StochasticForceInference 2.0.1__tar.gz → 2.0.3__tar.gz

{stochasticforceinference-2.0.1 → stochasticforceinference-2.0.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: StochasticForceInference
-Version: 2.0.1
+Version: 2.0.3
 Summary: Stochastic Force Inference for Langevin SDEs
 Author-email: Pierre Ronceray <pierre.ronceray@univ-amu.fr>
 License: MIT
@@ -202,6 +202,16 @@ together with the method paper(s) relevant to the features you use. See
 documentation for the full, feature-specific list (overdamped, underdamped/ULI,
 trapeze, and PASTIS references).
 
+## Funding
+
+Funded by the European Union (ERC, SuperStoc, 101117322). Views and opinions
+expressed are however those of the author(s) only and do not necessarily
+reflect those of the European Union or the European Research Council Executive
+Agency. Neither the European Union nor the granting authority can be held
+responsible for them. P. R. also acknowledges funding from the French National
+Research Agency (ANR-16-CONV-0001) and from the Excellence Initiative of
+Aix-Marseille University A\*MIDEX.
+
 ## License
 
 MIT — see [LICENSE](https://github.com/ronceray/StochasticForceInference/blob/main/LICENSE) for details.

{stochasticforceinference-2.0.1 → stochasticforceinference-2.0.3}/README.md

@@ -148,6 +148,16 @@ together with the method paper(s) relevant to the features you use. See
 documentation for the full, feature-specific list (overdamped, underdamped/ULI,
 trapeze, and PASTIS references).
 
+## Funding
+
+Funded by the European Union (ERC, SuperStoc, 101117322). Views and opinions
+expressed are however those of the author(s) only and do not necessarily
+reflect those of the European Union or the European Research Council Executive
+Agency. Neither the European Union nor the granting authority can be held
+responsible for them. P. R. also acknowledges funding from the French National
+Research Agency (ANR-16-CONV-0001) and from the Excellence Initiative of
+Aix-Marseille University A\*MIDEX.
+
 ## License
 
 MIT — see [LICENSE](https://github.com/ronceray/StochasticForceInference/blob/main/LICENSE) for details.

{stochasticforceinference-2.0.1 → stochasticforceinference-2.0.3}/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
         -------
@@ -1155,11 +1169,17 @@ class OverdampedLangevinInference(BaseLangevinInference):
         )
         D = self.get_diffusion_timeop("noisy")  # already a TimeOperand with alias
 
+        # Contraction correctness relies on D_inst being SYMMETRIC: the einsum
+        # pairs D_inst with the derivative axis (n) of the Jacobian and A_inv
+        # with the output axis (o), giving (D_inst . A_inv) : grad b.  This is
+        # the intended order only because D ("noisy") is symmetric; a future
+        # non-symmetric instantaneous-D estimator would need the index roles of
+        # D and A_inv swapped (and would differ whenever [A_inv, D] != 0).
         prog_grad = Integrand(
             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-2.0.3}/SFI/inference/parametric_core/covariance.py

@@ -242,7 +242,7 @@ def build_ud_blocks_exact(alpha_plus, alpha_zero, alpha_minus, qing, Lambda,
 
         A_n^{\rm proc} &= Q^{xx}_{\rm out}(n)
           + J^{xv}_{\rm out}(n)\,S_v(n)\,J^{xv}_{\rm out}(n)^{\!\top},\\
-        S_v(n) &= Q^{vv}_{\rm in} - N Q^{xv\top}_{\rm in} - Q^{xv}_{\rm in} N^{\!\top}
+        S_v(n) &= Q^{vv}_{\rm in} - N Q^{xv}_{\rm in} - (N Q^{xv}_{\rm in})^{\!\top}
                   + N Q^{xx}_{\rm in} N^{\!\top}\ \big|_n ,\\
         C_n^{\rm proc} &= \bigl[Q^{xv}_{\rm out}(n) - Q^{xx}_{\rm out}(n)\,N(n{+}1)^{\!\top}\bigr]
                           J^{xv}_{\rm out}(n{+}1)^{\!\top},

