Skip to content

CHANGELOG.md

Latest commit

 

History

History
553 lines (425 loc) · 26.6 KB

CHANGELOG.md

File metadata and controls

553 lines (425 loc) · 26.6 KB
Raw

bertini2 Changelog

All notable changes to this project will be documented in this file.

The format is based on CHANGELOG.md and this project adheres to Semantic Versioning.

[3.4.0] - 2026-07-16

A one-call way to build a linear slice that passes through a chosen point, an endgame-hardening sweep (power series and Cauchy both), and friendlier start-point handling in the Python layer.

Added

  • Slice.through_point(variables, point, dim=1, coefficients=None, real=False, orthogonal=True, homogeneous=False) — the single place to make a slice through a given point (#343). With coefficients=None (the default) it draws a random block of dim linear forms (complex, or real with real=True; orthonormalized when orthogonal) and sets the constant column so every form vanishes at point; pass coefficients (a bare, non-augmented block) to use exactly those directional coefficients. Backed by the new C++ primitives Slice::ThroughPoint and a through_point option on Slice::RandomReal / Slice::RandomComplex. homogeneous=True builds a projective slice through the point (rows orthogonal to it); it is random-only.
  • HomotopySolver accepts start points in any faithful numeric representation (#350, fixing #347): multiprecision scalars, Python/numpy numbers, and constant symbolic nodes (an ndarray of symbolics.Complex previously crashed with a raw eigenpy converter error). Start points are transported values — the tracker refines them — so lossy doubles are welcome here; expressions still containing variables are refused as the math errors they are, with an error that names the variables. complex_mp now constructs explicitly from a Python complex (still no implicit conversion).

Fixed

  • The power-series endgame's Hermite interpolation now evaluates the actual Hermite interpolant (#353). The Horner walk over the doubled node list advanced at half speed, evaluating a different (lower-order) interpolant — the limit was still correct, but convergence order was degraded. This changes computed results at agreeing inputs: approximations land measurably closer to the truth (the old test oracle values were themselves off and have been re-pinned exactly).
  • max_cycle_number is enforced as a ceiling, not a floor (#353). The bound was applied with max(), and a near-unity sample ratio could push the estimate through an unsigned conversion of infinity (UB). Clamped before conversion; cycle-number candidates now default sanely on degenerate samples.
  • NaN is a failure, never Converged (#354). IEEE comparison semantics made every NaN comparison false, so a NaN correction step exited the convergence loop as success, and the security valve (norm > max_norm) was blind to NaN norms. Both endgames now fail fast on NaN approximations (new bertini::ContainsNaN in eigen_extensions.hpp — component-wise, because multiprecision complex NaN compares equal to itself) and the valve is NaN-aware.
  • Slow divergers truncate honestly in the Cauchy endgame (#355). Paths diverging to infinity slower than the shrinking time zones could grind precision escalation for minutes before dying. The security valve now also arms below B1's cycle_cutoff_time and watches the loop floor — the minimum dehomogenized norm over the Cauchy loop samples — which exceeds max_norm only when the entire loop is beyond it (single-sample spikes on legitimate paths cannot trip it).
  • Singularity classification uses the endpoint's spectral-norm condition number (#344, the B1 CondNumThreshold spec), instead of a mixed-norm estimate that mislabeled borderline endpoints.
  • frequency_of_CN_estimation was inert (#345): the tracker's condition-number refresh counter was passed by value, so the estimate never refreshed at the configured cadence.
  • max_precision_used is harvested from failed endgames too (#349); previously a path that failed after escalating precision reported as if it had never left double.

Changed

  • Binomial start points are computed at working precision (#346) — about 3.2× faster on small total-degree solves, with start-point accuracy unchanged (start points are transported values).

[3.3.2] - 2026-07-13

A metadata_for robustness patch (feed a solver result straight back in and it resolves), plus a documentation showcase and a friendlier docs site.

Fixed

  • metadata_for(point) no longer reports a spurious "more than one distinct solution cluster" for a point taken from the solver's own results (#338). It now matches against the multiplicity REPRESENTATIVES (one per distinct solution) rather than the complete set of coincident copies, and defaults the match tolerance to final_tolerance (the accuracy each endpoint is computed to) instead of the looser same-point clustering tolerance. Because representatives are at least the same-point tolerance apart, a final_tolerance window holds at most one — so a solution fed straight back resolves to its representative (carrying its .multiplicity) and can never be called ambiguous, even at a tolerance far tighter than the clustering scale. This bit hardest on singular points. See ADR-0056.

