Maximum-Likelihood Magnitude-Based Amplitude and Noise Variance Estimation
Maximum-Likelihood Magnitude-Based Amplitude and Noise Variance Estimation
28, 2021
Abstract—Maximum likelihood (ML) amplitude and noise vari- and covariance E[n(k)n∗ (k − j)] = 2σ 2 δ(j), where δ is the
ance estimation without having to jointly estimate the frequency Kronecker delta. It is assumed that there is no prior knowledge
and the phase and based only on information from the noisy re- on the unknown nonrandom parameters A, ω and θ, i.e., pa-
ceived signal magnitude, is studied for a single sinusoid in complex rameter A can be any positive, real number, and the modulo
additive white Gaussian noise. This estimation problem is equiva-
lent to the classic problem of parameter estimation for the Rician
−2π reduced ω and the modulo −2π reduced θ can take any
distribution. While solving the likelihood equation is impossible in value over the interval [−π, π). With these assumptions, the
general, we propose a new approach based on a large argument most common estimation approach is to apply the theory of
approximation. For the case with known noise variance, a closed- maximum likelihood (ML) for joint amplitude, frequency and
form ML amplitude estimator is obtained, which outperforms phase estimation. However, due to the nonlinear nature of the
the conventional root-mean-square estimator. For the case with problem, there is so far no explicit, closed-form solution for the
unknown noise variance, the closed-form joint amplitude and noise absolute maximum of the likelihood function. This has led to
variance estimators obtained do not require prior knowledge of one the fast-Fourier-transform-based frequency domain solution [1]
another. which, through the sequential computation of [1, eqs. (29), (30)
Index Terms—Amplitude estimation, cramer-rao bound, and (28)], results in the estimates of ω, θ and A, in the order
maximum likelihood (ML) estimation, noise variance estimation, as mentioned. Hence, the optimal ML estimate  can only be
rician distribution. obtained after obtaining the optimal ML estimates ω̂ and θ̂. It
is also known that there are some obstacles in estimating ω
and θ, such as outliers, computational complexity and phase
I. INTRODUCTION unwrapping, etc [1]–[3]. Therefore, it would be desirable to find
STIMATING the unknown parameters A, ω and θ from the an estimator  that is independent of ω̂ and θ̂ for convenience.
E noisy, discrete-time, complex data samples r(k), where This can be achieved by making use of only the magnitude
information of the data samples, i.e. {|r(k)|}k .
r(k) = Aej(ωk+θ) + n(k), k = 0, 1, 2, . . . (1) The magnitude-based amplitude estimation problem is equiv-
is an important problem in communication engineering. The alent to the classic problem of parameter estimation for the
additive, white Gaussian noise (AWGN) {n(k)} is a sequence Rician distribution [4]. For example, in otoacoustic emission
of circularly symmetric, independent and identically distributed (OAE) measurement [5] with signal model r(k) = Aejθ +
(i.i.d.), complex, Gaussian random variables with mean zero n(k), the measurement noise n(k) is found to be reasonably
well approximated by a circularly symmetric Gaussian noise,
Manuscript received August 17, 2020; revised January 5, 2021; accepted
but the sound pressures are typically reported as magnitudes
January 13, 2021. Date of publication January 29, 2021; date of current version without knowledge of the phase, i.e. {|r(k)|}k [5]. Estimation
March 4, 2021. The work was supported by the National Natural Science of magnitude A using {|r(k)|}k is equivalent to estimation of
Foundation of China under Grant 61571316. The associate editor coordinating the line-of-sight component of the Rician distribution. Similarly,
the review of this manuscript and approving it for publication was Dr. Luis the problem lies in areas like magnetic resonance imaging [6]–
Antonio Azpicueta-Ruiz. (Corresponding author: Yan Li.)
