Release 2.4.1

🐛 Fixes

• Fix CircularFourierDistribution.from_distribution() grid generation to use endpoint-free linspace instead of floating-step arange, preventing off-by-one samples for certain odd grid sizes. (#2946)
• Preserve nested import errors from installed optional packages in require_optional_dependency, so real import-time failures inside an installed package are no longer incorrectly reported as the package being missing. (#2944)
• Fix CalibratedPairwiseAssociationModel to correctly handle stringified binary positive labels such as "1" and "true" when selecting the positive column from two-column predict_proba output. (#2675)
• Validate AbstractCircularFilter.association_likelihood_numerical PDF evaluations so empty or multi-element outputs raise an error instead of silently using only the first value. (#3183)
• Fix fit_time_offset to validate and normalize the metric parameter before use, so typos and unhashable inputs raise clear errors instead of producing all-NaN results. (#3182)
• Fix Metropolis-Hastings sampler to reject multi-element ln_pdf or proposal-correction outputs instead of silently using only the first value for acceptance decisions. (#3180)
• Fix per-particle prediction stacking so (n, 1) Euclidean particle matrices remain (n, 1) instead of being transposed to (1, n). (#3179)
• Validate EvidenceComputationMode.mode type before membership checks, so unhashable inputs like lists raise ValueError instead of a raw TypeError. (#3176)
• Fix Pareto constraint spec parsing to raise ValueError for malformed scalar/object/text specs instead of leaking a raw TypeError from len(spec). (#3172)
• Fix EOT shape-transform scalar conversion (_as_shapely_scalar) to reject multi-value arrays instead of silently truncating to the first entry, preventing incorrect target translations and rotations. (#3174)
• Fix NumPy backend vectorizer to collect mapped outputs with NumPy stacking instead of preallocating from the first output, preventing silent dtype truncation (e.g., [0, 1.5] becoming [0, 1]). (#3171)
• Fix shared NumPy convert_to_wider_dtype to use NumPy's result_type for mixed-type promotion, so int64 and float32 correctly promote to float64 instead of truncating to float32. (#3173)
• Fix BackendNotSupportedError to treat supported_backends="numpy" as a single backend name rather than iterating over individual characters. (#3169)
• Fix explicit NumPy backend zeros to normalize dtypes from shared state when another backend (e.g., PyTorch) was initialized first, preventing incompatible dtype values from reaching NumPy. (#3166)
• Fix set_default_dtype to normalize dtype-like inputs (e.g., jnp.dtype("float32")) before the complex-dtype lookup, and prevent partial state mutation when a non-floating dtype is rejected. (#3165)
• Fix get_backend to convert missing optional dependencies for explicit backends (e.g., "pytorch") into OptionalDependencyError with install guidance instead of leaking a raw ModuleNotFoundError. (#3157)
• Fix backend support Markdown table rendering to escape literal pipe characters in cells instead of replacing them with fullwidth lookalikes. (#3153)
• Fix Euclidean MTT mean extractor to use the public get_tracks() selection API when available, preventing hidden tentative or deleted tracks from contaminating evaluation summaries. (#3154)
• Fix backend support table rendering to escape table-sensitive characters in Markdown cells, preventing malformed table output for metadata containing separator characters. (#3145)
• Fix Euclidean MTT target orientation handling for ambiguous orientation representations. (#3147)
• Preserve explicit positional dynamic dtypes in backend operations so explicitly specified dtypes are not lost. (#3039)
• Preserve one-dimensional multi-target ground-truth shape so single-dimension scenarios are not incorrectly reshaped. (#3116)
• Fix empty file-backed measurement counting so empty files contribute a correct measurement count rather than being skipped. (#3105)
• Fix NumPy trace to correctly forward keyword arguments, resolving failures when calling trace with offset or dtype kwargs. (#3095)
• Fix virtual backend submodule imports so that backend-specific submodules resolve correctly when accessed through the virtual backend namespace. (#3087)
• Fix fixed-lag Viterbi filter to correctly use custom transition context instead of the default context. (#3055)
• Fix LinearDiracDistribution constructor to accept and flatten weight arrays regardless of their shape, improving compatibility with array-producing workflows. (#3030)
• Fix backend error diagnostics to remain available even when no active backend is configured. (#3013)
• Treat boolean track matrix cell values as missing/absent instead of interpreting True/False as numeric track data. (#2993)
• Treat missing diagnostic track IDs as absent rather than raising an error, improving robustness of diagnostic summaries with incomplete track data. (#3011)
