BUG: weighted quantile for some zero weights (numpy#27549)

This PR fixed weighted quantiles (and percentiles) for a corner case:

* at least one weight is zero
* q=0 (0-quantile equals minimum)
Then:
```
np.quantile(np.arange(3), 0, weights=[0, 0, 1], method="inverted_cdf")
```
should return 2, the minimum when neglecting zero weight values. Current main returns 0.

Author: lorentzenchr

numpy/lib/_function_base_impl.py

@@ -4821,6 +4821,13 @@ def _quantile(
         # returns 2 instead of 1 because 0.4 is not binary representable.
         if quantiles.dtype.kind == "f":
             cdf = cdf.astype(quantiles.dtype)
+        # Weights must be non-negative, so we might have zero weights at the
+        # beginning leading to some leading zeros in cdf. The call to
+        # np.searchsorted for quantiles=0 will then pick the first element,
+        # but should pick the first one larger than zero. We
+        # therefore simply set 0 values in cdf to -1.
+        if np.any(cdf[0, ...] == 0):
+            cdf[cdf == 0] = -1
 
         def find_cdf_1d(arr, cdf):
             indices = np.searchsorted(cdf, quantiles, side="left")

numpy/lib/tests/test_function_base.py

@@ -4008,6 +4008,17 @@ def test_quantile_with_weights_and_axis(self, method):
                 )
         assert_allclose(q, q_res)
 
+    @pytest.mark.parametrize("method", methods_supporting_weights)
+    def test_quantile_weights_min_max(self, method):
+        # Test weighted quantile at 0 and 1 with leading and trailing zero
+        # weights.
+        w = [0, 0, 1, 2, 3, 0]
+        y = np.arange(6)
+        y_min = np.quantile(y, 0, weights=w, method="inverted_cdf")
+        y_max = np.quantile(y, 1, weights=w, method="inverted_cdf")
+        assert y_min == y[2]  # == 2
+        assert y_max == y[4]  # == 4
+
     def test_quantile_weights_raises_negative_weights(self):
         y = [1, 2]
         w = [-0.5, 1]