ah

setup.py

@@ -7,7 +7,7 @@
 
 setup(
     name="winning",
-    version="1.0.1",
+    version="1.0.2",
     description="Fast ability inference from contest winning probabilities",
     long_description=README,
     long_description_content_type="text/markdown",

tests/test_state_prices_from_extended_offsets.py

@@ -147,8 +147,7 @@ def assert_split_race(offsets_1, offsets_2, n_huge=1, n_inf=1, n_neg_inf=3):
     offsets = offsets_1 + offsets_2 + [HUGE] * n_huge + [float('inf')] * n_inf + [float('-inf')] * n_neg_inf
     prices = state_prices_from_extended_offsets(density=density, offsets=offsets, max_depth=1)
 
-
-    # Use a different approximation
+    # Use a different (poor) approximation if the second group is small
     best_in_2 = min(offsets_2)
     group_1 = offsets_1+[best_in_2]
     # Try to approximate if we don't hit lattice issues

winning/lattice.py

@@ -629,7 +629,7 @@ def divide_offsets(centered_offsets, max_best=20):
         offset_divider = np.mean(centered_offsets)  # should be 0
     else:
         srt_offsets = sorted(centered_offsets)
-        max_best = min(20, int(n/6))
+        max_best = min(20, int(n/6+2))
         gaps = [ abs(a) for a in np.diff([srt_offsets[0]]+srt_offsets) ][:max_best] # [0, 4, 2, ...]
         ndx_gap = gaps.index(max(gaps))
         try:
@@ -682,18 +682,18 @@ def state_prices_from_extended_offsets(density, offsets, max_depth=3, unit_ratio
     # Otherwise, having reached this far we have only float abilities, though some might be extremely large
     L = implied_L(density)
     W = int(approximate_support_width(density))
-    really_bad_horse_offset = min(offsets) + 2*W # <--- No chance of winning
+    really_bad_horse_offset = min(offsets) + W # <--- No chance of winning
 
-    # If there are really bad but finite horses, set them to float('inf') and call again
-    # On the next call, hopefully the centering leaves us with a tractable problem
+    # If there are bad but finite horses, set them to float('inf') and call again
+    # On the next call, hopefully the centering leaves us inside the lattice
     extremely_bad_ndxs = [i for i, o in enumerate(offsets) if o > really_bad_horse_offset ]
     if any(extremely_bad_ndxs):
         augmented_offsets = [o if i not in extremely_bad_ndxs else float('inf') for i, o in enumerate(offsets)]
         return state_prices_from_extended_offsets(density=density,
                                                   offsets=int_centered(augmented_offsets),
                                                   max_depth=max_depth)
 
-    # If there is a solitary extremely good horse, short-circuit
+    # If there is a solitary standout horse, short-circuit
     diff_to_best = [ o - min(offsets) for o in offsets]
     is_walkover = min(diff_to_best) > W*math.sqrt(len(offsets))
     if is_walkover: