Peak at block length v.s. on-the-fly method

General discussion of the Cambridge quantum Monte Carlo code CASINO; how to install and setup; how to use it; what it does; applications.
Post Reply
Tomohiro Ichibha
Posts: 10
Joined: Sun May 24, 2015 8:14 am

Peak at block length v.s. on-the-fly method

Post by Tomohiro Ichibha »

Dear,

I have a question about reblocking analysis.
Data below seems to be converged at 131072
block length (peak of Std.Err.), but
on-the-fly says, on the other hand, the optimal
block length is 4096.
Which one should I trust?

Ichibha.
--------------------------------
Block len Std error Err in error
1 3.889309E-04 3.708314E-07
2 4.666462E-04 6.292267E-07
4 5.690810E-04 1.085199E-06
8 7.052431E-04 1.901913E-06
16 8.833990E-04 3.369200E-06
32 1.117819E-03 6.029245E-06
64 1.419440E-03 1.082769E-05
128 1.794182E-03 1.935645E-05
256 2.264997E-03 3.456148E-05
512 2.837687E-03 6.124990E-05
1024 3.567381E-03 1.089451E-04
2048 4.554457E-03 1.968862E-04
4096 5.749629E-03 3.521654E-04 BEST
8192 7.274676E-03 6.325151E-04
16384 9.200726E-03 1.139994E-03
32768 9.658311E-03 1.718971E-03
65536 1.156532E-02 3.007819E-03
131072 1.424515E-02 5.634266E-03
262144 1.137400E-02 7.675037E-03
Katharina Doblhoff
Posts: 84
Joined: Tue Jun 17, 2014 6:50 am

Re: Peak at block length v.s. on-the-fly method

Post by Katharina Doblhoff »

I would sort of not trust either. Yours not because the error in the variance at your "peak" is nearly as large as the variance and the authomatic reblocking one not because it is not a a plateau. Overall, this looks a bit strange to me. Is this a VMC or a DMC run? What kind of system is this? What time step are you using?
Tomohiro Ichibha
Posts: 10
Joined: Sun May 24, 2015 8:14 am

Re: Peak at block length v.s. on-the-fly method

Post by Tomohiro Ichibha »

Hi Katharina,

Thank you for your response.
The target system is iron(II) phthalocyanine,
and the calculation is DMC with all-electron CASSCF trial wave function.
Its timestep is quite small, 0.0001, which is determined following
"R.M. Lee et al., Phys. Rev. E 83, 066706 (2011)."

Ichibha
Katharina Doblhoff
Posts: 84
Joined: Tue Jun 17, 2014 6:50 am

Re: Peak at block length v.s. on-the-fly method

Post by Katharina Doblhoff »

Dear Ichibha!

In this case, I would guess that you simply need more steps. Your total (imaginary) propagation time is only 26 a.u. I always use an estimate of about 1 a.u. in correlation time. So you are not converged in reblocking yet. So what you see is that you have not reached a plateau yet. The "peak" you see is just because of too large errors in the estimate of sigma. For short propagations it is sometimes better to use the correlation time estimate (and its error estimate). I am not quit sure right now, but I think it is given in the output. If not, it is stated when calling reblock.

Are you sure you want(need) to do this all electron? This looks humongous to me: 1xFe, 8xN, 32xC plus some H makes about 300 electrons. Considering the tiny time step one needs for all electron Fe, I would expect this to become very expensive (but I have no experience with all electron calculations).
Tomohiro Ichibha
Posts: 10
Joined: Sun May 24, 2015 8:14 am

Re: Peak at block length v.s. on-the-fly method

Post by Tomohiro Ichibha »

Dear Katharina,

Thank you for your reply.
The calculation may be huge certainly.
I will reexamine the calculation setting considering what you suggested.

Ichibha
Katharina Doblhoff
Posts: 84
Joined: Tue Jun 17, 2014 6:50 am

Re: Peak at block length v.s. on-the-fly method

Post by Katharina Doblhoff »

If considering redoing the calculations with pseudo-potentials, you may want to consider reading our recent paper http://pubs.acs.org/doi/abs/10.1021/acs.jctc.6b00160, but also this one by Nazarov, Shulenburger, Morales and Hood: http://link.aps.org/doi/10.1103/PhysRevB.93.094111 where the use of PPs, especially also for 3d atoms is discussed.
Post Reply