Fix densities for new element order and adapt documentation.

README.rst

@@ -63,9 +63,7 @@ continuous whereas the second element is discrete:
 
 .. code-block:: python
 
-    import numpy as np
-    is_continuous = np.full((3), True, dtype=bool)
-    is_continuous[1] = False
+    is_continuous = [True, False, True]
 
 To fit a mixed vine to the samples:
 
@@ -83,15 +81,6 @@ calculate their density and estimate the distribution entropy in units of bits:
     logpdf = vine.logpdf(samples)
     (entropy, standard_error_mean) = vine.entropy(sem_tol=1e-2)
 
-Note that for the canonical vine, the order of elements is important.  Elements
-should be sorted according to the importance of their dependencies to other
-elements where elements with important dependencies to many other elements
-should come first.  A heuristic way to select the order of elements is to
-calculate Kendall's tau between all element pairs
-(see ``scipy.stats.kendalltau``), to obtain a score for each element by summing
-the tau's of the pairs the element occurs in and to sort elements in descending
-order according to their scores.
-
 To manually construct and visualize a simple mixed vine model:
 
 .. code-block:: python

docs/usage.rst

@@ -10,9 +10,7 @@ continuous whereas the second element is discrete:
 
 .. code-block:: python
 
-    import numpy as np
-    is_continuous = np.full((3), True, dtype=bool)
-    is_continuous[1] = False
+    is_continuous = [True, False, True]
 
 To fit a mixed vine to the samples:
 
@@ -21,20 +19,24 @@ To fit a mixed vine to the samples:
     from mixedvine import MixedVine
     vine = MixedVine.fit(samples, is_continuous)
 
-``vine`` is now a ``MixedVine`` object.  To draw samples from the distribution,
-calculate their density and estimate the distribution entropy in units of bits:
+``vine`` is now a ``MixedVine`` object.  Note that for the canonical vine, the
+order of elements is important.  Elements should be sorted according to the
+importance of their dependencies to other elements where elements with important
+dependencies to many other elements should come first.  A heuristic way to
+select the order of elements is to calculate Kendall's tau between all element
+pairs, to obtain a score for each element by summing the taus of the pairs the
+element occurs in and to sort elements in descending order according to their
+scores.  This is what the ``MixedVine.fit`` method does internally by default to
+construct an improved canonical vine tree.  This internal sorting is used to
+construct the vine tree only, so the order of elements is not changed in a user
+visible way.  To prevent this internal sorting, set the ``keep_order`` argument
+to ``True``.
+
+To draw samples from the distribution, calculate their density and estimate the
+distribution entropy in units of bits:
 
 .. code-block:: python
 
     samples = vine.rvs(size=100)
     logpdf = vine.logpdf(samples)
     (entropy, standard_error_mean) = vine.entropy(sem_tol=1e-2)
-
-Note that for the canonical vine, the order of elements is important.  Elements
-should be sorted according to the importance of their dependencies to other
-elements where elements with important dependencies to many other elements
-should come first.  A heuristic way to select the order of elements is to
-calculate Kendall's tau between all element pairs
-(see ``scipy.stats.kendalltau``), to obtain a score for each element by summing
-the tau's of the pairs the element occurs in and to sort elements in descending
-order according to their scores.

mixedvines/mixedvine.py

@@ -37,8 +37,6 @@ class MixedVine(object):
 
     Attributes
     ----------
-    vine_type : string
-        Type of the vine tree.
     root : VineLayer
         The root layer of the vine tree.
 
@@ -56,7 +54,7 @@ class MixedVine(object):
         Sets a particular marginal distribution in the mixed vine tree.
     set_copula(layer_index, copula_index, copula)
         Sets a particular pair copula in the mixed vine tree.
-    fit(samples, is_continuous, vine_type, trunc_level, do_refine)
+    fit(samples, is_continuous, trunc_level, do_refine, keep_order)
         Fits the mixed vine to the given samples.
     '''
 
@@ -228,7 +226,7 @@ def _marginal_densities(self, samples):
                     old_settings = np.seterr(divide='ignore')
                     logp[:, k] = np.log(np.maximum(0, cdfp[:, k] - cdfm[:, k]))
                     np.seterr(**old_settings)
-            logpdf = logp[:, 0]
+            logpdf = logp[:, self.output_layer.input_indices[0][0]]
             dout = {'logpdf': logpdf, 'logp': logp, 'cdfp': cdfp, 'cdfm': cdfm,
                     'is_continuous': is_continuous}
             return dout
@@ -783,7 +781,7 @@ def _heuristic_element_order(samples):
         Finds an order of elements that heuristically facilitates vine
         modelling.  For this purpose, Kendall's tau is calculated between
         samples of pairs of elements and elements are scored according to the
-        sum of absolute Kendall's tau of pairs the elements appear in.
+        sum of absolute Kendall's taus of pairs the elements appear in.
 
         Parameters
         ----------
@@ -802,8 +800,8 @@ def _heuristic_element_order(samples):
         for i in range(1, dim):
             for j in range(i):
                 tau, _ = kendalltau(samples[:, i], samples[:, j])
-                score[i] += tau
-                score[j] += tau
+                score[i] += np.abs(tau)
+                score[j] += np.abs(tau)
         # Get order indices for descending score
         order = score.argsort()[::-1]
         return order