Nans in quantile accumulators

[wf34]

This code is a slightly modified example from the Boost.Accumualators documentation, https://scicomp.ethz.ch/public/manual/Boost/1.55.0/accumulators.pdf , p.50.

It has two constants, n, c : sample size and cache size.
The purpose of this library, if I'm not mistaken, that the stats are computed for c << n.
However, this code yields nans and fails the test.
When the cache size is set equal to a sample size c = n; the test passes.

#include <iostream>

#include <boost/accumulators/accumulators.hpp> 
#include <boost/accumulators/statistics/stats.hpp>
#include <boost/accumulators/statistics/tail_quantile.hpp>
#include <boost/random/lagged_fibonacci.hpp>
#include <cassert>

using namespace boost::accumulators;

int main() {
  double epsilon = 1.e-2;
  std::size_t n = 100000; // number of MC steps
  std::size_t c = 10000; // cache size
  typedef accumulator_set<double, stats<tag::tail_quantile<right> > > accumulator_t_right;
  typedef accumulator_set<double, stats<tag::tail_quantile<left> > > accumulator_t_left;
  accumulator_t_right acc0( tag::tail<right>::cache_size = c );
  accumulator_t_left acc2( tag::tail<left>::cache_size = c );

  boost::lagged_fibonacci607 rng;

  for (std::size_t i = 0; i < n; ++i) {
    double sample1 = rng();
    acc0(sample1);
    acc2(sample1);
  }

  std::cout << quantile(acc0, quantile_probability = 0.95) << std::endl;
  std::cout << quantile(acc0, quantile_probability = 0.75) << std::endl;
  std::cout << quantile(acc2, quantile_probability = 0.25) << std::endl;
  std::cout << quantile(acc2, quantile_probability = 0.05) << std::endl;

  assert(std::fabs(quantile(acc0, quantile_probability = 0.95 )- 0.95) < epsilon);
  assert(std::fabs(quantile(acc0, quantile_probability = 0.975)- 0.975) <epsilon );
  assert(std::fabs(quantile(acc0, quantile_probability = 0.75)- 0.75) <epsilon );
  assert(std::fabs(quantile(acc0, quantile_probability = 0.99 )- 0.99) <epsilon );
  assert(std::fabs(quantile(acc0, quantile_probability = 0.999)- 0.999) <epsilon );
  assert(std::fabs(quantile(acc0, quantile_probability = 0.25)- 0.25) <epsilon );
  assert(std::fabs(quantile(acc2, quantile_probability = 0.05 )- 0.05) <epsilon );
  assert(std::fabs(quantile(acc2, quantile_probability = 0.025)- 0.025) <epsilon );
  assert(std::fabs(quantile(acc2, quantile_probability = 0.01 )- 0.01) <epsilon );
  assert(std::fabs(quantile(acc2, quantile_probability = 0.001)- 0.001)<epsilon );

  return 0;
}

The output:

0.949523
nan
nan
0.0516118
a.out: ./1.cpp:35: int main(): Assertion `std::fabs(quantile(acc0, quantile_probability = 0.75)- 0.75) <epsilon' failed.

[adimajo]

You cache 10000 samples among 100000, so if my understanding is correct, you cannot estimate past the 90th quantile (resp. before the 10th quantile) by design, see tail_quantile.hpp.