On the determination of the space-time slip distribution via Linear Programming Techniques B. Caldeira (1) , V. Bushenkov (2) , G. Smirnov (3) , J. Borges (1) , and M. Bezzeghoud (1) (1) CGE, Department of Physics, University of Evora, Evora, Portugal (2) CIMA, Department of Mathematics, University of Evora, Evora, Portugal (3) Centre of Physics, Department of Mathematics and Applications, University of Minho, Braga, Portugal Reference author: Bento Caldeira Geophysical Centre of Évora and Physics Department, ECT, University of Évora Colégio Luís António Verney Rua Romão Ramalho, 59 7002-554 Évora Tel.: +351 266 740 800 E-mail: bafcc@uevora.pt 1 Abstract The purpose of the seismic source investigation is to obtain an accurate description of the rupture from seismic and geodesic data. The success of this depends fundamentally on three factors: source models, informative content of the data, and the methods used. Currently, inversion techniques are considered better methods for obtaining rupture characteristics from seismic and geodesic data. In this paper we suggest to formulate a Linear Programming (LP) inversion in DUAL formulation, for reconstructing the kinematics of the rupture of large earthquakes through space-time seismic slip distribution on the faults plane. In its general setting, the algorithm can produce results from strong ground motion waveforms, but can also be used with teleseismic waveforms as well as with geodesic data (static deformation). To test the algorithm and examine its stability, and robustness we define a synthetic rupture model, based on a real space-time slip distribution scheme. We compare the proposed approach with others standard algorithms, rehearsing reconstructions with same synthetic waveforms through these algorithms. Green functions were calculated by a finite differences method with a 3D structure model. 1 - Introduction The description that one can have of the seismic source is the manifestation of an imagined model, obviously outlined from Physic Theories and supported by mathematical methods. In that context, the modelling of earthquake rupture consists in finding values of the parameters of the selected physics-mathematical model, through which it becomes possible to reproduce numerically the records of earthquake effects on the Earth’s surface. Actually, these effects are the elastic records at near field source, elastic records at far field source, and inelastic deformations recorded by geodetic techniques. The detail and accuracy level, with which the characteristic parameters for large earthquakes are computed, depends on the combination of two factor classes involved in the process - used methods and data. The kinematic model of finite seismic source consists in a spatiotemporal distribution of slip vectors on a fault plane, previously divided into a grid of sub-faults. The entire parameters set that characterize the vector slip of each sub-fault is as follows: init time, amplitude, direction and temporal evolution (or, alternatively, rise time). Currently, the most popular approaches to determine the slip distribution models are the inversion of near-source ground motion data (e.g. Asano and Iwata 2009; Suzuki et 2 al., 2009; Hartzel et al., 2007) and the joint inversion of near source and teleseismic waveforms (e.g. Delouis et al., 2009; Yagi, 2004; Mozziconacci et al., 2009). The reason to prefer near source data is that it allows to reconstitute the rupture kinematics with more details than when teleseismic waveforms are only used. The main difficulties when trying to use this kind of data are as follows: at first, the lack of accelerometer coverage in some relevant seismic zones, at second, models of Earth structures must be accurate, that don’t occurs for many zones, and, at last, very high computation power is required for such waveforms modelling. According to Tarantola (1987), the study method of any physical system, whether it is a quantum particle, a galaxy or the Earth, is developed following the methodological proposal of the Discrete Inverse Theory that involves forward problem, parameterization and inverse problem. In the problem discussed in this paper, the forward problem is implemented by the operational module that computes the soil movements (elastic displacement, velocity, acceleration or deformation) for all observation points (stations). The central trait of the forward problem is the calculus of the Green’s functions. They consist to find an approximate solution of the secondorder elastodynamic equation for a point in an elastic medium when perturbed by a point unitary source using the known structure model and source mechanism model. There exist some approaches to resolve the problem. By using 1D layered Earth structure models it is possible to find reasonable Green’s functions at low frequencies (<1Hz) applying such methods as Kennett and Kerry, 1979; Bouchon, 1981; Bernard and Madariaga, 1984; Spudich and Fraser, 1984; Sikia, 1994. The Green’s functions describing wave propagation through complex 3D anisotropic structures can be calculated by algorithms based on finite differences (e.g. Olsen and Archuleta, 1996; Pitarka et al., 1998; Olsen, 2000; Larsen & Schultz, 1995), finite element (e.g. Bao et al., 1998), or also spectral elements (Komatitsch 1997, Komatitsch & Villote, 1998; Komatitsch et al., 2002, etc). The first attempts to invert waveform to characterize spatial and temporal rupture