FAQ¶

Build and setup¶

`make edf-omp` fails with "cannot find -llapack"¶

The edf-omp target links against system LAPACK/BLAS. Install them:

sudo apt install liblapack-dev libblas-dev    # Ubuntu/Debian

If they are installed at a non-standard path, edit Makefile lines 104–106 to reflect the correct paths.

OpenMP doesn't activate — `edf-omp` still runs single-threaded¶

Check that you are actually running edf-omp, not edf. The static edf binary cannot activate OpenMP regardless of OMP_NUM_THREADS.

OMP_NUM_THREADS=4 ./edf-omp < input.edfinp > output.edfout

Verify from the timer section: if _linearsys wall time scales with thread count, OpenMP is working.

AOM normalization errors¶

`edf.f: AOM does not satisfy the sum rule`¶

The AOM fails the check \(\sum_A S^A_{ij} \approx \delta_{ij}\) for some orbital pair. Causes:

Incomplete integration: PROMOLDEN did not converge the atomic basin integrals — increase angular/radial grid in PROMOLDEN.
Wrong AOM file: The .aom file does not match the .wfn file (different geometry, basis, or orbital ordering).
Large tolerance needed: For fuzzy partitions, small numerical errors can accumulate. Loosen TOLAOM temporarily to diagnose.

Quick fix: add AOMNORM n (where n is the last atom) to adjust the last atom's AOM row so the sum rule holds exactly.

All probabilities are zero¶

This usually means the AOM and WFN are inconsistent — the AOM integrals correspond to a different wavefunction. Verify that: - The .aom file was generated from the same geometry as the .wfn file. - The basis set in PROMOLDEN matches the one used in GAMESS/Gaussian. - ioverlap is correct for the source of the .aom file.

CASSCF wavefunction issues¶

"The WFN file does not contain CASSCF determinants"¶

The .wfn file must come from a patched version of GAMESS10 that writes determinant coefficients and active-space occupation numbers in the extended aimpac format. Standard GAMESS10 outputs do not include this information.

Probability sum is significantly less than 1¶

With PROBCUT > 0, only configurations above the threshold are printed. The TOTAL SUM line (not the SUM line) gives the true total. If the TOTAL SUM itself is much less than 1: 1. Check TOLAOM — the AOM normalization may be poor. 2. Check EPSDET — if set too aggressively, many determinant pairs are skipped.

Calculation is very slow for large CASSCF¶

The cost scales as \(O(N_\text{det}^2 \times N_\alpha \times N_\beta)\). To speed up: - Use EPSDET 1e-8 to skip negligible determinant pairs. - Use EPSWFN 1e-5 to drop small-coefficient determinants before the EDF loop. - Use NDETS n to limit the expansion. - Build edf-omp and set OMP_NUM_THREADS to the number of physical cores (only effective for ngroup = 2 or with RECUR).

Threading and performance¶

Performance degrades with more threads¶

Most likely cause: BLAS over-subscription. If the system BLAS (OpenBLAS, MKL) uses internal threading, each of the N OpenMP threads may spawn M BLAS threads, giving N×M total threads on a machine with N cores.

Fix:

export OPENBLAS_NUM_THREADS=1   # for OpenBLAS
export MKL_NUM_THREADS=1        # for Intel MKL

No speedup for `ngroup > 2`¶

Correct. The parallelized binedf.f engine is used only for ngroup = 2 (or when RECUR forces it). For more groups, the other engines (calcedfd.f, rcalcedf.f, etc.) are not yet parallelized.

Output interpretation¶

Printed probability sum is well below 1 but `TOTAL SUM` is 1¶

This is expected when PROBCUT > 0. The SUM line counts only printed configurations; the TOTAL SUM includes all of them. Decrease PROBCUT (or set PROBCUT -1) to see more configurations.

Delocalization indices are negative or very large¶

Negative two-center DIs can appear for fragments that contain many electrons and are not physically bonded. Very large DIs between fragments that formally share a double or triple bond (e.g., N₂) are correct — \(\delta(\text{N,N}) \approx 3\) for a triple bond.

Self-DI (localization index) \(\delta(i,i)\) is always positive by construction.

The N5+ output shows different orbital assignments than the reference¶

The dsyev LAPACK eigensolver (used since the Tier 2A upgrade) may return degenerate eigenvectors in a different order than the old jacobi routine. Both orderings are mathematically valid. This is an acceptable difference and does not affect EDF probabilities or delocalization indices.

Input file questions¶

Can I put comments in the input file?¶

Yes. Any line whose first non-blank character is # is treated as a comment and ignored.

Does keyword order matter?¶

Keywords after Record 3 are mostly order-independent. However: - NGROUP must appear before the fragment lines (Records 5.i). - MAXPOP/MINPOP must appear after the PRSRS keyword they modify. - BONDING/ENDBONDING must form a complete block.

Can I omit `END`?¶

Yes. edf stops at end-of-file if END is not given.