** Next:** Tests on real data
** Up:** comparing IRLS and Huber
** Previous:** Tests on synthetic data:

In this example, we contaminated the data with a low frequency/low velocity linear event that simulates,
to some extent, groundroll (Figure 5, left).
Because the operator **A** models only hyperbolas, the simulated groundroll introduces some
inconsistent data. It is therefore interesting to test the sensitivity of our inversion schemes (*l*^{2} and *l*^{1})
for this new problem.
The *l*^{2} inversion of the data leads to a fairly noisy model where
the five expected events are difficult to pick (Figure 5). However, without
groundroll and only a few artifacts, the remodeled *l*^{2} data does not display so dramatic effects.
With a ``*l*^{1}'' inversion, the model space is more accurately resolved, as shown in Figure
7.
Again, there are few differences between IRLS and the Huber solver, and it looks like
the ``*l*^{1}'' norm copes more easily with inconsistent events than *l*^{2}
(groundroll in this case).
**gr-freq30-L2-HUBER
**

Figure 5 From left to right: 1) Input data. 2) velocity
domain, *l*^{2} inversion result. 3) remodeled data from the *l*^{2} result

**comp-data-freq30
**

Figure 6 From left to right: 1) Remodeled data with *l*^{2}.
2) Remodeled data with the Huber solver. 3) Remodeled data using IRLS

**comp-mod-freq30
**

Figure 7 From left to right: 1) *l*^{2} velocity space.
2) Huber velocity space. 3) IRLS velocity space

** Next:** Tests on real data
** Up:** comparing IRLS and Huber
** Previous:** Tests on synthetic data:
Stanford Exploration Project

4/27/2000