processes of finite seismic sources were made in theoretical studies (see Gilbert, 1975; Hartzell et al. 1978) or applying to real situations (for example, San Fernando 1971 earthquake - Trifunac, 1974; Langston, 1978; Heaton, 1982; or the Imperial Valley 1979 earthquake - Hertzell and Helmberger, 1982). These early works represent the rupture model in the form of a succession of slips at sections of the rectangular fault plane. The init time of each section (sub-fault) was commanded by a rupture front that 3 spreads over the fault plane with constant velocity in all directions from the hypocentre. The evolution of each slip was given by a temporal function of certain shape. This source model scheme is known as model of simple time window. There were two main aspects difficult to resolve in the original models of single time window. The first is related with the shape and duration of the source time functions that is equal for all subfaults; that limits the frequency range of the data modelled. The second is the improper assumption to impose a rupture with constant velocity, which also affects a correct data modelling. These two difficulties were partially resolved by the model of multiple time windows (eg. Olson and Apsel, 1982; Cohee and Beroza, 1994). In the multiple time windows model, the evolution of slip in each sub-fault is represented by a succession of elementary source time function, which imposes that the rupture of each sub-fault occurs in separate time intervals. The multiple time windows version ensures a more realistic simulation of the ruptures although with significant increase of calculation efforts. In the essential, the present framework of finite-source models (eg. Ammon et al, 2005, Asano and Iwata, 2009; Mozziconacci et al., 2009, Robinson and Cheung, 2010; Delouis et al., 2009) are not much different from previous works; the major transformations refer to the grown of the computation scale and techniques imported from optimization. The search of the source parameterization that ensures the best fitting between observed and synthetic waveforms constitutes an inverse problem, whose can be solved by different methods. When the problem is posed to estimate whole source model parameters to large earthquakes, constitutes a nonlinear configuration. Otherwise, if it is possible to connect synthetic data with a model of linear equations, the problem can solve using the techniques of linear inverse theory (eg. Tarantola, 1987, Menke, 1984). The local inversion methods, as least squares, when applied to linear systems of large dimensions as seismic waveform inversion, are unstable due the existence of many local minima. In these cases is requires the introduction of stabilization factors physically reasonable: the constraints. The positivity that prohibits negative seismic moment values, is a constraint naturally assumed when used the Non Negative Least Squares algorithm (NNLS) (Lawson and Hanson, 1974) to inverts seismic waveforms to slip distribution (eg. Hartzell and Heaton, 1983; Ide et al., 1996; Suzuki et al., 2009). Minimum norm, minimum roughness, fix the total moment and fix the rupture velocity are also constrains used to regularize the inverse problem. 4 When the formulation of the nonlinear problem is applied to earthquake source inversion, there are some techniques to solve it. Global search algorithms have been implemented to explore all solution domains. Two of these algorithms extensively used are the Simulated Annealing (SA), and the Genetic Algorithm (GA), both developed from inspiration from nature's processes that try mimicking. The SA algorithm uses the sophisticated Monte Carlo random method (Kirkpatrick et al., 1983) to simulate the annealing in thermodynamics systems. The GA (Holland, 1975) operates on analogy with the evolution of biological populations. Ihmlé (1998) uses SA to make inversions of seismic data for the distribution of slip; Hernandez (2001) obtains slip distributions applying GA. Another variant of these methods with ability to be used in nonlinear inversions is the Neighbourhood algorithm (Sambridge, 1999). The inversions of seismic data to kinematic finite-fault slip distributions can adopt either objective functions of L1-norm, as L2-norm. Das and Suhadolc (1996) and Hartzel et al. (2007) analyse the important differences existed between L1 and L2 norm inversions. In this paper we present and test a Linear Programming (LP) inversion in Dual form, for reconstructing the kinematics of the rupture of large earthquakes through spacetime seismic slip distribution on finite faults planes. The proposed method can be considered as a continuation of the work started in Das and Kostrov (1990). The proposed algorithm uses strong ground motion waveforms, but it can also used with other types of data as teleseismic waveforms as well as with geodesic data (static deformation). To test the method a synthetic model was defined to compute seismograms that were inverted using the same approach as for the real data. We compare the proposed approach with others standard algorithms, rehearsing reconstructions with same synthetic waveforms through these algorithms. Green functions were calculated by the finite differences method applied to a 3D structure model. 