|
|
|
| Blocky models via the L1/L2 hybrid norm | |
|
Next: UNKNOWN SHOT WAVEFORM
Up: Claerbout: Blocky models: L1/L2
Previous: PLANE SEARCH
Here we set out to find blocky functions, such as well logs.
We will do data fitting with a somewhat -like convex penalty function
while doing model styling with a more -like function.
We might define the composite norm threshold residual
for the data fitting at the 60th percentile, and that
for regularization seeking spiky models (with blocky integrals)
as at the 5th percentile.
The data fitting goal and the model regularization goal at each
is independent from that at all other values.
The fitting goal says the reflectivity should be equal to its measurement (the seismogram).
The model styling goal says the reflectivity should vanish.
These two goals are in direct contradiction to each other.
With the L2 norm the answer would be simply
.
With the L1 norm, the answer would be either or depending on the numerical choice of .
Let us denote the convex function and its derivatives
for data space at the residual as
and for model space as
.
Remember, and , while normally vectors,
are here scalars (independently for each ).
loop over all time points {
# These are scalars!
loop over non-linear iterations {
Get derivatives of hybrid norm and for data goal.
Get derivatives of hybrid norm and for model goal.
# Plan to find to update
# Taylor series for data penalty
# Taylor series for model penalty
#
} end of loop over non-linear iterations
} end of loop over all time points
To help us understand the choice of parameters , , and ,
We examine the theoretical relation between and implied by
the above code as a function of and at
,
in other words, when the data has normal behavior and we are mostly
interested in the role of the regularization drawing weak signals down towards zero.
The data fitting penalty is
and its derivative .
The derivative of the model penalty
(from equation (15))
is
.
Setting the sum of the derivatives to zero we have
|
(27) |
This says is mostly a little smaller than ,
but it gets more interesting near
.
There the slope
which
says an
will damp the signal (where small) by a factor of 5.
Moving away from we see the damping power of
diminishes uniformly as exceeds .
|
|
|
| Blocky models via the L1/L2 hybrid norm | |
|
Next: UNKNOWN SHOT WAVEFORM
Up: Claerbout: Blocky models: L1/L2
Previous: PLANE SEARCH
2009-10-19