Introduction to special issue on compressive sensing in acoustics

Peter Gerstoft, Christoph F. Mecklenbräuker, Woojae Seong, and Michael Bianco

Citation: The Journal of the Acoustical Society of America 143, 3731 (2018); doi: 10.1121/1.5043089
View online: https://doi.org/10.1121/1.5043089
View Table of Contents: http://asa.scitation.org/toc/jas/143/6
Published by the Acoustical Society of America
Introduction to special issue on compressive sensing in acoustics
Peter Gerstoft,1,a) Christoph F. Mecklenbra
€ uker,2 Woojae Seong,3 and Michael Bianco1
1
Noise Lab, University of California San Diego, La Jolla, California 92093-0238, USA
2
Institute of Telecommunications, Vienna University of Technology, 1040 Vienna, Austria
3
Department of Naval Architecture and Ocean Engineering, Seoul National University, Seoul, South Korea
(Received 25 May 2018; revised 1 June 2018; accepted 5 June 2018; published online 29 June 2018)
Compressive sensing (CS) in acoustics has received significant attention in the last decade, and thus
motivates this special issue. CS emerged from the signal processing and applied math community
and has since generated compelling results in acoustics. This special issue primarily addresses the
acoustics CS topics of compressive beamforming and holography. For a sound field observed on a
sensor array, CS reconstructs the direction of arrival of multiple sources using a sparsity constraint.
Similarly, in holography a sparsity constraint gives improved sound field reconstruction over
conventional ‘2-regularization. Other topics in this issue include sparse array configurations (as
co-arrays) and sparse sensing in acoustic communication. V C 2018 Acoustical Society of America.

https://doi.org/10.1121/1.5043089
[JFL] Pages: 3731–3736

I. INTRODUCTION measured waveforms are linearly related to a physical source

model x 2 CN via a sensing matrix A 2 CMN . Ignoring
Reconstructing acoustic fields from a series of signal
additive noise, the linear relation is
measurements is a standard task in acoustics. Without prior
knowledge, there is no way to reconstruct a signal between y ¼ Ax: (1)
the measured samples. With suitable prior knowledge, a sig-
nal can be perfectly reconstructed from measured samples. In sparse and CS models, the dimension of x is much larger
For example, the Nyquist and Shannon1,2 sampling theorem than the observational vector y, N  M. If x is sparse this
demonstrated that with constraints on the signal’s frequency means that only K of it elements are non-zero, where K  N.
content, fewer samples are needed for perfect reconstruction. If x is compressible, it means that it is approximately sparse,
Around 2006, Candes et al. proved that prior knowledge i.e., x can be well approximated by a sparse physical source
about a signal’s sparsity enables perfect reconstruction with model.
far fewer samples than the NyquistShannon sampling theo- Here, A is the sensing matrix, with the element Amn
rem requires.3,4 This result inspired the recent paradigm of relating the waveform at the mth receiver to the physical
compressive sensing (CS) and sparse sampling, in which sig- source model coefficient xn. The wave propagation environ-
nal processing techniques for efficiently acquiring and recon- ment and source model define the form of Amn. For example,
structing a signal are used. CS has received much attention the formulation of the element Amn may be based on a
for its ability to estimate sparse model parameters with great wave’s travel time smn between the nth source and the mth
accuracy using few observations.5,6 CS and related sparse receiver,
algorithms rely on tools such as signal processing, optimiza- 1
tion theory, probability theory, etc., to obtain sparse solutions Amn ¼ pffiffiffiffiffi ejxðsmn sm1 Þ : (2)
M
to linear systems, which are often highly underdetermined.
For further details, see the excellent books.7,8 In beamforming, the nth column an is a called the steering
CS has been an active field in the last decade, with many vector. Here, sm1 is the reference time, receiver 1 is a refer-
results in acoustics. A common topic is sparsity in different ence receiver. If the wave propagation environment is known
domains, e.g., time domain,9,10 frequency domain,11 spatial accurately, expressions for Amn can be derived from the cor-
domain, wavenumber domains,12,13 spatial beam domain responding Green’s function to the wave equation.
(beamforming).14–18 CS has also been used in recent develop- Usually, the sensing matrix A has many more columns
ments in matched field processing.19–21 An Ocham’s razor than rows, and is therefore not invertible. Thus, the recovery
approach to acoustic inversion is pleasing, thus the sparse of the physical source model x is ill-posed. However, the
sensing approach has found use in ocean acoustic inversion prior assumption that the signal x is sparse or compressible
for extracting sparse perturbations to a baseline model.22,23 renders the problem feasible, and thus x can be well
recovered.
II. COMPRESSIVE SENSING AND SPARSE MODELS To further illustrate, A can be written as a product of a
In seismic and acoustic applications, we measure the selection matrix W 2 CMN and a basis matrix U 2 CNN ,
waveform y 2 CM on M receivers. We assume that the as shown in Fig. 1. In frequency domain processing, the
basis matrix U relates the complex frequency weights x
2 CN due to a few active tones and the time domain wave-
a)
Electronic mail: [email protected] forms z 2 CN :

