void llvm_model::set_prediction(const ObsId & obs_id, boost::ptr_vector<theta::Function> & coeffs_, boost::ptr_vector<HistogramFunction> & histos_){
observables.insert(obs_id);
const size_t n = coeffs_.size();
if(n!=coeffs_.size()) throw invalid_argument("Model::setPrediction: number of histograms and coefficients do not match");
if(histos[obs_id].size()>0 || coeffs[obs_id].size()>0)
throw invalid_argument("Model::setPrediction: prediction already set for this observable");
coeffs[obs_id].transfer(coeffs[obs_id].end(), coeffs_.begin(), coeffs_.end(), coeffs_);
histos[obs_id].transfer(histos[obs_id].end(), histos_.begin(), histos_.end(), histos_);
for(boost::ptr_vector<theta::Function>::const_iterator it=coeffs[obs_id].begin(); it!=coeffs[obs_id].end(); ++it){
ParIds pids = (*it).get_parameters();
parameters.insert(pids.begin(), pids.end());
}
size_t nbins = 0;
double xmin = NAN, xmax = NAN;
bool first = true;
for(boost::ptr_vector<HistogramFunction>::const_iterator it=histos[obs_id].begin(); it!=histos[obs_id].end(); ++it){
if(first){
it->get_histogram_dimensions(nbins, xmin, xmax);
first = false;
}
else{
size_t nbins_tmp = 0;
double xmin_tmp = NAN, xmax_tmp = NAN;
it->get_histogram_dimensions(nbins_tmp, xmin_tmp, xmax_tmp);
if(nbins!=nbins_tmp || xmin!=xmin_tmp || xmax!=xmax_tmp){
throw invalid_argument("llvm_model::set_prediction: histogram dimensions mismatch");
}
}
const ParIds & pids = (*it).get_parameters();
parameters.insert(pids.begin(), pids.end());
}
}
FunctionInfo::FunctionInfo(const ParValues & start_, const ParValues & step_, const Ranges & ranges_, const ParValues & fixed_parameters): start(start_), step(step_), ranges(ranges_){
ParIds pids = start.get_parameters();
for(ParIds::const_iterator pit=pids.begin(), it_end = pids.end(); pit!=it_end; ++pit){
if(!step.contains(*pit)) throw invalid_argument("FunctionInfo: step does not contain all parameters from start");
// 1. check whether parameter is fixed by step and range:
const pair<double, double> & r = ranges.get(*pit);
if(r.first==r.second){
if(step.get_unchecked(*pit)>0.0){
throw invalid_argument("FunctionInfo: inconsistent range/step given: range empty but step > 0");
}
fixed_parids.insert(*pit);
}
else{
if(step.get_unchecked(*pit)<=0.0){
throw invalid_argument("FunctionInfo: step <= 0.0 for non-empty range given");
}
}
// 2. check whether parameter is fixed by fixed_parameters. Note that
// the value given in fixed_parameter overrides the value according to start.
if(fixed_parameters.contains(*pit)){
double val = fixed_parameters.get(*pit);
start.set(*pit, val);
ranges.set(*pit, make_pair(val, val));
step.set(*pit, 0.0);
fixed_parids.insert(*pit);
}
}
}
void theta::fill_mode_support(theta::ParValues & mode,
std::map<theta::ParId, std::pair<double, double> > & support, const theta::Distribution & d){
ParIds pids = d.get_parameters();
d.mode(mode);
theta_assert(mode.contains_all(pids));
for(ParIds::const_iterator p_it=pids.begin(); p_it!=pids.end(); ++p_it){
support[*p_it] = d.support(*p_it);
}
}
boost::shared_ptr<FunctionInfo> Minimizer::create_nll_function_info(const Model & model, const boost::shared_ptr<Distribution> & override_parameter_distribution, const ParValues & fixed_parameters){
const Distribution & dist = override_parameter_distribution.get()? *override_parameter_distribution: model.get_parameter_distribution();
ParValues start;
dist.mode(start);
Ranges ranges(dist);
ParIds pids = fixed_parameters.get_parameters();
for(ParIds::const_iterator pit = pids.begin(); pit!=pids.end(); ++pit){
double val = fixed_parameters.get(*pit);
start.set(*pit, val);
ranges.set(*pit, make_pair(val, val));
}
ParValues step = asimov_likelihood_widths(model, override_parameter_distribution);
return boost::shared_ptr<FunctionInfo>(new DefFunctionInfo(start, step, ranges, fixed_parameters));
}
MinimizationResult Minimizer::minimize2(const Function & f, const FunctionInfo & info, const ParValues & fixed_parameters){
dynamic_cast<const DefFunctionInfo&>(info); // throws bad_cast
ParIds pids = fixed_parameters.get_parameters();
if(pids.size()==0){
return minimize(f, info.get_start(), info.get_step(), info.get_ranges());
