. WLS
again outperforms LMMSE and ICM. It achieves the same C. Overall system performance assessment
performance as LMMSE with roughly half the iterations. ICM
In Fig. 8, the overall system (namely, the cascade of
again does not improve after two iterations.
CES/DDS) performance is assessed for various algorithmic
combinations and simulation scenarios. In all cases, a Rayleigh
B. Performance assessment in CES channel and 16-QAM are used, with the number of iterations
The performance of the LS IFO estimation is shown in Fig. fixed to 4. The curve corresponding to JCPCE/CPR-LS (𝜎𝜑2 =
6 for a Rayleigh channel, 16QAM, Gaussian PHN (a = 0.96) 10−3 , 𝛿 = 0) is provided in order to assess the performance
and IFO of 𝛿 = 0.5. The approximation of (36) follows quite degradation due to PHN by itself, using the best channel
accurately the performance of the proposed estimator in the estimator. As we see, the ICI induces strong performance
high-SNR region (above 20dB or so). The performance of the degradation for BER close to 10−2 . On the contrary, when
DAGRES and POLYDOROS: DECISION-DIRECTED LEAST-SQUARES PHASE PERTURBATION COMPENSATION IN OFDM SYSTEMS 4793

−1
10
Perfect CSI, zero PHN,IFO
IFO not estimated
IFO estimated (always)
IFO estimated (if δ > (36))
RFO
approximation (36)
SNR=20dB

−2
10
−2

BER
10
BER

2
WLS / WLS (σφ=0.001,δ=0)

WLS / WLS (σ2φ=0.001,δ=0.5)

Perfect CSI (σ2φ=0,δ=0) SNR=30dB

JCPCE / CPR LS σ2=0.001,δ=0)
φ
WLS / CPR LS (σ2φ=0 δ=0.5)

−3
JCPCE / WLS (σ2φ=0.001,δ=0)
10
2
JCPCE / WLS (σφ=0.001,δ=0.5)

JCPCE / DL−LMMSE (σ2φ=0.001,δ=0)

JCPCE / LMMSE (σ2φ=0.001,δ=0)

−3
10
0.01 0.1 0.5
15 20 25 30 35 40
IFO(δ)
SNR (dB)

Fig. 9. BER versus FO for two SNR values with WLS in both CES and
Fig. 8. Total-system performance for various combinations of CES/DDS DDS.
modes.

Perfect SNR,PHN − 1it

only IFO is present, the cascade of WLS/CPR-LS exhibits 4.5
Perfect SNR, PHN − 8it
Perfect PHN only − 1it
near-optimal performance even in high-SNR regions. These 4 Perfect PHN only − 8it

two scenarios are included in order to clarify that, for these SNR degradation for BER=10−2
m~0 (L=128) − 1it
m=0.2 (L ≈ 2Lo) − 1it
specific PHN and FO values, the performance degradation is
3.5
m=0.2 (L ≈ 2Lo) − 8it

mainly due to PHN. We note that there exist many algorithms 3

m>>0 (L=1) − 1it
m=5 (L ≈ 0.5Lo) − 1it
in the literature that can estimate and remove FO quite m=2 (L ≈ 0.7Lo) − 8it
effectively and with lower computational complexity [22], but
2.5

fail in the presence of strong PHN since they are not designed 2

for it. The performance of the cascade WLS/WLS is almost the

1.5
same as JCPCE/WLS when only PHN is present. We observe
a marginal difference in the high-SNR region when FO is 1

