Fix INDEPENDENT mechanism crash on duplicate cliques

dpsynth/discrete_mechanisms/independent.py

@@ -70,7 +70,9 @@ def run_mechanism(
 
   potentials = initial_potentials
   if potentials is not None:
-    potentials = potentials.expand([m.clique for m in measurements])
+    # De-duplicate cliques (order-preserving); `expand` requires unique cliques.
+    unique_cliques = list(dict.fromkeys(m.clique for m in measurements))
+    potentials = potentials.expand(unique_cliques)
 
   model = mbi.estimation.mirror_descent(
       data.domain,

dpsynth/discrete_mechanisms/independent_test.py

@@ -31,6 +31,35 @@ def test_fits_one_way_marginals(self):
       actual = synthetic.project([col]).datavector()
       np.testing.assert_allclose(actual, expected, atol=0.1)
 
+  def test_duplicate_one_way_cliques_with_initial_potentials(self):
+    # Regression test: when `initial_potentials` is provided (e.g. the empty
+    # CliqueVector for the no-constraints case) and `initial_measurements`
+    # already holds the one-way marginals, the cliques measured in the loop
+    # duplicate them. `CliqueVector.expand` rejects duplicate cliques, so the
+    # mechanism used to crash with "Cliques must be unique.".
+    domain = mbi.Domain(["a", "b", "c"], [3, 4, 5])
+    data = mbi.Dataset.synthetic(domain, N=1000)
+
+    initial_measurements = [
+        mbi.LinearMeasurement(data.project((col,)).datavector(), (col,))
+        for col in data.domain
+    ]
+    initial_potentials = mbi.CliqueVector(domain, [], {})
+
+    config = independent.IndependentConfig(pgm_iters=500)
+    synthetic = independent.run_mechanism(
+        data,
+        config,
+        zcdp_rho=10000,
+        initial_measurements=initial_measurements,
+        initial_potentials=initial_potentials,
+    )
+
+    for col in data.domain:
+      expected = data.project([col]).datavector()
+      actual = synthetic.project([col]).datavector()
+      np.testing.assert_allclose(actual, expected, atol=0.1)
+
 
 if __name__ == "__main__":
   absltest.main()