Professional Documents
Culture Documents
Nonsmooth Nonconvex Optimization For Low-Frequency Geosounding Inversion
Nonsmooth Nonconvex Optimization For Low-Frequency Geosounding Inversion
Nonsmooth Nonconvex Optimization For Low-Frequency Geosounding Inversion
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING 1
Abstract—A study of the application of nonconvex regulariza- A particular version of the technique that works at very low
tion operators to the electromagnetic sounding inverse problem is frequencies, at what is known as low induction numbers, is of
presented. A comparison is presented among three nonconvex reg- special interest from both practical and theoretical, e.g., [2] rea-
ularization algorithms: one smooth usually considered, two nons-
mooth, and a convex one, the total variation (TV) operator. One of sons. At low induction numbers, the skin depth of the currents
the nonsmooth nonconvex regularization methods is a novel imple- induced in the ground must be much greater than the separation
mentation based on the Legendre–Fenchel transform and the Breg- between the coils. When this happens the depth of penetration
man iterative algorithm. The nonconvex regularization operator is is independent of frequency. Depth discrimination is accom-
approximated by the convex dual, and the minimization is then plished by varying solely the separation between the coils: the
implemented considering the equivalence between the Bregman it-
eration and the augmented Lagrangian methods. The algorithm is greater the separation the deeper the penetration. Although in
simple and provides for better models when applied to synthetic practice this version of the electromagnetic technique is less ef-
data, than those obtained with TV, and other nonconvex smooth fective than a combined frequency-separation approach, it can
regularizers. Results of the application to field data are also pre- be shown that the information provided by soundings at low in-
sented, observing that NS2 recovers a model in better agreement duction numbers is sufficient for a unique recovery of a vertical
with the truth, compared to those obtained with additional mag-
netometric resistivity data by other researchers. conductivity profile [3]. Depth discrimination works very much
in the same way as in the familiar direct-current methods, where
Index Terms—Geosounding, nonconvex optimization, regular- the spacing between current and potential electrodes determines
ization.
the depth of penetration. The main difference, in practice, is that
electromagnetic induction allows for faster field surveys, since
I. INTRODUCTION
it eliminates the need for material contact with the ground.
O image the internal properties of the earth, geophysicists
T relay mostly on the interpretation of measurements made
on the surface. This applies for deep soundings of hundreds of
In one-dimensional (1-D), the surface measurements can be
exactly represented by a linear functional of the unknown con-
ductivity σ of the subsurface, e.g., [3]. In 2-D, the corresponding
kilometers as for shallow studies of merely a few meters below functional is necessarily nonlinear, but it can be approximated
the surface. The electrical conductivity of rocks is often the for many practical applications using a linear functional [4].
property of interest in these type of studies. For this reason, a Apparent conductivity σa , a normalized version of the measure-
great number of electrical techniques have been developed to ments, can be expressed in 2-D as
infer the conductivity structure of the subsurface on the basis of
∞ ∞
surface measurements. 2|x2 − x1 |
σa = A2D (x, z, x1 , x2 )σ(x, z)dxdz
One technique is based on electromagnetic induction by π 0 −∞
means of an alternating current that is made to flow in a trans- (1)
mitting coil. This current generates an alternating magnetic field where A2D is presented on Section IV and x1 , x2 are the coor-
in the surrounding environment, which in turn induces an elec- dinates of the transmitting and receiving coils, respectively.
tromotive force both in the conductive ground and in a receiving The inverse problem posed by this equation differs from im-
coil, e.g., [1]. The currents induced in the ground depend on the age enhancement in several ways. The electrical conductivity
conductivity distribution and affect the receiving coil accord- distribution within the integral sign represents the unknowns or,
ingly; thus, making it possible to obtain information about the equivalently, the output data. These output data make up the im-
subsurface on the basis of purely surface measurements. Con- age of the subsurface and ideally their size is infinite, although
trol on the depth penetration is accomplished by varying the for practical reasons, the size is considered finite. The input data
frequency of the current in the transmitting coil or the separa- on the left-hand side are the measurements taken on the surface
tion between the transmitting and receiving coils, or by varying of the ground, and typically, they are only a few. The inverse
both frequency and separation. problem in this case is more of constructing an image that does
not exist than of enhancing a preexisting one.
