home *** CD-ROM | disk | FTP | other *** search
- The files MINV.RAT and MINV.FOR contain a subroutine
- which inverts a matrix and calculates the determinant.
- These files contain a test program which demonstrates its
- use. The method used is Gauss-Jordan elimination with
- partial pivoting. On large systems this subroutine has
- demonstrated better accuracy and several times the speed of
- similar routines marketed by large software houses. It does
- not have any protection against the case where the equations
- are of wildly different orders of magnitude, but it is just
- as accurate as more complicated pivot algorithms in normal
- problem solving and on random matrices.
- The test case shown is a good test of accuracy but
- not of the pivoting algorithm. If you ran it in single
- precision on an IBM or TI system you probably wouldn't get
- even 4 decimal places correct.
- Subfunctions are employed which are of types which
- are often coded in assembler even on large systems to permit
- rapid solution of problems which are much larger than this.
- There are usually machine-dependent methods unknown to
- FORTRAN compilers which will make them much faster. The
- names employed are essentially those used by Floating Point
- Systems for Array Processor FORTRAN. Incidentally, there are
- actual cases where an 8080 micro has been used as the front
- end for an FP AP120B, thus making a system which is faster
- (for floating point vector calculations) than any IBM 370!
-