StochasticForceInference 2.0.1__tar.gz → 2.0.2__tar.gz

{stochasticforceinference-2.0.1 → stochasticforceinference-2.0.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: StochasticForceInference
-Version: 2.0.1
+Version: 2.0.2
 Summary: Stochastic Force Inference for Langevin SDEs
 Author-email: Pierre Ronceray <pierre.ronceray@univ-amu.fr>
 License: MIT

{stochasticforceinference-2.0.1 → stochasticforceinference-2.0.2}/SFI/inference/overdamped.py

@@ -1069,7 +1069,16 @@ class OverdampedLangevinInference(BaseLangevinInference):
 
               M_a^{\text{S}} = \tfrac{1}{2}\bigl\langle v_t \cdot A^{-1}
               \cdot \bigl[b_a(X_t) + b_a(X_{t+1})\bigr] \bigr\rangle
-              \;-\; \bigl\langle D_{\text{inst}} : (A^{-1} \cdot \nabla_x b_a) \bigr\rangle
+              \;-\; \bigl\langle (D_{\text{inst}}\,A^{-1})_{kl}\,
+              \partial_k (b_a)_l \bigr\rangle
+
+           where :math:`k` is the derivative index and :math:`l` the output
+           (force) index of :math:`\nabla_x b_a`.  The matrix acting on the
+           Jacobian is :math:`D_{\text{inst}}\,A^{-1}` in this order: contracting
+           :math:`D_{\text{inst}}` with the output index instead would give
+           :math:`A^{-1} D_{\text{inst}}`, which differs whenever
+           :math:`[A^{-1}, D_{\text{inst}}] \neq 0` (anisotropic, non-aligned
+           instantaneous diffusion with a non-trivial decorrelating Gram).
 
            The force coefficients solve :math:`G \cdot C = M`
            where :math:`G_{ab} = \langle b_a \cdot b_b \rangle` (with mode variants).
@@ -1087,11 +1096,16 @@ class OverdampedLangevinInference(BaseLangevinInference):
             trapezoid average over X and X_plus, same rhs contraction on each, then 0.5*(...+...)
 
         - Stratonovich gradient correction:
-            eq='imn,ioma,no->ia' with ops (D, G, A)
+            eq='imn,inoa,om->ia' with ops (D, G, A)
             shapes:
               D: (i, m, n)    instantaneous diffusion (N, d, d) e.g. noisy
-              G: (i, o, m, a) basis.d_x()(X) shape (N, d_deriv, d_force, F)
-              A: (m, n)       A_inv (d, d)
+              G: (i, n, o, a) basis.d_x()(X) shape (N, d_deriv, d_force, F)
+              A: (o, m)       A_inv (d, d)
+            The derivative axis n of G is contracted with D and the output
+            (force) axis o with A_inv, i.e. the correction is
+            (D_inst . A_inv) : grad b.  Contracting D with the output axis
+            instead computes A_inv . D_inst, which differs whenever D_inst and
+            A_inv do not commute.
 
         Masking
         -------
@@ -1159,7 +1173,7 @@ class OverdampedLangevinInference(BaseLangevinInference):
             exprs=[G],
             times=[D],
             consts=[A],
-            terms=[Term(eq="imn,ioma,no->ia", ops=(D.alias, "G", "A"))],
+            terms=[Term(eq="imn,inoa,om->ia", ops=(D.alias, "G", "A"))],
         )
         self._force_D_grad_b_average = integrate(
             self.data, prog_grad, reduce="sum", chunk_target_bytes=self._chunk_target_bytes

{stochasticforceinference-2.0.1 → stochasticforceinference-2.0.2}/StochasticForceInference.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: StochasticForceInference
-Version: 2.0.1
+Version: 2.0.2
 Summary: Stochastic Force Inference for Langevin SDEs
 Author-email: Pierre Ronceray <pierre.ronceray@univ-amu.fr>
 License: MIT

{stochasticforceinference-2.0.1 → stochasticforceinference-2.0.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "StochasticForceInference"
-version = "2.0.1"
+version = "2.0.2"
 description = "Stochastic Force Inference for Langevin SDEs"
 authors = [{name = "Pierre Ronceray", email = "pierre.ronceray@univ-amu.fr"}]
 readme = "README.md"