Complex fixes in SDEs

diffrax/_solver/milstein.py

@@ -211,7 +211,7 @@ def step(
         leaves_ΔwΔw = []
         for i1, l1 in enumerate(leaves_Δw):
             for i2, l2 in enumerate(leaves_Δw):
-                leaf = jnp.tensordot(l1[..., None], l2[None, ...], axes=1)
+                leaf = jnp.tensordot(jnp.conj(l1[..., None]), l2[None, ...], axes=1)
                 if i1 == i2:
                     eye = jnp.eye(l1.size).reshape(l1.shape + l1.shape)
                     with jax.numpy_dtype_promotion("standard"):
@@ -305,7 +305,7 @@ def _to_treemap(_Δw, _g0):
         def __dot(_v0, _ΔwΔw):
             # _v0 has structure (leaf(y0), leaf(Δw), leaf(Δw))
             # _ΔwΔw has structure (leaf(Δw), leaf(Δw))
-            _out = jnp.tensordot(_v0, _ΔwΔw, axes=jnp.ndim(_ΔwΔw))
+            _out = jnp.tensordot(jnp.conj(_v0), _ΔwΔw, axes=jnp.ndim(_ΔwΔw))
             # _out has structure (leaf(y0),)
             return _out
 

diffrax/_solver/srk.py

@@ -407,7 +407,7 @@ def _comp_g(_t):
 
             g0_g1 = _comp_g(jnp.array([t0, t1], dtype=complex_to_real_dtype(dtype)))
             g0 = jtu.tree_map(lambda g_leaf: g_leaf[0], g0_g1)
-            # g_delta = 0.5 * g1 - g0
+            # g_delta = 0.5 * (g1 - g0)
             g_delta = jtu.tree_map(lambda g_leaf: 0.5 * (g_leaf[1] - g_leaf[0]), g0_g1)
             w_kgs = diffusion.prod(g0, w)
             a_w = jnp.asarray(self.tableau.coeffs_w.a, dtype=dtype)
@@ -456,7 +456,7 @@ def sum_prev_stages(_stage_out_buff, _a_j):
             )
             # Sum up the previous stages weighted by the coefficients in the tableau
             return jtu.tree_map(
-                lambda lf: jnp.tensordot(_a_j, lf, axes=1), _stage_out_view
+                lambda lf: jnp.tensordot(jnp.conj(_a_j), lf, axes=1), _stage_out_view
             )
 
         def insert_jth_stage(results, k_j, j):

diffrax/_term.py

@@ -257,7 +257,7 @@ def _callable_to_path(
 # control: Shaped[Array, "*control"]
 # return: Shaped[Array, "*state"]
 def _prod(vf, control):
-    return jnp.tensordot(vf, control, axes=jnp.ndim(control))
+    return jnp.tensordot(jnp.conj(vf), control, axes=jnp.ndim(control))
 
 
 # This class exists for backward compatibility with `WeaklyDiagonalControlTerm`. If we

test/helpers.py

@@ -100,7 +100,7 @@ def path_l2_dist(
     # and the length of saveat). Also sum all the PyTree leaves.
     def sum_square_diff(y1, y2):
         with jax.numpy_dtype_promotion("standard"):
-            square_diff = jnp.square(y1 - y2)
+            square_diff = jnp.square(jnp.abs(y1 - y2))
         # sum all but the first two axes
         axes = range(2, square_diff.ndim)
         out = jnp.sum(square_diff, axis=axes)