To understand this concept more clearly, it is best to begin with a very simple operator: nearest neighbour interpolation () and its adjoint (binning). For this operator, the Hessian matrix, , is exactly diagonal, and it would be possible to calculate the diagonal values accurately with any reasonable choice of reference model in equation (). A particularly simple choice of reference model is to fill the model-space with ones. Since the operator is nearest-neighbour interpolation, this results in a data-space full of ones too. Binning this data vector gives the fold of the operator in model-space Claerbout (1998a); and its inverse can be used directly as a weighting function for inverse nearest neighbour interpolation,
(130) | ||
For an example that is slightly more complex than nearest neighbour interpolation, consider linear interpolation. The Hessian matrix for linear interpolation, is tridiagonal rather than exactly diagonal; therefore, unless we know the true solution, any diagonal operator that we produce will be an approximation. This leads to a conundrum: is it better to find the diagonal of , or another approximation that incorporates information about the off-diagonal elements?
Following the same approach as above, a vector full of ones seems a reasonable choice for a reference model; and as before, this generates a data vector full of ones. Applying the adjoint of linear interpolation to this data vector produces the model-space fold, ,that can be used in the approximation,
(131) |
Figures and strengthens this conclusion. The top panels show a simple regularly-sampled model consisting of a four events. The middle panels show data-points obtained by linear interpolation from this model, and the lower panels show the reestimated models after normalized binning with the adjoint of the linear interpolator. The thin solid-lines, the dashed-lines and the dotted-lines show the true model and the results of normalization by the matrix diagonal and operator fold, respectively.
For Figure , I sampled one hundred data points. Only where the model consists of an isolated single spike (at m5) does the diagonal normalization out-perform operator fold. Elsewhere the amplitude of the original function is recovered more accurately by operator fold. For Figure , I only sampled 30 data points, leading to a less well-conditioned system than Figure . The row-based fold normalization scheme shows itself to be more robust in areas of poor coverage than the scheme that considers only the diagonal elements of the matrix.
linterp100
Figure 1 Comparison of normalizing adjoint linear interpolation with the row-sum of as opposed to its diagonal. Top panel shows the model. Center panel shows 100 interpolated data points. Lower panel shows reestimated data: dashed-line after diagonal normalization, dotted-line after column-based normalization. The solid-line shows the true model. |
linterp30
Figure 2 Same as Figure , but with only 30 data points. The diagonal-based approach (dashed-line) is less robust in areas of poor coverage than the row-sum approach (dotted-line). |