Tikhonov’s regularization method [5] is the standard tech-
Manuscript received September 29, 2015; revised December 15, 2015, Febru- nique applied to obtain models of the subsurface conductiv-
ary 9, 2016, and March 2, 2016; accepted March 19, 2016.
The authors are with the Centro de Investigación Cientı́fica y Educación ity distribution from electric or electromagnetic measurements
Superior de Ensenada, Ensenada 22860, México (e-mail: hugo@cicese.mx; (e.g., [6], [7]). The model proposed for the conductivity distri-
egomez@cicese.mx). bution m is obtained by minimizing the functional
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/JSTARS.2016.2549942 UT (m) = F (m) − d2H + λP (m). (2)
1939-1404 © 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 inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
2 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
simpler approaches for the geosounding method in the lin- A. Smooth Algorithm
ear case of low-induction numbers. We consider the use of
A smooth nonconvex regularization algorithm was proposed
nonsmooth nonconvex regularizers presenting two algorithms,
by Charbonnier et al. [35], considering the following operator:
both using the convex dual approximation based on Legendre–
Fenchel transform. One optimization is based on the variable PE P R = φ[(Dx m)k ] + φ[(Dz m)k ] (10)
splitting scheme and the other on directly solving the convex k k
dual by coordinate descent. Both algorithms are evaluated com-
where (Dx m)k is the finite difference implementation of first-
paring modeling results with those of one nonconvex smooth
order derivative at point k, and φ is the potential function. This
algorithm and with TV for synthetic and real data. All of the
operator has been considered on the vision community under
nonconvex implementations include some kind of Bregman
the name of “edge-preserving regularization” (EPR) [35], and
iteration.
applied to the resistivity inverse problem by Hidalgo-Silva et al.
[36]. Charbonnier et al. developed some properties that a poten-
II. BREGMAN ITERATIONS
tial function has to comply in order to obtain an EPR operator.
Bregman iterative regularization was introduced by Osher A particularly important nonconvex potential function initially
et al. [19] in image processing. They extended the Rudin et al. proposed by Geman and McClure [37] for image restoration is
[17] model into an iterative regularization model by using the
x2
Bregman distance. φ(x) = (11)
The Bregman distance based on a convex functional J(·) x2 + β 2
between points u and v is defined as with β > 0. Charbonnier et al. apply an iterative algorithm de-
veloped by Geman and Yang [25] called “half quadratic reg-
DJp (u, v) = J(u) − J(v) − p, u − v (6)
ularization” for the minimization of a modified version of the
where p ∈ δJ(v) is some subgradient in the subdifferential of original functional (3) with P = PE P R . The algorithm is aimed
J at the point v. to minimize the dual functional
Yin et al. [31] observed that the algorithm can also be repre-
UE∗ P R (m, bx , bz ) = Am − d2
sented as the original problem (3) but with dk +1 = d + (dk −
Amk ) as input, instead of d, obtaining the “add-back residual” +λ [φ∗ [(Dx )k , (bx )k ] + φ∗ [(Dz )k , (bz )k ]] (12)
version k
4 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
TABLE III with ρ > 0. Vese and Chan [34] considered these functions as re-
BREGMAN EPR ALGORITHM
duced nonconvex approximations of Mumford–Shah functional.
