gridstate 2.4.0__tar.gz → 2.4.1__tar.gz

{gridstate-2.4.0 → gridstate-2.4.1}/CHANGELOG.md

@@ -9,6 +9,28 @@ types of the merged pull requests; breaking and behavior-changing items are
 listed separately under "Breaking changes" and "Behavior changes". A release is
 cut by tagging the merged main commit; the version is derived from that tag.
 
+## v2.4.1 (2026-10-18)
+
+### Behavior changes
+
+- `ipm_pd`: the proximal term on the box-variable diagonal adapts to the step
+  length: ×10 after a step αp < 0.1, up to `box_prox_max` (new, default 1e-2 of
+  the median weight), ÷10 after αp ≥ 0.5, down to `box_prox`. It damps the step
+  only; the fixed point is unchanged. Estimates on the test models move by
+  ≤ 1.6e-8 p.u. V with unchanged iteration counts.
+- `ipm_pd`: a second stopping test, `fit_tol` (new, default 1e-2): the step moves
+  no data row by more than that fraction of its sigma (the μ, feasibility and
+  equality tests still apply).
+
+### Fix
+
+- `ipm_pd` no longer runs into its iteration limit and the primal fallback when
+  box variables are seen only through near-zero-weight rows (loads with
+  effectively free injections): such directions have barrier curvature only,
+  their steps dwarf the box and the boundary rule cut every step to a tiny αp.
+  A state the data do not determine no longer keeps the V/δ step test from
+  passing (it could move, or cycle, without changing any estimate).
+
 ## v2.4.0 (2026-10-17)
 
 ### Behavior changes

{gridstate-2.4.0 → gridstate-2.4.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gridstate
-Version: 2.4.0
+Version: 2.4.1
 Summary: Power system State Estimation: pure-Python WLS/IPM on numpy/scipy, input via a versioned file contract (npz).
 Author-email: Evgeny Istomin <gena@alpisto.eu>
 License: MIT

{gridstate-2.4.0 → gridstate-2.4.1}/gridstate/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '2.4.0'
-__version_tuple__ = version_tuple = (2, 4, 0)
+__version__ = version = '2.4.1'
+__version_tuple__ = version_tuple = (2, 4, 1)
 
-__commit_id__ = commit_id = 'g81767e831'
+__commit_id__ = commit_id = 'g640125491'

{gridstate-2.4.0 → gridstate-2.4.1}/gridstate/algorithms/ipm_pd.py

@@ -50,6 +50,19 @@ corrector's second-order term may spoil descent, and then the centering
 direction is taken instead (one more solve on the same factor). The dual step
 keeps its own fraction-to-boundary length.
 
+Adaptive box regularisation. A box variable seen only through rows of
+negligible weight (a load whose injection row is a near-zero-weight pseudo
+measurement) or through a combination the data do not resolve (a reactive flow
+circulating between two such loads) has no curvature but the barrier's, and
+the Newton step along it is orders of magnitude longer than the box. The
+fraction-to-boundary rule then cuts every step to ``αp ≪ 1``, the
+complementarity stops decreasing and the iteration crawls to its limit. The
+proximal term on the box diagonal is therefore raised tenfold after every
+step shorter than 0.1 (up to ``box_prox_max``) and lowered tenfold after
+every step of 0.5 or longer (down to ``box_prox``). The term damps the step
+only, so the fixed point is unchanged; a well-scaled problem takes long steps
+and keeps the base term.
+
 Start on the central path: ``s`` = half-range of a two-sided constraint (the
 distance for a one-sided one), ``y = μ0/s``; the initial point need not satisfy
 ``g − l = s``.
@@ -96,6 +109,14 @@ def _blocking_constraint(
     return ku, False
 
 
+# Adaptive box regularisation (see the module docstring): a primal step shorter
+# than _PROX_SHORT raises the term by _PROX_GROW, one of _PROX_LONG or longer
+# lowers it by the same factor.
+_PROX_SHORT = 0.1
+_PROX_LONG = 0.5
+_PROX_GROW = 10.0
+
+
 Direction = tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray, np.ndarray, np.ndarray]
 