2 - The forward problem The description of the elastic displacement produced at the Earth surface, as consequence of applied body forces or slip discontinuities in a semi-infinite elastic medium, constitutes the fundamentals for formal development of the methods to seismic source study. The representation theorem in its integral form (e.g. Aki and Richards, 1980; Ben-Menahem and Singh, 1981; Udias, 1999) is the starting point for 5 the construction of physical-mathematical formalism that supports the kinematic methods: +∞ ๐ ๐ข๐ (๐ฅ๐ , ๐ก) = ∫−∞ ๐๐ ∫Σ ๐ท๐ (๐, ๐)๐ถ๐๐๐๐ ๐๐ ๐๐ ๐บ๐๐ (๐ฅ, ๐ก − ๐, ๐)๐Σ (1) ๐ where ๐ข๐ (๐ฅ๐ , ๐ก) represents the ith component of the seismic displacement at observation point ๐ฅ๐ , and time t; ๐ท๐ (๐, ๐) is the mth component of the slip produced at time the position on ๐ถ๐๐๐๐ is a tensor that depends of the elastic proprieties and the geometry of the fault plane; ๐๐ is a vector perpendicular to the fault plane; and the 2nd order tensor ๐บ๐๐ (๐ฅ, ๐ก − ๐, ๐) is the Green Function, that represents the temporal evolution of the i component of the displacement at the position ๐ฅ๐ , due to a unitary slip in m direction produced at source, in position and time . The discretization of the integral form of the representation theorem (1) through proper parameterization of the source is the tool that allows us to compute the synthetic seismograms. To do this, the fault plane of the finite-source is discretized into a set of N sub-faults defined by a grid covering the entire surface and disposed along a square orthogonal referential xOy, where Ox is the strike-axis and Oy is the dip-axis. Also the total time of rupture needs to be discretized into Nt steps of time. Each subfault l (l=1,N) constitutes a point source that at certain time step k (k=1,Nt) initiates slips ฬ (breaks), according to a source time function ๐๐,๐,๐ (๐ก) in direction m. The slip vector is defined through the magnitude of the two orthogonal components m; one in the strike direction (m=1) and the other in the dip direction (m=2). The rupture described by this model is a sequence of slips, each one characterized by own: a) position, b) initial time, c) amplitude, d) direction, and e) source time function. The adopted finite-source model allows yet that each subfault after break once, can be reactivated and break later, in different stages of the rupture. The complete parameterization of this model requires yet defining the geometry of the fault plane; hypocenter position; size of each sub-fault; as well as the interval of the time-step in which rupture was discretized. ๐ The ith component of displacement at station j, ๐ข๐ (๐ก), is calculated by ๐ ๐ ๐ก 2 ฬ ∑๐ ๐ข๐ = ∑๐=1 ๐=1 ∑๐=1(๐๐,๐,๐ (๐ก) ∗ ๐บ๐,๐,๐,๐,๐ (๐ก)) ๐ฅ๐,๐,๐ (2) where the indices i,j,k,l,m represent: i – direction of the displacement at observation point (1=X, 2=Y, 3=Z); j – observation point; k – time step were the time of rupture was discretized; l – sub-faults; m – components of the slip vector (1 – strike direction, 6 2 – dip direction); the asterisk (*) denotes the convolution and ๐ฅ๐,๐,๐ is the slip. We ฬ approximate the source time function ๐๐,๐,๐ (๐ก) within each cell along strike and dip by a triangular function with unitary area. The Green functions were calculated considering propagation of seismic waves in a 3D media. We used the code E3D, an explicit elastic finite-difference wave propagation code (Larsen & Schultz 1995), based on the work of Madariaga (1976). The system of linear equations (2) that allow compute the synthetic seismograms (forward problem) can be translated to matrix language by the multiplication of matrices, u = A๐ฑ (3) where u is the vector that contain all the seismograms; x the vector of the slips of all subfaults in whole time steps; and A matrix of numerical parameters to compute the synthetic seismograms - the matrix of synthetics. The matrix A contains Green Functions relating slips in each subfault at each time-step. The set of Green Functions at different time steps differs only on the init time; can be obtained, to subsequent time steps, merely by a shift in time (delay) applied in the first set of them. Fig. 1 summarises the formulation of the forward problem. Under the hypothesis of fix slip direction and constant rise time of individual source time functions, the problem of complete seismic space-time slip distribution reconstruction reduces to the solution of a system of linear equations. It is well-known that this inverse problem is ill-posed Error! Reference source not found.. The usual regularization techniques Error! Reference source not found. can hardly be applied in this case because of a very high dimension of this problem. The problem can be overcome by introducing some additional regularizing constraints. Some additional physical hypotheses, like no-backslip constraint, result in condition of nonnegativeness of solutions to the system of linear equations. This makes natural to consider the inversion of the slip time history and distribution in the frame of linear programming theory. The linear programming approach was first applied to solve this problem in Error! Reference source not found.. This seems to be a unique reasonable method since as it is well-known Error! Reference source not found. the least-squares method is unable to solve the inversion problem due to the fact that many negative values of moment rate, which did not exist in the forward problem, are produced. 