GitHub

Changelog

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

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

2.6.1 (unreleased)

Fixed

  • In-place broadcasting into a wrapped tangent vector that stores a matrix, such as X .+= Y, works on Julia 1.10.

2.6.0 (25/09/2026)

This release is the result of an LLM-based code analysis to find inaccuracies, typos and similar small bugs.

Added

  • check_geodesic takes a keyword name (default "geodesic") that appears in its message and as the title of its plot; plot_check_geodesic takes name as well.
  • DocumenterCodeBlocks.jl plugin added to the documentation
  • DocumenterLandingPage.jl enhances the start page with a short teaser for the package now.
  • export EfficientEstimator, like the other approximation methods.
  • get_component and set_component! on a ProductManifold accept a Colon and a vector of indices, so p[M, :] and p[M, [1, 2]] and the corresponding assignments work.
  • is_flat(::VectorSpaceFiber), so a CotangentSpace and any user-defined vector space fiber report the flatness Fiber documents; before only TangentSpace had a method.
  • manifold_dimension(::Fiber), so a CotangentSpace and any other fiber report their fiber_dimension; before only TangentSpace had a method.
  • rand(rng, M::ValidationManifold; vector_at=), mirroring the method without a random number generator.
  • scalar multiplication from the right for FVector and ZeroVector, mirroring the existing a * X methods.
  • the test suite runs JET.jl on the whole package with all extensions loaded, on released Julia versions.

Fixed

  • angle clamps the acos argument to [-1, 1], which rounding could exceed for nearly parallel vectors.
  • change_representer!, change_metric! and Weingarten! now have a PowerManifoldNestedReplacing method; before they reached _write, which that representation does not define.
  • check_geodesic compares the mean speed with ‖X‖/(N-1), the second check its docstring lists; a geodesic at the wrong speed is now rejected.
  • check_point on DefaultManifold accepts keyword arguments, so is_point(M, p; atol=) reaches it instead of the generic check.
  • check_vector of ManifoldsBase.Test.TestSphere takes an atol, defaulting to sqrt(prod(representation_size(M))) * eps.
  • coordinate_eltype for quaternionic manifolds floats the element type, so an integer-valued point gets QuaternionF64 coordinates like the real and complex cases.
  • @default_manifold_fallbacks generates the diff and embedded vector transport forwardings with their method argument.
  • distance(M, p, q, r) and norm(M, p, X, r) on a power manifold seed their reduction in real(float(number_eltype(p))) for r = 1 and r = Inf; on a complex power manifold the first errored and the second returned a complex value.
  • embed_project and embed_project! on a ValidationManifold run their checks within = embed_project, so ignore_functions entries for embed_project apply.
  • get_basis on a ProductManifold splits its point with submanifold_components(M, p), so point types that implement only the two-argument form work as well.
  • get_coordinates! and get_vector! default their basis to default_basis(M, typeof(p)), matching the allocating variants; before they hard-coded DefaultOrthonormalBasis(), so the two disagreed for any manifold specializing default_basis.
  • get_coordinates! and get_vector! on a manifold with an EmbeddedForwardingType call the generic implementation, as their allocating variants do.
  • get_coordinates/get_vector with a DefaultOrthogonalBasis now allocate and call their mutating orthogonal variant instead of forwarding to the orthonormal one; the orthogonal-to-orthonormal fallback now only happens in get_coordinates_orthogonal!/get_vector_orthogonal!.
  • get_forwarding_type_embedding no longer inverts the embedding directness, and the three AbstractEmbeddingType constructors default to DirectEmbedding(); a manifold declaring DirectEmbedding() is now forwarded directly, as documented.
  • get_vectors on a ProductManifold returns basis vectors with their own storage instead of sharing arrays with each other and with the cached basis.
  • injectivity_radius(M::AbstractPowerManifold, m) forwards m to the wrapped manifold.
  • _inverse_retract! for ShootingInverseRetraction accepts kwargs... and forwards them to the retraction of its loop.
  • is_default_connection(M::ConnectionManifold) compares connection(M.manifold) with M.connection instead of returning true.
  • log! on a TangentSpace copies through the base manifold, so on a nested power manifold the buffers of the destination are kept.
  • manifold_dimension on an AbstractDecoratorManifold that decorates nothing throws a MethodError instead of recursing into a StackOverflowError.
  • ManifoldsBase.Test.TestPowerManifoldMultidimensional defines _read, so functions on a power manifold with this representation work instead of raising a MethodError.
  • NLSolveInverseRetraction stores project_point and project_tangent in their own fields.
  • parallel_transport_to! on a power manifold calls parallel_transport_to! on the base manifold elementwise, instead of its default vector transport.
  • project! on a PowerManifoldNestedReplacing writes to the result instead of the point.
  • rand(::ValidationManifold; vector_at=) unwraps vector_at.
  • requires_caching(::VeeOrthogonalBasis) is false, as for the DefaultOrthogonalBasis it forwards to.
  • retract_embedded!, retract_embedded_fused! and inverse_retract_embedded! embed with embed(M, ...), not embed(get_embedding(M), ...).
  • retract_fused includes t in its allocate_result, as exp_fused does; before a fused retraction allocated in the element type of p and X alone.
  • retract_fused, the two _retract_fused methods, the SasakiRetraction and StabilizedRetraction layer-2 and layer-3 methods, and inverse_retract_embedded! accept and forward kwargs..., so RetractionWithKeywords and InverseRetractionWithKeywords reach the last layer instead of raising a MethodError.
  • sectional_curvature on a power manifold weights each component by its Gram determinant and normalizes by that of the power manifold, as on a ProductManifold.
  • sectional_curvature on a ProductManifold weights each factor by its Gram determinant and normalizes by that of the product; it summed them unweighted before.
  • sectional_curvature_min on a power manifold uses the minimum, not the maximum, of the wrapped manifold.
  • ShootingInverseRetraction runs max_iterations iterations; the loop guard was strict, so max_iterations = 1 shot not at all and n gave the accuracy of n - 1.
  • show of a CachedBasis on a power manifold prints the basis type instead of constructing it, which failed for any basis type with a field, for example DiagonalizingOrthonormalBasis.
  • the :Point context in ignore_contexts also covers the base point check within is_vector on a ValidationManifold.
  • the allocating project(M::EmbeddedManifold, p, X) allocates in the representation size of the base manifold, not of the embedding.
  • the allocating vector_transport_to wraps keywords in VectorTransportWithKeywords instead of forwarding them to the keyword-less vector_transport_to!.
  • the retraction, inverse retraction and vector transport methods that @default_manifold_fallbacks generates, and those of ManifoldsBase.Test for DefaultManifold, accept keyword arguments and pass them on; before, a keyword argument raised a MethodError for wrapped point types.
  • the scaled retractions retract_{project,polar,qr,softmax,pade}_fused!, exp_fused, exp_fused! and vector_transport_to_embedded! accept keyword arguments and pass them on.
  • ValidationCotangentVectors get their stored base point updated by the in-place functions of a ValidationManifold.
  • vector_transport_direction_embedded! embeds the direction as a tangent vector.
  • zero_vector! on a ValidationManifold no longer forwards validation keywords to the wrapped manifold.

Changed

  • inverse_retract and inverse_retract! accept keyword arguments, like retract and retract! already did.
  • plot_slope takes a keyword name (default "") that both plotting backends use as the title of the plot, and the check functions pass their name on.
  • vector_transport_direction_embedded! computes the end point with the keyword retraction_method=default_retraction_method(M, typeof(p)) instead of exp, and passes all other keywords to the vector transport in the embedding.

2.5.1 (02/09/2026)

Fixed

  • prepare_check_result no longer fits the slope through errors that are down to round-off from evaluating the function; these no longer decrease with t and flattened the estimate, the more so the higher the order.
  • prepare_check_result now passes slope on to find_best_slope_window, which before looked for its default slope of 2.0.
  • prepare_check_result now passes line_base to plot_slope as the log10-intercept it is documented to be, anchored at the first fitted sample.

2.5.0 (08/07/2026)

Added

  • a ZeroVector type, to allow for static dispatch and avoid allocating unnecessary zeros similar to the Identity on LieGroups.jl.

2.4.0 (03/06/2026)

Added

  • an extension to also perform the numerical checks plots with Makie.jl (#272).

Fixed

  • documentation of the injectivity_radius function now displays clearer instructions for the implementation of a radius linked to a user's own AbstractRetractionMethod. (#270)

2.3.5 (03/04/2026)

Changed

  • moved Julia package dependency action to using dependabot
  • Bump compat of RecursiveArrayTools to include 4
  • unify the TestUtils and TestSuite package into one ManifoldsBase.Test submodule the same way it is done in Manifolds.jl
  • setup a CI to check for typographical errors with typos.

2.3.4 (25/03/2026)

Changed

  • submanifold_component with an integer index on ArrayPartition now skips creating Val objects and directly accesses the component. This is a more efficient way to access components of an ArrayPartition.

2.3.3 (21/03/2026)

Fixed

  • allocate_result in certain cases on a ProductManifold no longer returns an object of an incorrect type.

2.3.2 (25/02/2026)

Changed

  • get_vector! on AbstractPowerManifold now works faster by constructing views of the vector of coordinates instead of slicing it.

2.3.1 (14/02/2026)

Added

  • allocate(::T) where {T<:Number} method to allocate numeric scalars, which is used as a fallback for allocate when the input is a number. The method returns a zero-index array with element of type T.

2.3.0 (09/12/2025)

Added

  • Moved MetricManifold, AbstractAffineConnection, ConnectionManifold, LeviCivitaConnection, DefaultMetric, types from Manifolds.jl
  • rand! method that defaults to Random.default_rng() source.

2.2.2 (02/12/2025)

Added

  • ApproximateExponentialRetraction and ApproximateLogarithmicInverseRetraction as types to specify approximation methods for exp and log, which due to their approximation are also only approximate retractions and inverse retractions, respectively. (#248)

2.2.1 (13/11/2025)

Added

  • check_geodesic function to numerically check whether geodesics have constant speed and are compatible with log and parallel transport.

2.2.0 (04/11/2025)

Fixed

  • bugfixes on get_embedding, get_embedding_type, and get_forwarding_type, which now all consistently take the type of a point as their second (third for the last case) argument (and no longer sometimes a point sometimes its type)
  • has_components no longer propagates to the embedding

2.1.0 (30/10/2025)

Added

  • a StabilizedInverseRetraction that improves numerical stability of another inverse retraction by projecting the resulting tangent vector onto its tangent space.

2.0.2 (23/10/2025)

Added

  • Reference to embed_project in project documentation.
  • Expanded docs of inner.

2.0.1 (17/10/2025)

Added

  • get_vector, get_coordinates, project and their mutating variants now use traits for propagation.

2.0.0 (02/10/2025)

While this release should be mostly backward compatible, especially when using defined manifolds from Manifolds.jl, some breaking changes were introduced. To be precise, defining and using traits e.g. to dispatch functions to the embedding changed internally. If you defined your own manifolds and used traits, please check the documentation of the new trait

Added

  • an interface for quotient manifolds. this also unified the naming a bit. Formerly differential_canonical_project is now diff_canonical_project.

Changed

  • refactor the trait system to no longer use a list of traits but single traits separately for the metric and the embedding specification
  • Switch to using Runic.jl as code formatter

Removed

  • ODEExponentialRetraction was removed in favor of solve_chart_exp_ode implemented in Manifolds.jl.

1.2.0 (08/05/2025)

Added

  • tangent_vector_type for converting point types to matching tangent vector types.

1.1.0 (29/04/2025)

Added

  • default_basis(M) to be more flexible than a fixed DefaultOrthonormalBasis default.
  • StabilizedRetraction, a retraction that improves numerical stability of another retraction by projecting the resulting point.

1.0.3 (08/04/2025)

Changed

  • VectorSpaceFiber no longer requires the number system to be consistent with the wrapped manifold.

1.0.2 (07/04/2025)

Changed

  • Fiber no longer requires the number system to be consistent with the wrapped manifold.

Added

  • allocate method that works with numeric scalars.

1.0.1 (05/02/2025)

Fixed

  • An issue with allocation type promotion in exp_fused.

