(#504) Add regression test for CBPS rank-deficient SVD handling (#507)

Summary:
- Add a focused regression test for rank-deficient CBPS design matrices (duplicate covariate column).
- Asserts CBPS completes and near-zero singular values are filtered before optimization.
- Remove the comment `TODO: add unittest` in `cbps.py`.

Closes #504

Pull Request resolved: #507

Test Plan:
- [x] `pytest tests/test_cbps.py::Testcbps::test_cbps_rank_deficient_svd_filters_near_zero_singular_values -ra`
- [x] `pytest tests/test_cbps.py -ra`

Differential Revision: D107679003

Pulled By: talgalili

fbshipit-source-id: 99e1b7b58f272e23b2831dd40c19ad91ae7ded16

Author: ashoksmore

CHANGELOG.md

@@ -1,3 +1,11 @@
+# 0.22.0 (Unreleased)
+
+## Tests
+
+- Add regression test for CBPS rank-deficient SVD handling when covariate
+  columns are collinear (near-zero singular values filtered during SVD
+  preprocessing).
+
 # 0.21.0 (2026-06-02)
 
 ## New Features

balance/weighting_methods/cbps.py

@@ -606,7 +606,6 @@ def cbps(  # noqa
     U, s, Vh = scipy.linalg.svd(X_matrix, full_matrices=False)
 
     # remove near-zero singular values to address the rank-deficiency of X_matrix
-    # TODO: add unittest
     singular_value_threshold = 1e-10
     U = U[:, s > singular_value_threshold]
     Vh = Vh[s > singular_value_threshold, :]

tests/test_cbps.py

@@ -376,6 +376,44 @@ def test__reverse_svd_and_centralization(self) -> None:
             np.around(np.matmul(U, beta), 7),
         )
 
+    def test_cbps_rank_deficient_svd_filters_near_zero_singular_values(self) -> None:
+        """Regression test for rank-deficient design matrices in CBPS."""
+        sample_df = pd.DataFrame({"a": [0, 1, 0, 1, 0, 1], "b": [0, 1, 0, 1, 0, 1]})
+        target_df = pd.DataFrame({"a": [0, 1, 0, 1], "b": [0, 1, 0, 1]})
+        sample_weights = pd.Series([1.0] * len(sample_df))
+        target_weights = pd.Series([1.0] * len(target_df))
+
+        result = balance_cbps.cbps(
+            sample_df,
+            sample_weights,
+            target_df,
+            target_weights,
+            variables=["a", "b"],
+            transformations=None,
+            store_fit_metadata=True,
+        )
+
+        self.assertIn("weight", result)
+        weights = result["weight"]
+        self.assertIsInstance(weights, pd.Series)
+        self.assertEqual(len(weights), len(sample_df))
+        self.assertTrue(np.all(np.isfinite(weights)))
+
+        model = result["model"]
+        self.assertIsInstance(model, dict)
+        svd_s = model["svd_s"]
+        n_design_columns = len(model["X_matrix_columns"])
+
+        self.assertLess(
+            len(svd_s),
+            n_design_columns,
+            msg="Near-zero singular values should be removed for rank-deficient X",
+        )
+        self.assertTrue(
+            np.all(svd_s > TOLERANCE),
+            msg="Retained singular values should exceed the CBPS filtering threshold",
+        )
+
     def test_cbps_consistency_with_default_arguments(self) -> None:
         """Test CBPS function consistency with default arguments on complex data.