7 The hypothesis of constant slip direction in general is not verified and the "real" seismic slip time history and distribution reconstruction becomes an hard nonlinear problem. In this work we suggest an algorithm for seismic slip time history and distribution reconstruction allowing to solve the problem in its general setting. The solution of an auxiliary linear programming problem is an essential part of the developed method and, from this point of view, our wor can be considered as a continuation of the research started in Error! Reference source not found.. To test the algorithm we use a synthetic displacement function for the fault model and perform the inversion. 3 - The inversion algorithm The slip determination problem can be formalized in the frame of mathematical programming in the following way ๏กc, x๏ฑ ๏ฎ min , A(๏ฌ ) x = b, x ๏ณ 0. Here x (3) is the unknown vector of amplitudes and residuals (see Error! Reference source not found.) and the vector ๏ฌ represents the unknown rakes. Note that the displacement field models can be different but the mathematical formalization is always the same. If we fix the rake vector ๏ฌ , problem (Error! Reference source not found.) becomes a linear programming problem. This observation is the key to an effective solution of problem (Error! Reference source not found.). It turns out that the gradient of the minimized functional ๏กc, x๏ฑ with respect to ๏ฌ can be calculated in terms of the solution to the linear programming problem dual to (Error! Reference source not found.). The following algorithm describes the process. Algorithm: Given ๏ฌ0 , ๏ > 0 , and ๏ฅ > 0 . for k = 0,1,2,๏ Step 1. Solve linear programming problem (Error! Reference source not found.) with ๏ฌ = ๏ฌk and obtain xk . Step 2. Obtain search direction ๏ฌ k and a step ๏ค k > 0 . 8 if ๏ค k P๏ฌk P < ๏ฅ break else Step 3. Set ๏ฌk ๏ซ1 = ๏ฌk ๏ซ ๏ค k ๏ฌk . end(for) The second step of the algorithm is not trivial and we give its detailed description. Choose any direction ๏ฌ and ๏ค > 0 . Since the matrix A is a smooth function of ๏ฌ , we have A(๏ฌ ๏ซ ๏ค๏ฌ ) = A(๏ฌ ) ๏ซ ๏คA (๏ฌ ) ๏ซ o(๏ค ) . Let x(๏ค ) be a solution to the following linear programming problem: ๏กc, x๏ฑ ๏ฎ min , ( A(๏ฌ ) ๏ซ ๏คA (๏ฌ )) x = b, x ๏ณ 0. (4) We assume that the matrix A(๏ฌ ) has full row rank. Without loss of generality we have the representations A(๏ฌ ) = [ B(๏ฌ ), D(๏ฌ )] and A (๏ฌ ) = [ B (๏ฌ ), D(๏ฌ )] , where B(๏ฌ ) is a basis matrix (see Error! Reference source not found.) for problem (Error! Reference source not found.) with ๏ค = 0 . By c B we denote the basic components of the vector c . If ๏ค is sufficiently small, the solution to problem (Error! Reference source not found.) has the form x(๏ค ) = [ xB (๏ค ),0] . Since B(๏ฌ ) xB (0) = b and ( B(๏ฌ ) ๏ซ ๏คB (๏ฌ ))( xB (0) ๏ซ ๏คx๏ฆ B (0) ๏ซ o(๏ค )) = b, We get x๏ฆ B (0) = ๏ญ( B(๏ฌ )) ๏ญ1 B (๏ฌ ) x B (0). Let y (๏ฌ ) be a solution to the dual linear programming problem: ๏กb, y๏ฑ ๏ฎ max , ( A(๏ฌ ))T y ๏ฃ c. (5) Then we obtain d ๏กc, x(๏ค )๏ฑ = ๏ญ๏กcB , ( B(๏ฌ )) ๏ญ1 B (๏ฌ ) xB (0)๏ฑ d๏ค ๏ค =0 ๏จ ๏ฉ = ๏ญ๏ก ( B(๏ฌ )) ๏ญ1 cB , B (๏ฌ ) xB (0)๏ฑ = ๏ก y(๏ฌ ), B (๏ฌ ) xB (0)๏ฑ. T This formula gives a possibility to calculate the gradient ๏V (๏ฌ ) of problem (5) value V (๏ฌ ) . Indeed, taking ๏ฌ equal to i th basis vector, ei , we get ๏ถV (๏ฌ )/๏ถ๏ฌi . 9 Thus, we obtain the search direction ๏ฌk = ๏ญ๏V (๏ฌk ) and choose the step ๏ค k > 0 as the maximal ๏ค ๏]0, ๏] satisfying the condition ( xk ) B ๏ซ ๏ค ( B(๏ฌk )) ๏ญ1 ๏จ๏ฅ B (e )๏ฉ( x ) i i k B ๏ณ 0. . Let ๏ด > 0 . The function g๏ฆ Q (๏) (elementary moment rate function) has the form shown in Fig. 1. 5 - Initial point for the simplex method Let x ๏ R n . By ( x)i we denote the i -th coordinate of x . By e p we denote the vector with ๏ฌ0, i ๏น p, (e p ) i = ๏ญ ๏ฎ1, i = p. Consider the problem ๏กc, x๏ฑ ๏ฎ min , Ax ๏ฃ 0, ๏ญ 1 ๏ฃ ( x) i ๏ฃ 1, i = 1, n, (1) where A is an (m ๏ด n) -matrix with m < n . Suppose that rank A = m . Obviously x = 0 is an admissible point for problem ((1)). To find an initial point for the simplex-method we use the following algorithm. Set m0 = m , P0 = ๏ , A0 = A , b0 = 0 , and x0 = 0 . Assume that we have already constructed number mk , a set of indices Pk , an (mk ๏ด n) -matrix Ak , a vector bk ๏ R mk , and a point xk ๏ R n , such that Axk = bk . If mk = n , then the point xk is a vertex of the polyhedron X = {x ๏ R n | Ax ๏ฃ 0 | ( x)i |๏ฃ 1} . If mk < n , we construct the next quintet (mk ๏ซ1 , Pk ๏ซ1 , Ak ๏ซ1 , bk ๏ซ1 , xk ๏ซ1 ) in the following way. Using the Gauss elimination algorithm we obtain a representation Ak = [ Bk , Ck ] , where Bk is an upper triangular (mk ๏ด mk ) -matrix and Ck is a rectangular (mk ๏ด (n ๏ญ mk )) - 10 matrix. Consider a non-zero vector zk ๏ R vector xk = ( yk , zk ) ๏ R mk ๏ดR n ๏ญ mk n ๏ญ mk and put yk = ๏ญ Bk๏ญ1Ck zk ๏ R mk . The satisfies Ak xk = 0 . Set tk = sup{t || ( xk ๏ซ txk )i |< 1, i ๏ ๏ฏ Pk } and xk ๏ซ1 = xk ๏ซ t k xk . Obviously we have Ak xk ๏ซ1 = Ak xk ๏ซ t k Ak xk = bk . Put Pk ๏ซ1 = {i1 ,๏il } = {i = 1, n || ( xk ๏ซ1 )i |= 1}. k ๏ซ1 Denote by mk ๏ซ1 the sum of m and the cardinality of Pk ๏ซ1 . Define the matrix Ak ๏ซ1 adding to the matrix A the rows e p , p ๏ Pk ๏ซ1 . The first m components of the vector bk ๏ซ1 are equal to zero and (bk ๏ซ1 )i = ( xk ๏ซ1 )i , i ๏ Pk ๏ซ1 . Therefore Ak ๏ซ1 xk ๏ซ1 = bk ๏ซ1 . By induction, after at most n ๏ญ m steps we get a vertex of X . 