OK, so I guess kADAPTIVE fails on a discontinuity at r = 1, while accuracy of MC integrators is defined by ncall parameter, not by tolerances.
@moneta Maybe it would be better to put some more info in documentation of IntegratorMultiDim or IntegrationMultiDim::Type ? Currently information
about tolerance vs ncall for MC integration is given in the description of GSLMCIntegrator ctor (with which we don’t interact normally)
regarding discontinuity is given in AdaptiveIntegratorMultiDim description “usually works best for smooth functions”, but again we don’t really interact with it directly and the statement seems to weak as here we get a bias not e.g. long convergence time