I sincerely thank My great amigo Zoli for being Zoli I sincerely thank this amigo and this amigo for the kind invite I wish to, humbly and most sincerely, thank the The Society that publishes 2 of the very best and holds meetings in exotic places for honoring me with this marvelous and unique adventure This tour would have been a rout without Judy Wall who organised it ALL !!! Imagine. Angels do exist in the sky. With additonal thanks to Tury Taner, what can I say?, he who has done it all. Enders Robinson, he was and is, numero uno. Sven Treitel, there are no words, except, Sven. Arthur Weglein, my friend, my teacher. Mauricio Sacchi, without whom Tad would be Tad who? Those marvelous friends, colleagues, students, who must assume full responsibility for making me who I have become. Membership has its Advantages Scholarly Journals in Print and Online Networking Opportunities Receive Membership Discounts on: Professional Development Courses Publications Workshops and Meetings Need more information about joining SEG? SEG Membership Brochures and Applications are available today! Join Online http://membership.seg.org and now the taaaalk The role of Amplitude and Phase in Processing and Inversion Tadeusz Ulrych I have chosen this title, because I can talk about ANYTHING !! This presentation was prepared while partying in the local bar, illustrated in the next slide Definitions Consider x = s + n n is " all other stuff" , and is generallycalled" noise" Letting F represent Fourier transformation X = F [ x] = A x e i x where A x is the amplitude spectrum x is the phase spectrum A brief story Doug Foster arranges a presentation for Monday Dr. Doug J. Foster This is Me Sunday evening is slightly brutal I cannot remember [1] How many participants? [2] Where is my presentation? I have a Canadian cell with enough credit for ONE question What question do I ask? How many participants? or Where is the presentation? The answer to HOW MANY? is AMPLITUDE (goodbye presentation and future invitation) The answer to WHERE? is PHASE (Oblivious to the number, I blindly carried on) WHERE ? Information encodedin x HOW BIG ? Information encodedin Ax Relative “Importance” of Ax and x ? Original x only, Ax = 1 INTRODUCTION Mathematics is Beautiful. However, it is tiresome to digest. Therefore, this talk contains as little of this beauty as possible. Please remember, that the magic of mathematics lies in its physical interpretation. For example …. Question Why is it true that (-1)1/2 x = Because, as is well known (-1)1/2 = i and i is an operator that rotates by 90o Amplitude & Phase in blind deconvolution The Enders example The Man Enders Robinson The canonical model for the seismogram x t = w t ¤ qt + n t xt is the seismogram signature = w t ¤ qist the + nsource t w t ¤ q t + nistthe Greens function, the reflectivity t qt + n t is ‘everything else’, the noise This equation, x t = w t ¤ qt + n t is 1 equation with 2 unknowns. This is akin to 7= a + b and what is a and b uniquely ? This, of course, is an impossible problem unless a priori constraints are known or, at least, assumed Some more thoughts regarding Phase OUTLINE for the next few slides POCS and only-phase reconstruction Phase and cepstral processing Summary POCS Projection onto convex sets POCS attempts to solve an underdetermined, generally nonlinear, inverse problem G[x]+n=d where G is a nonlinear operator A convex set, A, is one for which the line joining any two points, x and y, in the set, is totally within the set. In other words, a set A in a vector space is convex, iff x and y Є A λx + (1 - λy) Є A 0 ≤ λ ≤ 1 Illustrating convex and non-convex sets A convex set A non-convex set Alternating POCS Iterative projection onto convex sets Possible stagnation point when one of the sets is non-convex Application of alternating POCS to the problem of reconstruction from phase-only to obtain the only-phase image The image, of finite support , is a convex set. The set of constraints, the thresholded image, is also another convex set. Phase-only Only-phase Original Phase in Cepstral analysis Phase is fundamental in cepstral processing Phase must be unwrapped Phase must be detrended A serious problem is additive noise The cepstrum (complex) is defined as -1 C(n) = F {ln[A(ω)] + iΦ(ω)} where -1 F is the inverse Fourier transform Application of cepstral analysis to thin bed blind deconvolution Compute cepstrum for each trace Stack the cepstra Transform back to the time domain Deconvolve with estimated wavelet The original synthetic section The original reflectivity The recovered wavelet Usual approach to deconvolution with ‘known’ source wavelet R(f)=X(f)W(f)H/(W(f)W(f)H+k) f-domain deconvolution BUT, we can do better! By utilizing a concept which we, and particularly Jon Claerbout and Mauricio Sacchi, have championed for over a decade. The principle of PARSIMONY some details to follow The original reflectivity Sparse deconvolution f-domain deconvolution Summary thus far Phase contains the vital information about location Only-phase reconstruction demonstrates the flexibility of POCS in inverse problems Proper phase processing leads to useful cepstral decompositions Some details concerning PARSIMONY or SPARSENESS Thanks to Mauricio Sacchi for help with PARSIMONY The concepr of Sparseness I honour the sparse ones .. Nicholas Copernicus Pierre de Laplace Thomas Bayes Sir Harold Jeffreys Edwin Jaynes John Burg and, of course, the sparsest of them all … An hour-long recording in the night sky Processing pre-Burg Processing pre-Burg 1.0 3.0 Frequency (cycles/hour) 5.0 Why extend with 0’s ? Why not ? Is this not the least presumptive ? Only if the star lived for 1 hour Processing post-Burg ? ? Question? How does one turn a ? into mathematics? John Burg’s answer: ?