 
@@ -212,6 +233,7 @@ def solve_ipm_pd(
     armijo_max: int = 10,
     max_search_failures: int = 20,
     step_tol: float = 1e-4,
+    fit_tol: float = 1e-2,
     feas_tol: float = 1e-6,
     max_iter: int = 100,
     box_eps: float = 1e-3,
@@ -225,6 +247,7 @@ def solve_ipm_pd(
     kkt_solver: KKTSolver | None = None,
     eq_prox: float = 1e-9,
     box_prox: float = 1e-6,
+    box_prox_max: float = 1e-2,
 ) -> IPMResult:
     """Mehrotra predictor-corrector IPM (see the module docstring).
 
@@ -251,6 +274,11 @@ def solve_ipm_pd(
             outside ``box_idx`` (V/δ) is below it. Box variables may still move
             along directions the data do not see (the common level of a node's
             generation and load), which change neither the state nor the residuals.
+        fit_tol: alternatively, stop once that step moves no data row by more
+            than ``fit_tol`` of its sigma (``max √w·|H·αp·Δx|``). A state the data
+            do not determine (a node behind loads whose injection rows carry
+            near-zero weight) can keep moving, or cycle, without changing any
+            estimated measurement; the fit has converged all the same.
         feas_tol: admissible primal infeasibility of the inequalities (p.u.).
         max_iter: iteration limit (one factorization each).
         box_eps: margin from the bounds when projecting ``x_init``.
@@ -271,6 +299,9 @@ def solve_ipm_pd(
             has no curvature along that common level but the barrier's, so a
             cold start produces huge box steps cut to tiny ``αp`` by the boundary
             rule; the term bounds those steps without moving the fixed point.
+        box_prox_max: ceiling of the adaptive box term (same units as
+            ``box_prox``, see the module docstring); a value not above
+            ``box_prox`` keeps the term fixed.
 
     Returns:
         :class:`~gridstate.algorithms.ipm.IPMResult`; ``status`` is ``"kkt"`` when
@@ -325,10 +356,9 @@ def solve_ipm_pd(
     is_eq = np.zeros(m, dtype=bool)
     is_eq[eq_rows] = True
     rho = w_base[eq_rows]
-    prox = max(
-        eq_prox * float(np.max(rho)) if rho.size else 0.0,
-        box_prox * float(np.median(w_base[~is_eq])) if (~is_eq).any() else 0.0,
-    )
+    eq_prox_abs = eq_prox * float(np.max(rho)) if rho.size else 0.0
+    w_median = float(np.median(w_base[~is_eq])) if (~is_eq).any() else 0.0
+    box_scale = box_prox
 
     huber_mask_arr = (
         np.asarray(huber_mask, dtype=bool) if huber_mask is not None else np.zeros(m, dtype=bool)
@@ -378,6 +408,7 @@ def solve_ipm_pd(
     w = w_base
     mu = mu0 if n_pairs else 0.0
     step_inf = float("inf")
+    fit_inf = float("inf")
     eq_inf = 0.0
     grad_inf = float("inf")
     status = "stalled"
@@ -399,7 +430,7 @@ def solve_ipm_pd(
             and mu <= mu_min
             and feas_inf <= feas_tol
             and bool(np.all(np.abs(c) <= eq_tol_arr))
-            and step_inf <= step_tol
+            and (step_inf <= step_tol or fit_inf <= fit_tol)
         ):
             status = "kkt"
             break
@@ -413,7 +444,7 @@ def solve_ipm_pd(
         w_mat[eq_rows] = rho
         w_mat[mb_rows] += d_k[nb:]
         diag = np.zeros(n)
-        diag[box_idx] = d_k[:nb] + prox
+        diag[box_idx] = d_k[:nb] + max(eq_prox_abs, box_scale * w_median)
         # Stationarity residual of the current point (for reporting).
         y_dual = np.zeros(nb + mb_rows.size)
         y_dual[il] += y_l
@@ -490,7 +521,8 @@ def solve_ipm_pd(
             il=il,
             iu=iu,
         )
-        slope = merit.slope(lin, H @ dx, ds_l, ds_u)
+        hdx = H @ dx
+        slope = merit.slope(lin, hdx, ds_l, ds_u)
         if slope >= 0.0 and n_pairs:
             # The corrector's second-order term can spoil descent; the pure
             # centering direction is a Newton step on the merit and descends
@@ -498,9 +530,11 @@ def solve_ipm_pd(
             dx_c, ds_l_c, ds_u_c, dy_l_c, dy_u_c, dlam_c = lin.direction(
                 target - s_l * y_l, target - s_u * y_u
             )
-            slope_c = merit.slope(lin, H @ dx_c, ds_l_c, ds_u_c)
+            hdx_c = H @ dx_c
+            slope_c = merit.slope(lin, hdx_c, ds_l_c, ds_u_c)
             if slope_c < 0.0 and np.all(np.isfinite(dx_c)):
                 dx, ds_l, ds_u, dy_l, dy_u, dlam = dx_c, ds_l_c, ds_u_c, dy_l_c, dy_u_c, dlam_c
+                hdx = hdx_c
                 slope = slope_c
                 n_centering += 1
 
