Major upgrade - #182
Draft
oschulz wants to merge 75 commits into
Draft
Conversation
Shouldn't call Core.Compiler.return_type directly in many places.
Will be used a lot when bridging from Distributions to MeasureBase.
Currently unused and undocumented, can add it back later when needed.
Created by generative AI.
Assisted by generative AI.
Unused and untested. (cherry picked from commit 1d27188)
Not used currently. (cherry picked from commit 9844583)
A rebase can easily be written explicitly. (cherry picked from commit fb3c98c)
`mintegral` should be used instead to express posteriors. (cherry picked from commit 3c61180)
To be re-introduced in sub-module MeasureOperators. (cherry picked from commit 0cdca3d)
Dirac measures with equal points are equal measures; the default object identity failed for array-valued points. Created by generative AI.
Products of weighted measures (tuples, named tuples and arrays) now pull the total weight out, following the canonical measure nesting. Empty products are explicit unit measures (Dirac of the empty variate). Arrays of measures are no longer copied via broadcast asmeasure, and arrays of a single singleton measure type collapse to power measures without requiring a non-empty array. The static-weight optimization now stays static when the array length is static. Created by generative AI.
Pushforwards of point masses collapse to point masses and pushforwards commute with weighting. The identity shortcut moves from _pushfwd_impl to pushfwd itself and the two style-specific pushfwd/pullbck methods merge into single PushFwdStyle methods, keeping dispatch unambiguous. Created by generative AI.
Created by generative AI.
Matches the other smart constructors, which are documented and exported. Created by generative AI.
Covers the power/product/superpose/pushfwd/restrict simplifications. Also wires the previously orphaned superpose tests into the test suite, updated to current basemeasure collapse behavior. Created by generative AI.
Measure transport now falls back to a pivot through a flat vector of standard uniform variates, using the with-rest transport protocol, when the DOF of either side is not fast-computable. This enables transport between hierarchical measures (Bind) and between known-DOF and unknown-DOF measures, as already promised by the transport_def docstring. The mvstd gateway now also verifies that the produced variate length matches the target standard power measure, instead of silently returning a variate of the wrong length. Created by generative AI.
Superposition simplifications now only happen when measure equality is decidable from the measure types alone (via the new _static_isequal), instead of branching on runtime measure equality. As a consequence, value-equal measures of non-singleton type (e.g. equal Diracs) now stay superpositions instead of collapsing to weighted measures, and basemeasure of superpositions with equal-typed components is type stable. Created by generative AI.
…ith-rest logdensityof is now a single generic entry function; measure types specialize the new MeasureBase.logdensityof_impl instead of logdensityof itself. The new logdensityof_with_rest protocol evaluates densities of measures that live at the beginning of a flat variate stream (a vector for vcat-combined and a NamedTuple for merge-combined measures), returning the log-density, the consumed variate and the unconsumed rest of the stream. Its default implementation determines the variate size via mspace_elsize (falling back to testvalue). Bind and CombinedMeasure evaluate densities of vcat- and merge-combined variates through logdensityof_with_rest in a single pass now, without materializing intermediate transport measures and without splitting sub-variates twice. Variates that are too long now result in an informative exception, and so do binds whose value combination function does not support variate splitting. Created by generative AI.
The generic three-argument logdensity_def now descends the base measure chains of both measures in lockstep after equalizing their static depths. Since the members of a shared chain suffix have the same depth-from-root on both sides, shared suffixes cancel symbolically. The descent is fully unrolled at compile time and constructs only the base measures it actually visits. Specialized relative densities move from three-argument logdensity_def methods to the new extension point MeasureBase.logdensity_rel_def. The descent checks for an applicable specialization at each visited measure pair; availability is decided purely by dispatch via a sentinel return type instead of method-table introspection, so the checks are free at run time and defining new specializations behaves like any ordinary method definition. This removes basemeasure_sequence-based chain materialization, the commonbase search, schema and the static_hasmethod gate from the relative density code path. Root measure pairs without a specialization now throw an informative exception instead of warning and returning NaN. Created by generative AI.
Products over vectors of same-typed unknown-DOF marginals (e.g. binds)
now transport from multivariate standard measures via a typed loop
instead of accumulating into a Vector{Any}; marginal vectors with
abstract element type keep the untyped fallback.
Created by generative AI.
Fixes ambiguities between the static-size power density methods and powers of primitive measures (the existing disambiguation method did not cover them), between the massof methods generated by @useproxy and massof over intervals, and in the legacy kernel constructors. Package ambiguity testing (including the Aqua ambiguity check) is now enabled in the test suite. Created by generative AI.
Removes kernel.jl (AbstractTransitionKernel and its subtypes, the kernel/kleisli constructors), the parameterized-measure kernel constructors and the kernel-based productmeasure methods. MKernel and mbind supersede this functionality. The basekernel helper stays, it is independent of the kernel types. Created by generative AI.