J. Acoust. Soc. Am. 143 (6), June 2018 0001-4966/2018/143(6)/3731/6/$30.00 C 2018 Acoustical Society of America
V 3731
 The recovery of the sparse physical model may then be
formulated as a constrained minimization program. We seek
the solution x with least squared error to the linear relation
(1) subject to the constraint that at most K elements of x are
non-zero,

min ky  Axk22 subject to kxk0  K: (5)

x2CN

The minimization program (5) is non-convex and NP hard,

though many numerical approaches exist, see, e.g., Refs. 7
and 8. A simple approach, called thresholding, is based on
finding the K strongest peaks in the objective function of Eq.
(5). Another simple approach, orthogonal matching pursuit
(OMP), solves Eq. (5) greedily, by finding peaks one-by-one
and eliminating the associated responses from the observa-
tions. Alternatively, the ‘0-norm constraint Eq. (5) may be
replaced by the ‘1-norm constraint kxk1  g which is a con-
vex approximation to the ‘0-norm,

min ky  Axk22 subject to kxk1  g: (6)

x2CN

Here g is a user defined value related to the sum of the signal

amplitudes. This approach is known as the LASSO.24 Using
Lagrange multipliers, Eq. (6) can be reformulated

minN ky  Axk22 þ kkxk1 ; (7)

x2C

where k > 0 is related to g. The breakthrough of CS (Refs. 7

and 8) came with the proof that, for sufficiently sparse sig-
nals, K  N, and sensing matrices with sufficiently incoher-
ent columns, the ‘0-norm minimization problem Eq. (5) is
equivalent (at least in the noiseless case) to its convex relax-
ation, the ‘1-norm minimization problem Eqs. (6) and (7),
FIG. 1. (Color online) Basic compressive sensing (CS) model with (a) ran- Refs. 5 and 6. By replacing the ‘0-norm with the convex ‘1-
dom and (b) deterministic selection matrices. norm, the problem can be solved efficiently with convex
optimization, even for large dimensions.25–27
Since its introduction in 1996, the LASSO and its
z ¼ Ux: (3)
numerous descendants have become common techniques.
The LASSO has a well-known Bayesian interpretation,
The above equation represents an inverse Fourier transform.
where the ‘1-norm is obtained from a Laplacian prior on x,
In CS, only a few components of the frequency vector x are
together with a Type-I likelihood approach. In contrast,
non-zero. Mathematically, this corresponds to the constraint
sparse Bayesian learning (SBL) uses Type-II likelihood (also
kxk0  K, where k  k0 is the ‘0-norm which counts the num-
ber of non-zero elements. Using a random selection matrix called evidence) with a Gaussian prior.28,29
W, we purge unnecessary measurements,
III. EXAMPLES
y ¼ Ax ¼ WUx ¼ Wz: (4) We motivate the usefulness of CS in acoustics with two
simple examples.
Thus, y results from less dense sampling than Nyquist
demands. This sampling configuration is shown in Fig. 1(a). A. Challenging Nyquist sampling
In the sensor array interpretation, z represents virtual
measurement data at very dense spatial sampling. With prior information that an acoustic signal is sparse,
Measuring z is generally impractical due to hardware limi- it is possible to sample less than required by Nyquist sam-
tations, and the sensor processing should be based on the pling. In this example, the signal is sampled with 190/
practical sensor array. Hence, this gives a single realiza- 1024 ¼ 20% of Nyquist sampling. Consider a physical
tion of the selection matrix W. The selection matrix is then source model composed of K ¼ 3 sinusoids with frequencies
modeled as deterministic, as indicated in the example in (f1, f2, f3) ¼ (100, 250, 400)/1024 ¼ (0.097, 0.24, 0.39), as
Fig. 1(b). shown in Fig. 2(a). The time domain waveform becomes