added (using the modified Moose JCPCE), which suggests

0.5
the adoption of the proposed method in both estimation stages.
Performance using the JCPCE/LMMSE cascade is also shown 0
−4
10
for PHN only. Here, a marginal difference is observed in the Wiener PHN variance
SNR’s of interest, but the LMMSE fails to converge in the
high-SNR region. The application of diagonal loading to the Fig. 10. SNR degradation for attaining BER of 10−2 under model mismatch
(Rayleigh)
LMMSE (named DL-LMMSE) is also shown. We see that
convergence improves, but the WLS still exhibits superior
performance. In Fig. 9, the setup is the same as in Fig. 8,
restricting it to WLS in both CES and DDS. Also we fix possibly mismatched SNR and SINR parameters, a simple
the SNR to 20dB and 30dB. The BER is shown versus IFO, noise-variance-level estimator is adopted. At each CES iter-
with the IFO ranging from 0.01 to 0.5. As shown, whenever ation, the noise vector 𝒏 ˆ 𝑓 (𝑖) is estimated by subtracting the
𝑓 (𝑖)
the IFO is always estimated, the BER performance remains 𝒈
ˆ from the phase compensated 𝒚 ˆ𝑓 (𝑖) . The power ratio
invariant regardless of the prior IFO statistics. However, when of the estimated channel over the variance of 𝒏 ˆ 𝑓 (𝑖) in each
the IFO is small (below a threshold that we may determine iteration is employed for the determination of both the SINR
via the approximation of (36)), it is actually better not to and SNR. Specifically, the SINR becomes the estimated ratio
estimate the IFO. This is because the ambient PHN affects the value of the first iteration (namely, before the estimated TPP
estimator negatively versus its actual (small) value. However, is removed), while the SNR becomes the ratio value of
things reverse above that threshold value, as shown in Fig. 9, each iteration (i.e., the SNR estimate is updated with each
which means that (36) may be used to decide when to turn iteration). These values are then passed on to the DDS for
the IFO estimator ‘on’. the determination of the symmetric window. In Fig. 10, the
SNR degradation at a BER of 10−2 in Rayleigh, 64-QAM,
Wiener PHN and zero RFO is shown for the proposed scheme
D. Sensitivity to model mismatch with 4 iterations at the CES and either 1 or 8 iterations at the
The performance of the proposed scheme will naturally DDS. As illustrated, the performance with the right parameters
degrade when the parameters related to the optimal window is marginally affected by the adoption of the aforementioned
deviate from the assumed values, namely the values of SNR, SNR and SINR estimators for both iterative and non-iterative
PHN and FO statistics. In order to test robustness to such cases (when we assume perfect knowledge of the PHN statis-
4794 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 9, SEPTEMBER 2009

