Le 03/06/2013 12:19, Gilles a écrit : > Hello. Hi Gilles,
> > On Mon, 03 Jun 2013 09:04:40 -0000, l...@apache.org wrote: >> Author: luc >> Date: Mon Jun 3 09:04:40 2013 >> New Revision: 1488914 >> >> URL: http://svn.apache.org/r1488914 >> Log: >> Added midpoint integration method. >> >> Patch contributed by Oleksandr Kornieiev. >> >> JIRA: MATH-967 > >> [...] > >> + private double stage(final int n) >> + throws TooManyEvaluationsException { >> + >> + final double max = getMax(); >> + final double min = getMin(); >> + >> + if (n == 0) { >> + final double midPoint = 0.5 * (max - min); > ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ > Is this correct? No, it was wrong! The fact it the problem was not detected by the tests cases is because the contribution of this single wrong initial point is completely negectible after just a few iteration (each iteration adds 2^n new points). Solving the error was trivial, but finding a test case that triggers it was difficult! It involved setting the integration parameters to very low accuracy, so very few iterations are performed. Thanks for spotting this. Luc > >> + s = (max - min) * computeObjectiveValue(midPoint); >> + return s; >> + } else { >> + final long np = 1L << (n - 1); // number of >> new points in this stage >> + double sum = 0; >> + // spacing between adjacent new points >> + final double spacing = (max - min) / np; >> + double x = min + 0.5 * spacing; // the first new point >> + for (long i = 0; i < np; i++) { >> + sum += computeObjectiveValue(x); >> + x += spacing; >> + } >> + // add the new sum to previously calculated result >> + s = 0.5 * (s + sum * spacing); >> + return s; >> + } >> + } >> + >> + /** {@inheritDoc} */ >> + protected double doIntegrate() >> + throws MathIllegalArgumentException, >> TooManyEvaluationsException, MaxCountExceededException { >> + >> + double oldt = stage(0); >> + iterations.incrementCount(); >> + while (true) { >> + final int i = iterations.getCount(); >> + final double t = stage(i); >> + if (i >= getMinimalIterationCount()) { >> + final double delta = FastMath.abs(t - oldt); >> + final double rLimit = >> + getRelativeAccuracy() * (FastMath.abs(oldt) >> + FastMath.abs(t)) * 0.5; >> + if ((delta <= rLimit) || (delta <= >> getAbsoluteAccuracy())) { >> + return t; >> + } >> + } >> + oldt = t; >> + iterations.incrementCount(); >> + } >> + >> + } >> + >> +} >> >> [...] > > > --------------------------------------------------------------------- > To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org > For additional commands, e-mail: dev-h...@commons.apache.org > > --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org For additional commands, e-mail: dev-h...@commons.apache.org