3732 J. Acoust. Soc. Am. 143 (6), June 2018 Gerstoft et al.
 X
K yðtÞ ¼ cos 2pf0 t þ cos 2pf1 t: (11)
zðtÞ ¼ sin ð2pfk tÞ ; (8)
k¼1 We first solve for the frequency components in Eq. (11)
using a M ¼ 64 point Fast Fourier Transform (FFT) [Fig.
where z ¼ ½zðt1 Þ; …; zðtN ÞT contains the samples of the time 3(b)]. Clearly, the two frequencies are a half-bin width
domain waveform z(t) for t ¼ 1,…, N. We choose N ¼ 1024, apart—too close together to be resolved per Nyquist sam-
as shown in Fig. 2(b). Next, we choose a single realization pling. Classic zero padding does not increase the frequency
of the random selection matrix W with M ¼ 190 random resolution. To illustrate this, the number of bins are
entries. This down-selects the complex frequency weights by increased from M ¼ 64 to M ¼ 256 by adding zeros after the
forming y ¼ Wz using Eq. (3) [blue circles in Fig. 2(b)]. 64th sample. Even with twice the resolution required for the
From Eq. (3), we can then evaluate the least square inversion sinusoid frequencies (11), the FFT does not recover the true
using the MoorePenrose left inverse Aþ of A, spectrum.
With the sparse model (5), the true spectrum is well recov-
x^ ¼ Aþ y ¼ AH ðAAH Þ1 y: (9)
ered using the FFT matrix (10) and K ¼ 2. The FFT matrix is
This gives the least square inversion frequency estimate x^ in constructed with M ¼ 128 (64 samples spread evenly over fre-
Fig. 2(c), from which the time series ^z ¼ U^x is obtained in quencies f < 0.5). CS is sensitive to aliasing in this setup, thus
Fig. 2(d). We find a sparse estimate using OMP in Fig. 2(e), using upper frequency f ¼ 1 would have required at least 128
from which the time series in Fig. 2(f) is obtained. The samples. An easy approach to solving for the spectrum from
sparse solution resembles the original data better than least (5) is the greedy OMP method. OMP cannot easily find two
squares, both in frequency domain (Fig. 2, left column) and peaks with the same amplitude and thus estimates the peaks
poorly, see Fig. 3(c). The more sophisticated sparse Bayesian
time domain (Fig. 2, left column).
learning (SBL)28,30 recovers the solution perfectly, see Fig.
B. Fourier transform example 3(d). The amplitude error is due to the finite sample size.

In a simple Fourier transform example, the sensing IV. THIS ISSUE

matrix is
This special issue Compressive sensing in acoustics con-
Amn ¼ ej½2pmðn1Þ=M : (10) tains 26 papers, which we have divided into five main topics.
These topics are summarized in the following.
We consider a N ¼ 64 sample signal consisting of two
cosines ðf0 ; f1 Þ ¼ 12 ð18; 19Þ=2M, with M ¼ 64, as shown in A. Holography
Fig. 3(a),
In early acoustic holography31 and near field acoustic
holography (NAH)32 works, solutions were based on least-
squares. Subsequently, Tikhonov regularization was adopted
to overcome ill-conditioning of the HelmholtzKirchoff
integral equation. Modern sparse regularization has recently

FIG. 2. (Color online) CS and least squares recovery of sinusoidal signals (3

sinusoids f ¼ 100/M, 250/M, 400/M, with M ¼ 1024 samples). Only samples
200–400 are shown. (a) True frequency spectrum of time series in (b). (b)
From the full measurement (red line) a random selection of N ¼ 190 points FIG. 3. (Color online) Resolving Fourier components with FFT and CS. (a)
(blue circles) is performed. For the random selection points, (c) Fourier N ¼ 64 sample signal constructed from two cosines (11). (b) Frequency
spectrum obtained by least squares and (d) corresponding reconstructed full response obtained by FFT M ¼ 64 (blue) and with zero-padding to a total
signal. For the random selection points, (e) Fourier spectrum obtained by measurement vector M ¼ 256 (red). Frequency response obtained by (c)
OMP sparse processing, and (f) corresponding reconstructed full signal. orthogonal matching pursuit (OMP) and (d) sparse Bayesian learning (SBL).