}
else{
ParValues start(info.get_start());
ParValues step(info.get_step());
Ranges ranges(info.get_ranges());
const ParIds & info_fixed = info.get_fixed_parameters();
for(ParIds::const_iterator pit = pids.begin(); pit!=pids.end(); ++pit){
if(!info_fixed.contains(*pit)){
throw invalid_argument("fixed parameter in minimize2 which is not fixed in info. This is not allowed.");
}
double val = fixed_parameters.get(*pit);
start.set(*pit, val);
step.set(*pit, 0.0);
ranges.set(*pit, make_pair(val, val));
}
return minimize(f, start, step, ranges);
}
}
theta::ParValues asimov_likelihood_widths(const theta::Model & model, const boost::shared_ptr<Distribution> & override_parameter_distribution){
const Distribution & dist = override_parameter_distribution.get()? *override_parameter_distribution: model.get_parameter_distribution();
ParIds parameters = model.getParameters();
ParValues mode;
dist.mode(mode);
Data asimov_data;
model.get_prediction(asimov_data, mode);
std::auto_ptr<NLLikelihood> nll = model.getNLLikelihood(asimov_data);
//0 value has same semantics for NLLikelihood:
nll->set_override_distribution(override_parameter_distribution);
double nll_at_min = (*nll)(mode);
ParValues result;
int k=0;
for(ParIds::const_iterator it=parameters.begin(); it!=parameters.end(); ++it, ++k){
ParId pid = *it;
const double pid_mode = mode.get(pid);
std::pair<double, double> support = dist.support(pid);
assert(support.first <= pid_mode && pid_mode <= support.second);
if(support.first == support.second){
result.set(pid, 0.0);
continue;
}
nll_mode_pid f(mode, pid, *nll, nll_at_min + 0.5);
//if one end is finite, try to use it. Save whether the interval end is considered
// "fl0", i.e. the interval end itself is finite but the function value there is invalid (< 0).
bool low_is_fl0 = false, high_is_fl0 = false;
if(std::isfinite(support.second)){
double f2 = f(support.second);
if(f2==0.0){
result.set(pid, fabs(pid_mode - support.second));
continue;
}
if(!std::isfinite(f2) || f2 < 0){
low_is_fl0 = true;
}
else{
result.set(pid, fabs(pid_mode - secant(pid_mode, support.second, 0.0, -0.5, f2, 0.05, f)));
continue;
}
}
if(std::isfinite(support.first)){
double f2 = f(support.first);
if(f2==0.0){
result.set(pid, fabs(pid_mode - support.first));
continue;
}
if(!std::isfinite(f2) || f2 < 0){
high_is_fl0 = true;
}
else{
result.set(pid, fabs(pid_mode - secant(support.first, pid_mode, 0.0, f2, -0.5, 0.05, f)));
continue;
}
}
//the support was either infinite or the values at the borders were not sufficiently high.
// Treat second case first:
if(low_is_fl0 && high_is_fl0){
result.set(pid, support.second - support.first);
continue;
}
//Now, one of the interval ends has to be infinite, otherwise we would not be here.
//Scan in that direction:
assert(std::isinf(support.first) || std::isinf(support.second));
bool found = false;
for(double sign = -1.0; sign <= 1.001; sign+=2.0){
if(!std::isinf(support.first) && sign < 0) continue;
if(!std::isinf(support.second) && sign > 0) continue;
// as step size, try the parameter value, if it is not zero:
double step = fabs(pid_mode);
if(step==0) step = 1.0;
for(int i=0; i<1000; ++i){
double fval = f(pid_mode + sign * step);
if(isinf(fval)){
step /= 1.5;
continue;
}
step *= 2.0;
if(fval > 0){
double xlow, xhigh, flow, fhigh;
xlow = pid_mode; flow = -0.5;
xhigh = pid_mode + sign * step; fhigh = fval;
if(sign < 0){
std::swap(xlow, xhigh);
std::swap(flow, fhigh);
}
assert(xlow <= xhigh);
result.set(pid, fabs(pid_mode - secant(xlow, xhigh, 0.0, flow, fhigh, 0.05, f)));
found = true;
break;
}
}
if(found) break;
}
if(found) continue;
stringstream ss;
ss << "asimov_likelihood_widths: could not find width for parameter " << pid;
throw Exception(ss.str());
}
return result;
}
MinimizationResult root_minuit::minimize(const theta::Function & f, const theta::ParValues & start,
const theta::ParValues & steps, const std::map<theta::ParId, std::pair<double, double> > & ranges){
//I would like to re-use min. However, it horribly fails after very few uses with
// unsigned int ROOT::Minuit2::MnUserTransformation::IntOfExt(unsigned int) const: Assertion `!fParameters[ext].IsFixed()' failed.
