diff --git a/src/scgenerator/math.py b/src/scgenerator/math.py
index 4da0bf1..11d1bbe 100644
--- a/src/scgenerator/math.py
+++ b/src/scgenerator/math.py
@@ -100,6 +100,7 @@ def _power_fact_array(x, n):
     return result
 
 
+@numba.njit()
 def abs2(z: np.ndarray) -> np.ndarray:
     return z.real**2 + z.imag**2
 
diff --git a/src/scgenerator/physics/pulse.py b/src/scgenerator/physics/pulse.py
index 0a1733a..9356a33 100644
--- a/src/scgenerator/physics/pulse.py
+++ b/src/scgenerator/physics/pulse.py
@@ -672,23 +672,23 @@ def g12(values):
 
     Parameters
     ----------
-    values : np.ndarray, shape (..., m, n)
+    values : np.ndarray, shape (..., m, n, nt)
         complex values following sc-ordering
 
     Returns
     -------
-    np.ndarray, shape (..., n)
+    np.ndarray, shape (..., n, nt)
         coherence function
     """
 
     # Create all the possible pairs of values
     n = len(values)
     field_pairs = itertools.combinations(values, 2)
-    mean_spec = np.mean(math.abs2(values), axis=-2)
+    mean_spec = np.mean(math.abs2(values), axis=-3)
     mask = mean_spec > 1e-15 * mean_spec.max()
     corr = np.zeros_like(values[0])
-    for pair in field_pairs:
-        corr[mask] += pair[0][mask].conj() * pair[1][mask]
+    for left, right in field_pairs:
+        corr[mask] += left[mask].conj() * right[mask]
     corr[mask] = corr[mask] / (n * (n - 1) / 2 * mean_spec[mask])
 
     return np.abs(corr)