Prevent manifold EMA state aliasing

src/pyrecest/filters/manifold_exponential_moving_average.py

@@ -1,5 +1,6 @@
 """Exponential moving average for states on manifolds."""
 
+import copy
 from typing import Any, Callable
 
 import numpy as np
@@ -107,16 +108,16 @@ def filter_state(self):
 
     @filter_state.setter
     def filter_state(self, new_state):
-        self._filter_state = new_state
+        self._filter_state = copy.deepcopy(new_state)
 
     def update(self, sample):
         """Update the moving average with a new manifold-valued sample."""
         if self._filter_state is None:
-            self._filter_state = sample
+            self.filter_state = sample
             return
 
         tangent_update = self.alpha * asarray(self.phi_inv(self._filter_state, sample))
-        self._filter_state = self.phi(self._filter_state, tangent_update)
+        self.filter_state = self.phi(self._filter_state, tangent_update)
 
     def get_point_estimate(self):
         """Return the current manifold estimate."""

tests/filters/test_manifold_exponential_moving_average_aliasing.py

@@ -0,0 +1,44 @@
+"""Regression tests for manifold EMA state ownership."""
+
+import numpy as np
+import numpy.testing as npt
+
+from pyrecest.filters import ManifoldExponentialMovingAverage
+
+
+def _phi_euclidean(state, tangent):
+    return state + tangent
+
+
+def _phi_inv_euclidean(state_ref, state):
+    return state - state_ref
+
+
+def test_first_sample_is_copied_into_filter_state():
+    sample = np.array([2.0, 3.0])
+    ema = ManifoldExponentialMovingAverage(
+        initial_state=None,
+        alpha=0.5,
+        phi=_phi_euclidean,
+        phi_inv=_phi_inv_euclidean,
+    )
+
+    ema.update(sample)
+    sample[:] = -1.0
+
+    npt.assert_array_equal(ema.filter_state, np.array([2.0, 3.0]))
+
+
+def test_explicit_state_assignment_does_not_alias_caller_array():
+    ema = ManifoldExponentialMovingAverage(
+        initial_state=np.array([0.0, 0.0]),
+        alpha=0.5,
+        phi=_phi_euclidean,
+        phi_inv=_phi_inv_euclidean,
+    )
+    replacement = np.array([4.0, 5.0])
+
+    ema.filter_state = replacement
+    replacement[:] = 0.0
+
+    npt.assert_array_equal(ema.filter_state, np.array([4.0, 5.0]))