1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
|
.. _orthogonalization:
==================================
Orthogonalizing the wave functions
==================================
Let `\tilde{\Psi}_{nG}` be an element of a wave function matrix
holding the value of `\tilde{\psi}_{n}(\mathbf{r}_G)` (state number
`n` and grid point number `G`). Then we can write the
:ref:`orthogonality requirement <orthogonality>` like this:
.. math::
\Delta v
\tilde{\mathbf{\Psi}}^T \hat{\mathbf{O}} \tilde{\mathbf{\Psi}} =
\mathbf{1},
where `\Delta v` is the volume per grid point and
.. math::
\hat{\mathbf{O}} = \mathbf{1} +
\sum_a \tilde{\mathbf{P}}^a \mathbf{\Delta O}^a
(\tilde{\mathbf{P}}^a)^T
\Delta v
is the matrix form of the overlap operator. This matrix is very
sparse because the projector functions `\tilde{P}^a_{iG} =
\tilde{p}^a_i(\mathbf{r}_G - \mathbf{R}^a)` are localized inside the
augmentation spheres. The `\Delta O^a_{i_1i_2}` atomic PAW overlap
corrections are small `N_p^a \times N_p^a` matrices (`N_p^a \sim 10`)
defined :ref:`here <overlaps>`.
Gram-Schmidt procedure
======================
The traditional sequential Gram-Schmidt orthogonalization procedure is
not very efficient, so we do some linear algebra to allow us to use
efficient matrix-matrix products.
Let `\tilde{\mathbf{\Psi}}_0` be the non-orthogonal wave functions.
We calculate the overlap matrix:
.. math::
\mathbf{S} =
\Delta v
\tilde{\mathbf{\Psi}}_0^T \hat{\mathbf{O}} \tilde{\mathbf{\Psi}}_0,
from the raw overlap `\tilde{\mathbf{\Psi}}_0^T
\tilde{\mathbf{\Psi}}_0` and the projections `(\tilde{\mathbf{P}}^a)^T
\tilde{\mathbf{\Psi}}_0`.
This can be Cholesky factored into `\mathbf{S} = \mathbf{L}^T
\mathbf{L}` and we can get the orthogonalized wave functions as:
.. math::
\tilde{\mathbf{\Psi}} = \tilde{\mathbf{\Psi}}_0 \mathbf{L}^{-1}.
Parallelization
===============
The orthogonalization can be paralleized over **k**-points, spins,
domains, and bands.
**k**-points and spins
----------------------
Each **k**-point and each spin can be treated separately.
Domains
-------
Each domain will have its contribution to the overlap matrix, and these
will have to be summed up using the domain communicator. The dense
linear algebra can be performed in a replication fashion on all MPI
tasks using LAPACK or in parallel on a subset of MPI tasks using ScaLAPACK.
Bands
-----
Band parallelization is described at :ref:`Band parallelization <band_parallelization>`.
|