They demonstrated, using the Legendre–Fenchel transform, that
Bregman EPR Algorithm
φ(t) can be represented using a convex decreasing function ψ
Input: d such that
Initialization: k = 0, m 0 = 0, d 0 = 0
repeat φ(t) = inf (v 2 t + ψ(v)), ∀t ≥ 0 (19)
v
d k + 1 ← d + (d k − A m k )
with
B. Non Smooth Regularizers
n2 = (br + dˆr − br +1 − dˆr +1 + mr −1 + mr +1 )
Consider the minimization of (3) with P (m) = φ(|∇m|),
with φ a nonconvex nonsmooth function such as and for br
|t| bkr +1 = shrink(mr − mr −1 − dˆr , λvr /γ) (23)
φ1 (|t|) = (17)
1 + ρ|t|
with the shrink operator defined as shrink(x, α) = (|x| −
or α)+ sign(x). Solution for b is obtained from the dual convex ob-
1 tained applying Legendre–Fenchel transform (9). For the case
φ2 (|t|) = log(1 + ρ|t|) (18) of φ1 (·), last term in (21) is replaced by λ( i vi2 |Dx m|i +
ρ
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
TABLE IV TABLE V
NONSMOOTH NONCONVEX NS1 ALGORITHM NONSMOOTH NONCONVEX NS2 ALGORITHM
Input: d Input: d
Initialization: k = 0, m 0 = 0, d 0 = 0, dˆ0 = 0 Initialization: k = 0, m 0 = 0, d 0 = 0
repeat repeat
d k + 1 ← d + (d k − A m k ) d k + 1 ← d + (d k − A m k )
dˆk + 1 ← b + ( dˆk − ∇m k ) update m k + 1 using 25
update m k + 1 using 22 update v z , v x using 26 or 27
update b z , b x using 23 k ←k+1
k ←k+1 until (R M S < m in )
until (R M S < m in ) Output m k
Output m k
the subdifferential of |x| given by x1 is the location of the transmitting coil, x2 the loca-
sgn(x) if x = 0 tion of the receiving coil, r1 = (x − x1 )2 + y 2 and r2 =
ϕ(x) = (28) (x − x2 )2 + y 2 .
Co{−1, 1}, if x = 0.
The coordinate system is shown in Fig. 1.
A Bregman iterative procedure is obtained easily, by using (25), The conductivity variations along y, the strike direction, are
(26), and (27) to solve the minimization step for mr . The al- neglected. The kernel for vertical dipoles can be integrated ana-
gorithm is presented in Table V. A main disadvantage in this lytically along strike and evaluated in terms of elliptic functions
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
6 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
TABLE VII
EFFECT OF λ AND γ ON ITERATIONS REQUIRED TO ATTAIN RM S = 0.002,
ALONG WITH APPROXIMATION ERROR, SHOWN IN PARENTHESES, FOR NS1
TABLE VIII
PARAMETERS OF ALGORITHMS
Algorithm Parameters
TVAL3 μ, β ,tol
EPR λ, β ,min
NS1 λ, γ , ρ,min
NS2 λ, ρ,min
oped for lower values of λ, but one of the blocks attain a smaller
structure than the real one. On the other side, a smoother model
is recovered for λ > 0.01, failing to recover the two blocks in
Fig. 3. Convergence curves, RMS fitness (a) and Approximation Error, the structure.
(b) versus iteration number k. The performance for several values of λ and γ is presented
in Table VII for NS1. The numbers also represent the itera-
TABLE VI tions required and approximation error attained when algorithm
EFFECT OF λ ON ITERATIONS REQUIRED TO ATTAIN RM S = 0.002, ALONG run until RM S = 0.002. Results with (−) indicate that the al-
WITH APPROXIMATION ERROR, SHOWN IN PARENTHESES, FOR EPR AND NS2 gorithm stagnates at some RMS above the required, without
reaching the target. Regardless of the nonconvexity of opera-
λ EPR NS2 tors in EPR, NS1, and NS2, they did not showed a divergent
0.0001 9 (1.37) 9 (1.21)
behavior.
0.001 9 (0.82) 9 (0.73) All the models have a similar structure for a given γ, re-
0.01 8 (0.80) 10 (0.66) gardless the value of λ. The best performance is observed when
0.1 13 (0.84) 10 (0.77)
1.0 45 (0.84) 18 (0.82)
λ ∈ [0.001, 0.01], with more contrast observed when γ = 0.001.
10.0 203 (0.87) 71 (0.82) In this case, a more compact model is developed for lower
values of γ, and a smoother one is obtained for greater γ.
A summary of parameters required by the algorithms is pre-
sented in Table VIII. TVAL3 and NS1 need two regularization
iteration number k, with m∗ the real model used to generate
parameters [(μ, β) and (λ, γ), respectively] due to the augmented
the synthetic data. Notice that all algorithms behave similarly in
Lagrangian formulation. As for EPR and NS2, λ is the only reg-
RMS fitting, but NS2 present much better performance in terms
ularization parameter.
of approximation error. The best model is obtained at about
ten iterations for EPR and NS2, and eight iterations for NS1.
