Repository navigation
Fitbauer v5.2.4
Fitbauer v5.2.4
A corrective release. A closer comparison with NORMOS, job by job and with
NORMOS itself re-run, showed that importing a .JOB did not always rebuild
the same model that NORMOS had fitted. This release fixes that, and corrects
the validation figures given with v5.2.3.
If you imported NORMOS .JOB files with v5.2.3 or earlier, import them
again. The fits may have had more free parameters than in NORMOS, and the
hyperfine field was on a slightly different scale.
Fixes in the .JOB import
- Only what the
.JOBfrees is free. In NORMOS, every parameter without
…FIT=.TRUE.is fixed. The importer only carried over the flags of ISO,
QUA, BHF, WID and DEP. Everything else —relative intensities and widths
(D13, D23, W13, W23, D21, W21), ETA, THE, PHI, BEX, GAX and the background
slope (BKGFIT)— was left as it was in the session, which is free by
default. Fitbauer could end up fitting 11 variables where NORMOS fitted 6. - Intensity ratios by area. D13/D23/D21 are area ratios. They used to be
converted into depth ratios, so freeing W13/W23 also freed the intensities.
They now go straight into int1/int2 with the by-area intensity convention. - NORMOS conventions are switched on. The NORMOS line pattern (derived
from the nuclear moments), intensities by area and its relaxation rule are
activated on import, so the job gives NORMOS's values. With the α-Fe pattern
used before, the field came out 0.035 % lower: about 0.01 T at 33 T, more
than the error of a good fit. Loading another spectrum afterwards switches
back to Fitbauer's conventions, unless you chose them by hand in Fit ›
Advanced fit options › Conventions (NORMOS compatibility). - Constraints. On import, constraints on WID now take the W13 scaling into
account, and W13/W23 constraints of a sextet are mapped to the right width
parameters (they were swapped).
Fix in the .JOB export
- Constraints are written (
NDEX/FACTOR/CONST). They used to be lost.
NORMOS has been checked to accept them and to fit with the same variables.
Validation figures corrected
The v5.2.3 notes said that, on 216 real NORMOS jobs, Fitbauer reached the same
or a lower χ² in 211 (median reduced χ² 1.08 → 0.81). Those figures came from
flaws in the comparison itself: the import freed parameters that NORMOS kept
fixed, the refit used the α-Fe line pattern, Fitbauer folded the data
differently, and the χ² was computed on pair averages, giving half of NORMOS's
value.
Redone on the data exactly as NORMOS folds them, with the same model and the
same free parameters, NORMOS's χ² is reproduced to about 1 %. Of 407 distinct
real jobs, 146 are a valid reference. The other 261 are discarded:
- 214 because NORMOS's reduced χ² is above 2;
- 19 because the two programs do not see exactly the same data;
- 28 because they cannot be analysed (missing files, unsupported options).
On the 146:
- Fitbauer reaches the same or a lower χ² in all of them (median 1.40 → 1.38).
- The median agreement is 0.0006 mm/s in δ, 0.0015 mm/s in ΔEQ, 0.0034 T in
BHF and 0.0014 mm/s in Γ. That is 0.15–0.2 of NORMOS's own standard errors,
with no bias. - The few differences above three standard errors are fits where NORMOS had
stopped short of the minimum. Re-run from Fitbauer's solution, NORMOS
reaches the same χ².
Known limitations
These remain. Where marked, the import shows a warning when they apply.
- Relative widths of a sextet. NORMOS uses WID (lines 3–4) with W13 and
W23 relative to it; Fitbauer uses Γ1 (lines 1–6) with Γ2/Γ1 and Γ3/Γ1.- The free/fixed status maps exactly when WID, W13 and W23 are all fixed,
when only WID is free, or when all three are free. - With W13 free and WID or W23 fixed, Fitbauer has one or two more free
width parameters than NORMOS (warned). - Constraints on W13/W23 are not exported. On import, only a factor-1
constraint between W13 values is exact.
- The free/fixed status maps exactly when WID, W13 and W23 are all fixed,
- Area constraints (
NDEXon DEP/ARE) are exact at the values in the
.JOB. If the fit changes the widths, the area ratio drifts slightly
(warned). - Folding. Fitbauer folds at the refined folding point, interpolating
between channels, and drops dead edge channels. NORMOS adds whole channel
pairs at ⌊PFP⌋ + 0.5. Line widths can come out about 0.005 mm/s narrower
than in NORMOS. When NORMOS's folding point is more than about one channel
from the centre of the spectrum, its edge channels cannot be reproduced
exactly. - Not carried over on import, as before (warned):
- the
OMEconstraint of the Ising relaxation; - the Billard–Chamberod model (
METHOD=3); - several distribution blocks in one job;
- the absolute
LAMDAsmoothing parameter (use the L-curve).
- the
- Not written on export: Hamiltonian treatments, relaxation and
distribution shapes. The exported.JOBis a starting point for NORMOS, not
an exact translation of every model.
Full suite: 763 tests passing.