= Processing post-Burg 1.0 3.0 Frequency (cycles/hour) 5.0 and the actual fabricated star … Key points of this part q Importance of sparseness in the recovery of low/high frequencies Spectral Extrapolation Sparse Inversion Blind Deconvolution Methods (MED, ICA etc.,) q Assumptions for the recovery of missing frequency components A few words about the problem Recovery of Green’s function from band limited data * s(t) = w(t) r(t) + n(t) s = Wr+ n Seismogram = Source Impulse Response + Noise The required inversion is performed by J (m norm) (d constr.) We use: p(m | d) p(d | m) p(m) to obtain J Priors to model sparse signals Two well-studied priors for the solution of inverse problems where sparsity is sought: Laplace Cauchy These priors translate into regularization constraints for the solution of inverse problems The latter is done via the celebrated Bayes Theorem How does it work? Define a cost function (derived from Bayes) and minimize it If all the hyper-parameters of the problem were properly chosen, the minimization should lead to solutions that – a) honor the data – b) are simple (Sparse) A sparse solution is associated with a signal with high frequency content. This is why sparse solutions are often used for problems of bandwidth recovery. Some Math….. R(r) | rk | ql1 norm k R(r) ln(1 qCauchy Norm qBayesian Cost to minimize: k rk2 2 ) J || Wr d ||22 2 R(r) 2 J = Misfit + (Regularization term derived from prior) Solution e.g. for regularization using the Cauchy norm ∇J = ∇{| | Wr - d | |22 + 2 R (r )} = 0 r = [ W W + Q (r)] W T 2 -1 T 1 Q ii = 2 + ri2 The last equation is solved using an iterative algorithm to cope with the nonlinearity Damped LS: all the unknown samples are damped by the same amount Qii 1 Cauchy: adaptive damping 1 1 Qii 2 r 2 2 i 0 ri 1 Qii 2 r 0 2 i ri Adaptive damping is what leads to sparse solutions Example: Non-Gaussian Impulse Response model via a Gaussian Mixture More area under green curve SPARSENESS Sparsity is controlled by the mixing parameter Mixing Parameter p=0.8 Data True impulse response Predictd data Estimated impulse response Mixing Parameter p=0.2 Data True impulse response Predicte d data Estimate impulse response Key features for proper recovery of the impulse response # Sparseness # Bandwidth Source BW (p=0.5) Error = difference between true and estimated impulse response Source functions used in the simulation AR Prediction in the f-domain ? ? ? AR Gap-filling algorithm AR Gap-filling algorithm (contd.) AR Predictive Extension Time True Recovered Input BL signal Frequency Summary [1] The eye is attracted to the light, but the mystery lies in the shadows. [2] Gaussian pdf’s imply Least Squares. [3] The mystery, the ? , lies in the heavy tails of nonGaussian pdf’s. The role of Phase in the attenuation of Surface and Internal Multiples The next few slides have been supplied by Arthur Weglein a friend and mentor The 1D FS multiple removal algorithm Data without a free surface R() 1 Data with a free surface R f () 1 contains free-surface multiples. Free surface demultiple algorithm R( ) R() R() R2 ( ) = primaries and internal multiples R f ( ) = primaries, free surface multiples and internal 1 multiples R( ) Total upfield R f ( ) 1 R( ) R f ( ) and, R ( ) 1 R f ( ) R( ) R f ( ) R f ( ) R f ( ) 2 3 Free surface demultiple example t1 2t2 2t1 t1 + t2 t2 R f ( t ) R1 ( t t1 ) R2' ( t t 2 ) R12 ( t 2t1 ) R2'2 ( t 2t 2 ) R f ( ) R1e it1 R2' e it 2 R12 e 2 it1 R2'2 e 2 it 2 2 R1 R2' e i ( t1 t 2 ) 2 2 2 i t 1 f 1 R ( ) R e ' 2 2 i t 2 R e 2 2R R e ' 1 2 2 i ( t 1 t2 ) ... t1 + t2 2t1 2t2 So R f () R2f () precisely eliminates all free surface multiples that have experienced one downward reflection at the free surface. The absence of low frequencies (and in fact any other frequencies) plays absolutely no role in this prediction. Please note that this Inverse Scattering approach to the attenuation of both surface and internal multiples, does not require knowledge of the velocity structure of the subsurface Measurement surface subsurface Mississippi Canyon Water Bottom Top Salt Base Salt Internal multiple Water Bottom Top Salt Base Salt Internal multiple algorithm b3 ( k g , k s , q g q s ) 1 2 2 dk1dk2e iq1 ( e g e s ) iq2 ( e g e s ) e dz1e i ( q g q1 ) z1 b1 ( k g , k1 , z1 ) z1 z2 i ( q2 q s ) z3 i ( q1 q2 ) z 2 dz e b ( k , k , z ) dz e b1 ( k 2 , k s , z3 ) 1 1 2 2 3 2 where b1 ( k g , k s , q g q s ) 2iqs D( k g , k s , ) 2iqsG s ( k g , k s , ) and z3 z 2 , z1 z 2 Araújo and Weglein (1994) The role of phase is clear and is of central importance. q Surface multiple attenuation involves convolution. q Internal multiple attenuation involves both convolution and correlation. Amplitude is, of course, also important. However, it is much less crucial than Phase. The reason is that if the Location is wrong, multiple attenuation will give birth to more multiples. (perhaps with the correct amplitude) Mississippi Canyon Input Predicted Output multiples (2D) Input Predicted Output multiples (2D) 1.7 Water bottom Seconds Top salt Base salt Internal multiples 3.4 Common Offset Panel (1450 ft) Common Offset Panel (2350 ft) It is time to Fly away But, One last slide Thank you for your Patience