14.09.2014 Views

CASINO manual - Theory of Condensed Matter

CASINO manual - Theory of Condensed Matter

CASINO manual - Theory of Condensed Matter

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

25.3.5 Matrix manipulation (level-shifting)<br />

We can rewrite Eq. (210) as a simple eigenproblem:<br />

(S + S T ) −1 (H + H T )a = Ea. (227)<br />

If the (S + S T ) −1 (H + H T ) matrix had the form<br />

⎛<br />

⎞<br />

λ 0 0 0 0<br />

0 λ 1 0 0<br />

⎜<br />

⎝<br />

.<br />

0 0 ..<br />

⎟<br />

0 ⎠ , λ 0 < λ i ∀ i = 1 . . . p, (228)<br />

0 0 0 λ p<br />

then there would be no change at all in the variable parameters α. It is possible to bring any<br />

(S + S T ) −1 (H + H T ) matrix qualitatively closer to this ‘no-change’ condition using ‘level-shifting’<br />

[65, 66], in which a positive real constant L is added to all the diagonal elements <strong>of</strong> the matrix except<br />

the first:<br />

⎛<br />

⎞ ⎛<br />

⎞<br />

λ 0 • • • λ 0 • • •<br />

• λ 1 • •<br />

⎜<br />

⎝<br />

.<br />

• • ..<br />

⎟<br />

• ⎠ → • λ 1 + L • •<br />

⎜<br />

⎝<br />

.<br />

• • ..<br />

⎟<br />

• ⎠ . (229)<br />

• • • λ p • • • λ p + L<br />

The approximation in Eq. (208) fails when the parameter changes are too large. Level-shifting can<br />

be used to reduce the size <strong>of</strong> the parameter changes, stabilizing the optimization. In casino, the<br />

best value for the constant L is determined automatically by performing a line minimization <strong>of</strong> the<br />

VMC energy estimate with respect to L. (Where possible, correlated sampling is used, both to reduce<br />

computational effort and to increase the statistical precision <strong>of</strong> energy differences.) As noted by<br />

Umrigar et al. [60, 61], level-shifting can equally well be applied to the H + H T matrix.<br />

25.3.6 Using energy minimization<br />

Energy minimization is used very similarly to standard variance minimization. After selecting it by<br />

setting opt method to ‘emin’, the user must choose the number <strong>of</strong> optimization cycles (opt cycles),<br />

and the number <strong>of</strong> optimization configurations per cycle (vmc nconfig write). As for variance<br />

minimization, vmc decorr period should be chosen so as to generate approximately uncorrelated<br />

configurations, and the number <strong>of</strong> configurations required is usually similar to the number required<br />

for variance minimization.<br />

The keyword emin min energy is unique to energy minimization. During level-shifting (described<br />

above), some <strong>of</strong> the candidate wave functions may be so poor that they cause numerical errors resulting<br />

in an energy estimate which appears to be very low, but is actually spurious. To protect against<br />

this, any energy estimate which is lower than emin min energy is rejected as invalid. During<br />

diagonalization, eigenvalues are also thrown away if they are below emin min energy. The user<br />

may set emin min energy to an energy below the ground-state energy if it is know, else casino will<br />

calculate an estimate at each cycle based on the energy and variance <strong>of</strong> the initial set <strong>of</strong> local energies,<br />

which works well in practice.<br />

Two other keywords affect the stabilization procedures (semi-orthogonalization and level-shifting) but<br />

they are mainly intended for the use <strong>of</strong> developers and should rarely, if ever, be changed by users.<br />

emin xi value controls the value <strong>of</strong> ξ in Eq. (223), but the default value <strong>of</strong> 0.5 performs well in all<br />

tests.<br />

As mentioned in Sec. 25.3.3, parameters with respect to which the energy is a shallow function may not<br />

allow energy minimization to converge accurately to the best <strong>of</strong> its precision. Therefore casino fixes<br />

parameters internally flagged as ‘shallow’ for the second half <strong>of</strong> the optimization cycles if using energy<br />

minimization; the precise behaviour can be controlled using the input keyword opt noctf cycles.<br />

Finally, a brief list <strong>of</strong> possibilities to investigate if an energy minimization run appears to diverge or<br />

otherwise behave badly (in no particular order):<br />

• Are enough configurations being used? (Try increasing vmc nconfig write.)<br />

• Is the reported VMC correlation time significantly greater than 1? (If so, increase<br />

vmc decorr period.)<br />

162

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!