This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2691672, IEEE
Transactions on Signal Processing
1

Effective Low-Complexity Optimization Methods

for Joint Phase Noise and Channel Estimation in
OFDM
Zhongju Wang, Prabhu Babu, and Daniel P. Palomar, Fellow, IEEE

Abstract—Phase noise correction is crucial to exploit full model in [8]), respectively [8]–[10]. An OFDM block, con-
advantage of orthogonal frequency division multiplexing (OFDM) sisting of several symbols, is transmitted and received with
in modern high-data-rate communications. OFDM channel esti- orthogonal subcarriers. Due to the introduction of phase noise,
mation with simultaneous phase noise compensation has there-
fore drawn much attention and stimulated continuing efforts. however, the orthogonality among subcarriers is lost, which
Existing methods, however, either have not taken into account causes degradation in performance of OFDM systems. Such
the fundamental properties of phase noise or are only able to loss of performance has been well-documented and studied
provide estimates of limited applicability owing to considerable with detailed analyses of, e.g., signal-to-noise ratio (SNR)
computational complexity. In this paper, we have reformulated and bit error rate (BER); see [1]–[3], [8]–[13] and many
the joint phase noise and channel estimation problem in the time
domain as opposed to existing frequency-domain approaches, references therein. For instance, it has been reported in [13]
which enables us to develop much more efficient algorithms using that for OFDM with 2048 subcarriers, Wiener phase noise of
the majorization-minimization technique. In addition, we propose 3dB bandwidth ∆f3dB of 100 Hz can bring about an SNR
two methods based on dimensionality reduction and regulariza- degradation over 10 dB, which implies the acute sensitivity of
tion, respectively, that can adapt to various phase noise levels OFDM systems to the existence of phase noise.
and SNR and achieve much lower estimation errors than the
benchmarks without incurring much additional computational Common phase error (CPE) and inter-carrier interference
cost. Several numerical examples with phase noise generated (ICI) are two detrimental effects caused by phase noise. CPE
by free-running oscillators or phase-locked loops demonstrate causes phase rotation to each subcarrier and does not change
that our proposed algorithms outperform existing methods with within a transmitted OFDM block. In contrast, ICI introduces
respect to both computational efficiency and mean squared error different interference to different subcarriers in the same block,
within a large range of signal-to-noise ratios.
and thus exhibits noise-like characteristics [14]. Even with
Index Terms—Carrier frequency offset (CFO), channel esti- small phase noise, where ICI dominates over CPE, CPE-
mation, majorization-minimization (MM), orthogonal frequency only correction can provide a 5 dB gain in SNR [3]. To be
division multiplexing (OFDM), phase noise.
general, in this paper we will consider phase noise that is
not necessarily small. Also, even though a constant carrier
I. I NTRODUCTION frequency offset (CFO) may exist apart from phase noise, we
assume CFO has been fixed using readily available methods,
ROMINENT advantages such as higher spectral effi- e.g., [15].
P ciency, adaptability to severe channel environments, and
efficient implementation have brought orthogonal frequency
Many works, e.g., [8], [16], [17], have studied phase noise
estimation assuming channel information is given or in the
division multiplexing (OFDM) into wide applications in mod- context of transmit data detection. Phase noise estimation and
ern communications. To fully exploit these advantages in real- compensation can reduce ICI between data subcarriers and
ity, we have to resolve some demanding issues—sensitivity to directly facilitate information decoding. In practice, however,
frequency synchronization errors, high peak-to-average power phase noise exists throughout the channel training and data
ratios, to name a few. In this paper, we will focus on the transmission stages, and the effects of phase noise to channel
frequency synchronization issue stemming specifically from estimates also influence the next data detection. Joint esti-
phase noise. mation of phase noise and channel has thus been proposed
Phase noise is a random process caused by the fluctuation to improve the quality of channel estimates [18]–[20]. The
in phase within receiver and transmitter oscillators that are joint estimation problem has been investigated as early as in
deployed to generate carrier signals for up-down conversion [13]. The least-squares estimator of channel impulse response
[1]–[7]. In practice, free-running oscillators and phase-locked is computed first; then heuristically, a window function as a
loops are widely used, for which phase noise is described by filter is applied to the obtained channel estimator to reduce
Wiener process and Gaussian process (or Ornstein-Uhlenbeck its sensitivity from phase noise and CFO. To be statistically
justified, maximum a posteriori channel estimator in [18] has
This work was supported by the Hong Kong RGC 16206315 research grant. exploited the statistical properties of phase noise. But the
Zhongju Wang and Daniel P. Palomar are with the Hong Kong University of authors use a Taylor expansion to approximate the nonlinear
Science and Technology (HKUST), Hong Kong. E-mail: {zwangaq, palo-
mar}@ust.hk. Prabhu Babu is with CARE, IIT Delhi, Delhi, India. Email: optimization objective function, which works only for small
prabhubabu@care.iitd.ac.in. phase noise. A simple alternating optimization method for the

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2691672, IEEE
Transactions on Signal Processing
2

joint estimation problem can be found in [19]. The critical Scalars, vectors, and matrices are denoted by italic letters,
issue with that method is its failure to deal with the constraint boldface lower-case letters, and boldface upper-case letters,
of phase noise in each iterative sub-problem. respectively. The superscript (·)T denotes the transpose, (·)H
Supposedly, estimating phase noise and channel was hard to the conjugate transpose. The ℓ2 -norm and ℓ∞ -norm of a vector
disentangle as previous works claimed. In [21], a novel formu- is denoted by ∥ · ∥ and ∥ · ∥∞ , respectively. The identity matrix
lation is proposed with phase noise and channel estimations is denoted by In with size specified by the subscript n. R is
unraveled. To solve the resulting optimization problem, the the set of real numbers. 1n is an all-one vector of length n.
authors replace the unimodular constraint on phase noise in λmax (·) denotes the maximum eigenvalue of a matrix. O(·)
the time domain with a relaxation assuming the magnitude of denotes Big-O notation.
phase noise is relatively small. Nevertheless, their method is
computationally unstable with a singularity issue that renders II. S YSTEM M ODEL AND D ESCRIPTION OF P HASE N OISE
the already approximated solution even more inaccurate. And
A. OFDM Transmission Model
recently, a method craftily using the spectral property of phase
noise is provided for the frequency domain-formulated prob- Suppose there are Nc subcarriers and an OFDM block is de-
T
lem [22]. Based on [21], the separate phase noise estimation noted by s = [s0 , . . . , sNc −1 ] . The time-domain symbols can
problem is solved by semidefinite programming (SDP). This be obtained by the unitary inverse discrete Fourier transform
method works fine when the number of subcarriers deployed in (IDFT):
OFDM is not too large and phase noise arises in a small level. Nc −1
1 ∑ j2πnk
In reality, however, the number of subcarriers can be as large xn = √ sk e Nc , n = 0, 1, . . . , Nc − 1. (1)
as tens of thousands, e.g., in terrestrial television broadcasting Nc k=0
system (DVB-T2) [23]. Let F be the Nc ×Nc unitary discrete Fourier transform (DFT)
Regarding the joint phase noise and channel estimation, matrix, then (1) can be written as
there are basically two classes of approaches: time-domain
[13], [18]–[20], [24]–[26] and frequency-domain approaches x = FH s. (2)
[8], [21], [22]. In this paper, we formulate the optimization
Assume a slow-varying channel whose response does not
problem in the time-domain representation and our contribu-
change within the transmission of several OFDM blocks and
tions are as follows. First, we prove the equivalence of the T
is denoted by h = [h0 , h1 , . . . , hL−1 ] (Nc ≫ L). Let
frequency-domain approach and the time-domain approach to T T
x = [x0 , x1 , . . . , xNc −1 ] and y = [y0 , y1 , . . . , yNc −1 ] . With
the problem formulation. It allows us to separate the joint
the cyclic prefix appending and removal, we have the OFDM
estimation problem and to focus on estimating phase noise.
transmission model [27, Ch. 3.4.4]:
And using the majorization-minimization technique, we devise [ ]
more efficient algorithms as opposed to solving an SDP as h
y =x⊛ + v, (3)
in [22]. The efficiency and low complexity of our proposed 0
algorithms enable us to deal readily with much larger number
where ⊛ denotes the operation of circular convolution, and
of subcarriers. Moreover, we offer two adaptive methods for T
v = [v0 , v1 , . . . , vNc −1 ] is a zero-mean circularly symmet-
further reducing estimation errors considering that the joint
ric complex Gaussian channel noise vector with distribution
estimation problem is underdetermined per se irrespective of
CN (0, 2σ 2 I).
the approach of formulation. To achieve this, dimensionality
To obtain the frequency-domain representation of (3), take
reduction has been adopted in [19], [22] to address either the
the DFT to both sides and we have1
underdetermined nature of the problem or the computational √
complexity. But instead of adopting a fixed dimension, we r = Nc Hs + w, (4)
run our algorithms with different reduced sizes and opt for
the solution that yields the minimal Bayesian information where r is the unitary DFT of the received time-domain
criterion (BIC). Besides dimensionality reduction, we also symbols y, H is a diagonal matrix with the Nc -point unitary
propose a regularization to the original estimation objective to DFT of h as the diagonal, and w is the unitary DFT of the
prevent potential over-fitting. The extra adaptability to various time-domain channel noise v. Let F̌ be a semi-unitary( matrix
)
levels of phase noise and SNR comes without incurring much formed by the first L columns of F, then H = Diag F̌h .
computational burden as simulated examples demonstrate.
The structure of this paper is as follows. We give the B. OFDM Transmission with Phase Noise
system model of OFDM with a description of phase noise In general, phase noise is present in the local oscillators
in Section II. In Section III, the problem formulation is that generate carrier signals for up-down conversion for the
presented after a review of existing methods. We dedicate time-domain symbols. And the effect of phase noise can be
Section IV to developing algorithms solving the formulated represented mathematically by multiplying each time-domain
problem with extensions based on dimensionality reduction symbol with a complex exponential with a random phase.
and regularization. Simulation results are given in Section V, Although phase noise exists in both the transmitter and the
followed by a conclusion to summarize the paper in Section receiver, herein only the effect at the receiver side is studied.
VI.
1 Note
√
We use the following notation throughout this paper. that the factor Nc results from using the unitary DFT.