6 - Data to evaluation To evaluate the proposed algorithm, we applied it to a synthetic scenario of seismic rupture similar to real sources. This kind of evaluation is extremely important since it represents the only reliable way to analyse the performance of the methods where the expected results are known (Beresnev, 2003). The synthetic waveforms used were calculated by Eq. 3 for a defined rupture model, to a set of 13 seismic stations distributed around the source as the geometry represented in Fig. 3. To each considered station were calculated the 3 components (N, E and Z) velocity waveforms. The seismic rupture was defined by a spatiotemporal distribution of slips dominated by thrust with left-lateral strike-slip component on a rectangular fault plane of 10 x 10 km2, oriented towards the N and dipping to East (dip is equal to 45 deg.). The fault plane was divided into a grid of 36 square subfaults, with size of 2 x 2 km2. Each individual slip vector (placed at each node of the grid) is characterised by two orthogonal components, in strike and dip directions, of each source is =atan( 2/ 1). 1 and 2, where the rake angle The rupture starts at initial time at hypocentre node at 12 km depth and spreads in all directions with a variable velocity. The slip of each 11 node is specified by the initial time, two components of the slip vector and a triangular Source Time Function (STF) with rise time . The rupture time is discretized using a temporal gridding of 0.3s. The defined source model assumes that some subfaults slips more than once at different stages of the rupture. Table 1 and Fig. 4 represents the space time slip distribution used and the complete STF. The Green Functions were computed using a finite-difference spatial and temporal grid spacing scheme of 0.5 km and 0.03sec respectively. The velocity model considered is a fragment of 100kmx100kmx70km of the 3-D velocity model of SW Iberia (Grandin et al. 2007). Given that the Green Functions simulated by the finite-difference method are only reliable at a maximum frequency (fmax) propotional to the minimum velocity (vmin), and the inverse of the grid spacing (h) given by fmax < vmin/5h (Pitarka, 1999), with the parameterization chosen, and the used velocity model, we generating a maximum frequency less of 1.4Hz. Thus the Green Functions calculated were filtered with a low pass Butterworth Filter with a cut-off frequency of 1.3Hz. 7 - Inversion Results From the same data were performed inversions using three different algorithms. Two standard - the least-squares of Lawson and Hanson (1974) through the formulation of Hartzell and Heaton (1983) and the Linear Programing by Primal formulation developed by of Das and Kostrov (1990); and a Dual version of Linear Programming presented in this paper. All the procedures have been used with same Green functions and source parameterization. Due the phenomena of generation and propagation of seismic waves, the amplitudes of the waveforms vary depending on the distance and relative position from the source. To prevent data calculated/recorded with the same accuracy may produce different weights in the objective function, the amplitude of waveforms was normalized. The reconstructed rupture model calculated using the Dual version of Linear Programming algorithm developed (Fig. 5) is comparable with the synthetic base model (Fig. 4). The likeness between the two models is evident in both the spatial distribution of slip as in its spatial occupation over time, characterized by a nonuniform rupture front. The “Total” slip distribution of the synthetic and reconstructed models (comparison of Figs. 4 and 5) proves the spatial likeness; the evolution of the rupture, displayed in the sequence of snapshots and in STF of synthetic and reconstructed models, proves the suitable temporal reproduction. The comparison 12 between the observed and model-predicted waveforms is in many real situations the only way to validate the reconstructions. Sometimes the two seismograms are simply displayed together to visual comparison; other times they are applied quantification criteria to characterize differences (eg. Geller and Takeuchi, 1995; Kristekova et al. 2006). In this paper we use the he Normalized Root-Mean-Square Deviation (NRMSD) method to compare seismograms, ∑๐ก(๐ข(๐ก) − ๐ขREF (๐ก)) √ ๐๐ ๐๐๐ท = ๐๐ก(๐ขmax − ๐ขmin )2 where ๐ข(๐ก) is the tested seismogram, ๐ขREF (๐ก) is the reference seismogram, Nt is the number of elements of the seismogram, ๐ขmax the maximum amplitude of the seismogram and ๐ขmin the minimum amplitude. Fig. 6 shows the synthetic and reconstructed models, and respective predicted waveforms. The similarity of the two models explains the good fit between the two sets of waveforms (NRMSD=1,23%). The same parameterization was applied to inverts by Primal formulation of Linear Programming. It was implemented using the same inversion code (reference to algorithm) applied to the dual version. To this case the convergence was also obtained, to the same solution, but after a processing time about 100 times higher than necessary from Dual formulation (~12 min with DUAL and ~14H with PRIMAL) and with a great number of warnings. Inversions by the formulation of Hartzell and Heaton (1983) don’t estimate the starttime slip of sub-faults. Is a parameter fixed considering that exist a rupture front spreading with constant velocity from the hypocenter, and that the start-time slip at each sub-fault coincides with the arrives of the rupture front. It´s a constraint that promotes a high reduction of the number of parameters to estimate and consequently makes possible the utilization of local inversions as NNLS of Lawson and Hanson (1974). This method was used with a rupture velocity fixed at 3km/s and the other parameterizations were the same used in the formulations of linear programming. The reconstructed model by this method (Fig. 7) differs significantly from the synthetic origin model used (Fig. 4), both in total slip distribution as well as their distribution in time. Surprisingly this divergence between models occurs despite the good fit on the two sets of waveforms (NRMSD=4,73%, Fig. 8), where can prove that different models can produce very similar waveforms. 