Added

  • solver.same_point_tolerance() — the clustering tolerance (final_tolerance × same_point_tolerance_multiplier), the scale the solver uses to group coincident endpoints into multiplicities. default_point_match_tolerance() now returns final_tolerance.
  • metadata_for(..., representatives_only=False) — a debugging view that matches against every endpoint, including non-representative multiplicity copies.

Documentation

  • New "Showpieces" gallery (#341): beautiful renders generated entirely from real tracked data. The Monodromy Loom — a 3-D braid of solution paths as a parameter loops the discriminant, showing a family's monodromy (its Galois action). The Flight Recorder — one hard path to a multiplicity-35 singular point with the adaptive-precision tracker's full telemetry (endgame spiral, precision staircase into mpfr, condition blow-up, step-size sawtooth), plus a system-level view of all 35 paths converging.
  • bertini2.org lands on the current release (#340): the docs-site root now redirects straight to the latest version instead of a version chooser (which moves to /versions.html).

[3.3.1] - 2026-07-13

A bindings-robustness patch.

Fixed

  • Storing a multiprecision complex into a real numpy array no longer crashes the interpreter (#336). Assigning a complex_mp that carries a nonzero imaginary part into a float64 array — e.g. M = np.zeros(...); M[i, j] = solver.real_solutions()[k][0], whose solution coordinates carry ~1e-13 imaginary noise — routed through a complex_mp → double cast that threw a C++ exception ("Could not convert imaginary number to scalar.") out of numpy's C cast loop, calling std::terminate() → SIGABRT. The cast now mirrors numpy's builtin complex128 → float64 behavior (keep the real part, discard the imaginary part). See ADR-0055.

Internal

  • The eigenpy_numpy numpy-interop test suite was never collected by pytest (its filename matched neither test_*.py nor *_test.py); renamed to eigenpy_numpy_test.py so it runs, and added regression coverage for the mp → narrower-dtype casts.

[3.3.0] - 2026-07-12

A large pass on the symbolic and solve/result ergonomics: symbolic substitution and richer differentiation on the function tree, a friendlier solve/settings surface, and a typed result taxonomy — plus internal tolerance typing and a documentation refresh.

Added

  • Symbolic subs, richer differentiation, eval ergonomics (#327): leaf-level node.subs({var: expr}) symbolic substitution; differentiate(x, 2) / differentiate([x, x, y]) repeated / sequence differentiation; positional eval with an explicit coordinate ordering and a strict=False option; constant-power folding (so I**2 → -1); node.simplify(); and a Python __repr__ that renders **.
  • solver.solve() returns a SolveResult (#330), and solver.result() re-derives it: the bare ZeroDimSolver / HomotopySolver record themselves, so they hand back the same records-aware result that bertini.solve returns.
  • ZeroDimResult (#331): the typed answer of a zero-dimensional solve — the distinct finite solutions plus real / singular / nonsingular / at_infinity / nonsolutions views. SolveResult is now a records decorator around .answer; both ZeroDimSolver and HomotopySolver produce one.
  • Config settings at solve / construction (#330): solve(**settings) and ZeroDimSolver(..., **settings) (keywords or a settings={...} dict), plus get_settings(as_dict=True) for a flat {field: value} view.
  • Random symbolic constants (#330): random_real(symbolic=True) / random_complex(symbolic=True) / random_vector(..., symbolic=True) and symbolics.random_real() / random_complex() produce random constant nodes.
  • metadata_for(pt) needs no explicit tolerance (#330): it defaults to the solver's own same-point tolerance (default_point_match_tolerance()).

Changed

  • precision= is now an integer number of digits; mptype= is the precision model ('double' / 'multiple' / 'adaptive') everywhere — ZeroDimSolver, bertini.solve, HomotopySolver, user_homotopy (#330). A string precision= is still honored as the old model alias for one release, with a DeprecationWarning.
  • Clearer errors: a SymPy expression passed to add_function now points at bertini.sympy_bridge.from_sympy; bertini.solve on a positive-dimensional (under-determined) system raises a clear "needs numerical irreducible decomposition, not yet implemented" instead of a raw solver error (#330, #331).
  • NumErrorT for tolerance typing (#329): the error / tolerance type now lives in num_traits.hpp and types the tolerance parameters across the trackers, endgames, and point-comparison helpers (readability only — NumErrorT is double).
  • An informative repr(solver) replaces the default object line (#330); the CLI splash drops its stale primary-authors block (#332).