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2691672, IEEE
Transactions on Signal Processing
3

The reason for this simplified consideration is the assumption 2) Frequency-Domain Property: Let ϕ and ϕ be the unitary
that at the transmitter side, the bandwidth of phase noise DFT of ejθ and e−jθ , respectively. It is well-known that ϕ
is small [14] or high-caliber oscillators are employed [22]. and ϕ are conjugate symmetric, i.e.,
Therefore, the following signal model with phase noise is
considered [8]: ϕk = ϕ∗−k . (12)
[ ]
h Observing that ejθ ⊙ e−jθ = 1Nc and applying the DFT to
y = ejθ ⊙ (x ⊛ ) + v, (5)
0 both sides, we can obtain the following constraint for spectral
[ ]T phase noise:
where ejθ := ejθ0 , ejθ1 , . . . , ejθNc −1 denotes phase noise, ϕ ⊛ ϕ = N c δk , (13)
and ⊙ denotes the Hadamard product. Taking the unitary DFT
on both sides of (5), we can obtain the frequency-domain where δk is the Kronecker delta function, i.e., δ0 = 1, and
signal model: δk = 0 for k ̸= 0. Indeed, (13) is a necessary and sufficient
r = ϕ ⊛ (Hs) + w, (6) description of the autocorrelation of the spectral components
T of any unimodular complex exponential sequence, which can
where ϕ = [ϕ0 , ϕ1 , . . . , ϕNc −1 ] = Fejθ , called spectral be easily verified by Fourier transform and its properties.
phase noise vector, and w are the unitary DFT of ejθ and Equivalently, (13) can be written in a matrix form as
v, respectively. For each received frequency-domain symbol
rk , k = 0, 1, . . . , Nc − 1, we have ΦH
ϕ Φϕ = Nc INc , (14)
c −1
N∑
where Φϕ = circ (ϕ) is a circulant matrix defined in (9). This
rk = ϕ0 Hk,k sk + ϕk−l Hl,l sl + wk , (7)
is the main property exploited in [22], termed the spectral
l=0,l̸=k
geometry.
where the first term, subjected only to the scaling of factor ϕ0 ,
is called CPE, and the second term, combining effects from
III. L ITERATURE R EVIEW AND P ROBLEM F ORMULATION
other subcarriers, is ICI. With rk = rk mod Nc , (6) can be
rewritten in the following matrix form Phase noise contamination can be removed from the re-
ceived OFDM symbols if a reliable estimate of the instan-
r = Φϕ Hs + w, (8) taneous realization of phase noise process is accessible. In
in which practice, phase noise exists throughout channel estimation
  stage and data detection stage, where phase noise estimation is
ϕ0 ϕNc −1 ··· ϕ2 ϕ1 entangled with the unknown channel and unknown transmitted
 ϕ1 ϕ0 ··· ϕ3 ϕ2 
  data, respectively. Plenty of works are available for estimating
 .. .. .. 
Φϕ =  . . . , (9) transmitted data with phase noise compensated, e.g., [19], [24].
 
ϕNc −2 ϕNc −3 ··· ϕ0 ϕNc −1  In this paper we will focus on joint phase noise and channel
ϕNc −1 ϕNc −2 ··· ϕ1 ϕ0 estimation, where transmitted symbols (called pilot symbols
or training symbols) are assumed known to receiver [27, Ch.
denoted by Φϕ = circ (ϕ), is a circulant matrix formed by 3.5.2]. A motivation for studying this joint estimation problem
spectral phase noise ϕ. The off-diagonals of Φϕ close to the is that assuming channel is quasi-static or slowly-varying, the
main diagonal
( correspond
) to low-frequency components. With channel estimate can be reasonably used in the subsequent data
H = Diag F̌h , (8) can be rewritten as estimation [19]. In the literature, methods for this purpose can
be categorized into two classes: time-domain approach and
r = Φϕ SF̌h + w, (10)
frequency-domain approach. In particular, we assume in this
where S = Diag(s) is a diagonal matrix with s as the diagonal paper that phase noise θ and channel impulse response h are
and F̌ a semi-unitary matrix with the first L columns of F. constant parameters to estimate, and OFDM symbols S are
given and known to the receiver.
C. Properties of Phase Noise
Two canonical models of phase noise are Wiener process A. Time-Domain Approaches
and Gaussian process when free-running oscillators and phase- In [19], the authors formulate the least-squares problem with
locked loops are respectively employed [8]. The statistical (10)
2
properties of phase noise have also been studied in [8], [14]. minimize r − Φϕ SF̌h , (15)
Before introducing some existing formulations of the phase h,θ,ϕ=Fejθ
noise estimation problem, some useful properties of phase and solve for channel and phase noise estimates alternately.
noise are presented here. (i−1)
At the ith iteration, given the phase noise estimate ej θ̂ ,
1) Time-Domain Property: Obviously, phase noise ejθ is
the channel estimate is computed by
determined only by the phase variable θ, and phase noise at
each OFDM subcarrier is unimodular, i.e., ( )−1
(i−1) (i−1) (i−1)
 jθ  ĥ(i) = F̌H SH (Φϕ )H Φϕ SF̌ F̌H SH (Φϕ )H r
e n  = 1, n = 0, 1, . . . , Nc − 1. (11) (16)

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2691672, IEEE
Transactions on Signal Processing
4

(i−1)
(i−1)
with Φϕ = circ(Fej θ̂ ). Let c = ejθ , then the estimate Instead of solving (21), dimensionality reduction is introduced
for phase noise is updated as to alleviate the computation complexity by estimating only the
( )−1 H H low-frequency components, cf. [8]. To achieve this, the phase-
ĉ(i) = FH PH PF F P r, (17) noise-geometry preserving transformation is defined by
where P = circ(SF̌ĥ(i) ). Yet there are two issues with ϕ = Tϕ̌, (22)
their method: the unimodular property of phase noise vector
where ϕ̌ of a shorter length N is the reduced spectral phase
is not considered when updating ĉ(i) ; and the alternating
noise to be estimated. An example of T is piecewise-constant
optimization scheme suffers from slow convergence.
transformation (PCT). Then an alternative optimization prob-
Some other heuristic methods include approximating phase
lem is posed as follows:
noise by a Taylor expansion [18], applying filtering to channel
H
estimate with a noise-suppressing function [13], approximating minimize ϕ̌ TH MTϕ̌
with sinusoidal waveforms [24], and Monte Carlo methods ϕ̌ (23)
H
[20], [25], [26]. subject to Φ̌ϕ Φ̌ϕ = N IN , Φ̌ϕ = circ(ϕ̌).
To solve the above problem, the S-procedure is invoked to
B. Frequency-Domain Approaches rewrite (23) as a semidefinite program (SDP). The original
spectral phase noise vector ϕ can be recovered by (22). To
In [8], a phase noise correction method is proposed by
guarantee the constraint (14) still holds, the authors provide
estimating the spectral components, based on the assumption
a sufficient condition for the transformation matrix. Their
that phase noise process can be characterized by a low-pass
method, however, suffers from several limitations. When the
signal and thus only a few spectral components need to be
reduced length N is not small enough, SDP reformulation still
estimated. But it is necessary to find a proper number of
renders a solution failing to satisfy the spectral constraint of
spectral phase noise components in order to achieve reliable
phase noise; yet, it is prohibited to solve a large dimensional
estimation. Although [8] also exploits the statistical properties
SDP. Nowadays, the number of subcarriers can be up to
of ICI to obtain the MMSE estimate of phase noise, their
thousands and to use this method, the original dimension
method is subject to two main issues: the channel is assumed
needs to be greatly reduced, which can result in the loss of
known and the MMSE estimation has not taken into account
reliability and accuracy in the obtained estimate. Furthermore,
the constraint (13) of spectral phase noise.
the reduced spectral phase noise does not necessarily satisfy
Following the same idea of [8] to estimate the low-
the spectral constraint as imposed in problem (23); thus, this
frequency components of phase noise, [21] formulates the
method gives a tightened solution.
problem of joint phase noise and channel estimation based
on least-squares. To acquire separate estimators, instead of
alternately updating (16) and (17), they substitute the channel C. Problem Formulation
estimate into the least-squares objective and the resulting error Based on (5), the time-domain OFDM model with phase
function for phase noise can be derived as noise is given by
√ ( )
1 H y = Nc Diag ejθ FH SF̌h + v. (24)
E (ϕ) = rH r − r Φϕ BΦH ϕr (18)
Nc
1 H ( H )T Similar to (15), we propose the following optimization prob-
= ϕ J1 R R − RH BR J1 ϕ, (19) lem:
Nc
√ ( ) 2
( )−1 H H minimize y − Nc Diag ejθ FH SF̌h . (25)
where B = SF̌ F̌H SH SF̌ F̌ S and J1 a permutation h,θ
matrix, left multiplication by which keeps the first row and Solving (25) for h gives the least-squares channel estimate
reverses the orders of the remaining rows. Note that in [21],
the expression for E (ϕ) is further simplified assuming the 1 ( H H )−1 H H ( )H
ĥ = √ F̌ S SF̌ F̌ S FDiag ejθ y. (26)
transmitted symbols are of constant-modulus. When solving Nc
for the phase noise estimate, however, an approximation by a And the resulting least-squares error for phase noise is
Taylor expansion is applied, which leads to a relaxed constraint ( ) ( )H
on phase noise. In practice, this approximation works only for E(θ) = yH Diag ejθ FH (INc − B) FDiag ejθ y, (27)
small phase noise. ( )−1 H H
where B = SF̌ F̌H SH SF̌ F̌ S . The phase noise
In contrast, [22] incorporates the fundamental spectral con-
estimation problem is thus formulated as
straint (14) into the formulation proposed in [21]. Let
( ) ( )H
1 ( )T minimize yH Diag ejθ FH (INc − B) FDiag ejθ y.
M= J1 RH R − RH BR J1 , (20) θ
Nc (28)
Let us introduce V = FH (INc − B) F and u = e−jθ . We
then the problem is formulated as
can rewrite (28) as the following quadratic problem:
minimize ϕH Mϕ minimize uH Diag(y)H VDiag(y)u
ϕ (21) u (29)
subject to ΦH
ϕ Φϕ = Nc INc , Φϕ = circ (ϕ) . subject to |un | = 1, n = 0, 1, . . . , Nc − 1.

