CASINO manual - Theory of Condensed Matter
CASINO manual - Theory of Condensed Matter
CASINO manual - Theory of Condensed Matter
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
• Is emin min energy too high? This may be the case if casino rejects what appear to be<br />
perfectly good energies.<br />
• Is energy minimization being used to optimize cut<strong>of</strong>f parameters? It is recommended to fix them<br />
at some point to allow the rest <strong>of</strong> the parameters to fully converge.<br />
26 Alternative sampling strategies<br />
26.1 Summary<br />
In VMC we directly use Monte Carlo sampling to estimate expectation values defined as the quotient<br />
<strong>of</strong> two integrals over 3N dimensional space, such as the total energy expectation value,<br />
E tot = 〈Ψ|Ĥ|Ψ〉/〈Ψ|Ψ〉. Estimation <strong>of</strong> integrals using Monte Carlo methods can generally be performed<br />
using samples drawn from any <strong>of</strong> a wide range <strong>of</strong> distributions. The usual VMC choice <strong>of</strong><br />
P (R) = λ|Ψ 2 (R)| (with λ an unknown and un-required normalization prefactor) is an analytically<br />
convenient choice, but is not in any sense optimal [67]. This leaves open the possibility <strong>of</strong> choosing a<br />
different sampling distribution, such as one that is optimal in the sense <strong>of</strong> providing the minimum statistical<br />
error for a given number <strong>of</strong> samples, or one that is maximally efficient in the sense <strong>of</strong> providing<br />
a given accuracy for less computational cost than other choices.<br />
Unfortunately there are some strong limitations on the distribution functions that we can use. Firstly,<br />
we require the ‘estimate’ to be an estimate, that is, to converge to the true value with increasing sample<br />
size due to the law <strong>of</strong> large numbers (LLN) being valid. Secondly, we usually require meaningful error<br />
bars to be available for a finite sample size estimate, that is that the central limit theorem (CLT)<br />
is valid in some form. The validity <strong>of</strong> either <strong>of</strong> these conditions must be arrived at analytically, and<br />
if they are not true then the conventional formulae for the sample mean and standard error provide<br />
numerical results that are unrelated to the estimate and error. Examples <strong>of</strong> when this occurs are<br />
provided by force estimates [68] and parameter optimization [69].<br />
In what follows, estimates constructed using an arbitrary distribution are described. Following this,<br />
three alternative choices <strong>of</strong> sampling distribution are summarized. The usual VMC choice is referred<br />
to as ‘standard sampling’, whereas the alternative methods are referred to as ‘alternative sampling’.<br />
26.2 Alternative sampling<br />
For estimating the Hamiltonian expectation value using an arbitrary sampling distribution, P (R), the<br />
estimate is composed <strong>of</strong> the quotient <strong>of</strong> two sample means<br />
∑ w(R)EL (R)<br />
Ē tot = ∑ , (230)<br />
w(R)<br />
where w(R) = |Ψ 2 |/P (R). In the limit <strong>of</strong> large sample size, this is a value drawn from a normal<br />
distribution with mean E tot and variance<br />
σE 2 = 1 ∫<br />
|Ψ 2 (R)|/w(R)d 3 R ∫ w(R)|Ψ 2 (R)| 2 (E L (R) − E tot ) 2 d 3 R<br />
[∫<br />
N<br />
|Ψ2 (R)| 2 d 3 R ] 2<br />
, (231)<br />
whose estimate is<br />
σ¯<br />
E 2 = N<br />
∑ ( w(R)<br />
2<br />
E L (R) − Ētot) 2<br />
N − 1 [ ∑ w(R)] 2 (232)<br />
as long as the bivariate CLT is valid. For standard sampling as described below, the random error is<br />
normal for energy estimates, but is not normal for most other estimates such as the energy surface<br />
implicit in optimization. For the three alternative sampling strategies, the random error is normal for<br />
most estimates, including the energy surface implicit in optimization.<br />
Note that the estimates given above are not the sample mean and standard error <strong>of</strong> any set <strong>of</strong> independent<br />
identically distributed random variables. In the input file this form <strong>of</strong> alternative sampling<br />
is activated by setting the vmc sampling keyword to one <strong>of</strong> the values described below.<br />
163