tics). Now, in order to assess the performance degradation blocks that are always present in OFDM systems (CPR
due to PHN model mismatch, a deliberate uncertainty is estimator, FFT). Finally, the diagonal-loading approach has
added as follows: the modeled Wiener PHN variance (the been combined with iterations in order to enhance convergence
one taken into account at window computation) is equal to of the proposed algorithm. This technique can be successfully
the actual value, multiplied by a mismatch parameter 𝑚. Let applied to other decision-directed algorithms as well.
𝐿𝑐(𝑖) be the continuous-valued window length computed via
(30) by use of the estimated SNR and SINR parameters and A PPENDIX A
assuming the actual PHN variance (𝐿𝑐(𝑖) is near optimal as COMPUTATION OF RESIDUAL ICI APPROXIMATION
shown previously)./Then,
√ the corresponding mismatched value
will equal to 𝐿𝑐(𝑖) 𝑚 (which means that model mismatch After removing the TPP estimate from the received data, the
translates directly to window mismatch). The two extreme residual phase rotation at each sample can be approximated
window lengths that can be computed by (31) would be either as 𝑟𝑙𝑒 ≈ 1 + 𝑗𝑟𝑙 ; thus (3) becomes
1 or 𝑁 , corresponding to very high and low values of 𝑚.
𝒚 ≈ (1 + 𝑐˜(0))𝒈 𝑓 + 𝒏
˜ ICI + 𝒏𝑓 (43)
Using these two extreme window values for the entire range of
PHN represents lower performance bounds due to any model with 𝑛˜ ICI
𝑖 computed as in (4) using 𝑐˜(𝑚); 𝑐˜(𝑚) is given by
mismatch (for the assumed scenario) and are both included (5) replacing 𝑝𝑒𝑘 with 𝑗𝑟𝑘 . Using the same approximation and
in Fig. 10 (for one iteration). For 𝐿 = 𝑁 , the performance derivation steps as in [14] for the determination of the ICI
is the same as with pure CPR removal since the estimator power (E[∣ℎ𝑓𝑘 ∣2 ] = 1, uncorrelated parameters), and since the
now reduces to the LS-CPR. For 𝐿 = 1, the dB loss depends data are normalized in power, the residual ICI power can be
on the PHN level, and ranges from a fraction of a dB to approximated as
2dB. The specific demonstrated values of the parameter 𝑚
𝑁
∑ −1
have been selected so that they result in marginal performance 2 2
𝜎
˜ICI ≈ 𝐸[∣˜
𝑐(𝑚)∣ ]
degradation, and this is done for both the iterative and the non-
𝑚=1
iterative scenarios. We choose 𝑚 = 0.2 in both scenarios for 𝑁 𝑁 𝑁 −1
displaying cases where the PHN variance is under-modeled, 1 ∑∑ ∑
= 𝐸 [𝑟 𝑟
𝑘 𝑙 ] 𝑒𝑗2𝜋(𝑘−𝑙)𝑚/𝑁 (44)
which translates to roughly double the window length. For the 𝑁2 𝑚=1
𝑘=1 𝑙=1
non-iterative scenario, only a marginal shift in error floor is
seen. For the iterative scenario, a marginal loss is seen when since
𝑁