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2691672, IEEE
Transactions on Signal Processing
5

Consequently, the joint phase noise and channel estimation Consider the problem of
problem boils down to the phase noise estimation problem minimize f (x) subject to x ∈ X. (36)
(29) followed by computing the channel estimate with (26). x

The MM technique starts from a feasible point x(0) ∈ X , and

D. Equivalence of Time- and Frequency-Domain Approaches solves a series of simpler majorized problems:
( )
In this section, we show that our formulation of the joint es- minimize g x; x(t) subject to x ∈ X , (37)
x
timation problem (25) and the resulting phase noise estimation
problem (29) are equivalent to the existing approaches. t = 0, 1, . . . , each of which produces an updated point x(t+1) .
Lemma 1. (Let ϕ
√ ) = Fejθ and Φϕ = circ(ϕ), then Φϕ = Basically, the surrogate objective, known as the majorization
jθ H function for f (x), should satisfy the following conditions:
Nc FDiag e F .
( ) ( )
Proof: According to the eigenvalue decomposition of a g x(t) ; x(t) = f x(t) , (38)
circulant matrix [28], ( )
(√ ( )T ) H g x; x(t) ≥ f (x) ∀x ∈ X , (39)
Φϕ = FDiag Nc F eT1 Φϕ F ( ) ( )
(√ ( ) H )∗ H ∇d g x(t) ; x(t) = ∇d f x(t) ∀x(t) + d ∈ X , (40)
= FDiag Nc FH eT1 Φϕ F ( )
(√ )∗ where ∇d g x(t) ; x(t) is the directional derivative of g at x(t)
= FDiag N c FH ϕ FH in the direction of d. Consequently, a series of points that
(√ )∗ result in non-increasing objective values are obtained:
= FDiag Nc e−jθ FH ( ) ( ) ( ) ( )
√ ( ) f x(t+1) ≤ g x(t+1) ; x(t) ≤ g x(t) ; x(t) = f x(t) ,
= Nc FDiag ejθ FH ,
(41)
where e1 = [1 0 · · · 0]T . ■ And any limit point of thus generated sequence of points is a
stationary solution to the original problem (36).
With Lemma 1, we can prove that the objective function in To develop an efficient MM-based algorithm, the series of
problem (25) is the same as that of (15): problems (37) should all be simple enough to solve—ideally,
√ ( ) 2
each should be solved with a closed-form solution. Crucial to
y − Nc Diag ejθ FH SF̌h
2 achieve
( such
) a goal is to find a good majorization function
√ ( ) g x; x(t) , which requires to properly exploit the particular
= FH r − Nc Diag ejθ FH SF̌h (30)
structure of the specific problem. Some general and useful
√ ( ) 2
= r − Nc FDiag ejθ FH SF̌h (31) rules for majorization can be found in [31]. In the next section,
2 we will devise MM algorithms to solve our problem (29) with
= r − Φϕ SF̌h . (32) two different majorizing methods. Also it will be illustrated in
simulations that the majorization is critical for the convergence
Since
√ ( )H √ ( )H speed of the obtained algorithms.
ΦH
ϕr = Nc FDiag ejθ FH r = Nc FDiag ejθ y,
(33) B. The MM Algorithms for Phase Noise Estimation
(27) is equivalent to the frequency-domain phase noise error The following lemma is introduced first, which is useful for
function (18): finding majorization functions.
1 H Lemma 2. Given a matrix A, P = A(AH A)−1 AH is an
E (ϕ) = r Φϕ (INc − B) ΦH ϕr (34)
Nc orthogonal projection matrix, which is unitarily similar to a
( ) ( )H diagonal matrix with diagonal entries being either 1 or 0 [32,
= yH Diag ejθ FH (INc − B) FDiag ejθ y.
(35) Corollary 3.4.3.3].
1) Loose Quadratic Majorization (LQM): Let us write Ṽ =
In the next section, we will use the majorization-minimization Diag(y)H VDiag(y). The objective in (29) can be majorized
technique to develop efficient algorithms to solve problem by a quadratic function at u0 as follows [33, Lemma 1]:
(29). { ( ) }
uH Ṽu ≤ 2Re uH 0 Ṽ − λINc u + 2λ∥u∥2 − uH 0 Ṽu0 ,
IV. A LGORITHMS (42)
in which λINc ⪰ Ṽ for some constant λ. Note that the largest
A. The Majorization-Minimization Technique
eigenvalue of V is 1 by Lemma 2, then we have λmax (Ṽ) ≤
The majorization-minimization (MM) technique provides ∥y∥2∞ . Choosing λ = ∥y∥2∞ will thus satisfy the majorization
an approximation-based iterative approach to solving an opti- condition. At the step t, the following majorized problem with
mization problem of a generic form [29]–[31]. As the original the surrogate objective function is solved (since ∥u∥2 is just
problem is difficult to address directly, the MM technique a constant):
follows an iterative procedure—a simpler surrogate objective {( )H ( ) }
function is minimized in each iteration—to find a local opti- minimize −2Re u(t) ∥y∥2∞ INc − Ṽ u
u (43)
mum. subject to |un | = 1, n = 0, 1, . . . , Nc − 1.

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2691672, IEEE
Transactions on Signal Processing
6