Fixed

  • A Complex (constant) node is accepted as a coefficient (#326, #328), matching the existing complex_mp behavior.
  • Publishing downloads only the wheel artifacts (#325): the Doxygen docs-cpp artifact no longer leaks into the release upload.

Documentation

  • The classic continuation-cartoon figure now auto-fits the finite paths, uses a total-degree linear-product start so start points do not overlap, and marks divergences with ∞ (inside the axes) and the start system with a green play triangle (#329).

[3.2.0] - 2026-07-10

Multiprecision linear algebra and a more flexible randomization on the library side, plus a substantial CI / build-infrastructure pass that markedly shortens the release cycle.

Added

  • bertini.linalg (#317): multiprecision linear algebra — LU, QR, and SVD — exposed by instantiating eigenpy's own decomposition visitors on the real_mp / complex_mp matrix types (reuse, not re-export). One dtype-agnostic surface that also handles float64 / complex128 via NumPy. See ADR-0054.
  • bertini.precision(A, n) (#317): set the working precision of an entire vector or matrix in a single call.
  • Randomization to any codimension (#315): System.randomize(codimension=…) and the bertini.randomize(system, codimension) free function generalize the square case; all prior randomize() calls are unchanged. See the ADR-0025 amendment.

Changed

  • Windows drops the /WHOLEARCHIVE link workaround (#287): the ExplicitRKPredictor Butcher tables are now C++17 inline members in the header, so the Windows test executables link normally. See ADR-0052.
  • Faster CI: wheel builds now run concurrently with the C++ tests — they are independent full compiles that shared no artifacts — and the Doxygen build runs in parallel with the wheel in the docs workflow (#316, ADR-0053). The Windows C++ test build now runs through ccache (#322).
  • A wheel-free Python docstring lint (tools/py_doclint.py) now runs beside the C++ Doxygen lint as part of the cheap doc-lint gate (#316).

Fixed

  • Docstrings and tutorials: escaped absolute-value bars that reStructuredText misread as substitution references, which had broken the documentation build (#313).

[3.1.0] - 2026-07-09

Quality-of-life and correctness release on top of 3.0.0: a thorough NumPy interoperability pass for the multiprecision dtypes, a batch of Python UI ergonomics discovered while writing a real-cellular-decomposition notebook, a records-recall escape hatch, and CI / documentation-infrastructure work.

Added

  • NumPy interoperability for real_mp / complex_mp (#306): full ufunc coverage, sorting slots, safe reductions (sum, prod, …), and tolerance-based comparisons against float64; element access returns owned copies. See ADR-0051.
  • Python UI ergonomics (#293–#304, #305):
    • list form x, y, z = bertini.variables(['x', 'y', 'z']) and variadic System.add_variable_group(x, y, z);
    • random factories random_real(), random_complex(), random_vector(n, real=…), all visible under bertini.random;
    • bertini.is_distinct_up_to(p, q, tol) — infinity-norm point comparison, accepting multiprecision and double vectors alike;
    • merge_multiplicities= on every solution accessor (default True);
    • solver.metadata_for(point) with a call-shape-determined return type;
    • System.functions(), System.copy_functions(), System.clone();
    • a group= projection kwarg on every solution getter and to_dataframe(), plus the System.coordinates_of(point, group) primitive behind it;
    • a sympy auto-sympify bridge, so sympy.Matrix([...nodes...]).det() works;
    • node.eval(point | dict | array) (previously keyword-only);
    • variable_group @ coefficients dot-product sugar;
    • NumPy-native helpers bertini.real / imag / abs / conj / round / sum / norm / is_real.
  • ZeroDimConfig.recall (#308, default True): set False to force a fresh re-track even when an identical ask is already recorded — the escape hatch for path observers, benchmarking, and re-verification. Transient: it does not affect the run's identity digest.

Changed

  • Solution accessors now merge multiplicities by default; pass merge_multiplicities=False for the raw per-path endpoints (#299).
  • CI builds against prebuilt dependency artifacts — a custom manylinux image plus macOS Boost / eigenpy tarballs (ADR-0049, #282) — cutting build time and flakiness.
  • Documentation is served from a branch-source docs-store with per-version snapshots (ADR-0050, #291, #292).

