06.04.2013 Views

Monte Carlo Methods in Statistical Mechanics: Foundations and ...

Monte Carlo Methods in Statistical Mechanics: Foundations and ...

Monte Carlo Methods in Statistical Mechanics: Foundations and ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Hamiltonian Hl;1 has the same functional form as the \ ne-grid" Hamiltonian Hl: itis<br />

speci ed by the coe cients 0 , 0 , 0 y, A 0 y <strong>and</strong> h 0 y. Thestep \compute Hl;1" therefore<br />

means to compute these coe cients. Note alsothe importance of allow<strong>in</strong>g <strong>in</strong> (5.23)<br />

for ' 3 <strong>and</strong> ' terms <strong>and</strong> for site-dependent coe cients: even if these are not present <strong>in</strong><br />

the orig<strong>in</strong>al Hamiltonian H HM, they will be generated on coarser grids. F<strong>in</strong>ally,<br />

we emphasize that the coarse-grid Hamiltonian Hl;1 depends implicitly on the current<br />

value of the ne-lattice eld ' 2 Ul although our notation suppresses this dependence,<br />

it should be kept <strong>in</strong> m<strong>in</strong>d.<br />

Basic (smooth<strong>in</strong>g) iterations. We have already discussed the damped Jacobi iteration<br />

as one possible smoother. Note that <strong>in</strong>this method only the \old" values ' (n) are<br />

used on the right-h<strong>and</strong> side of (5.9)/(5.10), even though for someofthe terms the \new"<br />

value ' (n+1) may already have been computed. An alternative algorithm is to useat<br />

each stage on the right-h<strong>and</strong> sidethe \newest" available value. This algorithm is called<br />

the Gauss-Seidel iteration. 17 Note that the Gauss-Seidel algorithm, unlike the Jacobi<br />

algorithm, depends on the order<strong>in</strong>g of the grid po<strong>in</strong>ts. For example, if a 2-dimensional<br />

grid is swept <strong>in</strong> lexicographic order (1 1), (2 1), :::,(L 1), (1 2), (2 2), :::,(L 2), :::,<br />

(1L), (2L), :::,(L L), then the Gauss-Seidel iteration becomes<br />

' (n+1)<br />

x1x2<br />

1<br />

=<br />

4 ['(n) x1+1x2 + '(n+1) x1;1x2 + '(n) x1x2+1 + ' (n+1)<br />

x1x2;1 + fx1x2]: (5.32)<br />

Another convenient order<strong>in</strong>g isthe red-black (or checkerboard) order<strong>in</strong>g, <strong>in</strong>whichthe<br />

\red" sublattice r = fx 2 : x1 + + xd is eveng is swept rst, followed by the<br />

\black" sublattice b = fx 2 : x1 + + xd is oddg. Notethat the order<strong>in</strong>g ofthe<br />

grid po<strong>in</strong>ts with<strong>in</strong> each sublattice is irrelevant [forthe usual nearest-neighbor Laplacian<br />

(5.1)], s<strong>in</strong>ce the matrix A does not couple sites of the same color. This means that redblack<br />

Gauss-Seidel is particularly well suited to vector or parallel computation. Note<br />

that the red-black order<strong>in</strong>g makes sense with periodic boundary conditions only if the<br />

l<strong>in</strong>ear size Ll of the grid l is even.<br />

It turns out that Gauss-Seidel is a better smoother than damped Jacobi (even if the<br />

latter is given its optimal !). Moreover, Gauss-Seidel is easier to program <strong>and</strong> requires<br />

only half the storage space (no need for separate storage of \old" <strong>and</strong> \new" values).<br />

The only reason we <strong>in</strong>troduced damped Jacobi at allisthat itiseasiertounderst<strong>and</strong><br />

<strong>and</strong> toanalyze.<br />

Many other smooth<strong>in</strong>g iterations can be considered, <strong>and</strong> can be advantageous <strong>in</strong><br />

anisotropic or otherwise s<strong>in</strong>gular problems [32, Section 3.3 <strong>and</strong> Chapters 10{11]. But<br />

we shall stick toord<strong>in</strong>ary Gauss-Seidel, usually with red-black order<strong>in</strong>g.<br />

will consist, respectively,ofm1 <strong>and</strong> m2 iterations of the Gauss-<br />

Thus, S pre<br />

l<br />

<strong>and</strong> S post<br />

l<br />

Seidel algorithm. The balance between pre-smooth<strong>in</strong>g <strong>and</strong> post-smooth<strong>in</strong>g is usually<br />

not very crucial only the total m1 + m2 seems to matter much. Indeed, one (butnot<br />

17 It is amus<strong>in</strong>g to note that \Gauss did not use a cyclic order of relaxation, <strong>and</strong> :::Seidel speci cally<br />

recommended aga<strong>in</strong>st us<strong>in</strong>g it" [36, p.44n]. See also [37].<br />

33

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

Saved successfully!

Ooh no, something went wrong!