It is obvious that a closed-form solution to (43) is: subjected to reciprocal common phase rotations. Let ĥ and θ̂
[ (( ) )] be channel and phase noise estimates, respectively. The least-
u(t+1) = exp j arg ∥y∥2∞ INc − Ṽ u(t) (44) squares error of the estimates ĥ and θ̂ is the same as that
[ (( ) of ejθc ĥ and θ̂ − θc 1Nc . Since θc keeps unchanged among
= exp j arg ∥y∥2∞ 1 − |y|2 ⊙ u(t)
)] subcarriers, it acts like CFO. Many effective methods can be
+ Diag(y)H FH BFDiag(y)u(t) , (45) found for CFO correction; see, e.g., [15], [18]. Assuming CFO
has been eliminated before estimating phase noise, we can
where the exponential and the squared magnitude | · |2 are thus set θc = 0. Therefore, once Algorithm 1 converges to a
taken element-wise. We call this method a loose quadratic solution u⋆ , phase ambiguity can be removed by the rotation:
majorization (LQM) because the structure of the original u⋆ ← u⋆ /u⋆0 .
objective function could have been better exploited as shown
below, which leads to faster convergence.
2) Tight Quadratic Majorization (TQM): Similar to (42), C. Dimensionality Reduction
the original objective can be majorized as follows: Dimensionality reduction has been proposed in [22] to
H H alleviate the computational complexity when solving an SDP
u Diag(y) VDiag(y)u
of size Nc , the number of OFDM subcarriers. More important,
≤ λuH Diag(y)H Diag(y)u as noted in [19], estimation problem (15) and equivalent
{ }
(25) are essentially underdetermined. To obtain reasonable
+ 2Re uH0 Diag(y)H
(V − λI Nc ) Diag(y)u
estimates, the number of unknowns in the problem needs to be
0 Diag(y) (λINc − V) Diag(y)u0
+ uH H reduced and, hence, a reduced phase noise vector is estimated.
(46)
{ } Similar to the transformation (22) introduced in [22] for
= 2Re uH0 Diag(y) (V − λINc ) Diag(y)u
H
estimating reduced spectral phase noise, we apply dimension-
+ 2λ∥y∥2 − uH H ality reduction to our problem (29) in the time domain. Recall
0 Diag(y) VDiag(y)u0 , (47)
u = e−jθ for phase noise θ. We define
where λINc ⪰ V for some constant λ and the equality follows
from the unimodular property of un , n = 0, 1, . . . , Nc − 1. To û = TN ǔ = TN e−j θ̌ (51)
find a good majorization function, we can choose λ = 1 by
Lemma 2. At the step t, the following majorized problem can as a mapping from a low-dimensional phase noise θ̌ ∈ RN
be obtained: to the original phase noise with the transformation matrix
{( )H } TN ∈ RNc ×N (N < Nc ). Two instances of TN are suggested
minimize −2Re u(t) Diag(y)H FH BFDiag(y)u in [22]—piecewise-constant transformation (PCT) and random
u
subject to |un | = 1, n = 0, 1, . . . , Nc − 1, perturbator. And it has been demonstrated that PCT, albeit
(48) simple, achieves the best performance. PCT is defined as
which results in a closed-form solution:  
[ ( )] 1Ns 0 ... 0
u(t+1) = exp j arg Diag(y)H FH BFDiag(y)u(t) .  0 1Ns . . . 0 
 
(49) TN =  . . .. ..  , (52)
 .. .. . . 
It will be demonstrated later that this method converges faster
0 0 ... 1Ns
owing to its tighter majorization.
with Ns = Nc /N . In this case, the transformation matrix acts
Algorithm 1 Algorithm for Phase Noise Estimation with like a sample-and-hold circuit to recover the desired phase
TQM. noise. Another transformation matrix is provided in [19] based
( )−1 H H
1: Compute B = SF̌ F̌H SH SF̌ F̌ S , set t = 0, and on linear interpolation. For simplicity, we focus on PCT in this
initialize u(0) = ejθ0 . paper, but the methods proposed in the following also apply
2: repeat to the linear interpolation transformation and a numerical
[ ( )]
3: u(t+1) = exp j arg Diag(y)H FH BFDiag(y)u(t) example is provided later.
4: t←t+1 Introducing dimensionality reduction requires us to solve a
5: until convergence. different optimization problem. By substituting (51) into (29),
we can obtain an estimation problem of a lower dimension. A
The whole procedure is summarized in Algorithm 1 for similar procedure, however, can be followed when majorizing
TQM. Since main difference from TQM lies in the update the new objective function and developing the MM algorithms.
of u(t+1) , the algorithm for LQM is omitted here. Once the For TQM, the update (49) is modified accordingly as
algorithm converges to solution u⋆ , the phase noise estimate H H
FH BFDiag(y)TN ǔ(t) )
ǔ(t+1) = ej arg(TN Diag(y) . (53)
can be obtained by
For LQM, the majorization function needs to be recom-
θ̂ = − arg (u⋆ ) . (50)
puted as the condition λIN ⪰ TH N ṼTN involves TN .
Remark 1: There exists phase rotation ambiguity in problem Notice that λmax (Ṽ) = ∥y∥2∞ , and λ can be set to be
(25): it can be seen that phase noise and channel estimates are ∥y∥2∞ λmax (TH
N TN ) such that the majorization inequality

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2691672, IEEE
Transactions on Signal Processing
7

constraint is satisfied. As a result, the update for LQM is be an (Nc − 1) × Nc first-order difference matrix. As phase
obtained as follows: noise does not change significantly between consecutive sam-
ples, we can use the regularizer ∥Du∥2 and solve the following
ǔ(t+1) = ej arg((∥y∥∞ λmax (TN TN )IN −TN ṼTN )ǔ ) .
2 H H (t)
(54)
regularized estimation problem:
For our chosen PCT, λmax (TH
N TN ) = Ns and (54) simplifies minimize uH Diag(y)H VDiag(y)u + µ∥Du∥2
to u (59)
[ (( ) )] subject to |un | = 1, n = 0, . . . , Nc − 1,
ǔ(t+1) = exp j arg ∥y∥2∞ Ns IN − TH
N ṼTN ǔ
(t)
.
where µ is a given regularization parameter. To solve the
(55) regularized problem with the MM framework, the previous
procedure can be followed to obtain the modified MM update,
Algorithm 2 Phase Noise Estimation with TQM and the
which is given by
Optimal PCT Selected by BIC. [ (
( )−1 H H
1: Compute B = SF̌ F̌H SH SF̌ F̌ S . Choose T as a u(t+1) = exp j arg Diag(y)H FH BFDiag(y)u(t)
set of PCT matrices of different reduced length N . ( ) )]
+ µ λmax (DH D)INc − DH D u(t) . (60)
2: for each TN ∈ T do
3: set t = 0 and initialize ǔ(0) Compared with dimensionality reduction method, solving the
4: repeat regularized problem saves the trouble of choosing sampling
H H H (t)
5: ǔ(t+1) = ej arg(TN Diag(y) F BFDiag(y)TN ǔ ) length N , but still the value of regularization parameter µ
6: t←t+1 needs to be properly chosen. In practice, the obtained estimates
7: until convergence are robust to µ within a large range of value.
8: û = TN ǔ(t+1)
9: choose û with the minimal BIC (û).
E. Computational Issues and Analysis of Convergence
In the initialization, both LQM and TQM need to compute
The transformation matrix requires the reduced dimension B. Since S is diagonal, computing SH S needs Nc complex
N to be specified in advance. Previous works have assumed a multiplications. Notice that F̌ consists of the first L (L ≪ Nc )
fixed PCT with given N , which is hardly flexible to different columns of unitary DFT matrix F. And using FFT to compute
SNR. Here, we prescribe a set of values of N and run our Nc -point DFT needs (Nc /2) log2 Nc complex multiplications
algorithm for each of those values. In particular, N is chosen and Nc log2 Nc complex additions. Then F̌H SH SF̌, which
as a factor of Nc such that PCT is well-defined. To choose is L × L Toeplitz, can be calculated with less than Nc +
the optimal N , we employ the BIC rule [34], which has been Nc log2 Nc complex multiplications and 2Nc log2 Nc complex
demonstrated very effective in model order selection to avoid additions. Computing the inverse of this L × L matrix is an
over-fitting. For each estimate û, the corresponding BIC is easy task (L is the number of the channel taps, usually around
defined as the order of 10), which requires O(L2 ) arithmetical operations
BIC (û) = −2 ln p (y, ǔ) + N ln Nc , (56) [35]. Furthermore, less computation will be required when
the pilot symbols have constant modulus, e.g., using M -QAM
where p (y, ǔ) is the probability density function of y given ǔ. signals [8], [22], for which SH S = αIL for some constant α
With model (5) and transformation (51), (56) can be rewritten and B is thus much simplified. To compute B, we still need
as additional FFT for the left and right matrix multiplications,
E(θ̂) each of which involves Nc L + ((L/2) log2 L + Nc − L) L
BIC (û) = 2 + N ln Nc , (57)
σ complex multiplications and L2 log2 L complex additions.
where E(θ̂) is the least-squares error (27) of the phase noise In the update of TQM, Diag(y)H FH BFDiag(y) is fixed
estimate û. The optimal PCT is then defined as the one that and thus needs to be computed only once, which requires
produces the minimal BIC. In doing so, improved estimates 2Nc2 + Nc2 log2 Nc complex multiplications and 2Nc2 log2 Nc
are expected, compared with the traditional methods [19], [22]. complex additions. The summary of the computational com-
Furthermore, the computational efficiency of LQM and TQM plexity for computing phase noise estimate is given in Table
also guarantees an acceptable computational cost. The whole I. The overall complexity is at the order of O(Nc2 log2 Nc )
procedure is described in Algorithm 2. for the initialization and O(Nc ) for each update. For TQM,
the algorithm usually converges in practice within around
10 iterations. Faster convergence speed can also be achieved
D. Regularization
by applying accelerations; see Table II. Once a phase noise
Apart from dimensionality reduction, another approach to estimate is obtained, channel estimate (26) can be computed
addressing the potential over-fitting inherent in (25) is to with Nc complex multiplications.
impose a regularization to the objective in problem (29). Let Compared with some existing methods, our proposed al-
 
1 −1 0 · · · 0 gorithms requires almost the same or even less compu-
0 1 −1 · · · 0 tational complexity. For instance, the state-of-the-art PNC
D= 