J. Acoust. Soc. Am. 143 (6), June 2018 Gerstoft et al. 3733
been introduced in acoustic holography.33–35 The formula- spectrogram frames, resulting in a high dimensional and
tion of sound field reconstruction is diverse, and depends on sparse super-frame-based representation.
the source characteristic and data measuring system. Thus Ramirez et al.45 develop an acoustic analog to the sin-
the solution methods employing CS are also diverse. gle-pixel-imager.46 Incoming acoustic signals are filtered
Fernandez-Grande and Daudet36 examine block-sparse through a passive imaging screen and subsequently received
methods for reconstructing acoustic near-field sound fields at a single acoustic receiver. Then CS is applied to localize a
based on total variation regularization. Numerical and source in a two-dimensional (2-D) waveguide. Mutual
anechoic chamber experiments indicate this method is suit- coherence of the sensing matrix is utilized as a metric to pre-
able both for compact and even spatially extended sources. dict the performance.
Bai and Chung37 use CS with block sparsity constraints Haltmeier et al.47 employ CS to reduce the number of
with an equivalent source method (ESM) formulation for the measurements in photoacoustic tomography. To ensure spar-
source identification problem. They propose an iterative CS sity, the second time derivative is applied to the measured
algorithm based on Newton’s method. Accurate localization pressure.
is shown in an anechoic chamber experiment using a planar
microphone array and various sources (loudspeaker and C. Beamforming and DOA estimation
fans). Hald38 performs NAH with data from a planar array In beamforming, direction-of-arrival (DOA) estimation
with randomly placed sensors. Based on ESM, five iterative refers to the localization of one or more sources from noisy
CS algorithms are developed. The algorithms are applied to wavefield measurements on an array of sensors. DOA esti-
synthetic data from monopole sources, and a point driven mation can be expressed as a linear underdetermined prob-
plate (not sparse) showing improved accuracy and less com- lem with a sparsity constraint enforced on its solution. The
putation time than CVX (Ref. 25) implementations. CS framework asserts that this is solved efficiently and accu-
Verburg and Fernandez-Grande39 estimate the frequency rately with a convex optimization procedure that promotes
response at arbitrary points inside a room using a plane wave sparse solutions.
expansion. Assuming spatial sparsity (valid for damped Applications in DOA estimation are obvious as many
rooms), the plane wave amplitudes are inverted via CS. The arrays observe signals from distinct directions. Early CS
reconstruction accuracy is analyzed in an experimental study. DOA applications are based on processing on single snap-
At low frequencies, accurate reconstructions are possible shots.14–16,48 Further work using multiple snapshots,17 dem-
with limited measurements. onstrate a high resolution capability even with coherent
Koyama et al.40 propose a method for sound reproduc- sources.49 The CS beamforming approach assumes that the
tion above the spatial Nyquist limit, which decomposes the sources are located on the same grid as the weight vector.
wave field into near-field source and far-field plane wave While this CS beamforming formulation could be performed
components. Basis functions representing the spatially sparse on a fine grid, this shortcoming will always give an unknown
near-field are then found using CS. These basis functions error. This concern has lead to grid-free methods in
must satisfy the Helmholtz equation, which makes them acoustics.18
highly correlated. Conventional CS techniques suffer from the basis mis-
Attendu and Ross41 reconstruct transient acoustic fields match issue, which is addressed in grid-free CS.18 Park et al.50
by combining time-domain NAH with an alternating direc- present a grid-free CS DOA estimation technique for multiple
tion CS algorithm. Experimental measurements from an snapshot data. For stationary DOAs, multiple snapshots pro-
impacted plate are used for comparison with reconstructed vide stable estimates, and a generalized total variation norm
transient signal. imposes a common sparsity pattern in the continuous angular
domain. Yang et al.51 develop a grid-free compressive beam-
B. Acoustical physics forming strategy for two-dimensional DOA estimation with
planar microphone arrays.
Dictionary learning, a form of unsupervised machine Tabassum and Ollila52 propose DOA recovery from a
learning, learns sparse signal representations from measured single-snapshot by augmenting CS beamforming with a
signal examples.7 It can improve sound-speed profile resolu- complex-valued path-wise weighted elastic net. In an elastic
tion by generating a dictionary of shape functions for sparse net, an additional constraint kxk2 with an appropriate
processing (e.g., CS) that optimally compress ocean sound- Lagrange multiplier is included in Eq. (7). Paik et al.53
speed profiles.42 In Alguri et al.,43 a dictionary learning frame- develop a covariance fitting algorithm for DOA estimation
work is used to detect damage with surrogate information. The of multiple, spatially sparse incident signals.
framework combines wave propagation characteristics of a With conventional DOA estimators, strong acoustic
test structure with geometric information from surrogate struc- interference influences localization accuracy and may even
tures to create a synthetic damage-free baseline. mask the DOAs of interest. In Yang et al.,54 a DOA estimator
Thakur et al.44 use a multi-layer, alternating sparse- based on sparse spectrum fitting is proposed and analyzed to
dense framework for bird species identification. From audio resolve closely spaced wideband signals in the presence of
recordings of bird vocalizations, they produce compressed strong interference.
convex spectral embeddings. Temporal and frequency mod- Tuna et al.55 use CS to map urban soundscapes with a
ulations in bird vocalizations are combined by concatenating portable microphone array. As the array traverses the

