trial/guiding wave function optimisation for DMC

General discussion of quantum Monte Carlo in electronic structure theory

trial/guiding wave function optimisation for DMC

Postby Vladimir_Konjkov » Wed Jul 19, 2017 4:15 am

Hello, all.

It's well known that 'quality' of trial/guiding wave function in DMC calculations greatly decrease variance of DMC energy and thus reduce the time required to achieve desired accuracy. Guiding wave function defines the branching term as the energy difference between true ground-state energy and the local energy of the guiding function. The better guiding wave function approximates true ground-state the less DMC energy variates.

In general, the JASTROW VMC optimized trial wave function is a suitable choice of guiding wave function in DMC. But since there are several ways to do JASTROW optimization which one gives the best result in terms of the DMC energy variance?

In the dissertation QUANTUM MONTE CARLO. CALCULATIONS OF. ELECTRONIC EXCITATIONS. By. Andrew James Williamson. Robinson College, Cambridge. CHAPTER 4. OPTIMISING TRIAL WAVEFUNCTIONS It was stated that

The quality of the trial/guiding wave function does not directly affect the final DMC estimate of the total energy of a given system (apart from the fixed node approximation). However, the intrinsic variance of VMC energy determines the variance of the estimate of the total energy at each step of the diffusion process. Therefore, as in the VMC technique, the number of DMC moves required to achieve a specific variance of the mean, decreases linearly with the intrinsic variance of VMC energy.


So it is better use variance optimization of JASTROW wave function?

Best, Vladimir.
In Soviet Russia Casino play you.
Vladimir_Konjkov
 
Posts: 78
Joined: Wed Apr 15, 2015 3:14 pm

Re: trial/guiding wave function optimisation for DMC

Postby Ryan_Hunt » Wed Jul 19, 2017 6:51 am

Vladimir,

The quoted variance-related argument is true. However, DMC doesn't just begin and end, sampling the equilibrium distribution the whole time. That's why we have an equilibration period. If one uses an energy minimised wave function, it makes sense that this equilibration period is shorter (there are more low-energy contributions to the sum c_i \Psi_i than high energy ones). Energy minimisation maximises the amount of "c_0" in our trial wave function, and lessens the extent of the equilibration period. I would say that a very good wave function from energy minimisation (about the variance minimised starting point) is the best of both worlds - lowering both the length of the equilibration period and, in principle, the DMC variance.

It should be said that if you have highly optimised (and good!) wave functions, you might not really notice this difference very much at all. The energy and variance minima do usually occur at different points in wave function parameter space, however, if you are using unreweighted variance minimisation this difference shouldn't be massive.

Personally, I tend to do as follows:

  • Fix cut-offs, use varmin_linjas to do a first round of optimisation.
  • Run a final few cycles of emin.
  • Feed final (highly optimised, lots of configs, etc.) trial function to DMC.

Occasionally, for tricky systems (and for some of the model systems I work with) I use MADmin. MADmin is slowly climbing my list, and at some point I expect it will be a favourite.

Good luck,
Ryan.
Ryan_Hunt
 
Posts: 22
Joined: Thu Mar 10, 2016 10:58 am
Location: Lancaster, UK

Re: trial/guiding wave function optimisation for DMC

Postby Vladimir_Konjkov » Fri Jul 21, 2017 7:16 am

Hello Ryan.

As stated by J. R. Trail Heavy-tailed random error in quantum Monte Carlo VMC energy variance has Heavy-tailed distribution when using the trial wfn, which does not have an exact nodal surface. This is the usual case.
VMC energy variance, moreover, does not even have an mean value, which leads to non-consistency in the process of finding the optimal JASTROW parameters by MADmin. This is especially noticeable if we investigate systems consisting of a heavy atom and one hydrogen atom like HCl or AlH, SiH radicals.

So MADmin is no longer the preferred method for me.

