fix a few bits of pid stuff

dit/pid/iccs.py

@@ -67,21 +67,26 @@ def pmi(input_, output):
 
     pmis = {tuple(marg): pmi(marg[0], marg[1]) for marg in marginals[1:]}
 
+    inputs_dist = sub_dists[tuple(vars[:-1])]
+    output_dist = sub_dists[(vars[-1],)]
+    joint_pmis = {e: np.log2(d[e]/(inputs_dist[e[:-1]]*output_dist[(e[-1],)])) for e in d.outcomes}
+
+    coinfos = {e: np.log2(np.prod(
+        [sub_dists[sub_var][tuple(e[i] for i in flatten(sub_var))] ** ((-1) ** len(sub_var)) for sub_var in sub_vars]))
+               for e in d.outcomes}
+
     # fix the sign of things close to zero
     for pmi in pmis.values():
         for e, val in pmi.items():
             if np.isclose(val, 0.0):
                 pmi[e] = 0.0
-
-    coinfos = {e: np.log2(np.prod(
-        [sub_dists[sub_var][tuple(e[i] for i in flatten(sub_var))] ** ((-1) ** len(sub_var)) for sub_var in sub_vars]))
-               for e in d.outcomes}
     for e, val in coinfos.items():
         if np.isclose(val, 0.0):
             coinfos[e] = 0.0
 
     i = sum(d[e] * coinfos[e] for e in d.outcomes if
-            all(np.sign(coinfos[e]) == np.sign(pmi[tuple(e[i] for i in marg)]) for marg, pmi in pmis.items()))
+            all(np.sign(coinfos[e]) == np.sign(pmi[tuple(e[i] for i in marg)]) for marg, pmi in pmis.items()) \
+            and np.sign(coinfos[e]) == np.sign(joint_pmis[e]))
 
     return i
 

docs/measures/pid.rst

@@ -371,10 +371,10 @@ This decomposition also displays an interesting phenomena, that of *subadditive
    ╚════════╧════════╧════════╝
 
 .. py:module:: dit.pid.idep
-:math:`\I_{dep}`
-----------------
+:math:`\Idep{\bullet}`
+----------------------
 
-James et al have developed a method of quantifying unique information based on the :ref:`Dependency Decomposition`. Unique information from variable :math:`X_i` is evaluated as the least change in sources-target mutual information when adding the constraint :math:`X_i Y`.
+James et al :cite:`james2017unique` have developed a method of quantifying unique information based on the :ref:`Dependency Decomposition`. Unique information from variable :math:`X_i` is evaluated as the least change in sources-target mutual information when adding the constraint :math:`X_i Y`.
 
 .. ipython::
 
@@ -393,7 +393,7 @@ James et al have developed a method of quantifying unique information based on t
 :math:`\Ipm{\bullet}`
 ---------------------
 
-Also taking a pointwise view, Finn & Lizier's :math:`\Ipm{\bullet}` instead splits the pointwise mutual information into two components:
+Also taking a pointwise view, Finn & Lizier's :math:`\Ipm{\bullet}` :cite:`finn2017` instead splits the pointwise mutual information into two components:
 
 .. math::
 

docs/references.bib

@@ -379,6 +379,13 @@ @incollection{griffith2014quantifying
   publisher={Springer}
 }
 
+@article{banerjee2017computing,
+  title={Computing the Unique Information},
+  author={Banerjee, Pradeep Kr. and Rauh, Johannes and Mont{\'u}far, Guido},
+  journal={arXive preprint arXiv:1709.07487},
+  year={2017}
+}
+
 % i_wedge
 @article{griffith2014intersection,
   title={Intersection information based on common randomness},
@@ -403,6 +410,22 @@ @article{ince2017measuring
   publisher={Multidisciplinary Digital Publishing Institute}
 }
 
+% i_dep
+@article{james2017unique,
+  title={Unique Information via Dependency Constraints},
+  author={James, Ryan G. and Emenheiser, Jeffrey and Crutchfield, James P.},
+  journal={arXiv preprint arXiv:1709.06653},
+  year={2017}
+}
+
+% i_pm
+@article{finn2017,
+  title={},
+  author={Finn, Conor and Lizier, Joseph},
+  journal={},
+  year={2017}
+}
+
 %% supporting
 
 @inproceedings{bertschinger2013shared,