For applications with small numbers of unknowns, say less than dozens or hundreds, the damped inverse operator is generally the most appropriate, whereas for problems when the unknown model is a field of more than thousands of unknowns, say reflectivity or velocity as a function of two coordinates, then the conjugate operator and its various extensions are, in practice, the answer.

Between the conjugate operator and the inverse operator lies an operator (often overlooked) that is sometimes a guide to the most effective compromise. This operator is a certain unitary operator. The solution and the unitary operator are given by

(5) |

(6) |

More typically, in practice the square-root operator is not easily found, nor is it needed precisely. This leads to

(7) |

One of the most important linear operators in seismic exploration is velocity analysis (summing along hyperbolic trajectories of various velocities). After having defined such a transform, you can push data into velocity space and pull it back again with the operator .Unless by chance or design you have defined your operator to be already unitary, you will find, when comparing the original data to that the processing weakens events at various times, locations, or frequencies. The remedy for this is to redefine the operator with various scale factors, such as that you can uncover either experimentally, theoretically, or by looking at Claerbout 1992a All this activity amounts to looking for a unitary representation such as equation (7). In my experience in geophysics, such efforts are not as systematic as they should be and could be. We should make a greater effort to digest the literature on preconditioning. Further, the Fortran language itself inhibits segregation of numerical-analysis concepts from application-specific concepts. Thus successful applications of advanced numerical analysis concepts are many fewer than we would like to see. This motivates a C++ study in this report Nichols et al. (1992).

Finally, I wish to cite another example of operator factorization for the inverse needed by (7). That example is the two-dimensional prediction-error operator in Abma and Claerbout, Abma and Claerbout (1992). The confusion caused by that paper at the 1992 SEP sponsor meeting is one reason I prepared this short note. The other reason, is to bring up again our lack of systematic use of preconditioning when we attack huge inverse problems. To begin with, I believe that the practical search for unitary operators is closely related to the practical search for preconditioners, though I have not formally confirmed it.

11/17/1997