∑ −1 {
the mismatch is low (near-optimal performance), whereas a 𝑁 − 1, 𝑘 = 𝑙
𝑒𝑗2𝜋(𝑘−𝑙)𝑚/𝑁 = (45)
marginal gain appears when the model mismatch increases −1, 𝑘 ∕= 𝑙
𝑚=1
(because it will be chosen larger by the mismatch mistake).
For cases where the PHN variance is over-modeled, the values then
of 𝑚 = 5 and 𝑚 = 2 are selected for the non-iterative and ( 𝑁) 𝑁 𝑁
2 𝑁 −11 ∑ [ 2] 1 ∑ ∑
the iterative scenarios, respectively. For the first scenario, the 𝜎
˜ICI ≈ E 𝑟𝑘 − 2 𝐸[𝑟𝑘 𝑟𝑙 ]
𝑁 𝑁 𝑁
window length is roughly halved performing approximately as 𝑘=1 𝑘=1 𝑙=1,𝑘∕=𝑙
when it is doubled. For the second scenario, the value of 𝑚 (46)
is chosen to be 2 in order to demonstrate near-optimal perfor- The first term of (46) is the dominant one because the
mance (iterations are more sensitive in over-modeling). The correlation fades quickly for distant samples. For large 𝑁 ,
2
performance loss in this case is only a shift of the error floor. 𝜎
˜ICI can be further approximated by
We note that the above mismatched window values (resulting 𝑁
2 1 ∑ [ 2]
from a specific mismatch PHN model parameter), in essence 𝜎
˜ICI ≈ E 𝑟𝑘 (47)
represent any mismatch, either in the parametric sense or the 𝑁
𝑘=1
modeling sense, that map to these window parameter values. i.e. by the average post compensation phase error power.
For example PHN model could be erroneously assumed as When compensation is performed by de-rotating the received
Gaussian, or a non-existing FO could be modeled, leading samples in the time domain using the estimated phase, the
again to window over- or under-determination. residual phase error equals the error of the estimated phase.
X. C ONCLUSIONS
In this paper, an alternative linear (per sample) TPP de- A PPENDIX B
scription model is proposed for OFDM systems, leading COMPUTATION OF REQUIRED MOMENTS
( )
to the development of a low complexity, robust windowed Define the R.V. 𝑧 ≜ 1/𝑥 where 𝑥 ∼ 𝜒2 𝑛𝑥 , 𝜎 2 . The pdf
LS estimator. It outperforms similar algorithms proposed in of 𝑧 is
the literature, on both dimensions of BER performance and ( )𝑛𝑥 /2−1
complexity when applied in DDS, while near-CRB chan- 1 1 − 2𝜎12 𝑧 −1
𝑓𝑧 (𝑧) = 𝑒 (48)
nel estimation is achieved when applied in CES. This was Γ(𝑛𝑥 /2)2𝑛𝑥 /2 𝜎 𝑛𝑥 𝑧
demonstrated through simulations under scenarios of strong therefore
PHN, FO plus challenging channel conditions. An additional ∫∞
1
advantage of the proposed scheme besides the performance E𝑧 [𝑧] = 𝑧𝑓𝑧 (𝑧) = (49)
and complexity gains is that it can be realized via functional 𝜎2 (𝑛𝑥 − 2)
0
DAGRES and POLYDOROS: DECISION-DIRECTED LEAST-SQUARES PHASE PERTURBATION COMPENSATION IN OFDM SYSTEMS 4795

