The properties we have just discussed are strongly related to the
implicit nature of mid-point rule. As consequence of this implicit
nature, we have to solve the time-stepping
Eq. (4.33) for the unknown
at each
time step which amounts to solve the following system of
nonlinear equations in the unknowns
:
(4.43)
where
is the following vector function:
(4.44)
and where
eff
(4.45)
is the right-hand-side of the conservative LLG equation. It is
interesting to notice that the damping is present in only one term
in the function
and it introduce only a slight
modification of the function.
The solution of the system of equation (4.43)
can be obtained by using Newton-Raphson (NR) algorithm. To this
end, we derive the jacobian matrix
F of the
vector function
which, after simple algebraic
manipulations, can be written in the following form
Ff
(4.46)
where
f is the jacobian matrix associated to
. By using the Eqs.(4.25),
(4.45), one obtains:
f
(4.47)
The main difficulty in applying NR method is that the Jacobian
F of
is a full matrix, due to
the long-range character of magnetostatic interactions which
reflects in the full nature of the matrix
. In this
connection, let us observe that the damping term affect only a
small sparse component of the jacobian
F and
thus does not introduce any basic difficulty.
Anyhow, due to the full nature of
F, the use
of the plain NR method would require an unpractical computational
cost. However, as it is usual in solving field problems with
implicit time stepping, we can use a quasi-Newton method by
considering a reasonable approximation of the Jacobian. In order
to have a sparse Jacobian one can consider the following
expression
F in which magnetostatic interactions are
not included:
Ff
(4.48)
where the matrix
f is
fexan
(4.49)
Basically, the latter equation is obtained by substituting the
full matrix
with its sparse component
exan in Eq. (4.47). Thus, the
iterative procedure can be summarized as follows:
withF
(4.50)
At each iteration, the linear system defined by the matrix
(4.49) has to be inverted. Since this matrix is
non-symmetric, we have found appropriate to use generalized
minimum residual (GMRES) method [91]. The
iteration (4.50) is repeated until the norm
is under a prescribed tolerance.
The iterative technique we developed to solve
Eq. (4.43) belongs to the main category of
quasi-Newton methods. In this respect, it has been
proven [92] that this kind of iterative procedure is
convergent and the order of convergency is the first order,
provided that the initial guess is sufficiently close to the
`true' solution of the system.
Next:4.6 Accuracy tests for Up:4. Geometric integration of Previous:4.4.1.3 Preservation of HamiltonianContents
Massimiliano d'Aquino
2005-11-26