 (58)
··· in [22] solves an SDP, which requires a complexity of
0 0 0 1 −1 O(N 3.5 log2 ( 1ϵ )) using efficient solver SeDuMi [36]. An

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2691672, IEEE
Transactions on Signal Processing
8

TABLE I
C OMPUTATIONAL C OMPLEXITY OF P ROPOSED A LGORITHMS FOR C OMPUTING P HASE N OISE

Arithmetical Operations
Initialization Update
Algorithms × + × +

TQM Nc2 Nc (Nc − 1)

LQM <N(c + Nc log2 Nc + ) Nc2 + Nc Nc2
< 2Nc log2 Nc +
TQM + PCT 2L L log2 L + Nc − L + 2Nc L + N2 N (N − 1)
2 2L2 log2 L + 2Nc2 log2 Nc ¨
LQM + PCT 2Nc2 + Nc2 log2 Nc + O(L2 )¶ N 2 + 2N N2 + N
TQM + Reg Nc2 + Nc Nc2 + 3Nc − 2
¶ Usually L ≪ Nc . And LQM needs additional Nc complex multiplications and Nc real additions to compute ∥y∥2∞ 1 − |y|2 .
¨ When PCT is applied, additional Nc2 − Nc N 2 and Nc2 − Nc N 2 + Nc − N complex additions are needed for TQM and LQM,
respectively.

alternating optimization method proposed in [19] requires a that gives the minimal BIC. In particular, opt-PCT is selected
per-update complexity of O(Nc3 ). And a non-iterative method from PCTs with N ∈ {25 , 26 , . . . , 2log2 Nc }. TQM + Reg is the
in [21] requires a complexity of O(N 3 ) apart from the regularization-based method as an alternative to dimensional-
computation of B. ity reduction and to accelerate the convergence speed, we have
According to (41), the obtained series of points {u(t) } employed SQUAREM method [38]. For comparison, Ignore
of LQM and TQM is non-increasing. To further show the PHN and Exact PHN are also included, where phase noise is
proposed algorithms converge to a stationary point, we first ignored and the exact phase noise is used, respectively, when
define a first-order optimality condition for minimizing a estimating channel impulse response. Whenever necessary, an
smooth function subject to an arbitrary constraint [37]. algorithm is initiated with an all-one vector and regarded
Proposition 1. Let f : RN → R be a smooth function. A converged when the squared ℓ2 -norm of the difference between
point x⋆ is a local minimum of f within a subset X ⊂ RN if two consecutive iterates is no larger than 10−8 . The maximum
number of iterations allowed toward the convergence is 500.
∇f (x⋆ )T y ≥ 0 ∀y ∈ TX (x⋆ ), (61)
All simulations were run in MATLAB on a PC with a 3.20
where TX (x⋆ ) is the tangent cone of X at x⋆ . GHz i5-4570 CPU and 8 GB RAM.
With Proposition 1, the convergence of our proposed al-
gorithms is guaranteed, i.e., the limit point of {u(t) } is a A. Phase noise and channel estimation
stationary point. A similar proof can be found in [33, Th. In this section, we define phase noise θ as a Wiener process
5]. [8], [19], [22]. The baseband sampling rate is fs = 20 MHz;
3-dB bandwidth ∆f3dB of phase noise is 500 Hz or 5000 Hz;
V. S IMULATION R ESULTS assuming CFO has been fixed, i.e., θ0 = 0, phase noise is
In simulations, we consider a Rayleigh fading channel of generated with
( √ )
length L = 10, where each tap is independently distributed 2π∆f3dB
with exponentially decreasing power of rate 0.7 and channel θn − θn−1 ∼ N 0, , (62)
fs
noise is circularly symmetric complex Gaussian with σ = 0.1.
Transmitted OFDM symbols are generated randomly (assumed for n = 1, . . . , Nc − 1. We first show four instances of
known to receiver) with distribution CN (0, 2I). To apply phase noise estimates to provide an intuitive idea of how
dimensionality reduction, we use PCT with the reduced di- different algorithms perform, and then compare the resultant
mension indicated by the value of N (51)–(52). Throughout phase noise and channel estimation errors by Monte Carlo
this section we choose N = 32 if a fixed PCT is applied. simulations.
The following methods are considered in our simulations. Figs. 1(a)–1(d) show phase noise estimates under four dif-
PNC, as the benchmark, is the algorithm proposed in [22], ferent scenarios. In all cases, PNC, TQM, and LQM yield the
where phase noise is estimated in the frequency domain by same estimate when the a fixed PCT is applied. NonIt, the non-
solving an SDP (23). A non-iterative method NonIt proposed iterative method, solves the original phase noise estimation
in [21] is also considered. Method AltOpt refers to the alternat- problem, assuming that the magnitude of phase noise is small,
ing optimization algorithm proposed in [19]. A modified ver- with a first-order Taylor approximation. Also assuming that
sion AltMM, based on the MM, is also compared. Specifically, phase noise is low pass, it recovers phase noise using only
we have modified AltOpt to take into account the phase noise several (3 in our simulations) low-pass spectral components.
constraint in each phase noise estimate update as opposed to Obviously, the obtained estimates by NonIt are only rough
the original algorithm; see Appendix for details. TQM and approximations compared with the true phase noise. Due to
LQM are two MM-based algorithms we have proposed to this reason, PNC is proposed in [22] to both take spectral
solve the time-domain problem (29). From a set of specified phase noise constraints into account and also make use of
PCTs with different values of N , opt-PCT is defined as the one more spectral components.

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2691672, IEEE
Transactions on Signal Processing
9

0.8 0.5
0.05
0.4
0
0.4
0.6
0.2 -0.05

0 0.3 -0.1
0.4
-0.15
-0.2
0.2
-0.2
0.2 20 40 60 80 100 120 20 40 60 80 100 120
Radian

Radian
0.1

0
0

-0.2 PNC + PCT PNC + PCT

AltOpt -0.1 AltOpt
NonIt NonIt
TQM + PCT (proposed) TQM + PCT (proposed)
-0.4 LQM + PCT (proposed) LQM + PCT (proposed)
TQM (proposed) -0.2 TQM (proposed)
TQM + opt-PCT (proposed) TQM + opt-PCT (proposed)
True PHN True PHN
-0.6 -0.3
1 100 200 300 400 500 600 700 800 900 1000 1 100 200 300 400 500 600 700 800 900 1000
Phase noise sample index Phase noise sample index

(a) (b)

5 1.5
1 0.2

4 0.5
1 0

0
3
0.5 -0.2
-0.5

2
-1 -0.4
20 40 60 80 100 120 0 20 40 60 80 100 120
Radian

Radian

-0.5
0

PNC + PCT -1 PNC + PCT

-1 AltOpt AltOpt
NonIt NonIt
TQM + PCT (proposed) TQM + PCT (proposed)
LQM + PCT (proposed) -1.5 LQM + PCT (proposed)
-2 TQM (proposed) TQM (proposed)
TQM + opt-PCT (proposed) TQM + opt-PCT (proposed)
True PHN True PHN
-3 -2
1 100 200 300 400 500 600 700 800 900 1000 1 100 200 300 400 500 600 700 800 900 1000
Phase noise sample index Phase noise sample index

(c) (d)
Fig. 1. Four instances of the estimated phase noise θ. Nc = 1024. Where PCT is applied, N = 32. (a) SNR = 15 dB, ∆f3dB = 500 Hz, (b) SNR = 35
dB, ∆f3dB = 500 Hz, (c) SNR = 15 dB, ∆f3dB = 5000 Hz, (d) SNR = 35 dB, ∆f3dB = 5000 Hz2 .

1) Small Phase Noise and Low SNR: In the small phase 2) Small Phase Noise and High SNR: A particular example
noise case with ∆f3dB = 500 Hz, as Fig. 1(a) shows, for this case is shown in Fig. 1(b). With the given PCT, we
using the given PCT results in a staircase-like estimate; loose can see that PNC, TQM, and LQM are still able to produce
though it may seem, it is actually beneficial when SNR is the same good estimate by and large, the resulting MSE of
limited, owing to the fundamental underdetermined issue of which is around 6.98. AltOpt offers a relatively better estimate
the original problem. In fact, TQM with opt-PCT provides the with MSE 0.6849, which is comparable to TQM without
same estimate as that of the benchmark—opt-PCT in this case PCT whose MSE is 0.7489. TQM with opt-PCT (opt-PNC
is T32 . In contrast, TQM without PCT and AltOpt produce is T256 here) outperforms other opponents though, with the
estimates with many undesired peaks associated with larger corresponding MSE of 0.2362, which implies that dimension
MSE. It implies, therefore, that with small phase noise and should not be reduced too much with high SNR. It should be
low SNR, dimensionality reduction is recommended in order mentioned that this will pose a challenge to the benchmark
to achieve a relatively better performance. PNC because it needs to deal with an SDP with size of 512
or even larger. More details will be illustrated in Section V-B.
2 The non-iterative method NonIt [21] has been favored for slow-varying
phase noise with a small number of OFDM subcarriers. For moderately fast-
varying phase noise or when the number of subcarriers is large, the estimates 3) Large Phase Noise and Low SNR: A major difference
obtained by NonIt are much worse than other methods; also see [17]. in this scenario from the previous one is that TQM without