3734 J. Acoust. Soc. Am. 143 (6), June 2018 Gerstoft et al.
soundscape, far-field delay-and-sum beamforming power is demonstrate improved receiver performance with several
stacked for tomographic reconstruction. The multi-frequency proposed channel estimators. By modeling time varying
2-D beamforming is solved via group LASSO and multi- ocean acoustic channels as sparse with constant or time-
frequency SBL. High-resolution noise mapping is demon- varying supports, Jiang et al.67 show that the estimation of
strated on data from two concentric circular arrays on a the time-varying ocean acoustic channel is transformed into
moving vehicle. a dynamic compressed sensing problem. Their approach
Speech localization and enhancement involves sound leads to a decision-feedback equalizer for a communication
source mapping and reconstruction from noisy recordings of receiver.
speech mixtures with microphone arrays. Conventional Ocean acoustic communication channels are sparse with
beamforming methods suffer from low resolution in these coherent multipaths, which affects the estimation perfor-
tasks, especially with a limited number of microphones. In mance of phase coherent communication signals.68,69
practice, there are only a few sources compared to the possi- Gendron70 shows that sparse broadband time varying acous-
ble DOA. Xenaki et al.56 have formulated DOA estimation tic responses in horizontally stratified environments can be
as a sparse signal reconstruction problem and solved it with well estimated from an undersampled vertical array. The
SBL.30 For characterizing a reverberant room, Giri et al.57 model is applied to a three element vertical array during a
propose an SBL approach for estimating relative impulse set of broadband acoustic observations in shallow water.
responses using short, noisy, and reverberant recordings.
ACKNOWLEDGMENTS
D. Co-prime and nested arrays
This work was supported by the Office of Naval
The study of co-prime and nested arrays are concerned Research, Grant No. N00014-18-1-2118.
with obtaining minimally redundant arrays, i.e., arrays with
1
a minimum number of sensors. A uniformly spaced line C. E. Shannon, “Communication in the presence of noise,” Proc. IRE
array contains ample redundancy, which for low power sig- 37(1), 10–21 (1949).
2
H. Nyquist, “Certain topics in telegraph transmission theory,” Trans. Am.
nals, gives robustness in the array processing. For higher Inst. Elec. Eng. 47(2), 617–644 (1928).
powered signals, the number of sensors can be reduced using 3
E. J. Candes, J. K. Romberg, and T. Tao, “Stable signal recovery from
nested arrays,58 which consist of two arrays with different incomplete and inaccurate measurements,” Commun. Pure Appl. Math.
59(8), 1207–1223 (2006).
spacing. A promising nested array approach is co-prime 4
D. L. Donoho, “Compressed sensing,” IEEE Trans. Inf. Theory 52(4),
arrays,59,60 i.e., where the two arrays have spacings that are 1289–1306 (2006).
prime numbers of the basis spacing. A conventional uniform 5
R. G. Baraniuk, “Compressive sensing,” IEEE Signal Proc. Mag. 24(4),
line array of M sensor can find M DOAs. In contrast, co- 118–121 (2007).
6
E. J. Candès and M. B. Wakin, “An introduction to compressive
prime sensor arrays consisting of two arrays with N and M
sampling,” IEEE Signal Proc. Mag. 25(2), 21–30 (2008).
