[BUG] MSM bug fixes (#3121)

* minor msm bug fix

* msm bug fixes

* updated tests

* remove redundent distance call

Author: chrisholder

aeon/distances/elastic/_msm.py

@@ -246,12 +246,10 @@ def _msm_independent_cost_matrix(
     x_size = x.shape[1]
     y_size = y.shape[1]
     cost_matrix = np.zeros((x_size, y_size))
-    distance = 0
     min_instances = min(x.shape[0], y.shape[0])
     for i in range(min_instances):
         curr_cost_matrix = _independent_cost_matrix(x[i], y[i], bounding_matrix, c)
         cost_matrix = np.add(cost_matrix, curr_cost_matrix)
-        distance += curr_cost_matrix[-1, -1]
     return cost_matrix
 
 
@@ -267,19 +265,19 @@ def _independent_cost_matrix(
     for i in range(1, x_size):
         if bounding_matrix[i, 0]:
             cost = _cost_independent(x[i], x[i - 1], y[0], c)
-            cost_matrix[i][0] = cost_matrix[i - 1][0] + cost
+            cost_matrix[i, 0] = cost_matrix[i - 1, 0] + cost
 
-    for i in range(1, y_size):
-        if bounding_matrix[0, i]:
-            cost = _cost_independent(y[i], x[0], y[i - 1], c)
-            cost_matrix[0][i] = cost_matrix[0][i - 1] + cost
+    for j in range(1, y_size):
+        if bounding_matrix[0, j]:
+            cost = _cost_independent(y[j], x[0], y[j - 1], c)
+            cost_matrix[0, j] = cost_matrix[0, j - 1] + cost
 
     for i in range(1, x_size):
         for j in range(1, y_size):
             if bounding_matrix[i, j]:
                 d1 = cost_matrix[i - 1][j - 1] + np.abs(x[i] - y[j])
-                d2 = cost_matrix[i - 1][j] + _cost_independent(x[i], x[i - 1], y[j], c)
-                d3 = cost_matrix[i][j - 1] + _cost_independent(y[j], x[i], y[j - 1], c)
+                d2 = cost_matrix[i - 1, j] + _cost_independent(x[i], x[i - 1], y[j], c)
+                d3 = cost_matrix[i, j - 1] + _cost_independent(y[j], x[i], y[j - 1], c)
 
                 cost_matrix[i, j] = min(d1, d2, d3)
 
@@ -292,40 +290,44 @@ def _msm_dependent_cost_matrix(
 ) -> np.ndarray:
     x_size = x.shape[1]
     y_size = y.shape[1]
+
     cost_matrix = np.full((x_size, y_size), np.inf)
-    cost_matrix[0, 0] = np.sum(np.abs(x[:, 0] - y[:, 0]))
+    cost_matrix[0, 0] = _univariate_squared_distance(x[:, 0], y[:, 0])
 
     for i in range(1, x_size):
         if bounding_matrix[i, 0]:
             cost = _cost_dependent(x[:, i], x[:, i - 1], y[:, 0], c)
-            cost_matrix[i][0] = cost_matrix[i - 1][0] + cost
-    for i in range(1, y_size):
-        if bounding_matrix[0, i]:
-            cost = _cost_dependent(y[:, i], x[:, 0], y[:, i - 1], c)
-            cost_matrix[0][i] = cost_matrix[0][i - 1] + cost
+            cost_matrix[i, 0] = cost_matrix[i - 1, 0] + cost
+
+    for j in range(1, y_size):
+        if bounding_matrix[0, j]:
+            cost = _cost_dependent(y[:, j], x[:, 0], y[:, j - 1], c)
+            cost_matrix[0, j] = cost_matrix[0, j - 1] + cost
 
     for i in range(1, x_size):
         for j in range(1, y_size):
             if bounding_matrix[i, j]:
-                d1 = cost_matrix[i - 1][j - 1] + np.sum(np.abs(x[:, i] - y[:, j]))
-                d2 = cost_matrix[i - 1][j] + _cost_dependent(
+                d1 = cost_matrix[i - 1, j - 1] + _univariate_squared_distance(
+                    x[:, i], y[:, j]
+                )
+                d2 = cost_matrix[i - 1, j] + _cost_dependent(
                     x[:, i], x[:, i - 1], y[:, j], c
                 )
-                d3 = cost_matrix[i][j - 1] + _cost_dependent(
+                d3 = cost_matrix[i, j - 1] + _cost_dependent(
                     y[:, j], x[:, i], y[:, j - 1], c
                 )
-
                 cost_matrix[i, j] = min(d1, d2, d3)
+
     return cost_matrix
 
 
 @njit(cache=True, fastmath=True)
 def _cost_dependent(x: np.ndarray, y: np.ndarray, z: np.ndarray, c: float) -> float:
     diameter = _univariate_squared_distance(y, z)
-    mid = (y + z) / 2
+    mid = (y + z) / 2.0
     distance_to_mid = _univariate_squared_distance(mid, x)
 
-    if distance_to_mid <= (diameter / 2):
+    if distance_to_mid <= (diameter / 4.0):
         return c
     else:
         dist_to_q_prev = _univariate_squared_distance(y, x)

aeon/distances/elastic/tests/test_distance_correctness.py

@@ -60,7 +60,7 @@
     "wddtw": [38144.53125, 19121.4927, 1.34957],
     "twe": [4536.0, 3192.0220, 3030.036000000001],
     "msm_ind": [1515.0, 1517.8000000000004, 1557.0],  # msm with independent distance
-    "msm_dep": [1897.0, 1898.6000000000001, 1921.0],  # msm with dependent distance
+    "msm_dep": [190547.0, 190549.800000000020, 190589.0],  # msm with dependent distance
 }
 basic_motions_distances = {
     "euclidean": 27.51835240,
@@ -78,7 +78,7 @@
     # msm with independent distance
     "msm_ind": [84.36021099999999, 140.13788899999997, 262.6939920000001],
     # msm with dependent distance
-    "msm_dep": [33.06825, 71.1408, 190.7397],
+    "msm_dep": [192.24477562339214, 205.82382238128477, 277.7315058567359],
 }
 
 

aeon/testing/expected_results/expected_distance_results.py

@@ -166,7 +166,7 @@
     "msm": [
         [2.5687181410313746, 28.97170437295012],
         [2.5687181410313746, 28.97170437295012],
-        [2.5687181410313746, 28.657118461088324],
+        [1.0924085990342982, 13.072037194954508],
         [1.8756413986565008, 22.362537814430787],
     ],
     "adtw": [