Adds mspace_elsize methods for standard measures (scalar), Dirac and powers of scalar-variate measures (static where the axes are static), and NamedTuple-product variate names based on the marginal names alone. Flat vector streams now support scalar variates (consuming a single stream element, also in vcat variate splitting) and multi-rank variates (consumed in flattened form and reshaped). This enables densities and transport for binds with scalar-variate primary measures and density evaluation of multi-dimensional powers inside flat streams. Created by generative AI.
Relative densities between product measures were a specialization of logdensity_rel itself, bypassing the support-check layer at the product level. They are now logdensity_rel_def methods (with a type-stable tuple-marginals variant) evaluating marginals via unsafe_logdensity_rel, support checking happens for the products as a whole in logdensity_rel. Created by generative AI.
…sities The smf-based interval mass now fails with an informative exception for measures without a statistical measure function instead of a MethodError on NoSMF values. The generic non-real-variate log-density of primitive measures returns a static zero instead of false. test_smf now tolerates insupport results that are not booleans (NoFastInsupport) and the logdensity_def docstring points to logdensity_rel_def as the extension point for specialized relative densities. Created by generative AI.
Created by generative AI.
…riate sizes Densities and standard-measure transport for products of hierarchical measures (in vector, tuple and named-tuple marginal form) work through the with-rest machinery; keep it that way. Created by generative AI.
combinesets(f_c, α, β) combines two sets along the value combination semantics of mcombine, with specific set representations where possible: cartesian products concatenate under vcat and merge, one-dimensional cartesian powers of equal (singleton-typed) base sets concatenate under vcat, and implicit domains of measures combine into the implicit domain of the combined measure. CombinedMeasure now provides mdomain. Also fixes two latent bugs uncovered by the new tests: setcartprod for NamedTuple sets had an unbound type parameter and never matched, and membership tests for CartesianPower used reversed argument order in Base.in. Created by generative AI.
Created by generative AI.
logdensities(mu, X) computes the log-density of mu at each point in X, preserving the shape of X. Measure types specialize logdensities_impl. Power measures unwrap into power axes arguments of the internal machinery (ordered outermost first), so implementation methods never dispatch on nested PowerMeasure type signatures. Scalar-variate measures evaluate as a single flat broadcast; power batches with flat variate storage (ArrayOfSimilarArrays) additionally fuse the per-point reduction into a single segmented sum, keeping GPU-backed data on-device. Created by generative AI.
logdensityof_impl and logdensity_def for PowerMeasure now unwrap the power structure into the power axes arguments of the batched density machinery. Nested powers with flat variate storage evaluate as a single fused broadcast and segmented reduction, GPU-compatible. Replaces the per-level power density methods, including the static-size specialization and its disambiguation methods, and removes the now unused infer_logdensity_type. Created by generative AI.
Batched protocol functions pair with their scalar counterparts by name: batched_logdensityof_impl implements logdensities, further batched_* functions (with-rest, transport, rand) will follow the same scheme. Created by generative AI.
Reactant traced scalars subtype Number, not Real. Widens primitive leaf density kernels, batched density eltype gates, with-rest stream signatures, log-weight arguments and numtype conversion sources. Real stays where realness is semantic (domain membership, numtype request parameters). Created by generative AI.
Replaces insupport ternaries in the primitive and standard density kernels with _checksupport (ifelse-based) and the value branches in _combine_logd_with_ladj with nested ifelse. StdUniform's insupport uses non-short-circuiting comparisons. Required for traced values (Reactant), where control flow must not depend on runtime values. Created by generative AI.
The weight-shifted density of a support-safe base density is support-safe, so weighted measures need no explicit support check. Removes a redundant per-point insupport sweep over the base measure. Created by generative AI.
Widens Real argument types to Number in the cdf/quantile/affine transport machinery and makes the StdUniform transport gateways branch-free (out-of-support results are masked via ifelse, the quantile argument is clamped to keep eager evaluation valid). Distributions with standard-measure or affine transport origins now work with traced values; families that require inverse incomplete beta/gamma functions remain host-only. Created by generative AI.
Domain membership methods for traced numbers: Reactant's traced scalars subtype Number, not Real or Integer, so membership in RealValues and IntegerValues is decided by their value type parameter. Created by generative AI.
oschulz
marked this pull request as draft
August 23, 2026 14:30
Codecov Report❌ Patch coverage is Additional details and impacted files@@ Coverage Diff @@
## main #182 +/- ##
===========================================
+ Coverage 54.11% 68.46% +14.35%
===========================================
Files 44 64 +20
Lines 1253 2134 +881
===========================================
+ Hits 678 1461 +783
- Misses 575 673 +98 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.