ROMBERG is now MACSYM;ROMBRG FASL and is autoloading. Just call
ROMBERG and it will be there. - JPG
GJC@MIT-MC 01/12/81 13:50:27
To: (FILE [SHARE;ROMBRG USAGE]) at MIT-MC
The old conventions of what are efficient ways to call ROMBERG,
or what are ways which work in at all, are now repaired.
Given F(N):=ROMBERG(expression,VAR,0,N)$
If F is translated/compiled then the expression
will also be translated/compiled! Also, the
old "efficient" way, which could ONLY BE USED
ON TRANSLATED FUNCTIONS, (what a pain in the neck
right?), now works on all functions.
This change is made possible due to the increased
power of the macsyma->lisp translator, and some
reworking of the numerical code.
[Anonymous functions, i.e. LAMBDA expressions, are
now translated into lisp functions.]
The entry point to ROMBERG is now ROMBERG_SUBR,
which only takes 3 arguments, the first being a function.
[This is never seen by the user.]
ROMBERG(F,A,B) => ROMBERG_SUBR(F,A,B)
ROMBERG(X^2,X,A,B) => ROMBERG_SUBR(LAMBDA([X],X^2),A,B)
The arrow "=>" signifies a transformation which takes
place through the same mechanisms as the *new* macsyma
"macro" feature. To see how this feature can be
useful in numerical applications see SHARE;SIMPSN MACRO.
Proper use of MODE_DECLARE is still important of course.
p.s. These same extensions apply to the syntax of the
INTERPOLATE command. However, the restrictions on the
use of the PLOT2 command unfortunately still hold.
This takes lots of work on PLOT2 to fix.
GJC@MIT-MC 03/20/80 19:09:30
To: (FILE [SHARE;ROMBRG USAGE]) at MIT-MC
The source for ROMBERG is now MAXSRC;ROMBRG > and NUMER;ROMBRG MMACRO.
The three argument version of ROMBERG now works for macsyma functions,
not just translated functions. Problems with arguments not getting
properly floated, e.g. ROMBERG(X,X,1/10,2/10) have been fixed, problems
with the arguments to ROMBERG not getting evaluated in compiled code
have been fixed.
CFFK@MIT-MC 12 AUG 1979 1743-EDT
To: (FILE [SHARE;ROMBRG USAGE]) at MIT-MC
At SK's suggestion, ROMBERG (BROMBERG) will now exit if an absolute
error condition is satisfied. This is governed by the new variable
ROMBERGABS (BROMBERABS) - default value 0.0 (0.0B0).
Assuming that successive estimates produced by ROMBERG are Y[0], Y[1],
Y[2] etc., then ROMBERG will return after N iterations if
(roughly speaking)
(ABS(Y[N]-Y[N-1]) <= ROMBERGABS OR
ABS(Y[N]-Y[N-1])/(IF Y[N]=0.0 THEN 1.0 ELSE Y[N]) <= ROMBERGTOL)
is TRUE. (The condition on the number of iterations given by ROMBERGMIN
must also be satisfied.) (The old exit condition, was on the relative
error if ABS(Y[N]) > 1.0, and on the absolute error otherwise. Pretty
random!)
Thus if ROMBERGABS is 0.0 (the default) you just get the relative error
test. The usefulness of the additional variable comes when
you want to perform an integral, where the dominant contribution
comes from a small region. Then you can do the integral over
the small dominant region first, using the relative accuracy
check, followed by the integral over the rest of the region
using the absolute accuracy check.
Example: Suppose you want to compute
Integral(exp(-x),x,0,50)
(numerically) with a relative accuracy of 1 part in 10000000.