13 8- Discussion and conclusions The reconstructing of the kinematics of the rupture for large earthquakes from a set of their effects recorded on the Earth surface constitutes a scientific problem, whose solution consists to determinate the evolution of the slip vector over the fault area. The inversion formulations to produce physically consistent solutions must be stated in its global formulation, which evolves extensive models. The Linear Programming (LP) techniques using the PRIMAL formulation (Das Kostrov, 1990; Hartzel and Liu, 1995) revels an appropriate tool to make this since the scale of the problem does not reach large dimensions. Is a inversion method that explores full space of solutions. The easily to incorporate constraints that improves the convergence, since reduces the number of vertices in the hyperspace of the solutions, is an advantage of this kind of approaches relatively the iterative methods. However, if the problem of inversion is parameterized in order to involve a large number of equations, its resolution through the PRIMAL LP inversion techniques can become a problem difficult to solve with conventional computation means. To these cases, that occur when we want to reconstruct the kinematics of the rupture for great earthquakes, we suggest to formulate a Linear Programming inversion in DUAL formulation. It reduces the dimension of the variable space, put the observed data (u) in the objective function and consequently makes the computation process more stable. We show that the DUAL formulation presents clear advantages in both the convergence and computing time relatively to the PRIMAL formulation used by other authors. The ability that the slip inversion methods have to build detailed scenarios of rupture makes them one of the more attractive tools to study the seismic source. However, a detailed analysis of the solutions acquired by these methods, warns for a set of cautions that must be present before use of these tools. The first is the choice of method. Looking at the literature we found a number of different ways of obtain inversion scenarios. All of them consider equivalent levels of respect to physical and numerical requirements (eg Hartzel and Heaton, 1983, Hernandez et al. 2001; Valle and Bouchon, 2004). However, when apply to the same events, with same data, show different results, as can be compared from the works of Wald and Heaton (1994) and Cohee and Beroza (1994), for the Landers earthquake of 1992. Selected the method is necessary fix the parameterization: the geometry of the fault; size for the sub-faults; shape of source time functions; number of time windows in each subfault,... . Arise evidences that 14 different parameterizations lead to different results, but equally plausible. Because the complex nature of the problem that does not allow knowing the best solution, is not possible use real data to compare methods or investigate parameterization schemes, based on analysis of solutions (Beresnev, 2003). These studies can only be performed with synthetic data calculated from rupture models defined. Therefore we tested the stability and robustness of the alghoritm using synthetic waveforms computed from a defined slip distribution model. The results reveal the good likeness between the reconstructed model, using the DUAL LP inversion algorithm proposed and the known source model. The reconstruction tested with the same synthetic waveforms, but using the NNLS algorithm (considering the rupture propagation with constant velocity), differs from the model used, although showing a good fit of data (Fig?). We note that the two versions of LP inversion formulations (PRIMAL and DUAL) converge in the same solution, however with very different computing times. Using the same simplex inversion routine, the DUAL converge after about ~40min whereas the PRIMAL after ~14h. The method presented in this paper can be generalized to be used jointly with other data types (Geodetic and teleseismic waveforms). The results obtained in this paper encourage us to apply the algorithm proposed with real seismic and geodetic data. Acknowledgments This work has been developed with the support of the ‘Fundac¸a˜o para a Cieˆncia e a Tecnologia (FCT)’ (Science and Technology Foundation) of the ‘Ministe´rio da Cieˆncia, Tecnologia e do Ensino Superior (MCTES)’, through the project SISMOD/LISMOT (PTDC/CTE-GIN/82704/2006). 