Figure illustrates a process for factorizing a non-stationary filtering matrix. This approach is accurate if the filters vary smoothly in space, and it is the approach that I follow for the examples in this chapter.

The first step is to build the extrapolation matrix. Then we can extract the impulse response associated with each spatial location, and factor it into causal and anticausal components. By inserting the causal components into the column of a lower triangular matrix, and the anticausal components into a row of an upper triangular matrix, we can begin to build the two invertible matrices. The process has to be repeated for each spatial location.

Figure 1

In practice, we can save time by factoring the filters in advance, and storing them in a table. For every value of , we precompute the factors of the 1-D helical filter, , and store filter coefficients in a look-up table. We then extrapolate the wavefield by non-stationary convolution [equation ()], followed by non-stationary polynomial division with equation ().

Since we interpolate filters, not downward continued wavefields as in
phase-shift plus interpolation migration Gazdag and Sguazzero (1984), the number of
reference velocities used has minimal effect on the overall cost of
the migration.
Indeed, the cost of this process is proportional to the number of grid
nodes times the number of filter coefficients; so for a fixed number
of filter coefficients, the cost is linear in the number of grid
nodes [*O*(*N*)].

5/27/2001