next up previous print clean
Next: Linearization Up: Theory of wave-equation MVA Previous: Objective function

Gradient

Optimization of the objective function in Equation (19) requires computation of its gradient with respect to slowness. The objective function J can be rewritten using the inner product as:
\begin{displaymath}
J\left(s\right)= \frac{1}{2} \left< {\bf I}\left(\AA \u - {\...
 ...ght), {\bf I}\left(\AA \u - {\bf B}\mathcal T\right) \right\gt.\end{displaymath} (20)
A perturbation of the function J is related to a perturbation of the wavefield by the relation:
\begin{displaymath}
\delta J\left(s\right)= \left< {\bf I}\left(\AA \u - {\bf B}\mathcal T\right), {\bf I}\AA \delta \u \right\gt.\end{displaymath} (21)
If we replace $\delta \u$ from Equation (11) we obtain:
\begin{displaymath}
\delta J\left(s\right)= \left< {\bf I}\left(\AA \u - {\bf B}\mathcal T\right), {\bf I}\AA {\bf G}\delta s \right\gt,\end{displaymath} (22)
therefore the gradient of the objective function can be written as
\begin{displaymath}
\nabla_sJ= {\bf G}^* \AA^* {\bf I}^* {\bf I}\left(\AA \u - {\bf B}\mathcal T\right).\end{displaymath} (23)

Following the definition of the operator ${\bf G}$, we can write
\begin{displaymath}
{\bf G}^* = {\bf S}^* {\bf E}^* \left[\left({\bf 1}- {\bf E}...
 ...bf E}^* \left[\left({\bf 1}- {\bf E}\right)^{ *} \right]^{-1} .\end{displaymath} (24)

Finally, the expression for the gradient of the objective function with respect to slowness becomes  
 \begin{displaymath}
\nabla_sJ= 
 {\bf S}^* {\bf E}^* \left[\left({\bf 1}- {\bf E...
 ... \AA^* {\bf I}^* {\bf I}\left(\AA \u - {\bf B}\mathcal T\right)\end{displaymath} (25)
which takes special forms depending on our choice of the operators $\AA$ and ${\bf B}$:

WEMVA by TIF WEMVA by DSO
$ \nabla_sJ= {\bf S}^* {\bf E}^* \left[\left({\bf 1}- {\bf E}\right)^{ *} \right]^{-1}
 {\bf I}^* {\bf I}\left(\u - \mathcal T\right)$ $ \nabla_sJ= {\bf S}^* {\bf E}^* \left[\left({\bf 1}- {\bf E}\right)^{ *} \right]^{-1}
{\bf D}^* {\bf I}^* {\bf I}{\bf D}\u $

The gradient in Equation (25) is computed using the adjoint state method, which can be summarized by the following steps:

1.
Compute by downward continuation the wavefield
\begin{displaymath}
\AA^* {\bf I}^* {\bf I}\left(\AA \u - {\bf B}\mathcal T\right).\end{displaymath} (26)
2.
Compute by upward continuation the adjoint state wavefield
\begin{displaymath}
\mathcal W= \left[\left({\bf 1}- {\bf E}\right)^{ *} \right]...
 ...\AA^* {\bf I}^* {\bf I}\left(\AA \u - {\bf B}\mathcal T\right),\end{displaymath} (27)
i.e. solve the adjoint state system
\begin{displaymath}
\left({\bf 1}- {\bf E}\right)^{ *} \mathcal W= \AA^* {\bf I}^* {\bf I}\left(\AA \u - {\bf B}\mathcal T\right).\end{displaymath} (28)
3.
Compute the gradient
\begin{displaymath}
\nabla_sJ= {\bf S}^* {\bf E}^* \mathcal W.\end{displaymath} (29)

next up previous print clean
Next: Linearization Up: Theory of wave-equation MVA Previous: Objective function
Stanford Exploration Project
11/11/2002