I'd like to thank everyone who sent me suggestions.  These included
 using LET to force use of the persent structure for the FORCE
 computation, which pointed out that the program was starting with
 a much higher gradient than the previous run completed with.
 Which in turn made suggestions I improve the optimization most
 pertinent.  I was also advised to try TS;  I neglected to mention
 in my original posting that TS with this system seems always to
 lead to a structure with no negative eigenvalues.
 I was also advised to try RECALC=3.  I'm trying these suggestions,
 and seem to be making progress.
 Irene Newhouse