@@ -521,6 +555,7 @@ def solve_ipm_pd(
             a_p = tr_radius / free_inf
         # The stopping test measures the Newton step, not how far the search cut it.
         step_inf = a_p * free_inf
+        fit_inf = a_p * float(np.max(np.sqrt(w_data) * np.abs(hdx))) if m else 0.0
         a_full = a_p
         r_new = None
         # A direction that does not descend (inconsistent slacks of a nonlinear
@@ -563,6 +598,11 @@ def solve_ipm_pd(
             eq_inf,
             feas_inf,
         )
+        if box_prox_max > box_prox:
+            if a_p < _PROX_SHORT:
+                box_scale = min(box_scale * _PROX_GROW, box_prox_max)
+            elif a_p >= _PROX_LONG:
+                box_scale = max(box_scale / _PROX_GROW, box_prox)
         x = x + a_p * dx
         s_l = s_l + a_p * ds_l
         s_u = s_u + a_p * ds_u

{gridstate-2.4.0 → gridstate-2.4.1}/gridstate.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gridstate
-Version: 2.4.0
+Version: 2.4.1
 Summary: Power system State Estimation: pure-Python WLS/IPM on numpy/scipy, input via a versioned file contract (npz).
 Author-email: Evgeny Istomin <gena@alpisto.eu>
 License: MIT

gridstate-2.4.1/gridstate.egg-info/scm_version.json

@@ -0,0 +1,8 @@
+{
+  "tag": "2.4.1",
+  "distance": 0,
+  "node": "g640125491ba6b35d3beb20b4fd33688f9cceb117",
+  "dirty": false,
+  "branch": "HEAD",
+  "node_date": "2026-10-18"
+}

{gridstate-2.4.0 → gridstate-2.4.1}/tests/test_ipm_pd.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 from types import SimpleNamespace
+from typing import Any
 
 import numpy as np
 import pytest
@@ -148,6 +149,68 @@ def test_equality_rows_with_a_box_free_common_level() -> None:
         assert out.x[1] - out.x[2] == pytest.approx(0.7, abs=1e-5)
 
