Page 142 - Basic Structured Grid Generation
P. 142

Differential models for grid generation  131

                          It follows that
                                                T        T       T
                                               p r i+1 = p r i − λ i p Ap i = 0            (5.63)
                                                                 i
                                                i
                                                         i
                        when λ i satisfies eqn (5.60).
                          Suppose that, following a search for a minimum along the direction p i from the
                        ‘point’ u i , we have reached the point u i+1 . We can calculate r i+1 from eqn (5.62),
                        and now we seek the value of α i in eqn (5.61) which will yield the new direction of
                        search p i+1 .
                          Substituting u = u i+1 + λ i+1 p i+1 = u i+1 + λ i+1 (r i+1 + α i p i ) into eqn (5.55) gives,
                        in a similar manner to the derivation of eqn (5.57),
                              1 T
                                            T
                                                           r
                          E = u    Au i+1 − d u i+1 − λ i+1 r T i+1 i+1 − λ i+1 α i r T  p
                              2 i+1                                    i+1 i
                                                                     2 T
                                1
                                      2
                                                        T
                              + (λ i+1 ) {r T  Ar i+1 + 2α i (p Ar i+1 ) + (α i ) p Ap i }
                                2        i+1            i              i
                                                                          T
                                                                                       2 T
                            = E i+1 − λ i+1 r T  r  1 2  2  T i+1 Ar i+1 + 2α i (p Ar i+1 ) + (α i ) p Ap i },
                                         i+1 i+1 + (λ i+1 ) {r
                                                                                         i
                                                                          i
                        using eqn (5.63), where E i+1 is the value of E at u i+1 . To minimize E, we clearly
                        must choose α i so that the final term in brackets is minimized, and thus we obtain
                                                             T
                                                           p Ar i+1
                                                             i
                                                     α i =−        .                       (5.64)
                                                             T
                                                            p Ap i
                                                             i
                          The iterative scheme is now complete, but we make the observation that
                                                                   T
                                             p T  Ap i = r T i+1 Ap i + α i p Ap i = 0     (5.65)
                                                                   i
                                              i+1
                        when α i satisfies eqn (5.64). Thus the direction p i+1 is orthogonal to the direction of
                        the vector Ap i .Since A is symmetric, it is also the case that p i is orthogonal to the
                        vector Ap i+1 . The two directions p i , p i+1 are said to be conjugate with respect to the
                        symmetric matrix A.
                          To summarize, given a starting guessed solution u 0 , an initial residual r 0 = d−Au 0 ,
                        and an initial direction p 0 = r 0 , the conjugate gradient method is defined by the iterated
                        cycle of steps given by
                                       T
                                 T
                        • λ i = (r p i )/(p Ap i );
                                 i     i
                        • u i+1 = u i + λ i p i ;
                        • r i+1 = r i − λ i Ap i ;
                                   T
                                            T
                        • α i =−(p Ar i+1 )/(p Ap i );
                                   i        i
                        • p i+1 = r i+1 + α i p i .
                           5.6 Numerical solutions of Winslow equations
                        Obtaining a grid in two dimensions typically involves the following steps:
                        1. Set up a rectangular grid in the computational domain and obtain an initial guess for
                           the values of x and y at the grid points. This may be achieved by Transfinite Inter-
                           polation, or Hermite interpolation, or simply by a univariate interpolation between
                           two opposite boundaries.
   137   138   139   140   141   142   143   144   145   146   147