Repository navigation
time_change should integrate cumulative inter-arrival segments #82
Description
time_change(h, pp::InhomogeneousPoissonProcess) currently computes the compensator at each event time and at tmax via independent quadrature calls. Independent calls have independent ulp-level rounding errors, so the transformed times can come out of order or slightly past new_tmax. Switching to a cumulative integration over inter-arrival segments removes both failure modes by construction with no change in cost.
Metadata
Metadata
Assignees
Labels
No labels
Activity
Fixes #82 and #83 (#84)
JoseKling commentedon May 19, 2026
It occurred to me that this should actually be standard implementation for
AbstractUnivariateProcess, using theintegrated_ground_intensitymethod, instead ofintegrated_intensity.This should be the correct implementation for any process, only to be replaced if there is some benefit. We could even leave the
DomainErrormessage, although replaced by something more generic, like"""
time_change: integrated intensity over [$(edges[i_neg]), $(edges[i_neg + 1])] is negative. This could mean:
integrated_ground_intensity."""
rsenne commentedon May 19, 2026
So basically make this implementation the fallback i.e., if there is no analytical we can always just use quadrature? That sounds good to me! Want to open an issue with the proposal?