I seek to provide an electronic textbook with dozens of readily reproducible examples of application of geophysical inverse theory. On my sabbatical leave during the academic year '93-'94, I gathered four wide ranging geophysical data sets, depth sounding (Galilee), radar (Vesuvious), satellite altimetry (Madagasgar), and Seabeam (ocean ridge), and began preparing TDF, ``Three dimensional filtering: Environmental soundings image enhancement". The first time I taught this material I discovered that three of the four data sets had problems with severe noise and the fourth had missing values where someone had previously edited out the bad points. Turning a vice into a virtue, I realized that no guide to data fitting is satisfactory without a collection of useful tools for dealing with very noisy data.

I tried to iteratively reweight residuals by the inverse of the observed residual variance. Unfortunately, such reweighting is an art that is difficult to teach, difficult to reuse, and often fails. Iterative reweighting encourages ad-hoc solutions that defy the formal optimization literature. However, iterative reweighting can hardly be avoided since real data is noisy and its variance is usually nonstationary. We should be able to greatly reduce the need for ad hoc reweighting if we use robust methods.

Theoretically,
it is well known
(FGDP)
that the *L _{1}* norm approach
allows for a data set containing some infinite data values (noise).
A straightforward approach to geophysical inversion
(model fitting or regression)
is a mixed

(1) |

I did some preliminary tests on a regression like
equation (1) (but all *L _{1}*).
I was disappointed that twenty times more iterations than usual
were required and each iteration was significantly more costly.

We will see that with the Huber approach, the conjugute-direction method is only slightly changed. A big question is whether the convergence rate remains attractive. This is important for geophysical problems which are usually large.

11/12/1997