VI. APPLICATION TO FIELD DATA
Some artifacts appear on the model when the algorithm NS1 is
permitted to continue after that iteration. An application to field data is presented now, the data corre-
A study was conducted in order to observe the sensitivity spond to measurements taken at Las Auras, a site near the city
to the regularization parameters on the three nonconvex algo- of Tecate, México. The purpose of the surveys was to identify
rithms. Results are presented in Table VI for EPR and NS2, with sediment-filled faults affecting the construction of a dam. We
β = 1 and ρ = 1. are considering the vertical and horizontal dipole measurements
The Table shows how many iterations were needed and the taken at several points across strike for 10, 20, and 40 m separa-
corresponding approximation error obtained when the algo- tions as reported by Pérez-Flores et al. [4]. Applying maximum
rithms run until a RMS fitness of 0.002 was attained, for λ ∈ smoothness regularization, they observed the presence of two
[0.0001, 10.]. The best performance was obtained at λ = 0.01 main conductive zones, corresponding to sediment-filled faults.
with model reported in Fig. 2. A more compact model is devel- The two main conductors observed by them with the same data
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
8 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
VII. CONCLUSION
A new algorithm for nonconvex nonsmooth optimization
(NS2) is presented, observing better performance when com-
pared with synthetic data with TV and other two nonconvex
methods. A study is presented on a empirical evaluation of
nonconvex nonsmooth algorithms applied to the low-induction
number geosounding method. Results of the application to field
data are also presented, observing that NS2 recovers a model in
better agreement with the truth, compared to those obtained with
additional magnetometric resistivity data by other researchers.
ACKNOWLEDGMENT
The authors would like to thank S. Méndez Delgado for pro-
viding the field data.
REFERENCES
[1] F. S. Grant and G. F. West, Interpretation Theory in Applied Geophysics.
New York, NY, USA: McGraw-Hill, 1965.
[2] S. Méndez Delgado, E. Gómez-Treviño, and M. A. Pérez-Flores, “Forward
modelling of direct current and low-frequency electromagnetic fields using
integral equations,” Geophys. J. Int., vol. 137, pp. 336–352, 1999.
[3] E. Gómez-Treviño, F. Esparza, and S. Méndez-Delgado, “New theoreti-
cal and practical aspects of electromagnetic soundings at low induction
numbers,” Geophysics, vol. 67, pp. 1441–1451, 2002.
[4] M. A. Pérez-Flores, S. Méndez-Delgado, and E. Gómez-Treviño, “Imag-
ing low-frequency and dc electromagnetic fields using a simple linear
Fig. 4. Results for the field data (a) field data, (b) Bregman EPR result, approximation,” Geophysics, vol. 66, no. 4, pp. 1067–1081, 2001.
(c) TVAL3 result, (d) NS1 result, and (e) NS2 result. [5] A. Tikhonov and V. Arsenin, Solutions of Ill-Posed Problems. Washington,
DC, USA: V. H. Winston, 1977.
[6] R. L. Parker, Geophysical Inverse Theory. Princeton, NJ, USA: Princeton
Univ. Press, 1994.
[7] S. C. Constable, R. L. Parker, and C. G. Constable, “Occam’s inversion:
A practical algorithm for generating smooth models from electromagnetic
appear at 20–40 m and 90–100 m in x, extending from 0 to sounding data,” Geophysics, vol. 52, pp. 289–300, 1987.
5 m in depth. In the same paper, they incorporate magnetomet- [8] H. Hidalgo, J. L. Marroquı́n, and E. Gómez-Treviño, “Piecewise smooth
models for electromagnetic inverse problems,” IEEE Trans. Geosci. Re-
ric resistivity data jointly with the vertical and horizontal data, mote Sens., vol. 36, no. 2, pp. 556–561, Mar. 1998.
observing that the second conductive block extends in x from [9] I. M. Varentsov, “A general approach to the magnetotelluric data inversion
80 to 100 m. in a piecewise-continuos medium,” Izvestya Phys. Solid Earth, vol. 38,
pp. 913–934, 2002.
