gradient is actually complex conjugate of gradient

Author: kevinsung

python/ffsim/variational/orbital_optimization.py

@@ -170,7 +170,9 @@ def fun_jax(generator: np.ndarray) -> float:
     def fun(x: np.ndarray) -> tuple[float, np.ndarray]:
         generator = _generator_from_parameters(x, norb=norb, real=real)
         val, grad = value_and_grad(generator)
-        return val, _generator_to_parameters(grad, real=real)
+        # The complex conjugate of the gradient is actually returned
+        # See https://github.com/jax-ml/jax/issues/4891
+        return val, _generator_to_parameters(grad.conj(), real=real)
 
     if initial_orbital_rotation is None:
         initial_orbital_rotation = np.eye(norb)

tests/python/variational/orbital_optimization_test.py

@@ -86,6 +86,6 @@ def test_optimize_orbitals():
     np.testing.assert_allclose(energy, -108.58613393502857)
     assert np.iscomplexobj(orbital_rotation)
     assert len(result.x) == norb**2
-    assert result.nit <= 8
-    assert result.nfev <= 11
-    assert result.njev <= 11
+    assert result.nit <= 12
+    assert result.nfev <= 14
+    assert result.njev <= 14