1.0.0 (05/02/2025)

Changed

  • to avoid logical ambiguities to the forthcoming LieGroups.jl, the “fusing” variant exp(M, p, X, t) has been moved to its own name exp_fused(M, p, X, t) and similarly exp!(M, q, p, X, t) has been moved to its own name exp_fused!(M, q, p, X, t). Note that the new exp_fused! method is not exported and by default falls back to calling exp! with t*X. Actions to take
    • if you just implemented an own exp(M, p, X) or exp!(M, q, p, X) everything works as before.
    • if you implemented a fused variant exp!(M, q, p, X, t) you have to adapt two things
      1. move that implementation to ManifoldsBase.exp_fused!(M, q, p, X, t)
      2. Implement the default exp!(M, q, p, X) = ManifoldsBase.exp_fused!(M, q, p, X, one(eltype(p))),
      or an own specific implementation for the non-fused variant.
  • Similar to exp, the “fusing” variant retract(M, p, X, t, m) has been moved to its own name retract_fused(M, p, X, t, m) and similarly retract!(M, q, p, X, t, m) has been moved to its own name retract_fused!(M, q, p, X, t, m). Note that the new retract_fused! method is not exported and by default falls back to calling retract! with t*X. Actions to take
    • if you just implemented an own retract(M, p, X, m) or retract!(M, q, p, X, m) everything works as before.
    • if you implemented a fused variant retract!(M, q, p, X, t, m) you have to adapt two things
      1. move that implementation to ManifoldsBase.retract_fused!(M, q, p, X, t, m)
      2. Implement the default retract!(M, q, p, X, m) = ManifoldsBase.retract_fused!(M, q, p, X, one(eltype(p)), m), or an own specific implementation for the non-fused variant.
  • the TVector type has been renamed to AbstractTangentVector
  • the CoTVector type has been renamed to AbstractCotangentVector

Removed

  • parallel_transport_along(M, p, X, c), vector_transport_along(M, p, X, c, m) as well as their mutating variants are removed from the API for now. It was never specified how to actually specify a curve c and the method was only implemented for Euclidean in Manifolds.jl, where it is the identity.

0.15.24 (17/01/2025)

Added

  • extended support for allocate_on with ArrayPartition.

0.15.23 (09/12/2024)

Added

  • a field point to ValidationFibreVector to potentially store the point of the vector.
  • a field store_base_point to ValidationManifold to indicate whether for new fibre vectors the base point should be stored.
  • a keyword ignore_contexts to ValidationManifold to ignore certain contexts from validation, such as :Input, :Output, :Point, or :Vector.
  • a keyword ignore_functions to ValidationFibreVector to ignore certain contexts within a single function. This is provided as a dictionary with the key being the (allocating) function and the value is a context or vector of contexts.

Changed

  • the internal function array_value was renamed to internal_value and is now exported, since it can be also used on elements that store values different from arrays,

e.g. a ValidationMPoint storing a subtype of a ManifoldPoint. array_value is hence deprecated.

  • Minimum Julia version is now 1.10 (the LTS which replaced 1.6)

0.15.22 (15/11/2024)

Added

  • DefaultOrthonormalBasis() is now the default basis for get_vector, get_vector!, get_vectors, get_coordinates and get_coordinates!.

0.15.21 (12/11/2024)

Fixed

  • Coordinate allocation was improved to be more friendly with automatic differentiation.

0.15.20 (24/10/2024)

Changed

  • norm function on VectorSpaceFiber (such as TangentSpace) now needs to be called without the point. The passed point was already ignored before.

0.15.19 (20/10/2024)

Changed

  • make has_components introduced in the last version a decorator trait function.

0.15.18 (18/10/2024)

Added

  • distance(M, p, q, r) to compute r-norms on manifolds that have components.
  • distance(M, p, q, m, r) to compute (approximate) r-norms on manifolds that have components using an AbstractInverseRetractionMethod m within every (inner) distance call.
  • norm(M, p, X, r) to compute r-norms on manifolds that have components.

