 |
 |
 |
 | Wave-equation tomography by beam focusing |  |
![[pdf]](icons/pdf.png) |
Next: Numerical computation of search
Up: Gradient of the objective
Previous: Derivatives with respect to
The evaluation of the derivatives of the moveout parameters
with respect to slowness
follows a slightly different procedure from the one above
because the moveout parameters are solutions
of the optimization
problems 5 and 9.
We take advantage of the fact that
we need to evaluate the derivatives only at the solution
points, where the objective functions
are maximized and thus their derivatives with respect to the moveout
parameters are zero.
We can therefore write:
and
Using the rule for differentiating implicit functions,
and taking
advantage that the fitting problems are all independent
from each other (i.e. the cross derivatives with respect to the moveout
parameters are all zero), we can formally write:
 |
(A-20) |
and
 |
(A-21) |
Appendix A presents the analytical development of these expressions
to compute the derivatives of the moveout parameters with respect to slowness.
As for the derivatives of the main objective function
with respect to moveout parameters,
the final results for the special case of
and
have
a fairly simple analytical expression.
The derivative of the local moveout parameters are
(A-1):
 |
(A-22) |
and the derivative of the global moveout parameters are
(A-2):
 |
(A-23) |
in which
.
In both equations 22 and 23
the operator
represents a convolution with the recorded data,
whereas the operator
is the basic wave-equation tomography operator
that that links perturbations in the slowness model
to perturbations in the modeled data.
Combining the derivatives
in equation 22
with the derivatives in
equations 16-17
we can compute the gradient of the local objective
function 3 with respect to slowness as:
 |
(A-24) |
I will now examine the effects of each of the terms
in equation 24
starting from the rightmost one.
The third term (III) produces a scalar
for each local curvature parameter
.
This scalar multiplies the traces in each beam,
after they have been differentiated in time and scaled by
,
as described by the second term (II).
Notice that the phase introduced by the time derivative
of the correlation function in (II)
is crucial for the successful backprojection into the slowness model
that is accomplished by the first term (I).
In this term, first
projects the traces
of each individual beam into the space of the global array,
then the convolution with the recorded data
time shifts the correlation function by the time delay of the events.
Finally,
the adjoint of the operator
backprojects the perturbation in the wavefields at the receiver array
into the slowness model.
The expression of the gradient of the global objective
function 7 with respect to slowness
is similarly derived by combining the derivatives
in equation 23
with the derivatives in
equations 18-19
and is the following three-terms expression:
The structure of equation 25
is similar to the structure of equation 24
and the terms have similar explanations.
The only important difference is that
in term II the chain
performs the stack over the local arrays and the assemblage
of the stacked traces into the global array,
whereas its adjoint in term I
spreads the stacked traces back into the local arrays
reforming the local beams.
 |
 |
 |
 | Wave-equation tomography by beam focusing |  |
![[pdf]](icons/pdf.png) |
Next: Numerical computation of search
Up: Gradient of the objective
Previous: Derivatives with respect to
2010-05-19