Fixed

  • Random seeding: set_random_seed now governs every draw, and real projection directions come out actually real (#294).
  • NumPy 2.5 reduction use-after-free / uninitialized-slot hazards in the eigenpy bindings (#306).
  • Windows wheel builds no longer spuriously fail on a precompiled-header mtime race (-fno-pch-timestamp, #306).

[3.0.0] - 2026-07-07

The first stable release of the modernized Bertini 2: a rebuilt C++17 core and a much friendlier Python interface (import bertini), with the bertini2 CLI shipped inside the wheel. Wheels for CPython 3.10 – 3.14 on Linux (manylinux_2_34), macOS (arm64), and Windows. This release consolidates ~70 internal PRs; the full themed index, per-PR links, and the upstream issues it closes are in #238.

Added

  • Durable, resumable output with provenance. Every solve writes a plain-text structured output directory — content-addressed inputs, one results file per run, walkable provenance chains. Solves consult it first, so a killed run finishes on rerun instead of restarting. Content-identity digests mean seed=42 reproduces the exact same homotopy forever, across machines and versions.
  • Content-addressed function trees and systems — hash-consing / interning, symbolic Jacobian, Seal() — so equal objects share an identity.
  • Block-structured systems, with the multihomogeneous start system exposed to Python; user-defined homotopies with start points; automatic square-up and filtering of overdetermined inputs; first-class randomization.
  • sympy bridge — exact two-way conversion and round-trip solving.
  • to_dataframe() for pandas; Unicode / emoji identifiers and string-valued config fields.
  • The bertini2 CLI ships inside the wheel (bertini2 on your PATH after pip install), emitting Bertini 1.7-compatible solution files.
  • Executable, doctest-verified tutorials; a C++ documentation-lint gate; and ADRs for load-bearing design decisions. Docs at https://bertini2.org.

Changed

  • Flattened, discoverable public Python API; build systems from a list of functions (System.add_functions) and get solutions back in your own coordinates by default; a revamped configuration model.
  • Parallel-by-default solving on shared memory — no MPI required — with MPI serialization and reproducibility fixes for multihomogeneous solves at scale.
  • ~10× faster multiprecision evaluation (tiered SLP arithmetic, allocation-free eval, common-subexpression elimination, a stateful in-place multiprecision LU solver) and 5 – 7× faster well-conditioned adaptive-precision solves by staying in double precision where it is provably safe (cyclic-5: 11s → 3.5s).
  • Rewritten predict / correct (per-track condition probe, pure kernel); adaptive-numeric-type AMP endgames.

Fixed

  • Cauchy endgame security and pole-zone guards — never reports success at a non-root.

[2.0.2] - 2026-05-22

Packaging and CI maintenance following the 2.0.1 PyPI debut.

Changed

  • Factored out the documentation workflow and added a ref input for tag rebuilds (#223, #224).
  • Extended the Python support matrix in CI and updated cibuildwheel (#229).
  • Minor 2.0.1 follow-up fixes and MPI sync force-push handling (#226, #227).
  • Version bump to 2.0.2 and README Python-version refresh (#225, #230).

Full changelog: https://github.com/bertiniteam/b2/compare/v2.0.1...v2.0.2

[2.0.1] - 2026-05-16

First release under the bertini2 PyPI name. This is the consolidation of several months of cross-platform packaging work, dependency-compatibility fixes, documentation, and a final round of precision-handling fixes in the system / SLP path. Intel macOS is dropped from the supported platform list for this release (see Removed below).

Added

  • Wheel distributions on PyPI for Linux, macOS (Apple Silicon), and Windows across Python 3.9 – 3.13. Linux wheels use the manylinux_2_28 image; macOS and Linux wheels are produced via cibuildwheel; Windows wheels bundle required DLLs via delvewheel and a windows_dll_manager.py helper that registers DLL search paths before importing _pybertini.
  • GitHub Pages documentation site: C++ API via Doxygen, Python API via Sphinx.
  • Version numbers (b2, GMP, MPFR, Eigen, Boost) exposed from the Python package so users can query them at runtime.
  • CHANGELOG.md itself, following the Keep a Changelog format.

