//----------------------------------------------------------------------------- // Package Title ratpak // File fact.c // Author Timothy David Corrie Jr. (timc@microsoft.com) // Copyright (C) 1995-96 Microsoft // Date 01-16-95 // // // Description // // Contains fact(orial) and supporting _gamma functions. // //----------------------------------------------------------------------------- #include #include #include #if defined( DOS ) #include #else #include #endif #include #define ABSRAT(x) (((x)->pp->sign=1),((x)->pq->sign=1)) #define NEGATE(x) ((x)->pp->sign *= -1) //----------------------------------------------------------------------------- // // FUNCTION: factrat, _gamma, gamma // // ARGUMENTS: x PRAT representation of number to take the sine of // // RETURN: factorial of x in PRAT form. // // EXPLANATION: This uses Taylor series // // n // ___ 2j // n \ ] A 1 A // A \ -----[ ---- - ---------------] // / (2j)! n+2j (n+2j+1)(2j+1) // /__] // j=0 // // / oo // | n-1 -x __ // This was derived from | x e dx = | // | | (n) { = (n-1)! for +integers} // / 0 // // GregSte showed the above series to be within precision if A was chosen // big enough. // A n precision // Based on the relation ne = A 10 A was chosen as // // precision // A = ln(Base /n)+1 // A += n*ln(A) This is close enough for precision > base and n < 1.5 // // //----------------------------------------------------------------------------- void _gamma( PRAT *pn ) { PRAT factorial=NULL; PNUMBER count=NULL; PRAT tmp=NULL; PRAT one_pt_five=NULL; PRAT a=NULL; PRAT a2=NULL; PRAT term=NULL; PRAT sum=NULL; PRAT err=NULL; PRAT mpy=NULL; PRAT ratprec = NULL; PRAT ratRadix = NULL; long oldprec; // Set up constants and initial conditions oldprec = maxout; ratprec = longtorat( oldprec ); // Find the best 'A' for convergence to the required precision. a=longtorat( nRadix ); lograt(&a); mulrat(&a,ratprec); // Really is -ln(n)+1, but -ln(n) will be < 1 // if we scale n between 0.5 and 1.5 addrat(&a,rat_two); DUPRAT(tmp,a); lograt(&tmp); mulrat(&tmp,*pn); addrat(&a,tmp); addrat(&a,rat_one); // Calculate the necessary bump in precision and up the precision. // The following code is equivalent to // maxout += ln(exp(a)*pow(a,n+1.5))-ln(nRadix)); DUPRAT(tmp,*pn); one_pt_five=longtorat( 3L ); divrat( &one_pt_five, rat_two ); addrat( &tmp, one_pt_five ); DUPRAT(term,a); powrat( &term, tmp ); DUPRAT( tmp, a ); exprat( &tmp ); mulrat( &term, tmp ); lograt( &term ); ratRadix = longtorat( nRadix ); DUPRAT(tmp,ratRadix); lograt( &tmp ); subrat( &term, tmp ); maxout += rattolong( term ); // Set up initial terms for series, refer to series in above comment block. DUPRAT(factorial,rat_one); // Start factorial out with one count = longtonum( 0L, BASEX ); DUPRAT(mpy,a); powrat(&mpy,*pn); // a2=a^2 DUPRAT(a2,a); mulrat(&a2,a); // sum=(1/n)-(a/(n+1)) DUPRAT(sum,rat_one); divrat(&sum,*pn); DUPRAT(tmp,*pn); addrat(&tmp,rat_one); DUPRAT(term,a); divrat(&term,tmp); subrat(&sum,term); DUPRAT(err,ratRadix); NEGATE(ratprec); powrat(&err,ratprec); divrat(&err,ratRadix); // Just get something not tiny in term DUPRAT(term, rat_two ); // Loop until precision is reached, or asked to halt. while ( !zerrat( term ) && rat_gt( term, err) && !fhalt ) { addrat(pn,rat_two); // WARNING: mixing numbers and rationals here. // for speed and efficiency. INC(count); mulnumx(&(factorial->pp),count); INC(count) mulnumx(&(factorial->pp),count); divrat(&factorial,a2); DUPRAT(tmp,*pn); addrat( &tmp, rat_one ); destroyrat(term); createrat(term); DUPNUM(term->pp,count); DUPNUM(term->pq,num_one); addrat( &term, rat_one ); mulrat( &term, tmp ); DUPRAT(tmp,a); divrat( &tmp, term ); DUPRAT(term,rat_one); divrat( &term, *pn); subrat( &term, tmp); divrat (&term, factorial); addrat( &sum, term); ABSRAT(term); } // Multiply by factor. mulrat( &sum, mpy ); // And cleanup maxout = oldprec; destroyrat(ratprec); destroyrat(err); destroyrat(term); destroyrat(a); destroyrat(a2); destroyrat(tmp); destroyrat(one_pt_five); destroynum(count); destroyrat(factorial); destroyrat(*pn); DUPRAT(*pn,sum); destroyrat(sum); } void factrat( PRAT *px ) { PRAT fact = NULL; PRAT frac = NULL; PRAT neg_rat_one = NULL; DUPRAT(fact,rat_one); DUPRAT(neg_rat_one,rat_one); neg_rat_one->pp->sign *= -1; DUPRAT( frac, *px ); fracrat( &frac ); // Check for negative integers and throw an error. if ( ( zerrat(frac) || ( LOGRATRADIX(frac) <= -maxout ) ) && ( (*px)->pp->sign * (*px)->pq->sign == -1 ) ) { throw CALC_E_DOMAIN; } while ( rat_gt( *px, rat_zero ) && !fhalt && ( LOGRATRADIX(*px) > -maxout ) ) { mulrat( &fact, *px ); subrat( px, rat_one ); } // Added to make numbers 'close enough' to integers use integer factorial. if ( LOGRATRADIX(*px) <= -maxout ) { DUPRAT((*px),rat_zero); intrat(&fact); } while ( rat_lt( *px, neg_rat_one ) && !fhalt ) { addrat( px, rat_one ); divrat( &fact, *px ); } if ( rat_neq( *px, rat_zero ) ) { addrat( px, rat_one ); _gamma( px ); mulrat( px, fact ); } else { DUPRAT(*px,fact); } }