𝑥
Define the R.V. 𝑧 ≜ 𝑥+𝑦 where 𝑥, 𝑦 are statistically then
( ) ( )
independent R.V.’s with 𝑥 ∼ 𝜒2 𝑛𝑥 , 𝜎 2 , 𝑦 ∼ 𝜒2 𝑛𝑦 , 𝜎 2 . [( )2 ] [ [( )2 ]]
The pdf of 𝑧 is [23] E𝑛(:) ,𝑔(:) 𝑛LS
𝑙 = E𝑔(:) E𝑛(:) ∣𝑔(:) 𝑛LS
𝑙 (59)
Γ(𝑛′𝑥 + 𝑛′𝑦 ) 𝑛′ −1 ′
𝑓𝑧 (𝑧) = 𝑧 𝑥 (1 − 𝑧)𝑛𝑦 −1 (50) where
Γ(𝑛′𝑥 )Γ(𝑛′𝑦 )
/⎛ ⎞
were 𝑛′𝑥 = 𝑛𝑥 /2, 𝑛′𝑦 = 𝑛𝑦 /2. Thus, [( )2 ] ∑
⎝2 2
E𝑛(:) ∣𝑔(:) 𝑛LS
𝑙 = 𝜎𝑛2 ∣𝑔𝑘 ∣ ⎠ (60)
𝑛′ 𝑘∈𝐷(𝑙)
E𝑧 [𝑧] = ( ′ 𝑥 ′ ) (51)
𝑛𝑥 + 𝑛𝑦
For the[ ‘per-symbol’
] based [ scenario
] and since
and E𝑛(:) ∣𝑔(:) 𝑛LS = 0, Var 𝑛 LS
equals
𝑛′𝑥 (𝑛′𝑥 + 1) 𝑙 𝑛(:) ∣𝑔(:) 𝑙 [ (60).
] For
E𝑧 [𝑧 2 ] = ( ′ )( ) (52) the ‘per-frame’ scenario and since 𝐸𝑛(:) ,𝑔(:) 𝑛LS = 0,
𝑛𝑥 + 𝑛′𝑦 𝑛′𝑥 + 𝑛′𝑦 +1 [ LS ] ∑𝑙 2
/ Var𝑛(:) ,𝑔(:) 𝑛𝑙 is given by (59). Since ∣𝑔𝑘 ∣ ∼
∑ ∑ ( ) 𝑘∈𝐷(𝑙)
Define the R.V. 𝑤 ≜ 𝑥𝑖 𝑦𝑖 𝑥𝑖 , where 𝑥𝑖 and 𝑦𝑖
𝜒2 2(𝐿- + 𝐿+ + 1), 𝜎𝑔2 /2 , (29) follows from (59) ,(60) and
𝑖∈𝐷 𝑖∈𝐷 ( )
are statistically independent R.V.’s with 𝑥𝑖 ∼ 𝜒2 2, 𝜎 2 , (49). [ ]
E𝑦𝑖 [𝑦𝑖 ] = 𝑚𝑖 , E𝑦(:) [𝑦𝑖 𝑦𝑗 ] = 𝜌𝑖,𝑗 and 𝐷 the window of the For Var 𝜖LS 𝑙 ,
R.V.’s indices of length 𝑀 . Then
/ 𝜖LS
𝑙 = Im {ℒ𝒮 (𝑗𝒈 ⊙ 𝝐(𝑙)∣𝒈, 𝐷(𝑙))}
∑ ∑ /
E𝑦(:) ∣𝑥(:) [𝑤] = 𝑥𝑖 𝑚𝑖 𝑥𝑘 (53) ∑ 2
∑ 2
𝑖∈𝐷 𝑘∈𝐷 = ∣𝑔𝑘 ∣ 𝜖𝑘 (𝑙) ∣𝑔𝑘 ∣ (61)
/ 𝑘∈𝐷(𝑙) 𝑘∈𝐷(𝑙)
∑ 𝑥𝑖
( )
Also 𝑥𝑖 𝑥𝑘 = 𝑥𝑖 +𝑥′ with 𝑥′ ∼ 𝜒2 2𝑀 − 2, 𝜎 2 . Thus, [ ]
𝑘∈𝐷 Var𝑛(:) ∣𝑔(:) 𝜖LS
𝑙 can be
( computed
) from (61) using (53) and
from (51) and (53)
[ ] (55). Since 𝑔𝑘 ∼ 𝜒2 2, 𝜎𝑔2 /2 and 𝐷(𝑙) has length (𝐿- +
1 ∑ 𝐿+ + 1), (25) follows from (61) using (54),(57), the statistics
E𝑥(:) ,𝑦(:) [𝑤] = E𝑥(:) E𝑦(:) ∣𝑥(:) [𝑤] = 𝑚𝑖 (54) [ ]
𝑀 of 𝝐(𝑙) and after some algebra. The computation of Var 𝜃𝑙LS
𝑖∈𝐷
for the case of Wiener and first-order Gaussian PHN models
Additionally