• Fix Autograd quadratic assignment to correctly use default solver options when none are provided. (#2975)
• Fix NumPy backend choice to return elements in population order when shuffle=False, correcting previously unsorted output. (#2979)
• Fix JAX randint bounds handling to correctly validate and convert integer bounds. (#2982)
• Fix overwritten association probability gate regression so the gate threshold is correctly applied after probability updates. (#3007)
• Fix PyTorch QR decomposition mode contract to correctly handle mode='r' and other mode variants. (#3004)
• Fix shared NumPy QR mode signatures to correctly handle all supported mode variants. (#2999)
• Fix replay grid filter to handle nonfinite position inputs gracefully instead of propagating NaN/Inf values. (#2936)
• Fix Euclidean mean extraction to correctly handle raw array states (not wrapped in a distribution object). (#2938)
• Accept None evidence mode metadata as a valid absent-metadata indicator, fixing errors in evidence mode objects with no metadata. (#2925)
• Normalize overflowing assignment costs to a large finite sentinel value, preventing integer overflow from corrupting assignment solutions. (#2860)
• Fix out-of-sequence measurement accepted diagnostic parsing to correctly read boolean values from serialized records. (#2877)
• Fix evidence support boolean metadata parsing to correctly read boolean values from serialized metadata. (#2875)
• Fix masked measurement observed dimension handling to preserve dimension sets correctly when constructed from iterables. (#2858)
• Fix registration point validation to correctly import backend functions needed for coordinate checks. (#2853)
• Handle overflowing particle diagnostics scalar values by clamping to finite limits instead of propagating infinity. (#2844)
• Handle mixed result parameter types (e.g., integer followed by float) in result aggregation without dtype truncation. (#2843)
• Fix surface-band probability computation for sequence inputs, correctly aggregating probabilities over multiple points. (#2772)
• Normalize overflowing seed scalar values to a valid range instead of leaking raw overflow errors. (#2828)
• Handle nonfinite nearest-time queries gracefully instead of returning incorrect results. (#2825)
• Handle overflowing optional diagnostics scalar values by clamping instead of propagating overflow. (#2821)
• Fix ambiguous Euclidean MTT target orientation handling to resolve orientation correctly. (#2810)
• Fix fragment completion candidate observation validation to correctly accept valid candidate observations. (#2769)
• Handle string return_smoothed flag values explicitly, rejecting non-boolean strings instead of silently coercing them. (#2782)
• Fix GGIW tracker update to accept empty measurement scans gracefully without raising errors. (#2785)
• Fix selection fallback ranking to use the correct tie-breaking order. (#2788)
• Fix Pareto front computation for non-scalar metrics so all dimensions are checked correctly. (#2790)
• Fix axis labels for distributions on specific manifolds (e.g., torus, sphere) to display correct coordinate names. (#2799)
• Handle non-parametric plot result summaries gracefully in evaluation reporting. (#2805)
• Fix gated assignment to correctly handle optional matching when no assignment meets the gate threshold. (#2800)
• Fix duplicate session-label detection to work correctly with custom fill values. (#2783)
• Fix 1D replay grid position inputs to be handled correctly instead of being misinterpreted as higher-dimensional inputs. (#2780)
• Honor shuffle=False in random choice backends so unshuffled draws return elements in their original order. (#2774)
• Ensure candidate pruning penalty values remain large after normalization, preventing reduced penalties from causing incorrect candidate selection. (#2766)
• Fix fixed-lag Viterbi filter prefix context handling to correctly propagate prefix hypotheses. (#2765)
• Fix PyTorch backend NumPy integer dtype normalization so NumPy integer dtypes are correctly mapped to PyTorch equivalents. (#2760)
• Apply metadata correctly in no-op record smoothing passes so smoothed records retain their source metadata. (#2751)
• Skip disabled callable dimension attributes during dimension inference to prevent errors from non-callable or disabled attribute accessors. (#2752)
• Fix serialized innovation accepted flag parsing to correctly deserialize boolean values from stored records. (#2747)
• Handle scalar single-target MTT ground-truth inputs correctly, preventing shape errors for scalar state values. (#2745)
• Preserve singleton measurement axis when processing single-measurement updates, preventing dimension collapse. (#2743)
• Fix observation-cost assignment numeric gap handling for sessions with gaps in observation numbering. (#2729)
• Fix observation-cost gap handling for inferred session sizes when sessions have unequal observation counts. (#2734)
• Fix ...

