wiki:ticket/278/TicketSummary/CommentFromGordonAndMarek

Email correspondence with Gordon Inverarity and Marek Wlasak

From: Steinle, Peter
To: Inverarity, Gordon; Wlasak, Marek
Cc: Jin Lee
 
Thanks for tips on the extra diagnostics

– the spikes didn't bother me too much, I understand why some of them are there, and appreciate that any ideal values for any diagnostics require a correctly tuned B & R.( Been fun reacquainting myself with Desroziers & Ivanov etc. for the last couple of days  )

It was more the ratio of Jb(final) to Jo(final) that raised my eyebrows – but I think we have an explanation: This was from a tropical area dominated by satellite data: bkg error variance maybe ½ to 1/3 of the dominant obs error – especially since dominant obs are of temperature. 

Bkg error covariances might be a bit low, but possibly not outrageously so

Pete

From: Inverarity, Gordon [mailto:gordon.inverarity@metoffice.gov.uk] 
Sent: Wednesday, 22 June 2016 11:14 PM
To: Peter Steinle; Wlasak, Marek
Cc: Jin Lee
Subject: RE: [Access Development] #278: Diagnosis of global VAR inner loop convergence [SEC=UNCLASSIFIED]

I agree with Marek about the need to look at job.out files. These are where you can follow the detailed progress of the minimisation and see whether any problems are explicitly reported. Another useful diagnostic is to set totalpenstats=.true. in the observation namelist. This puts the Jo contribution for each observation type into the job.stats file and would show which instrument is associated with the spikes.
 
On the final total J statistic, it's worth noting that Tim Payne has shown that the final value of 0.5 * no. of obs depends on the R matrix being correct and that the inclusion of satellite observations with an artificially diagonal R matrix breaks this assumption and hence the result. VAR can generate these statistics for you using the namelist setting use_jdivp in the observation namelist.
 
Gordon
 
 
 
Gordon Inverarity  Scientific Manager of Data Assimilation Methods 
Met Office  FitzRoy Road  Exeter  Devon  EX1 3PB  United Kingdom 
Tel: +44 (0)1392 884656 
Email: gordon.inverarity@metoffice.gov.uk  Website: www.metoffice.gov.uk 
http://www.metoffice.gov.uk/research/people/gordon-inverarity 
 
 
Catch up with the latest #WeatherStory- http://www.metoffice.gov.uk/news/weatherstory
Our magazine Barometer is now available online at http://www.metoffice.gov.uk/barometer/
 
-----Original Message-----
From: Peter Steinle [mailto:P.Steinle@bom.gov.au] 
Sent: 20 June 2016 23:50
To: Wlasak, Marek
Cc: Inverarity, Gordon
Subject: RE: [Access Development] #278: Diagnosis of global VAR inner loop convergence [SEC=UNCLASSIFIED]
 
 
Hi Marek
 
Thanks - I had vague memories about some normalization by number of obs. Looks like a sign I have been focussing on operational and admin issues for way too long (or worse still ... old age)
 
Pete
-----Original Message-----
From: Wlasak, Marek [mailto:marek.wlasak@metoffice.gov.uk] 
Sent: Monday, 20 June 2016 11:40 PM
To: Peter Steinle
Cc: Inverarity, Gordon
Subject: RE: [Access Development] #278: Diagnosis of global VAR inner loop convergence [SEC=UNCLASSIFIED]
 
Hi there
 
The final J for a consistent system with correctly specified statistics it should be equal to 0.5 * number of observations. How many observations do you have? (This is a theoretical result makes a number of assumptions and is called the Bennett-Talagrand ratio)
 
I can see your ticket if I look at 
 
http://accessdev.nci.org.au/trac/ticket/278
 
I can view the pictures also if I look there.
 
I can't see your suite configuration u-ac651. Is that accessible? Are you using Hessian preconditioning? What is the cov file?
 
The pictures you have generated probably come from the stats files.  It is better to get the values from job.out file as they are the final true values.  The stats file holds the first guess of J and grad J for each minimisation; the value is dependent on how far you go in a certain direction.
 
Smaller background error variances will reduce the number of iterations it takes to get to suboptimal solution; it could also cause model drift.
 
I hope this helps
 
Kind regards
 
Marek
 
 
 
 
 
 
-----Original Message-----
From: Peter Steinle [mailto:P.Steinle@bom.gov.au] 
Sent: 20 June 2016 08:41
To: Inverarity, Gordon; Wlasak, Marek
Cc: Jin Lee
Subject: FW: [Access Development] #278: Diagnosis of global VAR inner loop convergence [SEC=UNCLASSIFIED]
 
Hi guys
 
We were looking at cost function convergence for various reasons, and I just realized, I don't have a simple answer ready for why the final J is so heavily dominated by Jo. The reduction in Jo seems quite small (and so of course the increase in Jb is also quite small). Having an order of magnitude difference between the final Jo and Jb seems a bit hard to explain.
 
Examples from our system are in the attached ticket - can you see this ticket? Anyway here are some pix.
 
Do either of you have any comments or suggestions?
 
Pete
 
-----Original Message-----
From: Jin Lee [mailto:jtl548@nci.org.au] 
Sent: Monday, 20 June 2016 4:39 PM
Subject: Re: [Access Development] #278: Diagnosis of global VAR inner loop convergence
 
#278: Diagnosis of global VAR inner loop convergence
Reporter:        jtl548
Priority:        major
Changes (by jtl548):
 
* cc: pjs548 (added)
 
 
-- 
Ticket URL: <http://accessdev.nci.org.au/trac/ticket/278#comment:2>
Access Development <http://accessdev.nci.org.au>

Last modified 4 years ago Last modified on Nov 4, 2016 8:21:22 AM