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2691672, IEEE
Transactions on Signal Processing
10

10 3
0
10
PNC + PCT
TQM + PCT (proposed)
LQM + PCT (proposed)
2 TQM (proposed) 10 -1
10 TQM + opt-PCT (proposed)
TQM + Reg (proposed)

-2
10
MSE of Phase Noise Estimate

10 1

MSE of Channel Estimate

10 -3

0
10

10 -4

10 -1
10 -5 PNC + PCT
TQM + PCT (proposed)
LQM + PCT (proposed)
TQM (proposed)
10 -2 TQM + opt-PCT (proposed)
10 -6
TQM + Reg (proposed)
Ignore PHN
Exact PHN

10 -3 10 -7
15 20 25 30 35 40 45 50 55 15 20 25 30 35 40 45 50 55
SNR SNR

(a) (b)

10 3 10
0

PNC + PCT
TQM + PCT (proposed)
LQM + PCT (proposed)
TQM (proposed) 10 -1
TQM + opt-PCT (proposed)
TQM + Reg (proposed)
2
10
10 -2
MSE of Phase Noise Estimate

MSE of Channel Estimate

10 -3
1
10

10 -4

10 -5 PNC + PCT
0 TQM + PCT (proposed)
10
LQM + PCT (proposed)
TQM (proposed)
TQM + opt-PCT (proposed)
10 -6
TQM + Reg (proposed)
Ignore PHN
Exact PHN

10 -1 10 -7
15 20 25 30 35 40 45 50 55 15 20 25 30 35 40 45 50 55
SNR SNR

(c) (d)
Fig. 2. Joint estimation of phase noise and channel under different values of SNR with Nc = 1024 and 2000 Monte Carlo simulations: (a) averaged
MSE of phase noise estimate, ∆f3dB = 500 Hz (b) averaged MSE of channel estimate, ∆f3dB = 500 Hz, (c) averaged MSE of phase noise estimate,
∆f3dB = 5000 Hz, (d) averaged MSE of channel estimate, ∆f3dB = 5000 Hz.

PCT yields a phase noise estimate with many undesired perform better, the resulting MSE of which are 0.6938 and
peaks and so does AltOpt; see Fig. 1(c). The reason is that 0.5352, respectively. Opt-PCT in this example is T512 .
low SNR renders the original problem more susceptible to Since phase noise estimate by NonIt is much worse than
the underdetermined issue—the same as the case of small other methods and AltOpt and AltMM have MSE of phase
phase noise and low SNR in Fig. 1(a). Nonetheless, TQM noise no better than TQM, we will only compare our proposed
with opt-PCT gives a very good estimate outperforming other algorithms with PNC with respect to MSE. For comparison of
opponents (opt-PCT in this case is T128 ). computation time for each algorithm, see Table II.
4) Large Phase Noise and High SNR: In this last example, Figs. 2(a)–2(d) show averaged MSE of phase noise and
except NonIt, all the other algorithms yield nearly good channel estimates with 500 Monte Carlo trials. From Fig. 2(a)
estimates as Fig. 1(d) shows. Fixed PCT for PNC, TQM, and 2(c), we see that PNC, TQM, and LQM are comparable
and LQM, however, still provide a relatively loose result to each other with dimensionality reduction when SNR is low.
that could have been improved with available high SNR, the In contrast, TQM without PCT can provide much better phase
MSE of which are 9.9862, 9.9862, and 9.9870, respectively. noise estimates for high enough SNR. A similar result can
AltOpt yields an estimate with smaller MSE 1.8928. We can be found for channel estimation in Fig. 2(b) and 2(d). But
also expect that TQM without PCT and TQM with opt-PCT TQM produces better estimates with opt-PCT than without

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2691672, IEEE
Transactions on Signal Processing
11

1 5
10 10
TQM
LQM
TQM + PCT
LQM + PCT
TQM + Reg (µ=50)
TQM + Reg (µ=100)
0 4
10 10

Phase noise estimation objective

MSE of Phase Noise Estimate

10 -1 10 3

-2 2
10 10
TQM + PCT
TQM + opt-PCT
TQM + Reg (µ=100)
TQM + Reg (µ=200)
TQM + LI
TQM + opt-LI

10 -3 10 1
15 20 25 30 35 40 45 50 55 10 0 10 1 10 2
SNR Iteration

(a)
Fig. 4. Convergence of TQM and LQM. Nc = 1024. ∆f3dB = 500 Hz.
0
10

10 -1 The authors only prove the equivalence (strong duality) be-

tween the reduced problem and its SDP reformulation. For
10 -2
the original problem, however, the strong duality has not
been established. From simulations, the resultant estimate of
MSE of Channel Estimate

10 -3
frequency-domain phase noise vector is not a reasonably good
solution that satisfies the spectral constraint. Also, solving
an SDP only gives an intermediate solution that requires an
10 -4
additional eigendecomposition step. PNC can thus easily fall
within infeasibility and singularity issues when dimension is
10 -5 TQM + PCT
TQM + opt-PCT not reduced enough. These issues, however, have not been
TQM + Reg (µ=100)
TQM + Reg (µ=200) addressed in [22].
TQM + LI
10 -6
TQM + opt-LI
Ignore PHN
Exact PHN
B. Algorithm convergence
10 -7
15 20 25 30 35 40 45 50 55
SNR In this section, we first present an example of convergence
properties of our proposed algorithms, and then give a compar-
(b)
ison of CPU time consumed by each algorithm. Convergence
Fig. 3. Joint estimation of phase noise and channel with different values of criteria defined previously apply here as well.
SNR, Nc = 1024, ∆f3dB = 500 Hz, and 500 Monte Carlo simulations: (a)
averaged MSE of phase noise estimate, (b) averaged MSE of channel estimate. Fig. 4 demonstrates convergence of four methods. TQM and
LQM converge to the same optimal solution with the same ini-
tialization. TQM, however, converges remarkably faster than
PCT. Even though the benchmark PNC can deal with larger LQM, within twenty iterations or fewer. This is because TQM
SNR, the computational burden is prohibitive, not to say the employs a much tighter majorization function than LQM.
additional computational issues; see the following remark and Thus we adopt TQM with opt-PCT in previous simulations
an illustration of CPU time consumed by each algorithm in to achieve the same performance with respect to estimation
Section V-B. Using regularization, TQM can further reduce error and to save computation time at the same time. On the
MSE compared with TQM with PCT. Here, we choose reg- other hand, as shown in previous examples, applying PCT in
ularization parameter µ = 50 for ∆f3dB = 500 Hz and the case of large phase noise and high SNR causes loss of
µ = 10 for ∆f3dB = 5000 Hz. But for a given ∆f3dB , the quality in the obtained estimates, which is substantiated by
performance of resulting estimates is robust to the choice of µ. the fact that without PCT, much lower objective value can be
See Figs. 3(a)–3(b) for more details. The LI is the interpolation achieved. Using different values of regularization parameters,
matrix proposed in [19] and opt-LI, similar to opt-PCT, is TQM converges to different objective values with comparable
obtained with BIC. For LI to be properly defined, the sampling computation time. The resulting estimation errors, however,
length N is chosen from {34, 94, 342, 1024}, which is more are almost the same as Figs. 2(a)–2(b) have shown.
restricted than PCT. For fixed LI, we choose N = 34. A further comparison in terms of the computation time
Remark 2: PNC is proposed in [22], where the original among our proposed algorithms and the benchmarks is pro-
problem is reformulated as an SDP by using the S-procedure. vided in Table II. In the case with PCT applied, our proposed

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2691672, IEEE
Transactions on Signal Processing
12

TABLE II
CPU TIME OF DIFFERENT ALGORITHMS WITH ∆f3dB = 5000 H Z , SNR = 35 D B, AND 500 M ONTE
C ARLO SIMULATIONS†

CPU Time (s)

Nc = 512 Nc = 1024
Algorithms Mean Min Max Mean Min Max
PNC + PCT (benchmark) 0.4782 0.4297 0.5952 0.9139 0.8478 1.2091
AltOpt + PCT (benchmark) 0.6427 0.4915 0.9084 3.1677 2.6017 5.3007
AltMM + PCT (modified benchmark) 0.2299 0.1723 0.3232 0.8926 0.6785 1.3401
TQM (proposed) 0.0169 0.0155 0.0263 0.0845 0.0758 0.1047
LQM (proposed) 0.2467 0.0468 0.3667 1.5480‡ 0.3011 2.1801
TQM + PCT (proposed) 0.0125 0.0121 0.0145 0.0491 0.0483 0.0542
LQM + PCT (proposed) 0.0159 0.0154 0.0211 0.0637 0.0626 0.0700
TQM + opt-PCT (proposed) 0.0784 0.0753 0.1068 0.3694 0.3596 0.4090
TQM + Reg, µ = 50 (proposed) 0.0198 0.0168 0.0303 0.1296 0.0915 0.2852
TQM + Reg, µ = 100 (proposed) 0.0215 0.0169 0.0386 0.1528 0.0930 0.4229
† CPU time measured for each algorithm includes only computing phase noise estimate. For our
proposed algorithms, Fast Fourier transform (FFT) is employed to improve computational efficiency.
‡ LQM costs more CPU time when Nc is large. But some efficient acceleration schemes can be used
to boost the convergence rate, e.g., the SQUAREM method [38].

