Re: Brent solver change.

Posted by Francis Duffy on
URL: http://quantlib.414.s1.nabble.com/Brent-solver-change-tp13747p13771.html

Hi Luigi,

 

I noticed that a number of the other 1D solvers had the same issue, i.e., the calculation of root_ is fine but the final calculation of the funtion f may have been performed with a value different from the return value of root_. Hence, if you are using the solver in an iterative routine, the underlying results may not be updated with the return value of root_.

 

The attached file, solvers1d.diff, shows the changes that I made to fix this. The files iterativesolvers.xpp give a unit test that exercises the fix for the solvers Brent, Newton, NewtonSafe and Ridder. The Bisection method passes the test with and without the fix. No changes were made to the Secant or False Position method.

Thanks and regards,

Francis.

On Fri, Nov 30, 2012 at 10:48 PM, Francis Duffy <[hidden email]> wrote:

Hi Luigi,

 

No problem, I will do that over the coming days and send it to you.

 

Thanks,

Francis.

On Fri, Nov 30, 2012 at 4:27 PM, Luigi Ballabio <[hidden email]> wrote:

Ok.  If you have time, could you also write a test case to exercise the fix?

Thanks,
    Luigi

On Fri, Nov 30, 2012 at 4:37 PM, Francis Duffy <[hidden email]> wrote:
> Hi Luigi,
>
> Yes, I think so.
>
> Regards,
> Francis.
>
>
> On Fri, Nov 30, 2012 at 2:25 PM, Luigi Ballabio <[hidden email]>
> wrote:
>>
>> Hi Francis,
>>     is your patch still needed after the changes?
>>
>> Luigi
>>
>> On Thu, Nov 29, 2012 at 10:44 PM, Francis Duffy
>> <[hidden email]> wrote:
>> > Hi,
>> >
>> > You are right. This change was made in the HEAD in response to a query
>> > to
>> > the quantlib-dev group on 26/07/2012. The change is shown below.
>> >
>> > Regards,
>> > Francis.
>> >
>> >>svn diff -r {2012-07-26}:HEAD QuantLib\ql\math\solvers1d\brent.hpp
>> > Index: QuantLib/ql/math/solvers1d/brent.hpp
>> > ===================================================================
>> > --- QuantLib/ql/math/solvers1d/brent.hpp        (revision 18300)
>> > +++ QuantLib/ql/math/solvers1d/brent.hpp        (revision 18376)
>> > @@ -46,11 +46,22 @@
>> >              Real min1, min2;
>> >              Real froot, p, q, r, s, xAcc1, xMid;
>> > -            // dummy assignements to avoid compiler warning
>> > -            Real d = 0.0, e = 0.0;
>> > -            root_ = xMax_;
>> > -            froot = fxMax_;
>> > +            // we want to start with root_ (which equals the guess) on
>> > +            // one side of the bracket and both xMin_ and xMax_ on the
>> > +            // other.
>> > +            froot = f(root_);
>> > +            ++evaluationNumber_;
>> > +            if (froot * fxMin_ < 0) {
>> > +                xMax_ = xMin_;
>> > +                fxMax_ = fxMin_;
>> > +            } else {
>> > +                xMin_ = xMax_;
>> > +                fxMin_ = fxMax_;
>> > +            }
>> > +            Real d = root_- xMax_;
>> > +            Real e = d;
>> > +
>> >              while (evaluationNumber_<=maxEvaluations_) {
>> >                  if ((froot > 0.0 && fxMax_ > 0.0) ||
>> >                      (froot < 0.0 && fxMax_ < 0.0)) {
>> >
>> >
>> > On Thu, Nov 29, 2012 at 7:11 PM, cf16r <[hidden email]> wrote:
>> >>
>> >> additional issue is that Brent method doesn't take in to account
>> >> initial
>> >> guess.
>> >> when you call Brent solveImpl method through solver1d.solve with guess
>> >> as
>> >> one of the arguments, root_ is set to this value in solve() but just
>> >> after
>> >> entering solveImpl() it is changed back to root_=xMax_ without any
>> >> checking
>> >> whether guess is correct.
>> >> in this scenario I think there is also need to call f(root_) before
>> >> returning solution.
>> >>
>> >> regards,
>> >> cf16
>> >>
>> >
>> >
>> >
>> > ------------------------------------------------------------------------------
>> > Keep yourself connected to Go Parallel:
>> > VERIFY Test and improve your parallel project with help from experts
>> > and peers. http://goparallel.sourceforge.net
>> > _______________________________________________
>> > QuantLib-users mailing list
>> > [hidden email]
>> > https://lists.sourceforge.net/lists/listinfo/quantlib-users
>> >
>
>

------------------------------------------------------------------------------
Keep yourself connected to Go Parallel:
DESIGN Expert tips on starting your parallel project right.
http://goparallel.sourceforge.net/
_______________________________________________
QuantLib-users mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/quantlib-users