[oschulz]

Description written by Claude Code (model Claude Fable 5.1, standard effort) on behalf of the maintainer; the commits carry the trailer Assisted-By: generative AI.

BAT 6 builds on the rebuilt MeasureBase (JuliaMath/MeasureBase.jl#182): BAT's own measure layer, distribution wrappers and transforms are gone, hierarchical priors are mbinds, walker densities are evaluated in batches, and sample storage is structured (StructArrays laid out like MeasureBase's rand, constants as Fill columns). The branch is a sequence of 18 commits over main; redesign.md (first commit, to be dropped before merging) records the design and open points.

The release notes below will go into the registrator comment of v6.0.0.

---

BAT.jl v6.0.0

Breaking changes

BAT's measures are now MeasureBase measures. BAT used to carry its own
measure layer (BATMeasure and friends), which duplicated what MeasureBase
now provides, and its own distribution-transform implementation. Both are
gone, BAT builds on MeasureBase 0.15 instead.

• BAT.BATMeasure is gone, BAT's measure types subtype
  MeasureBase.AbstractMeasure directly.

• The wrapper types BAT.BATDistMeasure, BAT.BATPwrMeasure,
  BAT.BATPushFwdMeasure, BAT.BATWeightedMeasure and
  BAT.BATSuperpositionMeasure are gone. Use MeasureBase.asmeasure,
  powermeasure (^), pushfwd, weightedmeasure and superpose.

• BAT.StandardUvUniform, BAT.StandardMvUniform, BAT.StandardUvNormal
  and BAT.StandardMvNormal are gone. Use MeasureBase.StdUniform(),
  MeasureBase.StdNormal() and their powers, like StdNormal()^n.

• BAT.DistributionTransform is gone. Use MeasureBase.transport_to.
  bat_transform and the transform intents (NormalBased, UniformBased,
  ToRealVector, DoNotTransform) are unchanged, but the transformation
  functions they return are MeasureBase.TransportFunctions and transformed
  measures are MeasureBase.PushforwardMeasures.

• HierarchicalDistribution is gone. Use MeasureBase.mbind.

• distprod now returns a MeasureBase.ProductMeasure instead of a
  ValueShapes.NamedTupleDist resp. a Distributions.Product. Marginals
  are specified as before (distributions, intervals, arrays of those, or
  constant values), constant marginals become MeasureBase.Dirac.

• lbqintegral and distbind are deprecated in favour of
  MeasureBase.mintegrate/mintegrate_exp and MeasureBase.mbind.

• DensitySampleVector stores the variates in the layout that rand(μ^n)
  produces for their measure μ, instead of a ShapedAsNTArray over flat
  storage: numbers in vectors, arrays as nested arrays (ArraysOfArrays),
  named tuples as StructArrays of their components and constant
  components as Fill columns. varshape(samples) is derived from this
  layout. Vectors of named tuples passed to the constructor are converted
  to it. The new constructor
  DensitySampleVector(X, μ; logd, weight, info, aux) takes a batch of
  variates of μ or a matrix of their flat free components. Constant
  components must match the point masses of μ, and all variates must
  share one layout.

• Statistics of samples have the structure of the variates: mean, var,
  std, quantile, median, minimum, maximum and mode return named
  tuples resp. arrays instead of ShapedAsNT values, and so do effective
  sample sizes and credible intervals. minimum and maximum are
  elementwise, they used to return the lexicographically smallest resp.
  largest variate. cov and cor work on the flat free components of any
  samples.

• unshaped.(samples) and bat_transform(Vector, samples) return samples
  with the flat free components of the variates. vcat and merge of
  samples allocate and promote the types of the weights and log-densities,
  so samples with integer and real weights can be combined (merge used to
  keep the types of its first argument). push!, append! and merge!
  require growable storage, like the flat storage of MCMC output, and throw
  an ArgumentError otherwise.

• Broadcasting a value shape over samples, vs.(samples), and
  bat_transform with a value shape on samples are deprecated. Use
  DensitySampleVector(X, μ) to get samples in the layout of a measure.

• BAT no longer defines mean, var, std and cov of measures.
  MeasureBase defines them for measures with exactly known moments and
  throws an ArgumentError for others. BAT keeps its median and mode
  methods.

• MeasureBase renamed insupport and params to insupportof and
  paramsof. BAT implements insupportof for its measures and
  likelihoods, custom measures and likelihoods used with BAT must add
  methods to the new names as well.

• batmeasure is BAT's canonicalization: it turns distributions into
  measures, named tuples of distributions and NamedTupleDists into product
  measures, and MeasureBase density measures into PosteriorMeasures.

• Densities of measures are -Inf outside their support instead of
  throwing, and transports are NaN there; the boundary of the support can
  transport to infinite variates. Accordingly,
  BAT.checked_logdensityof no longer throws for a NaN density at a
  non-finite variate, it returns a zero density there.

• GridSampler places its grid points at the cell midpoints of the unit
  hypercube instead of including its boundary.

• DistributionsAD and ForwardDiffPullbacks are no longer dependencies.

Further behavior changes

• MeasureBase.basemeasure of a posterior measure is its prior and
  MeasureBase.logdensity_def is the log-density of its likelihood, so a
  posterior behaves like the mintegrate_exp(likelihood, prior) it
  represents.

• MeasureBase.weightedmeasure no longer special-cases posterior measures,
  it yields a WeightedMeasure for them like for any other measure. BAT's
  own renormalization (e.g. bat_sample on an auto-renormalized target)
  still rescales the likelihood and keeps the posterior structure.

• ValueShapes.varshape and ValueShapes.unshaped are defined for
  MeasureBase measures and derived from the measure structure.

• truncate_batmeasure supports measures with scalar variates, via
  MeasureBase's truncate. Products of univariate distributions and
  posterior measures (via their prior) are truncated as before. It still
  doesn't renormalize.

• bat_write marks the layout of each object in HDF5 files with an
  attribute bat_layout (format version 6, see BATHDF5IO), so bat_read
  restores nested named tuples, array-valued and constant parameters
  exactly. Files of BAT 5.x are read as before, with the layout inferred
  from the values, now including matrix-valued parameters.

• Samples without parameter names, like flat samples, can be plotted.
  They are selected by position (among the free parameters, in the order
  of unshaped.(samples)) and labeled v_1, v_2 and so on.
  BAT.bat_marginalize keeps the sample weights.

• Sample reports work for samples without free parameters, they list the
  constant parameters only.

• EllipticalSliceMCMCSampling requires a posterior measure whose prior
  transforms to a standard multivariate normal, and says so instead of
  failing later.

• Algorithms that record the degrees of freedom of a measure (e.g. the
  importance samplers) record nothing for measures that declare none,
  instead of throwing.

• The MCMC samplers evaluate the target density for all walkers of a step in
  one batched call (MeasureBase.logdensities) instead of walker by walker.
  Measures with fused batched kernels may return results that differ from the
  per-walker ones in the last ulp, so MCMC runs with a fixed RNG seed can take
  a different path than before.

• Support for GPUs is provisional. Sample storage, its flat conversion and
  concatenation work with device arrays. Batches of variates are drawn in
  one call, rand(rng, μ^n), so on a GPU they come from a single device
  RNG instead of per-chain partitions of BAT's counter-based RNG. Reports,
  plots and HDF5 output require samples on the host.

https://claude.ai/code/session_01L5yPndh2K6u14z3CwF9YXf