sensors can find the DOA of O(MN) narrowband plane wave 7
M. Elad, Sparse and Redundant Representations: From Theory to
sources using only O(M þ N) sensors. Applications in Signal and Image Process. (Springer, New York, 2010),
Bush and Xiang61 apply model-based Bayesian infer- 8
pp. 1–359.
S. Foucart and H. Rauhut, A Mathematical Introduction to Compressive
ence to a broadband co-prime array in order to not only Sensing (Springer, New York, 2013), pp. 1–589.
accurately estimate DOA, but also to predict the number of 9
N. R. Chapman and I. Barrodale, “Deconvolution of marine seismic data
sound sources. using the l1 norm,” Geophys. J. Int. 72(1), 93–100 (1983).
10
Ramirez and Rogers62 demonstrate on acoustic data that C. Yardim, P. Gerstoft, W. S. Hodgkiss, and J. Traer, “Compressive geoa-
coustic inversion using ambient noise,” J. Acoust. Soc. Am. 135(3),
synthetic aperture processing reduces side-lobe and grating 1245–1255 (2014).
lobe levels, compared to the physical co-prime aperture. 11
H. Yao, P. Gerstoft, P. M. Shearer, and C. Mecklenbr€auker, “Compressive
Rooney et al.63 estimate power spectral densities for co- sensing of the Tohoku-Oki Mw 9.0 earthquake: Frequency-dependent rup-
prime arrays using multi-tapering methods in snapshot- ture modes,” Geophys. Res. Lett. 38(20), 1–5, https://doi.org/10.1029/
2011GL049223 (2011).
starved scenarios. Bautista and Buck64 compare the power 12
F. Le Courtois and J. Bonnel, “Compressed sensing for wideband wave-
spectral densities for co-prime array using conventional number tracking in dispersive shallow water,” J. Acoust. Soc. Am. 138(2),
beamforming and product processors. 575–583 (2015).
13
A. Drèmeau, F. Le Courtois, and J. Bonnel, “Reconstruction of dispersion
curves in the frequency-wavenumber domain using compressed sensing
E. Communication on a random array,” IEEE J. Oceanic Eng. 42(4), 914–922 (2017).
14
Localization algorithms for acoustic sources implicitly G. F. Edelmann and C. F. Gaumond, “Beamforming using compressive
sensing,” J. Acoust. Soc. Am. 130(4), EL232–EL237 (2011).
or explicitly depend on the relative delay between two or 15
S. Fortunati, R. Grasso, F. Gini, M. S. Greco, and K. LePage, “Single-
more received signals. Lim et al.65 propose two CS algo- snapshot DOA estimation by using compressed sensing,” EURASIP J.
rithms for relative delay estimation. They estimate the Adv. Signal Process. 120(1), 1–17 (2014).
16
A. Xenaki, P. Gerstoft, and K. Mosegaard, “Compressive beamforming,”
relative delay from the minimum eigenvalue estimated by
J. Acoust. Soc. Am. 136(1), 260–271 (2014).
log-sum penalized minor component analysis. 17
P. Gerstoft, A. Xenaki, and C. F. Mecklenbr€auker, “Multiple and single
Distributed compressed sensing can be applied to snapshot compressive beamforming,” J. Acoust. Soc. Am. 138(4),
enhance sparse channel estimation in ocean acoustic multi- 2003–2014 (2015).
18
A. Xenaki and P. Gerstoft, “Grid-free compressive beamforming,”
band communication systems. Zhou et al.66 uses receptions J. Acoust. Soc. Am. 137, 1923–1935 (2015).
from multiple acoustic frequency bands to enhance the 19
P. A. Forero and P. A. Baxley, “Shallow-water sparsity-cognizant source-
detection of channel tap positions. Recent field experiments location mapping,” J. Acoust. Soc. Am. 135(6), 3483–3501 (2014).

