Trial free-boundary methods
325
where
<b~k)(S),
etc., denote approximations at the point
X(k+l)(S),
obtained
by interpolation
if
the
point lies inside the seepage region and by
extrapolation
if
it is outside. Applications
of'
the method' often take
<b(k)(S)
to
be
the value of
<b(k)
at the grid point nearest to
X(k+l)(S).
Cryer
(1976b) reported a number of uses of this method. In the simple dam
problem, for example, an approximation to the velocity field, say
V~k),
can
be determined by differentiating the approximation to the velocity poten-
tial,
<b~k).
The free boundary
is
a streamline for the condition
C<b
=
<bn
= 0
and so an improved approximation,
r(k+l>,
can be obtained by integrating
the velocity field
V(k).
Cryer (1976b) concluded from the practical evidence available that trial
free-boundary methods as described so far are stable and converge
satisfactorily for porous
flow
problems. Any local method can
be
used
and smoothing
is
not required. Where there has been a suspicion of
instability (Taylor and Brown 1967; Neuman and Witherspoon 1970;
Kealy and Busch 1971) it seems to have occurred in the region where the
free boundary joins the seepage surface. The trouble may be associated
with the precise nature of the boundary conditions there, which can
include the
type of singularity discussed by Aitchison (1972) and in §8.3.
Shaw and Southwell (1941) asserted that in porous
flow
problems,
if
the
initial guess
r(O)
for the free boundary
is
horizontal then successive
trials
r(k)
converge monotonely downwards. They based a proof on the
maximum principle.
Cryer
(1976b), however, gives examples of instabilities in more general
free-boundary problems for which global methods may be preferred.
(iii) Global methods. In a global method the solution
is
computed for a
number of independent choices of
r(k)
and then, in principle, it
is
possible
to use inverse interpolation
or
some other technique to determine the
new trial boundary
r(k+l).
For example, the least-squares error in the
boundary condition on
r(k+l)
can be minimized on the evidence of the
dependence of this error on the parameters defining the set of chosen
r(k).
Even so, Fox and Sankar (1973) advise against local improvement of
mesh points by solving an inverse interpolation problem to find a new
position of just one mesh point on
r(k+l),
leaving all other points
untreated.
In
order to avoid the oscillations of the mesh points on
r(k+l)
which local improvement can produce, they propose that all mesh points
should be improved simultaneously. Their algorithm involves solving the
appropriate fixed boundary problem for
n + 1 different choices of r
initially but then only one new fixed-boundary problem has to be solved
per
iteration. Both the problem and the algorithm devised by Fox and
Sankar (1973) have features of general interest and are now described in
more detail.