Improvements to fractional Gaussian noise generation

[anntzer]

Quite a few changes to fGn generation.

• Fix https://github.com/crflynn/stochastic/issues/26, and use a better caching method, just based on lru_cache. If you want to also have this on Py2, I can e.g. add a dependency on https://pypi.org/project/backports.functools-lru-cache/.
• Fix Davies-Harte method: I don't think you need to append a zero to the first row of the covariance matrix (at least this matches the description of the DH paper). This also makes the covariance matrix, well, a real covariance matrix, and thus positive, so there's never a need of falling back on Hosking.
• Speedup Davies-Harte method:
  • Immediately generate random numbers with the right scale, skipping the final multiplication step.
  • Use irfft() instead of appending the reverted array to each array and then only taking the real part (compensate normalization accordingly).
  • Vectorize the eigenvalues times random-complex multiplication.
  • Always generate a power-of-two-plus-one number of items (to use power-of-two length FFTs), then crop to the requested number of points (the covariance matrix is clearly still going to be correct).

[anntzer]

I pushed a commit dropping support for Py27/34. I very quickly skimmed through Dieker's paper, it says "The λj are real by construction and they are non-negative in the fGn case when αN is set to γ(N), see [3]. However, we set αN = 0 for reasons that become clear in Section 3." so it looks like that's only relevant for approximation methods?

[codecov-commenter]

Codecov Report

Merging #27 into master will not change coverage. The diff coverage is 100.00%.

@@            Coverage Diff            @@
##            master       #27   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files           35        35           
  Lines         1222      1286   +64     
=========================================
+ Hits          1222      1286   +64     

Impacted Files	Coverage Δ
stochastic/continuous/brownian_motion.py	100.00% <100.00%> (ø)
stochastic/noise/fractional_gaussian_noise.py	100.00% <100.00%> (ø)
stochastic/base.py	100.00% <0.00%> (ø)
stochastic/discrete/moran.py	100.00% <0.00%> (ø)
stochastic/continuous/gamma.py	100.00% <0.00%> (ø)
stochastic/continuous/bessel.py	100.00% <0.00%> (ø)
stochastic/continuous/poisson.py	100.00% <0.00%> (ø)
stochastic/discrete/bernoulli.py	100.00% <0.00%> (ø)
stochastic/noise/colored_noise.py	100.00% <0.00%> (ø)
stochastic/discrete/random_walk.py	100.00% <0.00%> (ø)
... and 12 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update ddcb0b2...b620af9. Read the comment docs.