Efficient random sampling via linear interpolation.
When you have a density function, but you would like to create a set of sample points from that density function, you can use linear interpolate sampling. Using the evaluation of the density at the two endpoints of 1D interval, or the four corners of a 2D rectangle, or generally the lintsampler provides a Python implementation of this.
See the documentation for further details.
This package was also reviewed for the Journal of Open Source Software (JOSS). See the paper here and the review thread here.
Three ways of installing lintsampler:
pip:
pip install lintsampler
conda:
conda install -c conda-forge lintsampler
- Simply cloning this repository.
If you have a density function, such as this multi-modal 1d pdf with the bulk of the density between -7 and 7,
import numpy as np
from scipy.stats import norm
def gmm_pdf(x):
mu = np.array([-3.0, 0.5, 2.5])
sig = np.array([1.0, 0.25, 0.75])
w = np.array([0.4, 0.25, 0.35])
return np.sum([w[i] * norm.pdf(x, mu[i], sig[i]) for i in range(3)], axis=0)lintsampler can efficiently draw samples from it on some defined interval (here a 100-point grid between -7 and 7):
from lintsampler import LintSampler
grid = np.linspace(-7,7,100)
samples = LintSampler(grid,pdf=gmm_pdf).sample(N=10000)Making a histogram of the resulting samples and comparing to the input density function shows good agreement -- and we can do even better by increasing the resolution.
See this page of the documentation for a more detailed explanation of this example.
Complete documentation, including more example notebooks, is available at lintsampler.readthedocs.io/.
The lintsampler maintainers welcome contributions to software, examples, and documentation. The maintainers are actively monitoring pull requests and would be happy to collaborate on contributions or ideas. If you have any requests for additional information or find any bugs, please open an issue directly.
If using lintsampler for a research publication, please cite our paper.
lintsampler is available under the MIT license. See the LICENSE file for specifics.