J. Acoust. Soc. Am. 143 (6), June 2018 Gerstoft et al. 3735
20 47
K. L. Gemba, W. S. Hodgkiss, and P. Gerstoft, “Adaptive and compressive M. Haltmeier, M. Sandbichler, T. Berer, J. Bauer-Marschallinger, P.
matched field processing,” J. Acoust. Soc. Am. 141(1), 92–103 (2017). Burgholzer, and L. Nguyen, “A new sparsification and reconstruction
21
K. L. Gemba, S. Nannuru, P. Gerstoft, and W. S. Hodgkiss, “Multi-fre- strategy for compressed sensing photoacoustic tomography,” J. Acoust.
quency sparse Bayesian learning for robust matched field processing,” Soc. Am. 143, 3838–3848 (2018).
48
J. Acoust. Soc. Am. 141(5), 3411–3420 (2017). D. Malioutov, M. Çetin, and A. S. Willsky, “A sparse signal reconstruc-
22
M. Bianco and P. Gerstoft, “Compressive acoustic sound speed profile tion perspective for source localization with sensor arrays,” IEEE Trans.
estimation,” J. Acoust. Soc. Am. 139(3), EL90–EL94 (2016). Signal Process. 53(8), 3010–3022 (2005).
23 49
Y. Choo and W. Seong, “Compressive sound speed profile inversion using A. Das, W. S. Hodgkiss, and P. Gerstoft, “Coherent multipath direction-
beamforming results,” Remote Sens. 10, 704 (2018). of-arrival resolution using compressed sensing,” IEEE J. Oceanic Eng.
24
R. Tibshirani, “Regression shrinkage and selection via the lasso,” J. R. Stat. 42(2), 494–505 (2017).
50
Soc. Ser. B 58(1), 267–288 (1996), available at http://citeseer.ist.psu.edu/ Y. Park, Y. Choo, and W. Seong, “Multiple snapshot grid free compres-
viewdoc/summary?doi=10.1.1.35.7574. sive beamforming,” J. Acoust. Soc. Am. 143, 3849–3859 (2018).
25 51
M. Grant and S. Boyd, CVX: Matlab software for disciplined convex pro- Y. Yang, Z. Chu, G. Ping, and Z. Xu, “Resolution enhancement of two-
gramming, version 2.1. cvxr.com/cvx (Last viewed 9 Feb 2016). dimensional grid-free compressive beamforming,” J. Acoust. Soc. Am.
26
M. Grant and S. Boyd, “Graph implementations for nonsmooth convex 143, 3860–3872 (2018).
52
programs,” in Recent Advances in Learning and Control, Lecture Notes in M. N. Tabassum and E. Ollila, “Sequential adaptive elastic net approach
Control and Information Sciences, edited by V. Blondel, S. Boyd, and H. for single-snapshot source localization,” J. Acoust. Soc. Am. 143,
Kimura (Springer-Verlag, London, 2008), pp. 95–110. 3873–3882 (2018).
27 53
S. Boyd and L. Vandenberghe, Convex Optimization (Cambridge J. W. Paik, W. Hong, J.-K. Ahn, and J.-H. Lee, “Statistics on noise covari-
University Press, New York, 2004), pp. 1–684. ance matrix for covariance fitting-based compressive sensing direction-of-
28
M. E. Tipping, “Sparse Bayesian learning and the relevance vector arrival estimation algorithm: For use with optimization via regularization,”
machine,” J. Mach. Learn. Res. 1, 211–244 (2001). J. Acoust. Soc. Am. 143, 3883–3890 (2018).
29 54
D. P. Wipf and B. D. Rao, “An empirical Bayesian strategy for solving the Y. Yang, Y. Zhang, and L. Yang, “Wideband sparse spatial spectrum esti-
simultaneous sparse approximation problem,” IEEE Trans. Signal Proc. mation using matrix filter with nulling in a strong interference environ-
55(7), 3704–3716 (2007). ment,” J. Acoust. Soc. Am. 143, 3891–3898 (2018).
30 55
P. Gerstoft, C. F. Mecklenbr€auker, A. Xenaki, and S. Nannuru, C. Tuna, D. L. Jones, S. Zhao, and T. N. T. Nguyen, “Wideband compres-
“Multisnapshot sparse Bayesian learning for DOA,” IEEE Signal Proc. sive beamforming tomography for drive-by large-scale acoustic source
Lett. 23(10), 1469–1473 (2016). mapping,” J. Acoust. Soc. Am. 143, 3899–3911 (2018).
31 56
A. F. Metherell, H. M. El-Sum, J. J. Dreher, and L. Larmore, “Introduction A. Xenaki, J. B€unsow Boldt, and M. Græsbøll Christensen, “Sound source
to acoustical holography,” J. Acoust. Soc. Am. 42(4), 733–742 (1967). localization and speech enhancement with sparse Bayesian learning
32
E. G. Williams, Fourier Acoustics: Sound Radiation and Nearfield beamforming,” J. Acoust. Soc. Am. 143, 3912–3921 (2018).
57
Acoustical Holography (Academic, New York, 1999). R. Giri, T. A. Srikrishnan, B. D. Rao, and T. Zhang, “Empirical Bayes
33
G. Chardon, L. Daudet, A. Peillot, F. Ollivier, N. Bertin, and R. based relative impulse response estimation,” J. Acoust. Soc. Am. 143,
Gribonval, “Near-field acoustic holography using sparse regularization 3922–3933 (2018).
58
and compressive sampling principles,” J. Acoust. Soc. Am. 132(3), P. Pal and P. P. Vaidyanathan, “Nested arrays: A novel approach to array