// when calling SetFixedVariable(...).
//Using a "new" one every time seems very wastefull, but it seems to work ...
std::auto_ptr<ROOT::Minuit2::Minuit2Minimizer> min(new ROOT::Minuit2::Minuit2Minimizer(type));
//min->SetPrintLevel(0);
if(max_function_calls > 0) min->SetMaxFunctionCalls(max_function_calls);
if(max_iterations > 0) min->SetMaxIterations(max_iterations);
MinimizationResult result;
//1. setup parameters, limits and initial step sizes
ParIds parameters = f.get_parameters();
int ivar=0;
for(ParIds::const_iterator it=parameters.begin(); it!=parameters.end(); ++it, ++ivar){
std::map<theta::ParId, std::pair<double, double> >::const_iterator r_it = ranges.find(*it);
if(r_it==ranges.end()) throw invalid_argument("root_minuit::minimize: range not set for a parameter");
pair<double, double> range = r_it->second;
double def = start.get(*it);
double step = steps.get(*it);
stringstream ss;
ss << "par" << ivar;
string name = ss.str();
//use not the ranges directly, but a somewhat more narrow range (one permille of the respective border)
// in order to avoid that the numerical evaluation of the numerical derivative at the boundaries pass these
// boundaries ...
if(step == 0.0){
min->SetFixedVariable(ivar, name, def);
}
else if(isinf(range.first)){
if(isinf(range.second)){
min->SetVariable(ivar, name, def, step);
}
else{
min->SetUpperLimitedVariable(ivar, name, def, step, range.second - fabs(range.second) * 0.001);
}
}
else{
if(isinf(range.second)){
min->SetLowerLimitedVariable(ivar, name, def, step, range.first + fabs(range.first) * 0.001);
}
else{ // both ends are finite
if(range.first==range.second){
min->SetFixedVariable(ivar, name, range.first);
}
else{
min->SetLimitedVariable(ivar, name, def, step, range.first + fabs(range.first) * 0.001, range.second - fabs(range.second) * 0.001);
}
}
}
}
//2. setup the function
RootMinuitFunctionAdapter minuit_f(f);
min->SetFunction(minuit_f);
//3. setup tolerance
min->SetTolerance(tolerance_factor * min->Tolerance());
//3.a. error definition. Unfortunately, SetErrorDef in ROOT is not documented, so I had to guess.
// 0.5 seems to work somehow.
min->SetErrorDef(0.5);
//4. minimize. In case of failure, try harder. Discard all output generated in min->Minimize.
bool success;
{
theta::utils::discard_output d_o(true);
success = min->Minimize();
if(!success){
for(int i=1; i<=n_retries; i++){
success = min->Minimize();
if(success) break;
}
}
} // d_o is destroyed, output resumed.
//5. do error handling
if(not success){
int status = min->Status();
int status_1 = status % 10;
//int status_2 = status / 10;
stringstream s;
s << "MINUIT returned status " << status;
switch(status_1){
case 1: s << " (Covariance was made pos defined)"; break;
case 2: s << " (Hesse is invalid)"; break;
case 3: s << " (Edm is above max)"; break;
case 4: s << " (Reached call limit)"; break;
case 5: s << " (Some other failure)"; break;
default:
s << " [unexpected status code]";
}
throw MinimizationException(s.str());
}
//6. convert result
result.fval = min->MinValue();
ivar = 0;
const double * x = min->X();
const double * errors = 0;
bool have_errors = min->ProvidesError();
if(have_errors) errors = min->Errors();
for(ParIds::const_iterator it=parameters.begin(); it!=parameters.end(); ++it, ++ivar){
result.values.set(*it, x[ivar]);
if(have_errors){
result.errors_plus.set(*it, errors[ivar]);
result.errors_minus.set(*it, errors[ivar]);
}
else{
result.errors_plus.set(*it, -1);
result.errors_minus.set(*it, -1);
}
}
result.covariance.reset(parameters.size(), parameters.size());
//I would use min->CovMatrixStatus here to check the validity of the covariance matrix,
// if only it was documented ...
if(min->ProvidesError()){
for(size_t i=0; i<parameters.size(); ++i){
for(size_t j=0; j<parameters.size(); ++j){
result.covariance(i,j) = min->CovMatrix(i,j);
}
}
}
else{
for(size_t i=0; i<parameters.size(); ++i){
result.covariance(i,i) = -1;
}
}
return result;
}