0.15.17 (04/10/2024)

Changed

  • Mildly breaking: the number system parameter now corresponds to the coefficients standing in front of basis vectors in a linear combination instead of components of a vector. For example, DefaultOrthonormalBasis() == DefaultOrthonormalBasis(ℝ) of DefaultManifold(3, field=ℂ) now has 6 vectors, and DefaultOrthonormalBasis(ℂ) of the same manifold has 3 basis vectors.

0.15.16 (13/09/2024)

Changed

  • Adapt the traits, so that they also can be used when only using ManifoldsBase,

without importing internal structs like EmptyTrait and TraitList

0.15.15 (29/08/2024)

Changed

  • Refactored error message code when ProductManifold is used without RecursiveArrayTools.jl.

0.15.14 (27/08/2024)

Added

  • A helpful error message when ProductManifold is used without RecursiveArrayTools.jl.

Changed

  • representation_size for ProductManifold now returns nothing instead of a one-element tuple. This change makes it easier to notice errors caused by not having RecursiveArrayTools.jl loaded.

0.15.13 (10/08/2024)

Changed

  • Fixed a small bug that caused calling get_vectors on PowerManifolds

to sometimes cause an error, cf #199.

0.15.12 (03/08/2024)

Changed

  • Improved performance of power manifold creation and some cases of get_component on product manifold.

0.15.11 (28/07/2024)

Added

  • Function allocate_on to generically allocate point and tangent vectors on a manifold without a pre-existing instance but of a particular type.
  • Function default_type to get the default type of points and tangent vectors for a manifold.
  • Package extension for the Quaternions.jl package that handles allocation.

Changed

  • Default allocation method was made more robust to custom promotion functions.

0.15.10 (19/05/2024)

Added

  • Functions fill(p, N) and fill!(P, p, N) to fill values into a point on a power manifold N.
  • introduce a base_point(TpM) to access the base point of a tangent space
  • introduce TpM[i] to access tangent spaces of factors from an AbstractPowerManifold or a ProductManifold.

0.15.9 (02/05/2024)

Added

  • Tests now also use Aqua.jl to spot problems in the code such as ambiguities.
  • introduce a check_inverse_retraction function to numerically check whether an inverse retraction method is a (correct) inverse retraction.
  • introduce a check_retraction function to numerically check whether a retraction method is a (correct) retraction.
  • introduce a check_vector_transport function to numerically check whether a vector transport is a (correct) vector transport.

Changed

  • introduced a ManifoldsBaseTestUtils module to encapsulate common types and function definitions in different parts of the tests.

0.15.8 (13/03/2024)

Added

  • sectional_curvature , sectional_curvature_max and sectional_curvature_min functions for obtaining information about sectional curvature of a manifold.

0.15.7 (24/01/2024)

Fixed

  • is_point and is_vector can now more stably :info or :warn when they return false, since they employ showerror for these displays.

0.15.6 (15/12/2023)

Added

  • An AbstractApproximationMethod to specify estimation methods for other more general functions,

as well as a default_approximation_method to specify defaults on manifolds.

  • An EmbeddedVectorTransport to use a vector transport in the embedding and a final projection.

Fixed

  • number_eltype correctly returns scalar type for nested array types like number_eltype(Vector{Vector{Float64}}).

0.15.5 (13/12/2023)

Added

  • Compatibility with RecursiveArrayTools v3.

0.15.4 (25/11/2023)

Fixed

0.15.3 (17/11/2023)

Fixed

  • Pass kwargs in rand! for AbstractPowerManifold to appropriate methods on the wrapped manifold.

0.15.2 (8/11/2023)

Fixed

  • vee and hat now use real coefficient basis for complex manifolds.

0.15.1 (30/10/2023)

Added

  • zero_vector(TpM) to generate a zero vector in the tangent space
  • a GitHub CI action that errors, when this file was not updated on a PR

Fixed

  • is_point and is_vector for the tangent space now correctly forward to vector checks on the corresponding manifold. The same for both check_sizes
  • add [compat] entries for the standard libraries.

