1. 18 Jul, 2018 2 commits
2. 11 Dec, 2016 1 commit
3. 08 Dec, 2016 1 commit
• Licensing: remove Exhibit B (MPL 2.0) · d71f8b3a
```What I wanted to achieve by enacting Exhibit B of the Mozilla Public
License 2.0 is to prevent someone from just re-licensing the code under
The MPL 2.0 does not allow such a blatant re-licensing attempt
(I guess this is a consequence of Article 3.4) and this is explicitly
said in the MPL 2.0 FAQ, Q14, paragraph 2:
https://www.mozilla.org/en-US/MPL/2.0/FAQ/#mpl-and-lgpl```
4. 03 Dec, 2016 2 commits
5. 01 Dec, 2016 3 commits
6. 30 Nov, 2016 3 commits
• Src: manually provide a matrix decomposition · 9487615b
```Preserving the symmetry fixes the out-of-memory situations [on my
computer] with the test matrix bmwcra_1.```
• Src: subspace iteration for small search spaces · 7776693d
```This commit replaces Chebyshev polynomials with inverse iteration to
avoid accidentally wiping out a perfectly good approximate eigenvector
(cf. commit 33c1, Nov 28, 2016).

Inverse iteration updates only the lower part of the spectrum quite well
so the idea is to compute eigenpairs with Rayleigh-Ritz early and then
to improve eigenpairs selectively. The two major problems tackled in
this commit are missing eigenpairs and the forward error; the backward
error is usually not a problem, e.g., on vanbody, the backward error of
all eigenpairs after one inverse iteration is below the single-precision
epsilon 1.1*10^-7. We ensure finding all desired eigenpairs by applying
shift-and-invert to eigenpairs with eigenvalues above the cutoff. A
large forward error can have two causes:
- a large condition number, or
- the use of shifted matrices causing a large residual.
Hence, to reduce the forward error we must use inverse iteration without
shifts, especially after executing shift-and-invert.

With this commit, we find all 1265 eigenpairs below the cutoff 20 on
vanbody in 993 seconds (1780 seconds CPU time). This is a significant
improvement over the previous record with commit 3dce, July 27, 2016
(1063 seconds wall-clock time, 1890 second CPU time). Additionally, the
search space dimension decreased from 2064 to 1453.```
7. 28 Nov, 2016 6 commits
8. 27 Nov, 2016 6 commits
9. 26 Nov, 2016 7 commits
10. 25 Nov, 2016 1 commit
11. 23 Nov, 2016 2 commits
12. 22 Nov, 2016 6 commits