Get my PhD Thesis
Get my PhD Thesis
Get my PhD Thesis
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
074103-12 Thøgersen et al. J. Chem. Phys. 123, 074103 2005<br />
and accept a favorable configuration shift. A configuration<br />
shift may be recognized when an a orb min profile has an<br />
abrupt change where on the right-hand side a orb min is close to 1<br />
and on the left-hand side a orb min is close to 0. To maintain the<br />
high degree of control characteristic of the TRSCF method,<br />
the energy of the new configuration is checked before the<br />
shift is accepted, at the cost of an additional Kohn–Sham<br />
matrix build. As seen from Fig. 4a, this check is well worth<br />
the effort, saving more than ten iterations, and thus it is made<br />
an integrated part of our TRSCF implementation.<br />
VI. THE DIIS METHOD VIEWED<br />
AS A QUASI-NEWTON METHOD<br />
Since its introduction by Pulay in 1980, the DIIS method<br />
has been extensively and successfully used to accelerate the<br />
convergence of SCF optimizations. We here present a rederivation<br />
of the DIIS method to demonstrate that, in the iterative<br />
subspace of density matrices, it is equivalent to a quasi-<br />
Newton method. From this observation, we conclude that, in<br />
the local region of the SCF optimization, the DIIS steps can<br />
be used safely and will lead to fast convergence. The convergence<br />
of the DIIS algorithm in the global region is also<br />
discussed and is much more unpredictable.<br />
We assume that, in the course of the SCF optimization,<br />
we have determined a set of n+1 AO density matrices<br />
D 0 ,D 1 ,D 2 ,...,D n and the associated Kohn–Sham or Fock<br />
matrices FD 0 ,FD 1 ,FD 2 ,...,FD n . Since the electronic<br />
gradient gD is given by 11<br />
gD =4SDFD − FDDS,<br />
79<br />
we also have available the corresponding gradients<br />
gD 0 ,gD 1 ,gD 2 ,...,gD n . We now wish to determine a<br />
corrected density matrix,<br />
n<br />
D¯ = D 0 + c i D i0 , D i0 = D i − D 0 , 80<br />
i=1<br />
that minimizes the norm of the gradient gD¯ . For this purpose,<br />
we parameterize the density matrix in terms of an antisymmetric<br />
matrix X=−X T and the current density matrix<br />
D 0 as 11<br />
DX = exp− XSD 0 expSX.<br />
81<br />
With each old density matrix D i , we now associate an antisymmetric<br />
matrix X i such that<br />
D i = exp− X i SD 0 expSX i = D 0 + D 0 ,X i S + OX 2 i .<br />
82<br />
Introducing the averaged antisymmetric matrix,<br />
n<br />
X¯ = c i X i ,<br />
i=1<br />
we obtain<br />
83<br />
n<br />
DX¯ = D 0 + c i D 0 ,X i S + OX¯ 2 ,<br />
i=1<br />
84<br />
where we have used the S-commutator expansion of DX¯ <br />
analogeous to Eq. 82. Our task is hence to determine X¯ in<br />
Eq. 83 such that DX¯ minimizes the gradient norm<br />
gDX¯ . In passing, we note that, whereas D¯ is not in<br />
general idempotent and therefore not a valid density matrix,<br />
DX¯ is a valid, idempotent density matrix for all choices of<br />
c i .<br />
Expanding the gradient in Eq. 79 about the currentdensity<br />
matrix D 0 , we obtain<br />
gDX¯ = gD 0 + HD 0 X¯ + OX¯ 2 ,<br />
85<br />
where HD is the Jacobian matrix. Neglecting the higherorder<br />
terms, our task is therefore to minimize the norm of the<br />
gradient,<br />
n<br />
gc = gD 0 + c i HD 0 X i ,<br />
86<br />
i=1<br />
with respect to the elements of c. For an estimate of<br />
HD 0 X i , we truncate the expansion,<br />
gD i = gD 0 + HD 0 X i + OX i 2 ,<br />
and obtain the quasi-Newton condition,<br />
gD i − gD 0 = HD 0 X i .<br />
Inserting this condition into Eq. 86, we obtain<br />
n<br />
gc = gD 0 + <br />
i=1<br />
n<br />
c i gD i − gD 0 = c i gD i ,<br />
i=0<br />
87<br />
88<br />
89<br />
where we have introduced the parameter c 0 =1− n<br />
i=1 c i . The<br />
minimization of gc=gc may therefore be carried out as<br />
a least-squares minimization of gc in Eq. 89 subject to the<br />
constraint<br />
n<br />
c i =1.<br />
90<br />
i=0<br />
If we consider gD i as an error vector for the density matrix<br />
D i , this procedure becomes identical to the DIIS method.<br />
From Eq. 86 we also see that DIIS may be viewed as a<br />
minimization of the residual for the Newton equation in the<br />
subspace of the density matrix differences D i −D 0 , i=1, n,<br />
where the quasi-Newton condition is used to set up the subspace<br />
equations. Since the quasi-Newton steps are reliable<br />
only in the local region of the optimization, we conclude that<br />
the DIIS method can be used safely only in this region, when<br />
the electronic Hessian is positive definite.<br />
The optimal combination of the density matrices is obtained<br />
in the DIIS method, by carrying out a least-squares<br />
minimization of the gradient norm subject to the constraint in<br />
Eq. 90. However, since a small gradient norm in the global<br />
region does not necessarily imply a low Kohn–Sham energy,<br />
the DIIS convergence may be unpredictable. Furthermore,<br />
we may encounter regions where the gradient norms are<br />
Downloaded 23 Aug 2005 to 130.225.22.254. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp