Revert Reff

avehtari · Oct 31, 2017 · cac66c0 · cac66c0
1 parent bcf75ef
commit cac66c0
Show file tree

Hide file tree

Showing 2 changed files with 4 additions and 4 deletions.
diff --git a/m/psislw.m b/m/psislw.m
@@ -1,11 +1,11 @@
 function [lw,kss] = psislw(lw,Reff)
 %PSIS Pareto smoothed importance sampling
-%   
+%
 %  Description
 %    [LW,K] = PSISLW(LW,Reff) returns log weights LW
 %    and Pareto tail indeces K, given log weights and optional arguments:
 %      Reff - relative MCMC efficiency N_eff/N
-%    
+%
 %  Reference
 %    Aki Vehtari, Andrew Gelman and Jonah Gabry (2017). Pareto
 %    smoothed importance sampling. https://arxiv.org/abs/1507.02646v5
@@ -30,7 +30,7 @@
     % Divide log weights into body and right tail
     n=numel(x);
     xs=sort(x);
-    xcutoff=xs(end-ceil(min(0.2*n,3*sqrt(n*Reff))));
+    xcutoff=xs(end-ceil(min(0.2*n,3*sqrt(n/Reff))));
     if xcutoff<log(realmin)
         % need to stay above realmin
         xcutoff=-700;

diff --git a/py/psis.py b/py/psis.py
@@ -154,7 +154,7 @@ def psislw(lw, Reff=1.0, overwrite_lw=False):
     kss = np.empty(m)
 
     # precalculate constants
-    cutoff_ind = - int(np.ceil(min(0.2 * n, 3 * np.sqrt(n * Reff)))) - 1
+    cutoff_ind = - int(np.ceil(min(0.2 * n, 3 * np.sqrt(n / Reff)))) - 1
     cutoffmin = np.log(np.finfo(float).tiny)
     logn = np.log(n)
     k_min = 1/3