Merge pull request #597 from joergbuchwald/loguniform

loguniform scaling

docs/user_guide/advanced.rst

@@ -102,6 +102,7 @@ parameter bounds but sample-specific metadata.
 
 * unif: uniform distribution
     e.g. :code:`[-np.pi, np.pi]` defines the lower and upper bounds
+* logunif: logarithmic uniform with lower and upper bounds
 * triang: triangular with lower and upper bounds, as well as
      location of peak
      The location of peak is in percentage of width
@@ -110,6 +111,7 @@ parameter bounds but sample-specific metadata.
      A soon-to-be deprecated two-value format assumes the lower bound to be 0
      e.g. :code:`[3, 0.5]` assumes 0 to 3, with a peak at 1.5
 * norm: normal distribution with mean and standard deviation
+* truncnorm: truncated normal distribution with upper and lower bounds, mean and standard deviation
 * lognorm: lognormal with ln-space mean and standard deviation
 
 

src/SALib/util/__init__.py

@@ -134,6 +134,7 @@ def _nonuniform_scale_samples(params, bounds, dists):
     dists : list
         list of distributions, one for each parameter
             unif: uniform with lower and upper bounds
+            logunif: logarithmic uniform with lower and upper bounds
             triang: triangular with lower and upper bounds, as well as
                     location of peak
                     The location of peak is in percentage of width
@@ -205,6 +206,9 @@ def _nonuniform_scale_samples(params, bounds, dists):
             else:
                 conv_params[:, i] = params[:, i] * (b2 - b1) + b1
 
+        elif dists[i] == "logunif":
+            conv_params[:, i] = sp.stats.loguniform.ppf(params[:, i], a=b1, b=b2)
+
         elif dists[i] == "norm":
             if b2 <= 0:
                 raise ValueError("""Normal distribution: stdev must be > 0""")

tests/test_util.py

@@ -2,6 +2,7 @@
 from numpy.testing import assert_equal, assert_allclose
 
 import numpy as np
+from scipy import stats
 import pytest
 
 from SALib.util import (
@@ -203,3 +204,23 @@ def test_nonuniform_scale_samples_truncnorm():
         ]
     )
     np.testing.assert_allclose(actual, expected)
+
+
+def test_nonuniform_scale_samples_logunif():
+    """
+    Test the rescaling of samples for logarithmic uniform  distribution
+    """
+    problem = {
+        "num_vars": 1,
+        "dists": ["logunif"],
+        "bounds": [[1e-22, 1e-18]],
+        "names": ["x1"],
+    }
+    actual = latin.sample(problem, 10000)
+    rng = np.random.default_rng()
+    x = stats.loguniform.rvs(size=100000, a=1e-22, b=1e-18, random_state=rng)
+    res = stats.cramervonmises_2samp(actual[:, 0], x)
+    zero_hypothesis_rejected = True
+    if res.pvalue > 0.05:
+        zero_hypothesis_rejected = False
+    assert zero_hypothesis_rejected is False