Ming-Wei Wu is with the School of Information and Electronic Engineering, [9], speech processing [10], synthetic aperture radar [11], [12],
Zhejiang University of Science and Technology, Hangzhou 310023, China (e- sonar [13] and channel modeling [14], [15].
mail: [email protected]). ML amplitude estimation is proposed in [7]. However, due
Yan Jin is with the School of Information and Electronic Engineering, Zhe- to the complicated form of the derived ML condition, existing
jiang University of Science and Technology, Hangzhou 310023, China, and also
with the Center for Advanced Microelectronic Devices, NUS (Suzhou) Research
methods are not able to find explicit solutions. Therefore, numer-
Institute (NUSRI), Suzhou 215123, China (e-mail: [email protected]). ical methods are resorted to, such as the iterative Brent’s algo-
Yan Li is with the Department of Electrical and Computer Engineering rithm in [7] and search using numerical optimization in [8], both
(ECE), National University of Singapore (NUS), Singapore 117583 (e-mail: of which have high computational complexity. In addition, [6]–
[email protected]). [8] all require prior knowledge of the noise variance parameter
Tianyu Song was with the Department of Electrical and Computer Engineer-
ing (ECE), National University of Singapore (NUS), Singapore 117583, and σ 2 . For noise variance estimation, [16] proposed a magnitude-
also with the Center for Advanced Microelectronic Devices, NUS (Suzhou) based estimator using the method of moments with quadratic
Research Institute, Suzhou 215123, China. He is now with Huawei Device, sample mean and fourth-power sample mean, but it does not
Shenzhen 518129, China (e-mail: [email protected]). attempt to optimize the performance. A joint ML amplitude and
Pooi-Yuen Kam is with the School of Science and Engineering, Chinese
University of Hong Kong, Shenzhen 518172, China, and also with the Center for
noise variance estimator is proposed in [8] by maximizing the
Advanced Microelectronic Devices, NUS Research Institute (NUSRI), Suzhou two-dimensional log-likelihood function of A and σ 2 . Again,
215123, China (e-mail: [email protected]). its optimization cannot be solved analytically and numerical
Digital Object Identifier 10.1109/LSP.2021.3055464 optimization with high complexity must be resorted to.
1070-9908 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: National University Fast. Downloaded on October 31,2024 at 08:48:55 UTC from IEEE Xplore. Restrictions apply.
WU et al.: MAXIMUM-LIKELIHOOD, MAGNITUDE-BASED, AMPLITUDE AND NOISE VARIANCE ESTIMATION 415
Â2 M = A2 + 2σ 2 . (18)
where
where ÂSMm is given in (6) and Â2 M is given in (7). The solution
N −1
in (19) with the − sign results in a small value and the solution in 1
(20) with the + sign results in a large value, which are unrealistic. Â4 M = |r(k)|4 . (24)
N
Therefore, the real estimates with the + sign in (19) and the − k=0
sign in (20) are chosen, i.e. The CRLB for noise variance estimation is given by CRLBn =
σ 4 /N [8, eq. (33)]. Fig. 4 shows a gap between the two
2 1
Âunr = ÂSMm + 4Â2SMm − 3Â2 M , (21) magnitude-based noise variance estimators and CRLBn . Our
3 3 new estimator σ2 outperforms the conventional σ2
mr when the
r
2
2 4 2 2 2
SNR is greater than 0 dB.
σr = Â2 M − ÂSMm − ÂSMm 4ÂSMm − 3Â2 M .
3 9 9
(22) V. CONCLUSION
We have addressed the problem of estimating the amplitude
The joint ML amplitude estimator in (21) and the noise variance of a single sinusoid observed in Gaussian noise, based on mea-
estimator in (22) work independently and require only informa- surements of the received signal magnitude alone. While this
tion of the magnitude |r(k)|. estimation problem is simpler when the noise variance is known,
Fig. 3 shows that the MSE performance of the amplitude the joint estimation of the amplitude and the noise variance is
estimator Âunr in (21) is much better than that of the magnitude not difficult also when the noise variance is unknown. Consistent
sample mean estimator ÂSMm without the knowledge of σ 2 at application of the ML estimation principle together with accu-
all SNR values. rate analytical approximations of the likelihood equation leads
In comparison, a conventional noise variance estimator using to simple estimators that show better performance than those of
only knowledge of the received signal magnitude is obtained existing ad hoc estimators and are applicable to fields in [9]–[13]
by approximating the second and fourth absolute moments with where the SNR is typically above 0 dB.
Authorized licensed use limited to: National University Fast. Downloaded on October 31,2024 at 08:48:55 UTC from IEEE Xplore. Restrictions apply.
418 IEEE SIGNAL PROCESSING LETTERS, VOL. 28, 2021
Authorized licensed use limited to: National University Fast. Downloaded on October 31,2024 at 08:48:55 UTC from IEEE Xplore. Restrictions apply.