In the level-set approach, we generate a sequence of functions
k
(
x
) such that
D
k
!
D ;
where
@ D
k
=
f
x
:
k
(
x
) = 0
g
. Note that
k
may b e a continuous parameter, representing time, if
we consider evolution.
In an obstacle reconstruction problem, we can think of the variable
u
as the desired unknown.
The function
u
in (1b), can b e given a level-set description as follows. We intro duce an unknown
function
(
x
) which is tied to
u
in
u
(
x
) =
(
u
int
for
f
x
:
(
x
)
<
0
g
u
ext
for
f
x
:
(
x
)
>
0
g
:
(2)
Several advantages of this mo de of representing the unknown through the function
(
x
) b ecomes
apparent:
1. No a priori assumptions about the connectedness (topology) of
D
need to b e made.
That is,
D
could b e made up of several disconnected subregions. Moreover,
D
could
b e multiply connected.
2. No a priori assumptions on the nature of
D
need to b e made, e.g., one often nds in
the scattering literature the assumption that the unknown obstacle is star-shap ed.
On the other hand, one immediately sees that in using this description, a linear inverse problem
b ecomes nonlinear. This is due to the nonlinear dep endence of
u
on
.
Under this description, the inverse problem stated in (1) becomes:
Find
(
x
)
in
u
(
x
) =
(
u
int
f
x
:
(
x
)
<
0
g
u
ext
f
x
:
(
x
)
>
0
g
;
such that
A
(
u
) =
g :
We will next describ e two approaches for nding
(
x
) in such a problem. The rst one is based on
a time evolution, leading to a Hamilton-Jacobi equation. The second is based on optimization. We
note that our motivation for prop osing this approach is due to its exibility in describing obstacles.
We do not exp ect this formulation to help in making an illposed problem b etter b ehaved.
3. Formal calculation of variations.
In order to nd the dep endence of the forward map
on small changes on the obstacle b oundary, we need to calculate the variation of
u
caused by a
variation in
. To facilitate this calculation, let
x
b e a p oint on the curve
@ D
=
f
x
:
(
x
) = 0
g
.
Supp ose
(
x
) is p erturb ed by a small variation
(
x
). Let
x
b e the resulting variation of the p oint
x
. The variation results in the region
D
to become a new region denoted by
D
0
.
By taking the variation of the equation
(
x
) = 0, we nd
+
r
x
= 0
:
(3)
Observe that the unit outward normal at
x
is
n
(
x
) =
r
(
x
)
jr
(
x
)
j
:
3