algorithms outperform PNC, AltOpt, and AltMM by saving 0.6

0.1
much time and at the same time achieve the same MSE as
0.5 0.05
shown in the previous examples. TQM gives as the same
estimate as LQM but is much more efficient owing to the 0.4
0

tighter majorization function employed. TQM with regular- -0.05

0.3
ization is also computationally efficient with an even further -0.1

improvement in the resulting estimation errors. 0.2 -0.15

20 40 60 80 100 120
Radian

0.1
C. Gaussian Phase Noise Estimation
0
Phase noise generated in a phase-locked loop is modeled
as a Gaussian process [18]. In this example, the number of -0.1
PNC + PCT
AltOpt
subcarriers is Nc = 512 with baseband sampling rate fs = 20 -0.2
NonIt
TQM + PCT (proposed)
MHz. The standard deviation θrms of phase noise generated by LQM + PCT (proposed)
TQM (proposed)
a phase-locked voltage controlled oscillator is 2 degrees. The -0.3
TQM + opt-PCT (proposed)
True PHN
single-pole butterworth filter with 3-dB bandwidth ∆f3dB = -0.4
100 Hz is adopted so that the covariance matrix of phase noise 1 100 200 300 400 500
Phase noise sample index
is ( )2
πθrms −2π∆f3dB |i−j|
Ci,j = e fs . (63)
180 Fig. 5. Gaussian phase noise estimation with Nc = 512 and SNR = 35 dB.
An instance of Gaussian phase noise and its estimates is
shown in Fig. 5. As its name indicates, Gaussian phase noise
will not drift away too much like Wiener phase noise. Similar
devised based on the majorization-minimization technique
to Fig. 1(b) and 1(d), large dimensionality reduction induces
and apply to two canonical models of phase noise—Wiener
considerable loss in the obtained estimates; with PCT, MSE
process and Gaussian process. To properly address the un-
for PNC, TQM, and LQM are 1.4049, 1.4049, and 1.4024,
derdetermined nature in the original estimation problem, di-
respectively. TQM without PCT and TQM with opt-PCT
mensionality reduction and regularization have been proposed
achieve the best performance in this example, both of which
with similar MM algorithms provided. The simulation results
have the same MSE of 0.1556. And opt-PCT is just an identity
have shown that when the same dimensionality reduction is
matrix, which causes no dimension to be reduced. Figs. 6(a)–
employed, our proposed algorithms achieve the same MSE
6(b) show MSE of the obtained estimates of Gaussian phase
as that of the benchmark but consume much less time. By
noise and channel, respectively. Still, our proposed algorithms
further selecting the optimal dimensionality reduction with
can provide much better estimates as in the previous examples
BIC or imposing appropriate regularization, our proposed
with Wiener phase noise.
algorithms produce significantly better estimates for moderate
SNR without demanding much additional computation time.
VI. C ONCLUSION It is expected that in modern applications of OFDM, where
We have proposed efficient algorithms for the joint phase a large number of subcarriers are deployed, the advantage of
noise and channel estimation in OFDM. The algorithms are our methods should be outstanding.

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2691672, IEEE
Transactions on Signal Processing
13

10
2
Algorithm 3 Phase Noise Estimation by Alternating Mini-
PNC + PCT mization and the MM.
TQM + PCT (proposed)
LQM + PCT (proposed) 1: Set i = 1 and initialize ĉ(0) = ejθ 0 .
TQM (proposed)

10
1 TQM + opt-PCT (proposed) 2: repeat
TQM + Reg (proposed) (i−1)
3: Φϕ = circ(Fĉ(i−1) )
( )−1 H H ( (i−1) )H
MSE of Phase Noise Estimate

4: ĥ = Nc F̌H SH SF̌
(i)
F̌ S Φϕ r
10 0
( (i)
)
5: P = circ SF̌ĥ
2
6: λ = Nc FH (SF̌ĥ(i) ) ∞
-1
7: t = 0, and initialize c(0)
10
8: repeat
9: a(t+1) = FH PH r + λc(t) − FH PH PFc(t)
(t+1)

10 -2
10: c(t+1) = ej arg a
11: until convergence
12: ĉ(i) = c(t+1)
10 -3
13: t←t+1
15 20 25 30 35
SNR
40 45 50 55
14: until convergence.

(a)
where P = circ(SF̌ĥ(i) ). Instead of updating c by the least-
0
10
squares solution (17), the MM method can be used to solve
problem (65). The majorization can be obtained as follows
10 -1
[33, Lemma 1]:
{ }
10 -2
∥r − PFc∥2 = rH r − 2Re rH PFc + cH FH PH PFc
MSE of Channel Estimate

(66)
{ H }
10 -3 ≤ r r − 2Re r PFc + λc c
H H
{( ) ( ) }
H
+ 2Re c(t) FH PH PF − λI c
10 -4
( )H ( )
+ c(t) λI − FH PH PF c(t) . (67)
10 -5 PNC + PCT
TQM + PCT (proposed)
LQM + PCT (proposed)
To obtain a good majorization function, we can choose λ as
TQM (proposed)
( )
10 -6 TQM + opt-PCT (proposed)
TQM + Reg (proposed) λ = λmax FH PH PF (68)
Ignore PHN ( ) 2
Exact PHN
10 -7
= Nc FH SF̌ĥ(i) , (69)
15 20 25 30 35 40 45 50 55 ∞
SNR
where Lemma 1 is applied to compute the maximum eigen-
(b) value of P. Minimizing (67) results in the update of c:
Fig. 6. Joint estimation of Gaussian phase noise and channel with different (t+1)
values of SNR, Nc = 512, and 500 Monte Carlo simulations: (a) averaged c(t+1) = ej arg a , (70)
MSE of phase noise estimate, (b) averaged MSE of channel estimate.
where a(t+1) = FH PH r + λc(t) − FH PH PFc(t) . The
complete procedure is described in Algorithm 3. Algorithm
A PPENDIX 3 can also be readily modified to incorporate PCT, for which
a similar majorization approach can be followed.
A LTERNATING O PTIMIZATION WITH THE MM
In the following, the constraint of phase noise is taken into R EFERENCES
account and the alternating optimization scheme is correspond-
[1] T. Pollet, M. V. Bladel, and M. Moeneclaey, “BER sensitivity of OFDM
ingly modified. systems to carrier frequency offset and Wiener phase noise,” IEEE Trans.
( With the unimodular constraint for c = ejθ , we have Commun., vol. 43, no. 2/3/4, pp. 191–193, Feb. 1995.
(i−1) )H (i−1) [2] L. Tomba, “On the effect of Wiener phase noise in OFDM systems,”
Φϕ Φϕ = Nc I (14). And the channel estimate is IEEE Trans. Commun., vol. 46, no. 5, pp. 580–583, May 1998.
updated by [3] A. G. Armada, “Understanding the effects of phase noise in orthogonal
)−1 H H ( (i−1) )H
frequency division multiplexing (OFDM),” IEEE Trans. Broadcast.,
( vol. 47, no. 2, pp. 153–159, June 2001.
ĥ(i) = Nc F̌H SH SF̌ F̌ S Φϕ r. (64) [4] C. Muschallik, “Influence of RF oscillators on an OFDM signal,” IEEE
Trans. Consum. Electron., vol. 41, no. 3, pp. 592–603, Aug. 1995.
Substitute (64) into the objective of problem (15), and the [5] T. H. Lee and A. Hajimiri, “Oscillator phase noise: A tutorial,” IEEE J.
following problem is obtained: Solid-State Circuits, vol. 35, no. 3, pp. 326–336, Mar. 2000.
[6] L. Piazzo and P. Mandarini, “Analysis of phase noise effects in OFDM
2 modems,” IEEE Trans. Commun., vol. 50, no. 10, pp. 1696–1705, Oct.
minimize ∥r − PFc∥ , (65) 2002.
c:|c|n =1,n=1,...,Nc

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2691672, IEEE
Transactions on Signal Processing
14

