[spraharsh]

test_generalized_dimer::test1 starts from an unseeded random configuration and failed intermittently (e.g. in CI on the meson branch) with:

TypeError: LineSearchError.__init__() missing 1 required positional argument: 'value'

Cause

The dimer's rotation step (FindLowestEigenVector) minimizes the curvature along a direction with MYLBFGS(rel_energy=True), which accepts a step only if (E_new - E_old) / |E_old| <= maxErise (1e-4). When the lowest curvature approaches zero (seen at ~1e-7), ordinary finite-difference noise (~0.1% relative) exceeds that, every step is rejected, and after 10 failed line searches _lbfgs_py.py does raise LineSearchError without its required argument, hence the TypeError.

Change

• LBFGS_CPP: get_state() / set_state() for the LBFGS memory (s, y, rho, H0, k), in C++ and Cython, with a size check
• FindLowestEigenVector(cpp_lbfgs=True): use LBFGS_CPP and keep its memory across update_coords (as MYLBFGS did). GeneralizedDimer turns it on. findTransitionState is unchanged (still MYLBFGS); it hasn't been evaluated the same way.
• LineSearchError is raised with a message.
• TODO in LowestEigPot on its Python overhead: with the C++ LBFGS each curvature evaluation costs ~34 us, of which 2.5 us is the LJ18 gradient; the rest is Python/numpy/orthogopt. A C++ LowestEigPotential exists in source/pele/lowest_eig_potential.hpp but lacks rotational orthogonalization and a Cython wrapper.
• TODO in the test with its remaining failure rate and seeds.

Evaluation

To avoid tuning on the seeds used to choose the optimizer (0-299), everything below uses held-out seeds, paired (same seeds for both rotators). A result counts as a transition state only if the exact Hessian has exactly one negative eigenvalue (rigid-body modes excluded) and rms gradient < 1e-3; "test fails" means res.success is False.

system	start	rotator	runs	verified TS (95% CI)	crashes	test fails
LJ18	random box (the test)	MYLBFGS	2000	96.2% (95.3–97.0)	66	71
LJ18	random box (the test)	LBFGS_CPP	2000	97.0% (96.2–97.7)	0	3
LJ18	minimum + noise	MYLBFGS	2000	95.0% (94.0–95.9)	98	102
LJ18	minimum + noise	LBFGS_CPP	2000	97.3% (96.5–97.9)	0	2

Not tuned on: LJ13, LJ38, Morse13 (rho=6) and binary LJ18, 500 seeds per start each. No crashes anywhere with LBFGS_CPP (301 → 0 over all 8000 runs), test-failure rate 334 → 67. Verified-TS rates are equal or higher within the confidence intervals, except binary LJ18 from a random box (89.8% vs 91.4%, overlapping CIs).

Known limitations, unchanged by this PR:

• The 3 remaining LJ18 failures (seeds 10089, 10735, 11584) all lose an atom and converge to a minimum plus a free atom: the dimer has no guard against dissociation.
• res.success only checks that the lowest curvature found is negative: 56 of the 2000 LJ18 runs "succeed" at points that are not first-order saddles (mostly two negative modes). With MYLBFGS it was 22, largely because runs that crashed now continue.

Testing

• pytest pele/: 591 passed (built with the current setup.py), including 2 new tests: restarting LBFGS_CPP from get_state() takes the same step as the original, and set_state with the wrong size raises.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 91.66667% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
pele/optimize/_lbfgs_py.py	0.00%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!