Strange error using SuperBayes

[Tom Varley]

Thanks for the prompt response to my last question!

I can only reward you with another (slightly strange) problem I have
been having. I am running SB v1.13 compiled with the latest intel
compilers on my local cluster and have been having a recurring
problem. Any time I try do the full scan (8 parameters) my chain
terminates with the error: "Too many interations in MHiggs", after a
few hours or so at random. The offending point looks like (trimmed
after I ran it in ./tester to reproduce the error):


This is SuperBayeS tester v1.11, October 2007
Getting Likelihood...
>>>> Input Params
m0 param : 1588.16040039062 (GeV)
mhalf param : 557.411437988281 (GeV)
a0 param : -4087.96484375000 (GeV)
tanb param : 43.8475532531738
sgnmu0 param : 1
>>>> Nuisance Params
mb(mb)^MSbar : 4.06332969665527 (GeV)
m top : 159.512466430664 (GeV)
alpha_s^MSbar : 0.122880749404430
alpha_em^MSbar : 7.810163079580099E-003
or al_em^-1: 128.038299560547
>>>> Fixed Nuisance Params
ftp_s : 0.118000000715256
ftp_n : 0.118000000715256
... Now calling softsusy interface ...
>>> Softsusy spectrum
Lights Chargino mass : 466.6876 (GeV)
Lights Higgs mass : 67.94718 (GeV)
Lights neutralino mass : 239.7419 (GeV)
>>> Softsusy: new sparticle masses
selectron R mass : 1600.976 (GeV)
smuon R mass : 1598.166 (GeV)
stau 1 mass : 383.5670 (GeV)
tau sneutrino mass : 1200.961 (GeV)
stop 1 mass : 405.0683 (GeV)
sbottom 1 mass : 1136.931 (GeV)
Lights squark mass : 1922.164 (GeV)
Gluino mass : 1363.736 (GeV)
>>> Error flags from SoftSusy (0=OK, 1=not OK)
soft no good REWSB : 0
soft problem with mu : 0
soft tachyonic Higgs : 0
soft tachyonic sferm : 0
soft Landau pole : 0
soft no EWSB : 0
soft no EW minimum : 0
soft Landau pole 2 : 0
soft no convergence : 0
... Now calling DS interface ...
>>> DS output
>>> Dark matter properties
Relic density : 1.2670
Gaugino fraction : 0.99920
>> Direct signatures
SI nn cross section: 0.60675E-10 (pb)
SD np cross section: 0.10796E-07 (pb)
SD nn cross section: 0.10578E-07 (pb)
>>> Collider quantities
MW : 80.34373 (GeV)
sin2eff : 0.23156
g-2 amplitude : 0.31245E-09
b -> s gamma : 0.28783E-03
Bs -> mu+ mu- : 0.16966E-07
BRR : 0.0000
sgn MSSM/SM : 1
>>> Error flags from DarkSusy (0=OK, 1=not OK)
Neutralino not LSP: 0
oh2 problem : 0
... Now calling BDecay interface...
Delta(MB): 11.877 (ps^-1)
>>> Error flags from Bdecay (0=OK, 1=not OK)
BD problem: 0
... Now calling FH interface...
FHHiggsCorr: error computing Mh1

>>> Error flags from FeynHiggs (0=OK, 1=not OK)
FH problem: 1
... Computing likelihood ...
-lnlike CDM = 0.4193877E+02
-lnlike nuisance = 0.2925820E+02
Too many interations in MHiggs


So I found the relevant line of code in calclike.f90 in the ./source
directory, and what seems odd is that even though the Higgs mass is
low enough to be discarded out of hand, SB seems to still be trying to
calculate the likelihood....


Any ideas? I am a little bit stuck...

Cheers!

[Roberto Trotta]

Hi Andy

I've tried your point with v1.11 compiled with Intel Fortran 8.1-5386.
I get precisely the same spectrum and observables values, but it also
runs through the like calculation seamlessly, giving:

... Computing likelihood ...
-lnlike CDM = 0.4193213E+02
-lnlike nuisance = 0.2925820E+02
-lnlike M_higgs = 0.1000000E+31
-lnlike sparticles = 0.0000000E+00
-lnlike g-2 = 0.4644840E+01
-lnlike Bs -> gamma = 0.2019793E+01
-lnlike Bs -> mumu = 0.0000000E+00
-lnlike M_W = 0.1731891E+01
-lnlike sineff = 0.1025733E-01
-lnlike Bdecay = 0.7532721E+00
New -lnlike: 0.10000E+31

This correctly gives a large value to the -ln like from the Higgs
mass, hence the point is effectively discarded, as it should.

Now in the Higgs likelihood routine there is an integral used to
smooth out the LEP limit incorporating the theoretical uncertainty in
the Higgs mass (which is this case ought to be irrelevant). This
integral is what is giving you troubles and why it stops. I strongly
sospect it is a "feature" of your v10 (I assume) compiler. Can you try
downgrading to v9 and tell us if this still happens?

Another useful test to run is to try and reproduce the output of
tester for the supplied point in the original version (which has a
reasonable Higgs mass) and make sure that at least in this region of
param space you get the right numbers. If not, then it's definitely
something to do with your local implementation.

Roberto

[Roberto Trotta]

Errata - I meant "Hi Tom" :-)

[Roberto Trotta]

Hi again

we seem to have nailed the problem down:

In calclike.f90 the smoothed 2D like for the Higgs mass is computed by
calling

call qtrap(HiggsLikeIntegrand, IntMin, IntMax, lnlike)
[...]
call qtrap(HiggsLikeIntegrand, LeftIntMin, LeftIntMax, Leftlnlike)

which for a very small mass give

lnlike = 0
Leftlnlike = 0

so

Leftlnlike/lnlike = "NaN"

on Line 500.

Apprently the ifc v8 takes Leftlnlike/lnlike to be 0 and everything
works fine while v9 and v10 give NaN which makes the loop go on
forever.

So I suggest to patch it by replacing Line 500 by the following

if ((Leftlnlike/lnlike < tolerance) .or. (Leftlnlike .eq. 0.d0)) then

This should do the trick for now. We'll patch it in a more solid way
in future releases.

Pls let us know if this works.

Roberto

[Tom Varley]

Just tried this and it seems to do the trick!

Once I adjusted it as you said on line 500 and made a similar
replacement on line 512, the tester returned the point with sensible
output (i.e. a large -lnlike).

I will test it in the wild and see what happens..I will let you know.

Thanks again and a very Merry Christmas!

Tom