{stochasticforceinference-2.0.1 → stochasticforceinference-2.0.3}/SFI/inference/underdamped.py

@@ -984,10 +984,17 @@ class UnderdampedLangevinInference(BaseLangevinInference):
 
               M_a = \bigl\langle \hat a_t \cdot A^{-1} \cdot b_a(\hat x_t, \hat v_t)
               \bigr\rangle
-              \;+\; w \,\bigl\langle -D_{\text{inst}} : (A^{-1}\cdot\partial_v b_a)
-              \bigr\rangle
-
-           where :math:`w = (1+2\ell)/3`, with :math:`\ell=1` (symmetric),
+              \;-\; w \,\bigl\langle (D_{\text{inst}}\,A^{-1})_{kl}\,
+              \partial_{v_k} (b_a)_l \bigr\rangle
+
+           where :math:`k` is the velocity-derivative index and :math:`l` the
+           output (force) index of :math:`\partial_v 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` (this is the overdamped
+           gradient-correction convention, mirrored here).  The weight is
+           :math:`w = (1+2\ell)/3`, with :math:`\ell=1` (symmetric),
            :math:`\ell=0` (early), :math:`\ell=-\tfrac{1}{2}` (anticipated).
 
         Contractions
@@ -997,7 +1004,12 @@ class UnderdampedLangevinInference(BaseLangevinInference):
 
         - correction term:
             eq='imn,no,imoa->ia' with ops (D_inst, A_inv, ∂_v b)
-            (the third axis of ∂_v b is the derivative index; we label it 'o' above).
+            (∂_v b has axes (i, m, o, a): m is the velocity-DERIVATIVE index
+            — contracted with D_inst — and o is the OUTPUT/force index —
+            contracted with A_inv; a is the feature.  This yields
+            (D_inst · A_inv) : ∂_v b.  Per stateexpr.py, d_v inserts the
+            derivative axis immediately before the rank block, so it is the
+            second axis, not the third.)
         """
         if not hasattr(self, "A_inv"):
             raise RuntimeError("A_inv not available. Compute diffusion first.")
@@ -1049,6 +1061,12 @@ class UnderdampedLangevinInference(BaseLangevinInference):
             alias="Gv",
         )
 
+        # Contraction correctness relies on D_inst being SYMMETRIC: D_inst (m,n)
+        # pairs with the velocity-derivative axis (m) of Gv and A_inv (n,o) with
+        # the output axis (o), giving (D_inst . A_inv) : grad_v b.  Correct only
+        # because D ("noisy") is symmetric; a non-symmetric instantaneous-D
+        # estimator would need the D/A_inv index roles swapped (they differ
+        # whenever [A_inv, D] != 0).  Mirrors overdamped._force_moments.
         prog_w = Integrand(
             exprs=[Gv],
             times=[D_op],

{stochasticforceinference-2.0.1 → stochasticforceinference-2.0.3}/SFI/statefunc/structexpr.py

@@ -428,6 +428,40 @@ def _feature_count(idx: Any, n_features: int) -> int:
     raise TypeError(f"Unsupported index type: {type(idx).__name__}")
 