∑ ∑ can be performed similarly for both scenarios.
𝑥𝑖 𝑥𝑘 𝜌𝑖,𝑘
𝑖∈𝐷 𝑘∈𝐷
E𝑦(:) ∣𝑥(:) [𝑤2 ] = ( )2 (55)
∑ A PPENDIX D
𝑥𝑖
𝑖∈𝐷
We prove in this Appendix that the symmetric window
Using (55) and reformulating properly, parameterized by 𝐿 has the minimum unconstrained PCEV
E𝑥(:) ,𝑦(:) [𝑤2 ] = among all windows with the same length, i.e. among all
⎡ ⎤ 𝐷(𝑙) = {𝑙 − 𝐿- , ..., 𝑙 + 𝐿+ } with 𝐿- + 𝐿+ = 2𝐿.
∑ 2 ∑ ∑
⎢ 𝑥𝑖 𝜌𝑖,𝑖 + 12 ((𝑥𝑖 + 𝑥𝑘 )2 − 𝑥2𝑖 − 𝑥2𝑘 )𝜌𝑖,𝑘 ⎥ Proof: For 𝐿- < 𝐿+ (the same can be proved for 𝐿+ > 𝐿- )
⎢ 𝑖∈𝐷 𝑖∈𝐷 𝑘∈𝐷,𝑘∕=𝑖 ⎥ there exists an integer 𝑝 with 0 < 𝑝 < 𝐿 such that 𝐿 = 𝐿 − 𝑝
𝐸𝑥(:) ⎢ ( )2 ⎥ -
⎣ ∑ ⎦ and 𝐿 = 𝐿 + 𝑝. Then, substituting 𝐿 , 𝐿 in (30) and after
𝑥𝑖 + - +
𝑖∈𝐷 some algebra
(56) [ ] [ ]
/ Var 𝜖LS = Var 𝜖LS 𝑙 (𝐿) +
∑ ′′ ( ) ( )2 2
𝑙
Also (𝑥𝑖 + 𝑥𝑘 ) 𝑥𝑚 = 𝑥′′𝑥+𝑥′′′ with 𝑥′′ ∼ 𝜒2 4, 𝜎 2 , 2𝜋 𝜎𝜀 12(𝐿 + 2)𝑝2 + (4𝐿𝑝 + 2𝑝)2 [ ]
( 𝑚∈𝐷 ) ⩾ Var 𝜖LS𝑙 (𝐿)
′′′ 2 2
𝑥 ∼ 𝜒 2𝑀 − 4, 𝜎 . Thus, from (56) using (52) 𝑁 12 (2𝐿 + 1) (2𝐿 + 2)
∑ ∑ ∑ (62)
𝜌𝑖,𝑖 + 𝜌𝑖,𝑘 where 𝜖LS (𝐿) is the corresponding term to the symmetric
𝑖∈𝐷 𝑖∈𝐷 𝑘∈𝐷
𝑙 [ ]
E𝑥(:) ,𝑦(:) [𝑤2 ] = (57) window. The same can be proven for Var 𝜃𝑙LS , i.e.,
𝑀 (𝑀 + 1)
[ ] [ ]
A PPENDIX C Var 𝜃𝑙LS ⩾ Var 𝜃𝑙LS (𝐿) (63)
COMPUTATION OF PCEV TERMS
[ LS ] for Wiener and first-order Gaussian PHN. Also
For Var 𝑛𝑙 ,
[ ] 1 [ ]
𝑛LS
𝑙 = Im {ℒ𝒮 (𝒏∣𝒈, 𝐷(𝑙))}
Var 𝑛LS 𝑙 = = Var 𝑛LS 𝑙 (𝐿) (64)
∑ 4𝛾𝑠𝑦𝑠 𝐿
(Re{𝑔𝑘 } Im{𝑛𝑘 } − Im{𝑔𝑘 } Re{𝑛𝑘 })
𝑘∈𝐷(𝑙)
= ∑ 2 (58) for all 𝐿- , 𝐿+ . Since the unconstrained PCEV of sample 𝑙 is
∣𝑔𝑘 ∣ the sum of (62),(63) and (64), it is thus minimized by the
𝑘∈𝐷(𝑙) symmetric window parameterized by 𝐿.
4796 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 9, SEPTEMBER 2009

