fixes issue #65

pkg/DESCRIPTION

@@ -14,14 +14,14 @@ Description: Implements an approximate string matching version of R's native
     implementation of soundex is provided as well. Distances can be computed between
     character vectors while taking proper care of encoding or between integer
     vectors representing generic sequences.
-Version: 0.9.4.6
+Version: 0.9.4.7
 Depends:
     R (>= 2.15.3)
 Imports:
     parallel
 URL: https://github.com/markvanderloo/stringdist
 BugReports: https://github.com/markvanderloo/stringdist/issues
-Date: 2017-07-26
+Date: 2018-03-05
 Suggests:
     testthat
 RoxygenNote: 6.0.1

pkg/NEWS

@@ -1,3 +1,7 @@
+version 0.9.4.7
+- Fixed edge case where cosine distance with q=1, between strings of repeating characters
+  yielded Inf (Thanks to Markus Dumke)
+
 version 0.9.4.6
 - Fixed argument passing error in lower_tri (thanks to Kurt Hornik)
 

pkg/src/qgram.c

@@ -381,8 +381,9 @@ double qgram_dist(
         dist[0] =  0.0;
       } else {
         // there are several ways to express the rhs (including ones that give 0L 
-        // at equal strings) but this has least chance of overflow.
-        dist[0] = 1.0 - dist[0]/(sqrt(dist[1]) * sqrt(dist[2]));
+        // at equal strings) but this has least chance of overflow
+        // fabs is taken to avoid numerical -0.
+        dist[0] = fabs(1.0 - dist[0]/(sqrt(dist[1]) * sqrt(dist[2])));
       }
       break;
     case 2:

pkg/tests/testthat/testStringdist.R

@@ -276,6 +276,8 @@ test_that("cosine distance computes correctly",{
     round(stringdist("aaa","abc",method="cosine",q=2),8),
     1.0
   )
+  # see issue #65
+  expect_equal(stringdist("abc","abcabc",method='cosine',q=1),0)
   # numerical accuracy test (thanks to Ben Haller)
   # note that 1 - 2/(sqrt(2)*sqrt(2)) != 0, so this used to give ~2.2E-16. 
   expect_equal( stringdist("ab","ab",method="cosine"),0.0,tolerance=0.0 )