Best, Vladimir.
In Soviet Russia Casino play you.
Vladimir_Konjkov
 
Posts: 78
Joined: Wed Apr 15, 2015 3:14 pm

Re: trial/guiding wave function optimisation for DMC

Postby Ryan_Hunt » Fri Jul 21, 2017 7:36 am

Vladimir,

I think you are correct, however, it is nearly always the case that after an initial parameter optimisation (varmin/madmin) one would want to (before DMC) run a few cycles of Emin.

I'm not suggesting to use either variance minimisation or mad minimisation when forming your final DMC trial function.

Out of ingerest: Have you had any problems using madman in practice?

Regards,
Ryan.
Ryan_Hunt
 
Posts: 22
Joined: Thu Mar 10, 2016 10:58 am
Location: Lancaster, UK

Re: trial/guiding wave function optimisation for DMC

Postby Vladimir_Konjkov » Sat Jul 22, 2017 9:48 am

Ryan_Hunt wrote:Vladimir,

I think you are correct, however, it is nearly always the case that after an initial parameter optimisation (varmin/madmin) one would want to (before DMC) run a few cycles of Emin.

I'm not suggesting to use either variance minimisation or mad minimisation when forming your final DMC trial function.

Out of ingerest: Have you had any problems using madman in practice?

Regards,
Ryan.


Ryan, since the most difficult part of JASTROW optimization is a fitting of cut-offs values in case of pure MADmin optimization I often observed how the cut-off to e-e term decreases to unphysically small value <1 au, which in turn led to an unphysical value of energy.
Default limits for cut-offs are quite broad and if you narrow them, then MADmin works quite stably.

Another problem is to find such a way of JASTROW optimization, functional bases and cut-off functions that would give the minimum DMC stderr for a given number of DMC STATISTIC-ACCUMULATION steps.
In this case the pure MADmin is not the best choice. I have some speculation about this, but no more.

Vladimir.
In Soviet Russia Casino play you.
Vladimir_Konjkov
 
Posts: 78
Joined: Wed Apr 15, 2015 3:14 pm

Re: trial/guiding wave function optimisation for DMC

Postby Neil Drummond » Sat Jul 29, 2017 6:44 pm

Dear Vladimir,

Optimisation of the trial wave function by energy minimisation theoretically makes DMC more efficient than any other optimisation method. See David Ceperley's argument in J. Stat. Phys 43, 815 (1986).

Best wishes,

Neil.
Neil Drummond
 
Posts: 68
Joined: Fri May 31, 2013 10:42 am
Location: Lancaster

Re: trial/guiding wave function optimisation for DMC

Postby Vladimir_Konjkov » Sun Jul 30, 2017 8:42 am

Neil Drummond wrote:Dear Vladimir,

Optimisation of the trial wave function by energy minimisation theoretically makes DMC more efficient than any other optimisation method. See David Ceperley's argument in J. Stat. Phys 43, 815 (1986).

Best wishes,

Neil.


Hello Neil.
Thanks, this is a very valuable article. I can say more that VMC variance is equal to DMC uncorrelated data variance.
VMC variance was taken from JASTROW optimized VMC calculations from Sample variance
DMC variance is stderr^2 * walkers * steps / N_corr
Even for different ways of JASTROW optimization the dependence is the same.

On the basis of this article I will try to write down the analytical dependence for N_corr, and plot a similar graph for it. I think that in the linear approximation it is not very complicated.

UPD.
For the eq. (15) form the article I've got:
V = 1.71+/- 0.07*Ev/(tP) for madmin
V = 1.73+/- 0.03*Ev/(tP) for emin

Best wishes,

Vladimir.
Attachments
variance.png
variance.png (21.55 KiB) Viewed 216 times
In Soviet Russia Casino play you.
Vladimir_Konjkov
 
Posts: 78
Joined: Wed Apr 15, 2015 3:14 pm


Return to Quantum Monte Carlo

Who is online

Users browsing this forum: No registered users and 1 guest

cron