Hi Richard,
Thanks very much for your reply. I expect the difference I’m seeing between the
versions when there’s a normalization factor and a pressure shift is due to the
calculation of the normalization factor on the unshifted frequencies. I don’t
know enough about the physics to comment which is more appropriate.
However, I still don’t understand the differences I’m seeing with the mirror
lines, as they are there even with no cutoff and no forefactor. In v2.3 I get
identical results if I use:
1. abs_lines_per_speciesAddMirrorLines with a single line with no pressure
shift @ H2O-161 183310107053.569 0.0 2.3119066399931e-16 296
2.70486699225234e-21 30326.8955785838 153557.646880829 0.71 0.71 296 0 3000000
0.1 0.05 0.05 -1 -1 -1
2. a manual mirror line: @ H2O-161 183310107053.569 0.0 2.3119066399931e-16
296 2.70486699225234e-21 30326.8955785838 153557.646880829 0.71 0.71 296 0
3000000 0.1 0.05 0.05 -1 -1 -1
@ H2O-161
-183310107053.569 0.0 2.3119066399931e-16 296 2.70486699225234e-21
30326.8955785838 153557.646880829 0.71 0.71 296 0 3000000 0.1 0.05 0.05 -1 -1 -1
In v2.5 these two line files give different results (when the first is applied
with mirroring_option=”None” and the second with mirroring_option=”Manual”)
Note that I’m doing these tests with a Lorentz line shape, so the differences
in the Voigt function shouldn’t matter.
Thanks for your time,
Stuart
From: Richard Larsson <ric.lars...@gmail.com>
Sent: 09 June 2021 17:01
To: Fox, Stuart <stuart....@metoffice.gov.uk>
Cc: arts_dev.mi@lists.uni-hamburg.de
Subject: Re: ARTS line absorption v2.3 vs v2.5
This email was received from an external source. Always check sender details,
links & attachments.
Hi Stuart,
All the differences between current 2.5 version and old version that I can
think of are:
The Doppler broadening is computed on shifted frequencies instead of on
unshifted frequencies. This should be more accurate as far as I am concerned.
The normalization factor is computed on the unshifted frequencies instead of on
the shifted frequencies. If I have not misunderstood the normalization
factors, this should also be closer to the physics since you shouldn't really
be scaling the line strengths with the line frequency but by the atmospheric
frequency (as in the new LM code). This is potentially an issue.
The line shape can no longer be a speed-up Voigt function but you have to use
the full calculations. The old code was wrong at a scale of 1e-6 if I
understand the old comments.
The new code does the mirroring inline unless manually added to the catalog.
Before you had to manually add a new line with negative line center. It is
possible there's a bug here, since I haven't tested this so accurately before.
This changes the cutoff frequency since the old code cutoff at CUTOFF - F0
instead at F0 + CUTOFF for the inline mirrored lines. (Thinking about it
practically, this makes the line shape non-symmetric, so I will have to add a
check to the lbl-check that there's no cutoff used when mirrored lines are
there by inline instead of by copy-pasta.) I would think that the difference
in cutoff here is the main difference you are seeing.
I think that's about it. There's of course entire different algorithm
involved, so some additional differences may be from there.
With hope,
//Richard
Den ons 9 juni 2021 kl 17:21 skrev Fox, Stuart
<stuart....@metoffice.gov.uk<mailto:stuart....@metoffice.gov.uk>>:
Hello all,
Thanks for all the work on the line cutoff behaviour in the latest version of
ARTS! However,I’m still struggling to reproduce the absorption line behaviour
between v2.3.1277 and v2.5 even after the latest work on the line cutoff from
Richard. In fact, even without a line cutoff there still seems to be
differences caused by (i) the interaction between a line normalization factor
(e.g. VVH) and the line pressure shift, and (ii) the treatment of mirror lines.
A simple example of (i) using only a single water vapour line with a Lorentz
profile and VVH pre-factor can be found here, with differences up to about
0.08K:
https://github.com/stuartfox/arts-cutoff-problem/tree/single-line
For (ii), if I set the normalization to None but add the mirror lines (with the
SameAsLineShape option) I get differences of 0.001K in this test, but much
larger values in a more realistic case with more lines (0.5-0.6K below 100GHz).
I’ve no idea which version is more correct, but it would be nice to understand
the causes of these variations!
Thanks for all your help,
Stuart