 
+def _selected_feature_indices(idx: Any, n_features: int) -> list[int]:
+    """Normalise a ``__getitem__`` index to a list of non-negative feature ids.
+
+    Mirrors the validation in :func:`_feature_count`; assumes ``idx`` has
+    already passed it (so range/type errors have been raised).
+    """
+    if isinstance(idx, int):
+        return [idx % n_features]
+    if isinstance(idx, slice):
+        return list(range(*idx.indices(n_features)))
+    # list/tuple of ints (validated by _feature_count)
+    return [int(i) % n_features for i in idx]
+
+
+def _flatten_concat_features(node: _StructNode) -> list[_StructNode]:
+    """Flatten a (possibly nested) ``_ConcatOp`` into its per-term sub-nodes.
+
+    Feature concatenation via ``&`` builds ``_ConcatOp`` trees whose children
+    are the individual feature terms.  Returns one entry per structural
+    feature; a non-concat node contributes itself (treated as a single
+    feature).  The caller must check ``len(result) == n_features`` before
+    trusting the decomposition — that equality is what distinguishes a
+    genuine ``&``-concatenated field (where feature selection can be resolved
+    structurally) from an opaque/parametric node whose feature axis is not
+    expressed as concat structure.
+    """
+    if isinstance(node, _ConcatOp):
+        out: list[_StructNode] = []
+        for c in node.children:
+            out.extend(_flatten_concat_features(c))
+        return out
+    return [node]
+
+
 # =====================================================================
 # StructuredExpr
 # =====================================================================
@@ -727,6 +761,28 @@ class StructuredExpr:
             lbls = tuple(self.labels[idx]) if self.labels else ()
         else:
             lbls = ()
+
+        # Feature selection is a structural (concat-level) concept: a compiled
+        # inner value carries no feature axis, so `_SliceOp`'s pointwise
+        # `v[..., idx]` would slice a *spatial* axis, not a feature (it errors
+        # for scalar sectors and silently returns the wrong component for
+        # vector/tensor sectors).  When the field is `&`-concatenated — the
+        # normal way multi-feature structured fields are built — resolve the
+        # selection by picking the corresponding `_ConcatOp` term(s), so it
+        # compiles through the working feature-concat path in `_compile_sector`
+        # and never reaches `_SliceOp`.
+        feats = _flatten_concat_features(self._node)
+        if len(feats) == self.n_features:
+            sel = _selected_feature_indices(idx, self.n_features)
+            picked = [feats[i] for i in sel]
+            new_node: _StructNode = (
+                picked[0] if len(picked) == 1 else _ConcatOp(tuple(picked))
+            )
+            return self._new(n_features=n, labels=lbls, node=new_node)
+
+        # Fallback: the feature axis is not expressed as concat structure
+        # (opaque/parametric node).  Preserve the prior algebra behaviour;
+        # such nodes are not produced by the SPDE embed path.
         return self._new(
             n_features=n,
             labels=lbls,

{stochasticforceinference-2.0.1 → stochasticforceinference-2.0.3}/StochasticForceInference.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: StochasticForceInference
-Version: 2.0.1
+Version: 2.0.3
 Summary: Stochastic Force Inference for Langevin SDEs
 Author-email: Pierre Ronceray <pierre.ronceray@univ-amu.fr>
 License: MIT
@@ -202,6 +202,16 @@ together with the method paper(s) relevant to the features you use. See
 documentation for the full, feature-specific list (overdamped, underdamped/ULI,
 trapeze, and PASTIS references).
 
+## Funding
+
+Funded by the European Union (ERC, SuperStoc, 101117322). Views and opinions
+expressed are however those of the author(s) only and do not necessarily
+reflect those of the European Union or the European Research Council Executive
+Agency. Neither the European Union nor the granting authority can be held
+responsible for them. P. R. also acknowledges funding from the French National
+Research Agency (ANR-16-CONV-0001) and from the Excellence Initiative of
+Aix-Marseille University A\*MIDEX.
+
 ## License
 
 MIT — see [LICENSE](https://github.com/ronceray/StochasticForceInference/blob/main/LICENSE) for details.

{stochasticforceinference-2.0.1 → stochasticforceinference-2.0.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "StochasticForceInference"
-version = "2.0.1"
+version = "2.0.3"
 description = "Stochastic Force Inference for Langevin SDEs"
 authors = [{name = "Pierre Ronceray", email = "pierre.ronceray@univ-amu.fr"}]
 readme = "README.md"