The data and the models obtained by EPR, NS1, NS2, and [10] F. J. Esparza and E. Gómez-Treviño, “Electromagnetic sounding in the
TVAL3 are presented in Fig. 4 for the joint (vertical + horizon- resistive limit and the Backus–Gilbert method for estimating averages,”
tal dipoles) inversion. The first three rows in Fig. 4(a) are the Geoexploration, vol. 24, pp. 441–454, 1987.
[11] F. Esparza and E. Gómez-Treviño, “1-D inversion of resistivity and in-
horizontal dipole data, and the lower three correspond to the duced polarization data for the least number of layers,” Geophysics,
vertical dipole data. vol. 62, pp. 1724–1729, 1997.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
[12] H. Hidalgo, E. Gómez-Treviño, and M. A. Pérez-Flores, “Linear programs [33] A. Blake and A. Zisserman, Visual Reconstruction, vol. 2. Cambridge,
for the reconstruction of 2-D images from geophysical electromagnetic MA, USA: MIT Press, 1987.
measurements,” Subsurface Sens. Technol. Appl., vol. 5, pp. 79–96, 2004. [34] L. Vese and T. Chan, “Reduced non-convex functional approximations
[13] D. C. Dobson, “Recovery of blocky images in electrical impedance to- for image restoration & segmentation,” Math. Dept., Univ. California Los
mography,” in Inverse Problems in Medical Imaging and Nondestructive Angeles, Los Angeles, CA, USA, Tech. Rep. CAM97-56, 1997.
Testing, Engl, Rundell, Louis, Ed. New York, NY, USA: Springer, 1997, [35] P. Charbonnier, L. Blanc-Féraud, G. Aubert, and M. Barlaud, “Determin-
pp. 43–64. istic edge-preserving regularization in computed imaging,” IEEE Trans.
[14] C. Vogel and M. Oman, “Fast, robust total variation-based reconstruction Image Process., vol. 6, no. 2, pp. 298–311, Feb. 1997.
of noisy, blurred images,” IEEE Trans. Image Process., vol. 7, no. 6, [36] H. Hidalgo-Silva, E. Gómez-Treviño, J. L. Marroquı́n, and F. J. Esparza,
pp. 813–824, Jun. 1998. “Piecewise continuous models for resistivity soundings,” IEEE Trans.
[15] F. Lenti, F. Nunziata, C. Estatico, and M. Migliaccio, “Analysis of re- Geosci. Remote Sens., vol. 39, no. 12, pp. 2725–2728, Dec. 2001.
constructions obtained solving-penalized minimization problems,” IEEE [37] S. Geman and D. McClure, “Statistical methods for tomographic image
Trans. Geosci. Remote Sens., vol. 53, no. 9, pp. 4876–4886, Sep. 2015. reconstruction,” in Proc. 46th Sess. Internat. Statist. Inst. Bulletin ISI, vol.
[16] T. Lin, Y. Zhang, L. Wan, Y. Qu, and J. Lin, “A para-whole space model for 52, pp. 5–21,1987.
underground magnetic resonance sounding studies,” IEEE J. Sel. Topics [38] H. Hidalgo-Silva and E. Gómez-Treviño, “Bregman iterative algorithms
Appl. Earth Observations Remote Sens., vol. 9, no. 1, pp. 264–271, Jan. for 2-D geosounding inversion,” Inverse Problems Sci. Eng., vol. 23,
2016. pp. 1085–1099, 2015.
[17] L. Rudin, S. Osher, and E. Fatemi, “Nonlinear total variation based noise [39] E. Esser, “Applications of Lagrangian based alternating direction methods
removal algorithms,” Physica D, vol. 60, pp. 259–268, 1992. and connections to split Bregman,” Comput. Appl. Math., Univ. California
[18] M. Benning, Singular Regularization of Inverse Problems, Ph.D. disser- Los Angeles, Los Angeles, CA, USA, Tech. Rep. CAM Rep. 09-31, 2009.
tation, Dept. Comput. Sci. Math., Univ. Muenster, Münster, Germany,
2011.