[7] K. Nikitopoulos and A. Polydoros, “Phase-impairment effects and [31] Y. Sun, P. Babu, and D. Palomar, “Majorization-minimization algorithms
compensation algorithms for OFDM systems,” IEEE Trans. Commun., in signal processing, communications, and machine learning,” IEEE
vol. 53, no. 4, pp. 698–707, Apr. 2005. Trans. Signal Process., vol. PP, no. 99, pp. 1–1, 2016.
[8] D. Petrovic, W. Rave, and G. Fettweis, “Effects of phase noise on OFDM [32] R. A. Horn and C. R. Johnson, Matrix analysis. Cambridge, U.K.:
systems with and without PLL: Characterization and compensation,” Cambridge Univ. Press, 2012.
IEEE Trans. Commun., vol. 55, no. 8, pp. 1607–1616, Aug. 2007. [33] J. Song, P. Babu, and D. P. Palomar, “Optimization methods for
[9] P. Mathecken, T. Riihonen, S. Werner, and R. Wichman, “Performance designing sequences with low autocorrelation sidelobes,” IEEE Trans.
analysis of OFDM with Wiener phase noise and frequency selective Signal Process., vol. 63, no. 15, pp. 3998–4009, Apr. 2015.
fading channel,” IEEE Trans. Commun., vol. 59, no. 5, pp. 1321–1331, [34] P. Stoica and Y. Selen, “Model-order selection: A review of information
May 2011. criterion rules,” IEEE Signal Process. Mag., vol. 21, no. 4, pp. 36–47,
[10] P. Mathecken, T. Riihonen, N. N. Tchamov, S. Werner, M. Valkama, July 2004.
and R. Wichman, “Characterization of OFDM radio link under PLL- [35] W. F. Trench, “An algorithm for the inversion of finite Toeplitz matrices,”
based oscillator phase noise and multipath fading channel,” IEEE Trans. J. Soc. Ind. Appl. Math., vol. 12, no. 3, pp. 515–522, 1964.
Commun., vol. 60, no. 6, pp. 1479–1485, June 2012. [36] J. F. Sturm, “Using SeDuMi 1.02, a MATLAB toolbox for optimization
[11] P. Robertson and S. Kaiser, “Analysis of the effects of phase-noise over symmetric cones,” Optim. Meth. Softw., vol. 11-12, pp. 625–653,
in orthogonal frequency division multiplex (OFDM) systems,” in 1995 1999.
IEEE Int. Conf. Commun. (ICC), vol. 3, Seattle, WA, June 1995, pp. [37] D. P. Bertsekas, A. Nedi, A. E. Ozdaglar et al., Convex analysis and
1652–1657. optimization. Belmont, MA.: Athena Scientific, 2003.
[12] K. Sathananthan and C. Tellambura, “Performance analysis of an OFDM [38] R. Varadhan and C. Roland, “Simple and globally convergent methods
system with carrier frequency offset and phase noise,” in IEEE 54th Veh. for accelerating the convergence of any EM algorithm,” Scandinavian
Technol. Conf., vol. 4, 2001, pp. 2329–2332. J. Statist., vol. 35, no. 2, pp. 335–353, 2008.
[13] S. Wu and Y. Bar-Ness, “OFDM channel estimation in the presence of
frequency offset and phase noise,” in 2003 IEEE Int. Conf. Commun.
(ICC), vol. 5, Anchorage, AK, May 2003, pp. 3366–3370.
[14] ——, “OFDM systems in the presence of phase noise: Consequences
and solutions,” IEEE Trans. Commun., vol. 52, no. 11, pp. 1988–1996, Zhongju Wang received the B.Eng. degree in com-
Nov. 2004. munication engineering from the University of Elec-
[15] P. H. Moose, “A technique for orthogonal frequency division multiplex- tronic Science and Technology of China (UESTC),
ing frequency offset correction,” IEEE Trans. Commun., vol. 42, no. 10, Chengdu, China, in 2013. He is currently pursuing
pp. 2908–2914, Oct. 1994. the Ph.D. degree in electronic and computer engi-
[16] V. Syrjala, M. Valkama, N. N. Tchamov, and J. Rinne, “Phase noise mod- neering at the Hong Kong University of Science
elling and mitigation techniques in OFDM communications systems,” in and Technology (HKUST), Hong Kong SAR. His
2009 Wireless Telecommun. Symp., Apr. 2009, pp. 1–7. research interests include applications of convex
[17] P. Mathecken, T. Riihonen, S. Werner, and R. Wichman, “Constrained optimization to signal processing, machine learning,
phase noise estimation in OFDM using scattered pilots without decision and financial engineering.
feedback,” IEEE Trans. Signal Process., vol. 65, no. 9, pp. 2348–2362,
May 2017.
[18] D. D. Lin, R. A. Pacheco, T. J. Lim, and D. Hatzinakos, “Joint estimation
of channel response, frequency offset, and phase noise in OFDM,” IEEE
Trans. Signal Process., vol. 54, no. 9, pp. 3542–3554, Sept. 2006.
[19] Q. Zou, A. Tarighat, and A. H. Sayed, “Compensation of phase noise in
OFDM wireless systems,” IEEE Trans. Signal Process., vol. 55, no. 11, Prabhu Babu received the Ph.D. degree in electrical engineering from the
pp. 5407–5424, Nov. 2007. Uppsala University, Sweden, in 2012. From 2013 to 2016, he was a Post-
[20] R. Carvajal, J. C. Aguero, B. I. Godoy, and G. C. Goodwin, “EM- Doctorial Fellow with the Hong Kong University of Science and Technology.
based maximum-likelihood channel estimation in multicarrier systems He is currently with the Center for Applied Research in Electronics (CARE),
with phase distortion,” IEEE Trans. Veh. Technol., vol. 62, no. 1, pp. Indian Institute of Technology Delhi, New Delhi, India.
152–160, Jan. 2013.
[21] P. Rabiei, W. Namgoong, and N. Al-Dhahir, “A non-iterative technique
for phase noise ICI mitigation in packet-based OFDM systems,” IEEE
Trans. Signal Process., vol. 58, no. 11, pp. 5945–5950, Nov. 2010.
[22] P. Mathecken, T. Riihonen, S. Werner, and R. Wichman, “Phase noise Daniel P. Palomar (S’99-M’03-SM’08-F’12) re-
estimation in OFDM: Utilizing its associated spectral geometry,” IEEE ceived the Electrical Engineering and Ph.D. degrees
Trans. Signal Process., vol. 64, no. 8, pp. 1999–2012, Apr. 2016. from the Technical University of Catalonia (UPC),
[23] Frame structure channel coding and modulation for a second generation Barcelona, Spain, in 1998 and 2003, respectively.
digital terrestrial television broadcasting system (DVB-T2), ETSI Std. He is a Professor in the Department of Electronic
EN 302 755 V1.3.1, 2009. and Computer Engineering at the Hong Kong Uni-
[24] R. A. Casas, S. L. Biracree, and A. E. Youtz, “Time domain phase noise versity of Science and Technology (HKUST), Hong
correction for OFDM signals,” IEEE Trans. Broadcast., vol. 48, no. 3, Kong, which he joined in 2006. Since 2013 he is
pp. 230–236, Sept. 2002. a Fellow of the Institute for Advance Study (IAS)
[25] F. Septier, Y. Delignon, A. Menhaj-Rivenq, and C. Garnier, “OFDM at HKUST. He had previously held several research
channel estimation in the presence of phase noise and frequency offset appointments, namely, at King’s College London
by particle filtering,” in Proc. 2007 IEEE Int. Conf. Acoust. Speech (KCL), London, UK; Stanford University, Stanford, CA; Telecommunica-
Signal Process. (ICASSP 2007), vol. 3, Honolulu, HI, Apr. 2007, pp. tions Technological Center of Catalonia (CTTC), Barcelona, Spain; Royal
289–292. Institute of Technology (KTH), Stockholm, Sweden; University of Rome “La
[26] ——, “Monte Carlo methods for channel, phase noise, and frequency Sapienza”, Rome, Italy; and Princeton University, Princeton, NJ. His current
offset estimation with unknown noise variances in OFDM systems,” research interests include applications of convex optimization theory, game
IEEE Trans. Signal Process., vol. 56, no. 8, pp. 3613–3626, Aug. 2008. theory, and variational inequality theory to financial systems, big data systems,
[27] D. Tse and P. Viswanath, Fundamentals of wireless communication. and communication systems.
Cambridge, U.K.: Cambridge Univ. Press, 2005. Dr. Palomar is an IEEE Fellow, a recipient of a 2004/06 Fulbright Research
[28] R. M. Gray, “Toeplitz and circulant matrices: A review,” Found. Trends Fellowship, the 2004 and 2015 (co-author) Young Author Best Paper Awards
Commun. Inf. Theory, vol. 2, no. 3, pp. 155–239, 2006. by the IEEE Signal Processing Society, the 2015-16 HKUST Excellence
[29] D. R. Hunter and K. Lange, “A tutorial on MM algorithms,” Amer. Research Award, the 2002/03 best Ph.D. prize in Information Technologies
Statist., vol. 58, no. 1, pp. 30–37, 2004. and Communications by the Technical University of Catalonia (UPC), the
[30] A. Beck and M. Teboulle, “Gradient-based algorithms with applications 2002/03 Rosina Ribalta first prize for the Best Doctoral Thesis in Information
to signal recovery,” in Convex Optimization in Signal Processing and Technologies and Communications by the Epson Foundation, and the 2004
Communications, D. Palomar and Y. Eldar, Eds. Cambridge, U.K.: prize for the best Doctoral Thesis in Advanced Mobile Communications by
Cambridge Univ. Press, 2009, pp. 42–88. the Vodafone Foundation and COIT.

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.