R EFERENCES [18] J. Liu and J. Li,“Parameter estimation and error reduction for OFDM-
based WLANs,” IEEE Trans. Mobile Computing, vol. 3, no. 2, pp. 152-
163, Apr. 2004.
[1] T. Pollet, M. V. Bladel, and M. Moeneclaey, “BER sensitivity of OFDM [19] S. Wu and Y. Bar-Ness, “OFDM systems in the presence of phase noise:
systems to carrier frequency offset and Wiener phase noise,” IEEE Trans. consequences and solutions,” IEEE Trans. Commun., vol. 52, no. 11, pp.
Commun., vol. 43, no. 2, pp. 191-193, Feb. 1995. 1988-1996, Nov. 2004.
[2] L. Tomba, “On the effect of Wiener phase noise in OFDM systems,” [20] S. M. Kay, Fundamentals of Statistical Processing, Vol. I: Estimation
IEEE Trans. Commun., vol. 46, no. 5, pp. 580-583, May 1998. Theory. Prentice Hall, 1993.
[3] A. Armada, “Understanding the effects of phase noise in orthogonal [21] D. G.Luenberger, Linear and Nonlinear Programming, 2nd ed. Addison
frequency division multiplexing (OFDM),” IEEE Trans. Broadcast., vol. Wesley, 1984.
47, no. 2, pp. 153-159, June 2001. [22] P. H. Moose, “A technique for orthogonal frequency division multiplex-
[4] K. Nikitopoulos and A. Polydoros, “Phase-impairment effects and ing frequency offset correction,” IEEE Trans. Commun. vol. 42, no. 10,
compensation algorithms for OFDM systems,” IEEE Trans. Commun., pp. 2908-2914, Oct. 1994.
vol. 53, no. 4, pp. 698-707, Apr. 2005. [23] A. Papoulis and S. Unnikrishna, Probability, Random Variables and
[5] L. Piazzo and P. Mandarini, “Analysis of phase noise effects in OFDM Stochastic Processes, 4th ed. McGraw-Hill, 2002.
modems,” IEEE Trans. Commun., vol. 50, pp. 1696-1705, Oct. 2002.
[6] M. S. El-Tanany, Y. Wu, and L. Hazy, “Analytical modeling and Ioannis Dagres was born in Athens, Greece, in
simulation of phase noise interference in OFDM-based digital television 1974. He received a BS degree in Computer En-
terrestrial broadcasting systems,” IEEE Trans. Broadcast. vol. 47, no. 3, gineering and an MS degree in Signal Processing
pp. 20-31, Mar. 2001. from the Technical University of Patras, Greece, in
[7] P. Smulders, “Exploiting the 60GHz band for local wireless multimedia 1997 and 1999 respectively. He is currently working
access: prospects and future directions,” IEEE Commun. Mag., vol. 40, towards his Ph.D. degree at the same university.
pp. 140-147, Jan. 2002. Since 1999 he has participated in several EU and
[8] C. H. Doan, S. Emami, D. A. Sobel, A. M. Niknejad, and R. W. Greek research projects and works as a Research
Brodersen, “Design considerations for 60GHz CMOS Radios,” IEEE Associate in the University of Athens. His research
Commun. Mag., vol. 42, pp. 132-140, Dec. 2004. interests lie in the area of signal processing for
[9] G. Fettweis, M. Lohning, D. Petrovic, M. Windisch, P. Zillmann, and communications, and in particular adaptive signal
W. Rave, “Dirty RF: a new paradigm,” in Proc. PIMRC’05, Berlin, pp. design and synchronization problems for multicarrier communication systems.
1340-1344, Sept. 2005.
Andreas Polydoros was born in Athens, Greece, in
[10] K. Nikitopoulos and A. Polydoros, “Compensation schemes for phase 1954. He was educated at the National Technical
noise and residual frequency offset in OFDM systems,” in Proc. University of Athens, Greece (Diploma EE, 1977),
GLOBECOM, San Antonio, TX, pp. 331-333, Nov. 2001. State University of New York at Buffalo (MSEE,
[11] P. Robertson and S. Kaiser, “Analysis of the effects of phase noise in 1979) and the University of Southern California–
OFDM systems,” in Proc. ICC, 1995. USC (Ph.D., EE, 1982). He was a faculty member
[12] V. S. Abhayawardhana and I. J. Wassell, “Common phase error cor- at USC in the Electrical Engineering Department
rection with feedback for OFDM in wireless communication,” in Proc. and the Communication Sciences Institute (CSI) in
GLOBECOM , Taipei, Taiwan, pp. 651-652, Nov. 2002. 1982-1997, becoming a Professor in 1992. He co-
[13] S. Wu, P. Liu, and Y. Bar-Ness, “Phase noise estimation and mitigation directed CSI in 1991-93. In 1997 he was elected
for OFDM systems,” IEEE Trans. Wireless Commun., vol. 5, no. 12, pp. Professor in the Department of Physics, University
3616-3625, Dec. 2006. of Athens, Greece. He directed the Electronics Laboratory of the Division
[14] D. Petrovic, W. Rave, and G. Fettweis, “Effects of phase noise on OFDM of Applied Physics in 1997-2007. The areas of his scientific interests are
systems with and without PLL: characterization and compensation,” stochastic digital communication theory, spread-spectrum systems and syn-
IEEE Trans. Commun., vol. 55, no. 8, pp. 1607-1616, Aug. 2007. chronization, digital mobile radio links and networks, detection-estimation-
[15] D. D. Lin, R. A. Pacheco, T. J. Lim, and D. Hatzinakos, “Joint estimation classification in uncertain environments and, more recently, flexible-radio
of channel response, frequency offset and phase noise in OFDM,” IEEE concepts. He is a co-inventor of the Per-Survivor Processing US patent (1995)
Trans. Signal Processing, vol. 54, no. 9, pp. 3542-3554, Sept. 2006. and co-author of the related paper. He has served as Associate Editor for
[16] D. D. Lin and T. Joon, “The variational inference approach to joint data Communications of the IEEE T RANSACTIONS OF I NFORMATION T HEORY
detection and phase noise estimation in OFDM,” IEEE Trans. Signal (1987-88), Guest Editor of various Special Issues and has been on the
Process., vol. 55, no. 5, pp. 1862-1874, May 2007. organizing committees of many conferences and workshops. He was elected
[17] H. L. Van Trees, Optimum Array Processing: Detection, Estimation and Fellow of the IEEE in 1995. He co-founded Trellisware Technologies in San
Modulation Theory, Part IV. John Wiley & Sons, 2002. Diego, CA, in 2000.