1521–1534 (2012). processing with enhanced degrees of freedom,” IEEE Trans Signal Proc.
34
J. Hald, “Fast wideband acoustical holography,” J. Acoust. Soc. Am. 58(8), 4167–4181 (2010).
59
139(4), 1508–1517 (2016). P. P. Vaidyanathan and P. Pal, “Sparse sensing with co-prime samplers
35
E. Fernandez-Grande, A. Xenaki, and P. Gerstoft, “A sparse equivalent and arrays,” IEEE Trans Signal Proc. 59(2), 573–586 (2011).
60
source method for near-field acoustic holography,” J. Acoust. Soc. Am. K. Adhikari, J. R. Buck, and K. E. Wage, “Extending coprime sensor
141(1), 532–542 (2017). arrays to achieve the peak side lobe height of a full uniform linear array,”
36
E. Fernandez-Grande and L. Daudet, “Compressive acoustic holography EURASIP J. Adv. Signal Process. 2014, 148.
61
with block-sparse regularization,” J. Acoust. Soc. Am. 143, 3737–3746 D. Bush and N. Xiang, “A model-based Bayesian framework for sound
(2018). source enumeration and direction of arrival estimation using a coprime
37
M. R. Bai, C. Chung, and S-S. Lan, “Iterative algorithm for solving acous- microphone array,” J. Acoust. Soc. Am. 143, 3934–3945 (2018).
62
tic source characterization problems under block sparsity constraints,” J. Ramirez, Jr. and J. S. Rogers, “Exploiting platform motion for passive
J. Acoust. Soc. Am. 143, 3747–3757 (2018). source localization with a co-prime sampled large aperture array,”
38
J. Hald, “A comparison of iterative sparse equivalent source methods for J. Acoust. Soc. Am. 143, 3946–3958 (2018).
63
near-field acoustical holography,” J. Acoust. Soc. Am. 143, 3758–3769 I. M. Rooney, Y. Liu, and J. R. Buck, “Spatial power spectral density esti-
(2018). mation using a multitapered coprime sensor array minimum processor,”
39
S. A. Verburg and E. Frenandez-Grande, “Reconstruction of the sound field J. Acoust. Soc. Am. 143, 3959–3971 (2018).
64
in a room using compressive sensing,” J. Acoust. Soc. Am. 143, R. Bautista and J. R. Buck, “Processor dependent bias of spatial spectral esti-
3770–3779 (2018). mates from coprime sensor arrays,” J. Acoust. Soc. Am. 143, 3972–3978
40
S. Koyama, N. Murata, and H. Saruwatari, “Sparse sound field decomposi- (2018).
65
tion for super-resolution in recording and reproduction,” J. Acoust. Soc. J.-S. Lim, H.-S. Pang, and K. Lee, “Time delay estimation based on log-
Am. 143, 3780–3795 (2018). sum and lp-norm penalized minor component analysis,” J. Acoust. Soc.
41
J.-M. Attendu and A. Ross, “Sparse regularization for reconstructing tran- Am. 143, 3979–3984 (2018).
66
sient sources with time domain nearfield acoustical holography,” Y. Zhou, A. Song, F. Tong, and R. Kastner, “Distributed compressed sens-
J. Acoust. Soc. Am. 143, 3796–3806 (2018). ing based channel estimation for underwater acoustic multiband trans-
42
M. Bianco and P. Gerstoft, “Dictionary learning of sound speed profiles,” missions,” J. Acoust. Soc. Am. 143, 3985–3996 (2018).
67
J. Acoust. Soc. Am. 141(3), 17491758 (2017). W. Jiang, S. Zheng, Y. Zhou, F. Tong, and R. Kastner, “Exploiting time
43
K. S. Alguri, J. Melville, and J. B. Harley, “Baseline-free guided wave varying sparsity for underwater acoustic communication via dynamic
damage detection with surrogate data and dictionary learning,” J. Acoust. compressed sensing,” J. Acoust. Soc. Am. 143, 3997–4007 (2018).
68
Soc. Am. 143, 3807–3818 (2018). W. Li and J. C. Preisig, “Estimation of rapidly time-varying sparse
44
A. Thakur, V. Abrol, P. Sharma, and P. Rajan, “Local compressed convex channels,” IEEE J. Oceanic Eng. 32(4), 927–939 (2007).
69
spectral embedding for bird species identification,” J. Acoust. Soc. Am. C. R. Berger, S. Zhou, J. C. Preisig, and P. Willett, “Sparse channel esti-
143, 3819–3828 (2018). mation for multicarrier underwater acoustic communication: From sub-
45
J. Ramirez, Jr., J. S. Rogers, and G. F. Edelmann, “A performance metric space methods to compressed sensing,” IEEE Trans. Signal Proc. 58(3),
for screen selection with the acoustic single pixel imager,” J. Acoust. Soc. 1708–1721 (2010).
70
Am. 143, 3829–3837 (2018). P. J. Gendron, “Sparse broadband time varying acoustic response model-
46
M. F. Duarte, M. A. Davenport, D. Takhar, J. N. Laska, T. Sun, K. F. ing and estimation from an undersampled vertical array with application
Kelly, and R. G. Baraniuk, “Single-pixel imaging via compressive to underwater communications,” J. Acoust. Soc. Am. 143, 4008–4017
sampling,” IEEE Signal Process. Mag. 25(2), 83–91 (2008). (2018).

3736 J. Acoust. Soc. Am. 143 (6), June 2018 Gerstoft et al.