johanjoensson / impurityModel / 33170189467

91%

Build Type

push

github

Committed by johanjoensson

Commit Message

Localize and tolerate a self-energy causality violation instead of only raising

The Mn self-energy run that motivated this also died mid-calculation with
UnphysicalGreensFunctionError -- a separate problem from the Lanczos
reporting bug (confirmed independent: the solver's ~7.5e-10 relative change
cannot produce the observed ~5e-3 absolute Im(Sigma) artifact). The saved
imag-sig-0-realaxis-Mn.dat shows a real, localized artifact: 20 of 2001
points violate causality in a ~2*delta-wide window around omega=-0.266,
coincident with a sharp hybridization resonance -- worst per-orbital ratio
(max_w Im(Sigma_ii) / max_w |Im(Sigma_ii)|, the causality-relevant measure)
computed directly from that file is 7.62e-3.

check_greens_function raised unconditionally on any positive diagonal Im,
with a message carrying no numbers -- so a 128-rank job's final step threw
away its own diagnosis along with the run.

- sigma.py: check_greens_function gains an optional relative `tol` (default
  0.0 -- exact original behaviour, used unchanged for the Green's function
  itself) and a localization report: violating-point count/fraction,
  contiguous omega window(s), the worst point and its per-orbital ratio (not a
  whole-block max|Im|, which would let a wide high-weight orbital mask a
  violation on a narrow one), and -- when get_sigma's new
  return_components=True hands back (G0^-1, G^-1) per block -- the
  Sigma = G0^-1 - G^-1 breakdown at the worst point, so the mechanism (the
  interacting G narrower than the bare hybridized G0) is visible in the
  traceback itself.
- config.py: SIGMA_CAUSALITY_TOL (default 1e-3) -- comfortably above
  roundoff, and below the 7.62e-3 this run actually measured, so the new
  default still raises on it rather than silently rescuing it.
- selfenergy.py: _self_energy_on_mesh passes SIGMA_CAUSALITY_TOL and the two
  components through; below tolerance the run now warns (with the report)
  and continues instead of aborting.

Tests: tolerance be... (continued)

Coverage Stats

120 of 122 new or added lines in 4 files covered. (98.36%)

73 existing lines in 3 files now uncovered.

30602 of 33643 relevant lines covered (90.96%)

0.91 hits per line