Changed

  • Package renamed: pybertini → bertini → bertini2. The current PyPI name is bertini2. Update your pip install and import statements accordingly; the Python module still imports as import bertini.
  • Version is now read from pyproject.toml by CMake, removing the two-place manual sync that previously drifted.
  • Build system: scikit-build-core configures CMake for the wheel build; pyproject.toml is the single source of truth for build inputs. The Linux wheel build rebuilds Boost.Python and eigenpy per target Python version so each wheel ships a matching libboost_python3X.
  • CI matrix expanded to cover Ubuntu, macOS (Apple Silicon), and Windows for Python 3.9 – 3.13. PRs targeting develop and pushes to develop / main run the full matrix; other branches run a fast Ubuntu + macos-14 + Python 3.11 matrix.
  • CI build now uses Boost 1.90 (was 1.87). Boost 1.90 pre-seeds thread_default_precision from the global default and guards against 0, removing a class of MPFR abort hazards.
  • boost_system is now conditionally linked for Boost < 1.89 only (Boost 1.89 dropped the separate library).
  • Eigen requirement updated to 3.3...3.4 with the macOS install switched to Homebrew's eigen@3 formula.
  • Tests across platforms now use a unified precision-handling pattern, eliminating per-platform skips and conditional precision rituals.
  • MACOSX_DEPLOYMENT_TARGET co-varies with the runner version so wheels produced on macos-14 are loadable on the same OS family.
  • TestPyPI publishes on every push to develop; PyPI publishes on tagged releases (v*.*.*) with Sigstore signing and a GitHub Release.

Fixed

  • System::operator+= and System::operator*= now invalidate the cached derivatives flag (is_differentiated_), so a subsequent eval rebuilds the SLP instead of evaluating against stale derivatives. The companion mutators Reorder, Simplify, and ClearVariables also invalidate the cache.
  • SLPCompiler::Compile now seeds the mpfr default precision from the SLP's own precision_ immediately before growing the mpfr_complex memory block. This prevents a mpfr_init2(x, 0) abort when thread_default_precision() is left at 0 on a fresh thread under Boost ≥ 1.87.
  • Python module init seeds thread_default_precision so the first MPFR construction on the interpreter thread is always valid.
  • eval overload registration order in the Python bindings changed so the double-precision overload is tried before the multi-precision overload for ambiguous inputs (e.g. NumPy int64 arrays); this avoids the eigenpy Vec<mpfr> extractor probing MPFR construction with precision 0.
  • EIGEN_MAKE_ALIGNED_OPERATOR_NEW moved to the public section of the System class (was incorrectly placed in a non-public section).
  • A second find_package(Boost) no longer clobbers Boost_LIBRARIES.
  • The _pybertini target is always created (previously it was only created when bertini was the top-level CMake project, breaking out-of-tree builds).
  • Various MSVC-specific build fixes: /bigobj, /EHsc, explicit-type workarounds for template instantiation issues, Release-config library linking.

Removed

  • Intel macOS (macos-15-intel) is not built or tested in this release. A SIGABRT in pytest involving MPFR/Boost surfaces only on Intel runners and cannot be reproduced on the maintainer's development hardware. Intel wheels will return once a reproducer or upstream fix is in hand. Users on Intel Macs should pin to a 1.0.x release or build from source.
  • Stale build artifacts and outdated Python-binding documentation removed from the repo.

[1.0.3] - 2025-05-16

Preparation for pypi release with github workflow

Changed

  • make it compatible for Windows

[1.0.2] - 2025-05-07

Preparation for pypi release with github workflow

Added

  • github workflow for pypi and github release

Changed

  • publish-to-test-pypi.yml for handling the comments correctly

Changed

  • merged the pull request for github ci release by @hkmoon in hkmoon#1
  • windows release preparation
    • size_t is translated into unsigned long in linux, mac while unsigned long long in windows 10: core/include/bertini2/eigen_extensions.hpp and core/test/classes/start_system_test.cpp are modified
    • use clang of LLVM in Windows since MSVC has different compiling way for template
    • use --no-isolation for scikit-build in Windows
  • For linux wheel naming convention, we cannot use x86_64, x86_i386 anymore for pypi repository. https://peps.python.org/pep-0600/
    • use auditwheel for it

New Contributors

  • @hkmoon made their first contribution in hkmoon#1

[1.0.1] - 2025-05-06

This is the initial version of the project.

Added

  • The base project