/* Define the function. N is a counter, so we can see how many
function evaluations were needed. */
F(X):=(MODEDECLARE(N,INTEGER,X,FLOAT),N:N+1,EXP(-X))$
TRANSLATE(F)$
/* First of all try doing the whole integral at once */
BLOCK([ROMBERGTOL:1.E-6,ROMBERABS:0.],N:0,ROMBERG(F,0,50)); ==> 1.00000003
N; ==> 257 /* Number of function evaluations*/
/* Now do the integral intelligently, by first doing
Integral(exp(-x),x,0,10) and then setting ROMBERGABS to 1.E-6*(this
partial integral). */
BLOCK([ROMBERGTOL:1.E-6,ROMBERGABS:0.,SUM:0.],
N:0,SUM:ROMBERG(F,0,10),ROMBERGABS:SUM*ROMBERGTOL,ROMBERGTOL:0.,
SUM+ROMBERG(F,10,50)); ==> 1.00000001 /* Same as before */
N; ==> 130
So if F(X) were a function that took a long time to compute, the
second method would be about 2 times quicker.
CFFK@MIT-MC 05/04/78 16:52:43
To: (FILE [SHARE;ROMBRG USAGE]) at MIT-MC
There is a new option ROMBERGMIN whose default value is 0, that
govern the minimum number of function evaluations that ROMBERG will
make. ROMBERG will evaluate its first arg. at least 2^(ROMBERGMIN+2)+1
times. This is useful for integrating oscillatory functions, when
the normal converge test might sometimes wrongly pass.
There are 2 ways of using this function:
1) An inefficient way that looks like the definite integral version of
INTEGRATE:
ROMBERG(,,,
);
Examples:
ROMBERG(SIN(Y),Y,1,%PI);
TIME= 39 MSEC. 1.5403023
F(X):=1/(X^5+X+1);
ROMBERG(F(X),X,1.5,0);
TIME= 162 MSEC. - 0.75293843
2) An efficient way that is more like RJF's ROMBERG function:
ROMBERG(,,);
Example:
F(X):=(MODEDECLARE([FUNCTION(F),X],FLOAT),1/(X^5+X+1));
TRANSLATE(F);
ROMBERG(F,1.5,0);
TIME= 13 MSEC. - 0.75293843
The first argument must be a TRANSLATEd or compiled function, which
returns a floating point number. (If it is
compiled it must be declared to return a FLONUM.) If the first argument
is not already TRANSLATEd, ROMBERG will not attempt to TRANSLATE it but
will give an error.
The accuracy of the integration is governed by the global variables
ROMBERGTOL (default value 1.E-4) and ROMBERGIT (default value 11).
ROMBERG will return a result if the relative difference in successive
approximations is less than ROMBERGTOL. It will try halving the
stepsize ROMBERGIT times before it gives up.
ROMBERG may be called recursively and thus can do double and triple
integrals.
Example:
INTEGRATE(INTEGRATE(X*Y/(X+Y),Y,0,X/2),X,1,3);
13/3 (2 LOG(2/3) + 1)
%,NUMER;
0.81930233
DEFINE_VARIABLE(X,0.0,FLOAT,"Global variable in function F")$
F(Y):=(MODE_DECLARE(Y,FLOAT), X*Y/(X+Y) )$
G(X):=ROMBERG('F,0,X/2)$
ROMBERG(G,1,3);
0.8193023
The advantage with this way is that the function F can be used for other
purposes, like plotting. The disadvantage is that you have to think up
a name for both the function F and its free variable X.
Or, without the global:
G1(X):=(MODE_DECLARE(X,FLOAT), ROMBERG(X*Y/(X+Y),Y,0,X/2))$
ROMBERG(G1,1,3);
0.8193023
The advantage here is shortness.
Q(A,B):=ROMBERG(ROMBERG(X*Y/(X+Y),Y,0,X/2),X,A,B)$
Q(1,3);
0.8193023
It is even shorter this way, and the variables do not need to be declared
because they are in the context of ROMBERG.
Use of ROMBERG for multiple integrals can have great disadvantages,
though. The amount of extra calculation needed because of the geometric
information thrown away by expressing multiple integrals this way can
be incredible. The user should understand and use the ROMBERGTOL and
ROMBERGIT switches.