0.15.0 (21/10/2023)

Added

  • ProductManifold type was migrated from Manifolds.jl.
  • Fiber, VectorSpaceFiber and TangentSpace types. TangentSpace is a generalized version of TangentSpaceAtPoint from Manifolds.jl.
  • A keyword to ValidationManifold which error= mode to use. This is by default the previous :error mode.
  • change_representer!, change_metric! and Weingarten! methods added to PowerManifold.
  • × now also works for retractions, inverse retractions, and vector transports to create their product versions
  • retract, inverse_retract, and vector_transport_to (and _dir) now also accept arbitrary retractions on the product manifold. These act the same as the n-fold product of a retraction.

Changed

  • retract now behaves like exp in the sense that it allocates early, which reduces the amount of code to dispatch through levels 1-3 twice
  • inverse_retract now behaves like log in the sense that it allocates early
  • Requires.jl is added as a dependency to facilitate loading some methods related to ProductManifolds on Julia 1.6 to 1.8. Later versions rely on package extensions.
  • Documenter.jl was updated to 1.0.
  • PowerManifold can now store its size either in a field or in a type, similarly to DefaultManifold. By default the size is stored in a field.
  • The signature of is_point was changed to be consistent with isapprox.. The error positional symbol (third argument) is now a keyword argument. We left the boolean shortcut in place. That means
    • is_point(M, p, true) works the same as before (false is the default anyways)
    • is_point(M, p, :warn) has to be changed to is_point(M, p; error=:warn)
  • The signature of is_vector was changed to be consistent with isapprox and is_point. The error positional symbol (fourth argument) is now a keyword argument. The error positional boolean (fourth argument) hence moved to fifth place (after check_base_point) This means
    • is_vector(M, p, X, true) should now be is_vector(M, p, X; error=:error)
    • is_vector(M, p, X, err, base) for two booleans err, base should now be is_vector(M, p, X, base, err)
    • is_vector(M, p, X, err, base) for a symbol err should now be is_vector(M, p, X, base; error=err)

Removed

  • Julia 1.0 is no longer supported. From now on, the earliest supported Julia version is 1.6.

0.14.12 (23/09/2023)

Changed

  • Introduce a thorough way to allocate tangent vectors for rand

0.14.11 (25/08/2023)

Added

  • Make the Weingarten map a decorator capable function.

0.14.10 (17/08/2023)

Added

  • introduce the Weingarten map and its in place variant Weingarten!.

0.14.9 (03/08/2023)

Added

  • Introduce an interface that allows the static size of a manifold to be a field as well.

0.14.8 (07/07/2023)

Changed

  • Improve show for cached bases and make it more robust

0.14.7 (07/07/2023)

Changed

  • the tutorial is now written in Quarto.

0.14.6 (10/06/2023)

Added

  • export the inplace random function rand!

0.14.5 (03/05/2023)

Added

  • Allow to specify an AbstractManifold when converting points or tangent vector types.

0.14.4 (10/04/2023)

Changed

  • Fix copy to work properly when copying Numbers

0.14.3 (16/03/2023)

Changed

  • Fix an allocation bug in nested power manifolds

0.14.2 (16/03/2023)

Added

  • adds a DependaBot workflow.

Changed

  • Fix an allocation issue with exp(M, p, X, t) that did not respect the type of t.

0.14.1 (18/02/2023)

Note that this release did not trigger a TagBot, so it appears within 0.14.2 in the tagged/created releases

Added

  • Introduce change_representer already in ManifoldsBase.

0.14.0 (15/02/2023)

Added

  • Type restriction for t in scaled retractions relaxed to Number.
  • embed_project(M::AbstractManifold, p) and embed_project(M::AbstractManifold, p, X) that are like projections onto a manifold or a tangent space but are guaranteed to be idempotent.

Changed

  • Retractions for scaled vectors no longer dispatch to non-scaled retractions. It is now reversed for performance reasons. Please either just define exp!(::MyManifold, q, p, X, t::Number) or both this and exp!(::MyManifold, q, p, X).
  • DefaultManifold now stores size in a field instead of the type itself to reduce the amount of compilation needed.
  • Fixed typo in inverse_retract_caley (now inverse_retract_cayley) and retract_caley (now retract_cayley).
  • retract_pade and retract_pade! now receive PadeRetraction objects instead of just n.