johanjoensson / impurityModel / 33192921138
92%

DEFAULT BRANCH: master
Ran
Jobs 1
Files 245
Run time 1min
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst
coverage: 90.958% (-0.003%) from 90.961%
33192921138

push

github

johanjoensson 
Document the Sigma causality violation as a discretization artifact (B3)

Runs the plan's B3 root-cause experiments on the real Mn archive that produced
the original 128-rank crash, reproduced serially via
test/support/real_workload.load_workload + calc_selfenergy(comm=None) on a
narrow omega slice: the rerun matches the archived violation to 4 significant
figures.

Four bounded experiments, each isolating one variable:
- Tightening the GF Lanczos tolerance 1000x leaves the violation unchanged
  (rules out non-convergence, confirming Part A and Part B are independent).
- u4=None (non-interacting) collapses the violation to roundoff (rules out a
  Sigma-construction bug: index alignment, delta convention, DC handling --
  none of those depend on U and would still show up).
- Tightening excitation_budget 4->3 makes the violation 60x worse.
- Doubling delta (broadening) 0.010->0.020 eliminates it entirely.

Conclusion: a discrete excited-sector many-body basis approximating a
continuous bath near a sharp Im(Delta) resonance, evaluated at a broadening
too small to smooth over that discreteness -- the standard finite-resolution
artifact of an ED impurity solver, not a code defect. Written up in
doc/known_issues.md alongside the SIGMA_CAUSALITY_TOL knob it justifies.

Loosening excitation_budget past 4 (to see the violation shrink from the
other direction) was attempted but killed as measured-infeasible on a single
core; recorded as such rather than silently dropped -- the tightening and
delta experiments already establish the direction and magnitude.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01D2dt2AJ97Aa4nhp2S7RtdR

30601 of 33643 relevant lines covered (90.96%)

0.91 hits per line

Coverage Regressions

Lines Coverage ∆ File
1
99.42
 -0.58% src/impurityModel/test/lanczos/test_block_lanczos_blowup.py
Jobs
ID Job ID Ran Files Coverage
1 33192921138.1 245
90.96
 GitHub Action Run
Source Files on build 33192921138