[19] S. Osher, M. Burger, D. Goldfarb, J. Xu, and W. Yin, “An iterated reg-
ularization method for total variation based image restoration,” SIAM
J. Multiscale Model. Simul., vol. 4, pp. 460–489, 2005.
[20] H. Hidalgo-Silva and E. Gómez-Treviño, “Bregman iterative algorithms
for 2-D geosounding inversion,” Inverse Problems Sci. Eng., vol. 23, no. 6, Hugo Hidalgo-Silva (M’15) received the B.Sc. de-
pp. 1085–1099, 2015. gree in electronics engineering from the Instituto
[21] Y. Hu and M. Jacob, “Higher degree total variation (HDTV) regularization Tecnológico Regional de Chihuahua, Chihuahua,
for image recovery,” IEEE Trans. Image Process., vol. 21, no. 5, pp. 2559– México, in 1980, the M.Sc. degree in electronics and
2571, May 2012. communications from the Centro de Investigación
[22] C. Li, W. Yin, and Y. Zhang, “User’s guide for TVAL3: TV minimization Cientı́fica y Educación Superior de Ensenada (CI-
by augmented lagrangian and alternating direction algorithms,” CAAM CESE), Ensenada, México, in 1985, and the Doc-
Rep., vol. 20, pp. 46–47, 2009. tor en Ciencias degree in computer science from the
[23] S. Geman and D. Geman, “Stochastic relaxation, Gibbs distributions, and Centro de Investigación en Matemáticas, Guanajuato,
the Bayesian restoration of images,” IEEE Trans. Pattern Anal. Mach. México, in 1999.
Intell., vol. PAMI-6, no. 6, pp. 721–741, Nov. 1984. He has been a Faculty Member of CICESE since
[24] M. Nikolova, “Analysis of the recovery of edges in images and signals 1985. His interests include inverse problems and pattern recognition.
by minimizing nonconvex regularized least-squares,” Multiscale Model.
Simul., vol. 4, no. 3, pp. 960–991, 2005.
[25] D. Geman and C. Yang, “Nonlinear image recovery with half-quadratic
regularization,” IEEE Trans. Image Process., vol. 4, no. 7, pp. 932–946,
Jul. 1995.
[26] D. Geman and G. Reynolds, “Constrained restoration and the recovery of
discontinuities,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 14, no. 3,
pp. 367–383, Mar. 1992.
[27] M. Nikolova, “Markovian reconstruction using a GNC approach,” IEEE
Trans. Image Process., vol. 8, no. 9, pp. 1204–1220, Sep. 1999.
[28] M. Nikolova, M. Ng, S. Zhang, and W. Ching, “Efficient reconstruction Enrique Gómez-Treviño received the B.Sc. degree
of piecewise constant images using nonsmooth nonconvex minimization,” in physics from the Universidad Autónoma de Nuevo
SIAM J. Imag. Sci., vol. 1, no. 1, pp. 2–25, 2008. León, San Nicols de los Garza, México, in 1974, and
[29] M. Nikolova, M. Ng, and C. Tam, “Fast nonconvex nonsmooth mini- the M.Sc. and Ph.D. degrees both in geophysics from
mization methods for image restoration and reconstruction,” IEEE Trans. the University of Toronto, Toronto, ON, Canada, in
Image Process., vol. 19, no. 12, pp. 3073–3088, Dec. 2010. 1977 and 1981, respectively.
[30] M. Jung and M. Kang, “Efficient nonsmooth nonconvex optimization for He performed postdoctoral research at the Univer-
image restoration and segmentation,” J. Sci. Comput., vol. 62, pp. 336– sity of Toronto during 1981–1982. He was a Member
370, 2014. of the faculty of the Universidad Autónoma de Nuevo
[31] W. Yin, S. Osher, D. Goldfarb, and J. Darbon, “Bregman iterative León, during 1982–1983, and since 1984, he has been
algorithms for l1 -minimization with applications to compressed sensing,” a Member of the faculty of Centro de Investigación
SIAM J. Imag. Sci., vol. 1, no. 1, pp. 143–168, 2008. Cientı́fica y Educación Superior de Ensenada, Ensenada, Mexico. His interests
[32] S. Kirkpatrick, C. D. Gelatt, and M. P. Vecchi, “Optimization by simulated include modeling and inversion of potential and electromagnetic data.
annealing,” Science, vol. 220, no. 4598, pp. 671–680, 1983. Dr. Gómez-Treviño is a Member of SEG, EAEG, UGM, and AGU.