 
+def test_pd_adaptive_box_prox_keeps_the_fixed_point(monkeypatch: pytest.MonkeyPatch) -> None:
+    # Every step counts as short, so the box term climbs to its ceiling and stays
+    # there: the iteration is damped, the solution is not. Rows 1-2 see the box
+    # variables (g, l) only through near-zero weights, as a free load does.
+    import gridstate.algorithms.ipm_pd as pd_mod
+
+    A = np.array([[1.0, 0.0, 0.0], [1.0, -1.0, 1.0], [0.0, 0.0, 1.0], [1.0, 0.0, 0.0]])
+    z = np.array([0.7, 0.0, 3.0, 0.75])
+    w = np.array([1.0, 1e-8, 1e-8, 1.0])
+    res, jac = _linear(A, z)
+    kw: dict[str, Any] = {
+        "box_idx": np.array([1, 2]),
+        "box_lo": np.array([-50.0, -50.0]),
+        "box_hi": np.array([50.0, 50.0]),
+        "step_tol": 1e-9,
+        "max_iter": 300,
+    }
+    fixed = solve_ipm_pd(np.zeros(3), res, jac, w, box_prox_max=0.0, **kw)
+    monkeypatch.setattr(pd_mod, "_PROX_SHORT", 2.0)
+    damped = solve_ipm_pd(np.zeros(3), res, jac, w, box_prox_max=1.0, **kw)
+    assert fixed.success and damped.success
+    assert damped.x[0] == pytest.approx(0.725, abs=1e-6)
+    np.testing.assert_allclose(damped.x[0], fixed.x[0], atol=1e-6)
+
+
+def test_pd_adaptive_box_prox_is_inactive_on_long_steps() -> None:
+    # A well-scaled problem takes long steps: the ceiling never comes into play.
+    rng = np.random.default_rng(4)
+    A = rng.normal(size=(8, 3))
+    z = rng.normal(size=8)
+    res, jac = _linear(A, z)
+    kw: dict[str, Any] = {
+        "box_idx": np.array([2]),
+        "box_lo": np.array([-5.0]),
+        "box_hi": np.array([5.0]),
+    }
+    base = solve_ipm_pd(np.zeros(3), res, jac, np.ones(8), box_prox_max=0.0, **kw)
+    adaptive = solve_ipm_pd(np.zeros(3), res, jac, np.ones(8), **kw)
+    assert base.iterations_outer == adaptive.iterations_outer
+    np.testing.assert_array_equal(base.x, adaptive.x)
+
+
+def test_pd_fit_tol_stops_once_the_estimates_settle() -> None:
+    # x1 enters only a row of negligible weight: once the measured x0 has
+    # settled, the remaining steps move x1 alone and no estimate noticeably.
+    z = np.array([1.0, 1.2, 3.0])
+    w = np.array([1.0, 1.0, 1e-12])
+
+    def residual_fn(x: np.ndarray) -> np.ndarray:
+        return z - np.array([x[0], x[0], np.sinh(x[1])])
+
+    def jacobian_fn(x: np.ndarray) -> csr_matrix:
+        return csr_matrix(np.array([[1.0, 0.0], [1.0, 0.0], [0.0, np.cosh(x[1])]]))
+
+    kw: dict[str, Any] = {"step_tol": 1e-12, "max_iter": 60}
+    by_fit = solve_ipm_pd(np.zeros(2), residual_fn, jacobian_fn, w, **kw)
+    by_step = solve_ipm_pd(np.zeros(2), residual_fn, jacobian_fn, w, fit_tol=0.0, **kw)
+    assert by_fit.success and by_step.success
+    assert by_fit.iterations_outer < by_step.iterations_outer
+    assert by_fit.x[0] == pytest.approx(1.1, abs=1e-9)
+
+
 def test_pd_merit_search_stops_gauss_newton_overshoot() -> None:
     # h = atan(x), z = 0 from x0 = 1.5: the Gauss-Newton step −(1 + x²)·atan(x)
     # overshoots and the full-step iteration diverges; the merit search does not.
@@ -157,7 +220,9 @@ def test_pd_merit_search_stops_gauss_newton_overshoot() -> None:
     def jacobian_fn(x: np.ndarray) -> csr_matrix:
         return csr_matrix(np.array([[1.0 / (1.0 + x[0] ** 2)]]))
 
-    kw = {"tr_radius": 1e6, "step_tol": 1e-10, "max_iter": 50}
+    # fit_tol off: far out on the plateau of atan a diverging step barely moves
+    # the estimate, and this test is about the search, not the stopping rule.
+    kw = {"tr_radius": 1e6, "step_tol": 1e-10, "fit_tol": 0.0, "max_iter": 50}
     no_search = solve_ipm_pd(
         np.array([1.5]), residual_fn, jacobian_fn, np.ones(1), armijo_max=0, **kw
     )

gridstate-2.4.0/gridstate.egg-info/scm_version.json

@@ -1,8 +0,0 @@
-{
-  "tag": "2.4.0",
-  "distance": 0,
-  "node": "g81767e831d3e3cddd3b312090767b2568d2c26bd",
-  "dirty": false,
-  "branch": "HEAD",
-  "node_date": "2026-10-17"
-}