Minuit2 gives -nan results

[Zhiyhep]

Dear ROOT experts:

I am trying to minimize an object function like this:

f(lambda) = Σ_i=1~n(2ln(Σ_j=1~mL_ij*lambda[j])+(fX2[i]+fY2[i])/(Σ_j=1~mL_ij*lambda[j])).

Where lambda is a 1000 dimension array and L, fX2, fY2 are parameters stored in TTree.

I defined a class named Likelihood_FCN with reloaded operator() as this:

double Likelihood_FCN::operator()(const double* lambda)
{
double result = 0;
int nentry = t_ml->GetEntries();
for(int i =0; i < nentry; i++)
{
 t_ml->GetEntry(i);
 double temp = 0;
 for(int j=0; j < fNVoxel; j++)
 {
   int index = fIndex[j];
   temp+=fLength[j]*lambda[index];
   std::cout << i<<": Length, lambda: " << fLength[j] << "," << lambda[index] << std::endl;
 }
 //std::cout<<"temp: "<<temp <<std::endl;
 result+=2*TMath::Log(temp)+(fScatt_Ang_X2+fScatt_Ang_Y2)/temp;
}
std::cout << result << std::endl;
return result;
}

And the minimizer is defined as below:

 ROOT::Math::Minimizer* minim = ROOT::Math::Factory::CreateMinimizer("Minuit2","Migrad");
 minim->SetMaxFunctionCalls(10000);
 minim->SetMaxIterations(10000);
 minim->SetTolerance(1);
 minim->SetPrintLevel(3);
 Likelihood_FCN lf = Likelihood_FCN("PoCA_Result.root","ML_Info","PoCA");
 int nvariable = lf.GetNVariable();
 ROOT::Math::Functor f(lf, nvariable);
 double *init_value = lf.Init();
 lf(init_value);
 minim->SetFunction(f);
 for(int i=0;i<nvariable;i++)
 {
   const string name = "x" + std::to_string(i);
   minim->SetLowerLimitedVariable(i,name,init_value[i],0.1,2.3);
 }
 minim->Minimize();

Segmentation fault occurs when I set the minimization strategy other than “Combined”.
The minimization result is an array of -nan value.

Besides, the evaluations of the FCN during minimization are also -nan, which should not
occur unless all members of lambda is 0.

Could you give some advice on the -nan result ?

The related files are here:
MLEM.tar.gz (123.4 KB)

Best,
Zhi Yu

---

ROOT Version (6.10/02):
Platform, compiler (e.g. Ubuntu 14.04, gcc4.8):

---

[Wile_E_Coyote]

Try something like (note: some Valgrind functionality requires that the source code, that you want to analyse, is compiled with debug symbols):

valgrind --tool=memcheck --leak-check=full [--show-reachable=yes] [--track-origins=yes] [--num-callers=50] [--vgdb=full] --suppressions=`root-config --etcdir`/valgrind-root.supp `root-config --bindir`/root.exe -l -q 'MLEM.cc++g'
valgrind --tool=exp-sgcheck [--num-callers=50] [--vgdb=full] --suppressions=`root-config --etcdir`/valgrind-root.supp `root-config --bindir`/root.exe -l -q 'MLEM.cc++g'

and especially carefully study messages that appear in the beginning of the output.
(Note: the --show-reachable=yes option will give you too many warnings, I believe.)

BTW. The above Valgrind commands include some [something_optional] options (if you want to use any of these options, REMOVE the “square brackets”).

See also:

AddressSanitizer memory error detector

The default valgrind’s “memcheck” memory error detector will not find various {stack,global}-buffer overflow bugs. Valgrind does provide a dedicated complementary “exp-sgcheck” experimental stack and global array overrun detector but, for the time being, it does not work, as soon as ROOT related libraries are used and it seems that it is possible to use it. On the other hand, gcc 4.8 and newer compilers provide their own AddressSanitizer (a.k.a. ASan) fast memory error detector (note: you may n…

[Zhiyhep]

Hi,

Thanks for your suggestion. I have no experience on valgrind, so i spent some time familiarizing the tool.

When I use the memcheck tool (after include some header files), I got an error message saying that : ‘to_string’ is not a member of ‘std’.
The full output message is as below

$: sudo valgrind --tool=memcheck --leak-check=full --suppressions=`root-config --etcdir`/valgrind-root.supp `root-config --bindir`/root.exe -l -a 'MLEM.cc++g'
==9462== Memcheck, a memory error detector
==9462== Copyright (C) 2002-2017, and GNU GPL'd, by Julian Seward et al.
==9462== Using Valgrind-3.13.0 and LibVEX; rerun with -h for copyright info
==9462== Command: /my_cache/ROOT6.10.02/bin/root.exe -l -a MLEM.cc++g
==9462== 
root [0] 
Processing MLEM.cc++g...
Info in <TUnixSystem::ACLiC>: creating shared library /home/cihenp/g4proj/MuImag/build/analysis/./MLEM_cc.so
In file included from /home/cihenp/g4proj/MuImag/build/analysis/MLEM_cc_ACLiC_dict.h:34:0,
                 from /home/cihenp/g4proj/MuImag/build/analysis/MLEM_cc_ACLiC_dict.cxx:17:
/home/cihenp/g4proj/MuImag/build/analysis/./MLEM.cc: In function ‘int MLEM()’:
/home/cihenp/g4proj/MuImag/build/analysis/./MLEM.cc:39:31: error: ‘to_string’ is not a member of ‘std’
     const string name = "x" + std::to_string(i);
                               ^
g++-4.8: error: /home/cihenp/g4proj/MuImag/build/analysis/MLEM_cc_ACLiC_dict.o: no such file or directory
Error in <ACLiC>: Compilation failed!
Error: Function MLEM() is not defined in current scope  :0:
*** Interpreter error recovered ***

How could this error occur ?

Best,
Zhi Yu

[Zhiyhep]

And thanks for pointing out a potential memory error.

I find that I defined a dynamic array in the Likelihood_FCN::Init() method.
I redefine the array as a private member of the class and the minimizer seems to be giving reasonable results now, at least the -nan value has disappeared.

Best,

Zhi Yu