BUG: Fix for rolling with uneven nanosecond windows have wrong results

doc/source/whatsnew/v1.4.0.rst

@@ -506,6 +506,7 @@ Groupby/resample/rolling
 - Bug in :meth:`GroupBy.apply` with time-based :class:`Grouper` objects incorrectly raising ``ValueError`` in corner cases where the grouping vector contains a ``NaT`` (:issue:`43500`, :issue:`43515`)
 - Bug in :meth:`GroupBy.mean` failing with ``complex`` dtype (:issue:`43701`)
 - Fixed bug in :meth:`Series.rolling` and :meth:`DataFrame.rolling` not calculating window bounds correctly for the first row when ``center=True`` and index is decreasing (:issue:`43927`)
+- Fixed bug in :meth:`Series.rolling` and :meth:`DataFrame.rolling` for centered datetimelike windows with uneven nanosecond (:issue:`43997`)
 - Bug in :meth:`GroupBy.nth` failing on ``axis=1`` (:issue:`43926`)
 - Fixed bug in :meth:`Series.rolling` and :meth:`DataFrame.rolling` not respecting right bound on centered datetime-like windows, if the index contain duplicates (:issue:`#3944`)
 

pandas/_libs/window/indexers.pyx

@@ -62,6 +62,14 @@ def calculate_variable_window_bounds(
     if closed in ['left', 'both']:
         left_closed = True
 
+    # GH 43997:
+    # If the forward and the backward facing windows
+    # would result in a fraction of 1/2 a nanosecond
+    # we need to make both interval ends inclusive.
+    if center and window_size % 2 == 1:
+        right_closed = True
+        left_closed = True
+
     if index[num_values - 1] < index[0]:
         index_growth_sign = -1
 

pandas/tests/window/test_rolling.py

@@ -1276,6 +1276,23 @@ def test_rolling_decreasing_indices_centered(window, closed, expected, frame_or_
     tm.assert_equal(result_dec, expected_dec)
 
 
+@pytest.mark.parametrize(
+    "window,expected",
+    [
+        ("1ns", [1.0, 1.0, 1.0, 1.0]),
+        ("3ns", [2.0, 3.0, 3.0, 2.0]),
+    ],
+)
+def test_rolling_center_nanosecond_resolution(
+    window, closed, expected, frame_or_series
+):
+    index = date_range("2020", periods=4, freq="1ns")
+    df = frame_or_series([1, 1, 1, 1], index=index, dtype=float)
+    expected = frame_or_series(expected, index=index, dtype=float)
+    result = df.rolling(window, closed=closed, center=True).sum()
+    tm.assert_equal(result, expected)
+
+
 @pytest.mark.parametrize(
     "method,expected",
     [