15 References Aki, K., and P. G. Richards, Quantitative Seismology: Theory and Methods (2 Volumes), Freeman and Company, San Francisco, 1980. Asano, K., e T. Iwata (2009), Source Rupture Process of the 2004 Chuetsu, MidNiigata Prefecture, Japan, Earthquake Inferred from Waveform Inversion with Dense Strong-Motion Data, BULLETIN OF THE SEISMOLOGICAL SOCIETY OF AMERICA, 99(1), 123-140, doi:10.1785/0120080257. Bao, H., J.Bielak, O.Ghattas, L.F.Kallivokas, D.R.O'Hallaron, J.R.Shewchuk, and J.Xu (1998). Large-scale simulation of elastic wave propagation in heterogeneous media on parallel computers, Comput. Methods Appl. Mech. Engrg. Vol.152, 85-102. Ben-Menahem, A. e Singh, S. J. (1981). Seimic Waves and Sources, Springer-Verlag, New York. Beresnev, I. A. (2003). Uncertainties in Finite-Fault Slip Inversions: To What Extent to Believe? (A Critical Review), Bull. Seismol. Soc. Am., 93(6), 2445-2458, doi:10.1785/0120020225. Bernard, P., and R. Madariaga, (1984). A new asymptotic method for modeling of near-field accelerograms, Bull. Seismol. Soc. Am., 74, 539-557. Bouchon, M., (1981). A simple method to calculate Green's functions for elastic layered media, Bull. Seismol. Soc. Am., 71, 959-971. Das, S., and B. V. Kostrov (1990). Inversion for seismic slip rate history and distribution with stabilizing constraints: application to the 1986 Andreanof Islands earthquake, J. Geophys. Res. 95, 6899–6913. Das, S., and P. Suhadolc (1996). On the inverse problem for earthquake rupture: The Haskell-type source model, J. Geophys. Res. 101, 5725-5738. 16 Delouis, B., M. Pardo, D. Legrand, e T. Monfret (2009), The Mw 7.7 Tocopilla Earthquake of 14 November 2007 at the Southern Edge of the Northern Chile Seismic Gap: Rupture in the Deep Part of the Coupled Plate Interface, Bull. Seismol. Soc. Am., 99(1), 87-94, doi:10.1785/0120080192. Gilbert, F. (1973), Derivation of source parameters from low-frequency spectra, Phil. Trans. R. Soc. Lond. Ser. A. 274, 369-371. Grandin, R., Borges, J.F., Bezzeghoud, M., Caldeira, B., Carrilho, F. (2007). Simulations of strong ground motion in SW Iberia for the 1969 February 28 (MS = 8.0) and the 1755 November 1 (M ~ 8.5) earthquakes – I. Velocity model, Geophys. J. Int., 171, 1144–1161, doi:10.1111/j.1365-246X.2007.03570.x Hartzell, S.H., and Heaton, T.H., 1983, Inversion of strong ground motion and teleseismic waveform data for the fault rupture history of the 1979 Imperial Valley, California, earthquake, Bull. Seismol. Soc. Am., vol. 73, no. 6, 1153-1184. Hartzell, S., P. Liu, C. Mendoza, C. Ji, e K. M. Larson (2007), Stability and Uncertainty of Finite-Fault Slip Inversions: Application to the 2004 Parkfield, California, Earthquake, Bull. Seismol. Soc. Am., 97(6), 1911-1934, doi:10.1785/0120070080. Hartzell, S., Frazier, G. e Brune, J. (1978). Earthquake Modeling in a Homogeneous Half-Space, Bull. Sismol. Soc. Am. 68, 301—316. Hernandez, B., Shapiro, N. M., Singh, S., Pacheco, J., Cotton, F., Campillo, M., Iglesias, A., Cruz, V., Gómez, J. e Alcántara, L. (2001). Rupture History of September 30, 1999 Interplate Earthquake of Oaxaca, Mexico (Mw=7.5) from Inversion of Strong-Motion Data, Geophys. Res. Lett. 28, 363—366 Holland, J. H. (1975). Adaptation in Natural and Artificial Systems, University of Michigan Press, Ann Arvbor, Michigan. 17 Ide, S., M. Takeo, e Y. Yoshida (1996), Source process of the 1995 Kobe earthquake: Determination of space-temporal slip distribution by Bayesian modeling, Bulletin of the Seismological Society of America, 86(3), 547-566. Ihmlé, P. F. (1998). On the Interpretation of Subevents in Teleseismic Waveforms: The 1994 Bolivia Deep Earthquake Revisited, J. Geophy. Res. 103, 17919-17932. Kennet, t, B. L. N, and N. J. Kerry, Seismic waves in a stratified half space, Geophys. J. R. Astron. Soc., 57, 557-583, 1979. Kirpatrick, S., C. D. Gelatt, and M. P. Vecchi (1983). Optimization by simulated annealing, Science 220, 671-680. Komatitsch, D. (1997). Me´thodes spectrales et e´le´ments spectraux pour l’e´quation de l’e´lastodynamique 2D et 3D en milieu he´te´roge`ne (Spectral and spectral-element methods for the 2D and 3D elastodynamics equations in heterogeneous media), Ph.D. thesis, Institut de Physique du Globe, Paris, France. Komatitsch, D., and J. P. Vilotte (1998). The spectral-element method: an efficient tool to simulate the seismic response of 2D and 3D geological structures, Bull. Seism. Soc. Am. 88, no. 2, 368–392. Komatitsch, D., and J. Tromp (2002a). Spectral-element simulations of global seismic wave propagation, I. Validation, Geophys. J. Int. 149, 390–412. Larsen, S.C. , Schultz, C.A., 1995. ELAS3D, 2D/3D Elastic Finite-Difference Wave Propagation Code, Lawrence Livermore National Laboratory, UCRLMA-121792, 18 pp. Larsen, S.C. & Schultz, C.A., 1995. ELAS3D, 2D/3D Elastic Finite-Difference Wave Propagation Code, Lawrence Livermore National Laboratory, UCRLMA-121792, 18 pp. 18 Lawson, C. e Hanson, R. (1974). Solving Least Squares Problems, Prentice-Hall (series in automatic computation), New Jersey Madariaga, R., 1976. Dynamics of an expanding circular fault, Bull. Seism. Soc. Am., 66 (3), 639–666. Miyake, H., T. Iwata, e K. Irikura (2003), Source Characterization for Broadband Ground-Motion Simulation: Kinematic Heterogeneous Source Model and Strong Motion Generation Area, Bulletin of the Seismological Society of America, 93(6), 2531-2545, doi:10.1785/0120020183. Mozziconacci, L., B. Delouis, J. Angelier, J. Hu, e B. Huang (2009), Slip distribution on a thrust fault at a plate boundary: the 2003 Chengkung earthquake, Taiwan, Geophysical Journal International, 177(2), 609-623, doi:10.1111/j.1365- 246X.2009.04097.x. Olsen, K. B. (2000). Site amplification in the Los Angeles Basin from 3D modeling of ground motion, Bull. Seism. Soc. Am. 90, S77–S94. Olsen, K. B., and R. J. Archuleta (1996). Three-dimensional simulation of earthquakes on the Los Angeles Fault System, Bull. Seism. Soc. Am. 86, 575–596. Pitarka, A., K. Irikura, T. Iwata, and H. Sekiguchi (1998). Three-dimensional simulation of the near-fault ground motion for the 1995 Hyogoken Nanbu (Kobe), Japan, earthquake, Bull. Seism. Soc. Am. 88, 428–440. Pitarka, A. (1999). 3D elastic finite-difference modeling of seismic motion using staggered grids with nonuniform spacing, Bull. Seism. Soc. Am. 89, 54–68. Robinson, D. P., e L. T. Cheung (2010), Source process of the Mw 8.3, 2003 TokachiOki, Japan earthquake and its aftershocks, Geophysical Journal International, 181(1), 334-342, doi:10.1111/j.1365-246X.2010.04513.x. 19 Saikia, C. K., Modified frequency-wave-number algorithm for regional seismograms using Filon's quadrature-modeling of L(g) waves in eastern North America, Geophys. J. Int., 118, 142-158, 1994. Spudich, P., and Frazer, L. N., Use of ray theory to calculate high frequency radiation from earthquake sources having spatially variable rupture velocity and stress drop, Bull. Seismol. Soc. Am., 74, 2061-2082, 1984. Suzuki, W., S. Aoi, e H. Sekiguchi (2009), Rupture Process of the 2008 Northern Iwate Intraslab Earthquake Derived from Strong-Motion Records, BULLETIN OF THE SEISMOLOGICAL SOCIETY OF AMERICA, 99(5), 2825-2835, doi:10.1785/0120080331. Tarantola, A. (1987). Inverse Problem Theory, Elseiver, Amesterdam. Udias, A. (1999). Principles of Seismology, Cambridge University Press, London. Yagi, Yuji. (2004) Source rupture process of the 2003 Tokachi-oki earthquake determined by joint inversion of teleseismic body wave and strong ground motion data, Earth Planet and Space, 56, 311-316. 20 FIGURES Fig. 1 – A) The 3 components (X,Y,Z) of synthetic seismograms to 2 stations (Station 1 and 2) produced by a slips in two directions from three subfaults (green, orange and blue) on a space time discretized fault plane are computed B) by the matrix product between Green Function’s of each subfault/station at each time-step, convolved with source time function, and the slips of correspondent subtault in same time-step. Fig. 2 – Geometric representation of DUAL formulation Solution Fig 3 – Geometry of the situation tested. At the surface the 13 triangles represent the stations; at the interior of the volume the fault plane. Fig. 4 – Individual panels (1 to 5) show the slip amplitude cumulative distribution (colorscale) at time windows of 0.78s. The panel (Total) shows the final slip distribution (red arrows) and the colorized contours show rupture time in 0.6-sec contours; The panel identified as “Complete STF” represents the rate of moment libertação Fig. 5- Comparison the original used source model (left) and the reconstructed by inversion of DUAL simplex method (right). The top trace is the stack power spatially integrated over the grid as a function of time, with red lines indicating the times of the spatial snapshots. Snapshots from the continuous back-projection (see Supplementary movie 1) indicating intervals in which the normal fault radiation dominates (during the first 50 s, typified by the 25 s frame), and when secondary events on the megathrust or in the upper plate occur (frames at 52 s, 91 s and 118 s). At 52 s, there is simultaneous radiation from the normal fault and the megathrust. The event at 118 s is probably in the upper plate, and although it is strong in the short-period radiation, it is not strong at long periods. The full animations for F-Net and five other networks are shown together in the Supplementary Information. Fig. 6- At left - Comparison the seismograms produced from the original used source model (blue) and reconstructed model (red); at right – total slip distribution of the two models: up- the original; bottom- the reconstructed from simplex DUAL. 21 Fig. 7- Comparison the original used source model (left) and the reconstructed by the NNLS method (right). Fig. 8- Left - Comparison the seismograms produced from the original used source model (blue) and reconstructed model by the NNLS method (red); right – total slip distribution of the two models: up- the original; bottom- the reconstructed from NNLS. Fig. 9- Comparison of the temporal evolution of the seismic moment rate (STF) of the three models: the original (blue); the reconstructed by Linear Programming (red); the reconstructed by NNLS (green). 22 Fig. 1 zj 1 u -1 1 zi Solution space -1 Fig. 2 23 Fig. 3 Fig. 4 24 Fig. 5 25 Fig. 6 26 Fig. 7 27 Fig. 8 Fig. 9 28 Tabeles Table 1 Position Position Slip in strike in dip strike in Slip in -dip axis axis direction direction (km) (km) 0 0 1.70 1.41 0 -2 4.83 -2 0 2 Init time Position Position Slip in strike in dip strike in Slip in -dip time axis axis (km) (km) 0.0 -4 6 2.20 3.31 2.1 8.05 0.6 4 6 2.72 3.28 2.1 0.54 0.43 0.6 -6 0 0.11 0.07 2.4 0 3.19 2.93 0.6 -6 4 2.08 2.21 2.4 0 2 2.79 3.49 0.6 0 8 2.06 2.56 2.4 4 2 4.46 6.68 0.6 2 8 0.12 0.09 2.4 -4 0 4.83 6.38 1.2 4 8 2.68 3.75 2.4 4 0 0.11 0.14 1.2 -4 4 4.13 2.54 2.7 -2 4 3.16 3.97 1.2 -6 2 4.92 3.41 2.7 0 4 4.16 3.47 1.2 -4 8 0.32 0.22 2.7 -2 -2 1.48 2.08 1.5 -2 8 1.24 1.27 2.7 2 -2 3.11 4.48 1.5 -6 8 0.63 0.37 3.0 -2 2 0.89 1.26 1.5 2 8 1.23 1.45 3.0 2 4 4.06 2.93 1.5 -4 8 1.70 1.41 3.0 4 -2 1.12 1.09 1.8 4 0 1.11 1.45 3.0 -4 2 0.15 0.12 1.8 -6 0 2.21 2.56 3.3 -6 -2 2.29 2.06 1.8 4 4 4.84 5.37 3.6 -2 6 4.43 2.79 1.8 -6 6 3.13 2.73 3.6 0 6 0.46 0.66 1.8 -4 2 2.10 2.35 3.6 2 6 2.73 2.54 1.8 -4 -2 1.96 2.12 3.9 2 2 0.09 0.09 2.1 2 2 2.12 3.12 3.9 -4 -2 0.06 0.04 2.1 1 (m) 2 (m) (sec.) direction direction Init 1 (m) 2 (m) 29 (sec.)