Release 2.4.0

✨ Features

• Add point-set distance and overlap metrics to pyrecest.evaluation, expose SCGP measurement reliability example wrappers from pyrecest.examples, fix JAX backend compatibility in SCGP measurement Jacobian assembly, and bump version to 2.4.0. (#2579)
• Add implicit scalar-field surface protocols to pyrecest.protocols, implicit surface evaluation helpers to pyrecest.evaluation, DVS normal-flow polarity helpers to pyrecest.experimental.dvs, and domain-neutral protected-tail selection utilities to pyrecest.evaluation. (#2578)
• Add DVS event contour likelihood utility functions under pyrecest.experimental.dvs and Pareto frontier and equal-quality selection helpers under pyrecest.evaluation. (#2577)
• Add support-point sampling helpers to pyrecest.sampling, non-mutating SCGP measurement scoring to pyrecest.filters, and SCGP contour sampling utilities used by the experimental DVS point-process tracker. (#2575)
• Add shared measurement reliability helpers (weight normalization, active masks, combined active-index selection, per-measurement covariance stacking) to pyrecest.filters, reused by FullSCGPTracker. (#2572)
• Add a shared filter update diagnostics helper exported from pyrecest.filters that records standard metadata keys during updates, applied to the GPRHM tracker update path. (#2570)

🐛 Fixes

• Resolve issue where the PyTorch backend divide operation incorrectly masked non-zero-denominator positions when ignore_div_zero=True; inputs are now promoted to tensors and placed on the correct device with widened dtypes before division. (#2564)

📦 Dependencies

• Bump astropy-iers-data from 0.2026.6.1.17.39.59 to 0.2026.6.8.17.49.5 (updated IERS Earth rotation and leap second tables). (#2574)
• Bump filelock from 3.29.0 to 3.29.1 (fixes soft lock symlink follow security issue). (#2573)
• Bump tqdm from 4.67.3 to 4.68.1 (fixes monitor thread atexit deadlock and adds itertools extras). (#2569)
• Bump platformdirs from 4.9.6 to 4.10.0 (adds user_projects_dir, user_publicshare_dir, user_templates_dir, user_fonts_dir, and user_preference_dir). (#2568)
• Bump idna from 3.16 to 3.18 (75% memory reduction, new 1024-character input length cap, and a lenient display decode option). (#2565)
• Bump setuptools from 81.0.0 to 82.0.1 (removes pkg_resources; projects relying on it should pin to an older setuptools version). (#2567)
• Bump astropy-iers-data from 0.2026.5.18.1.11.28 to 0.2026.6.1.17.39.59 (updated IERS Earth rotation and leap second tables). (#2566)

Release 2.3.0

✨ Features

• Add JAX custom gradient support via jax.custom_vjp, preserving keyword and default argument binding for differentiable functions. (#2559)
• Add DelayedOutputMixin shared interface for online fixed-lag smoothers, exported from pyrecest.smoothers. (#2560)
• Add ellipse geometry conversion, canonicalization helpers, and covariance projection utilities to the tracking module. (#2552)
• Add residual multi-hypothesis selector for discrete edit candidates with threshold fallback and conflict handling, exported from pyrecest.tracking. (#2547)
• Add reusable innovation/NIS diagnostic dataclasses and summary helpers, exposed through pyrecest.tracking. (#2537)
• Add dataframe-oriented model-comparison evidence margin helpers and hypothesis replay ranking utilities to the tracking module. (#2534)
• Add generic track fragment completion path utilities exported from pyrecest.utils. (#2530)
• Add generic evidence computation mode helpers, tracking event/record helpers, track-edit what-if scoring utilities, and sparse second-order evidence diagnostics. (#2532)
• Implement pyrecest.backend.signal.fftconvolve for the PyTorch backend using torch.fft, with SciPy-compatible mode={full,same,valid} semantics. (#2531)
• Add evidence support metadata helpers, generic record smoothing utilities, and weak/masked linear measurement model helpers. (#2529)
• Add reusable sparse transition row cache utilities and wire sparse second-order grid evidence through the cache. (#2527)
• Add sparse second-order grid evidence utility, exposed through filter lazy exports. (#2522)

🐛 Fixes

• Resolve issue where GaussianDistribution.from_distribution now raises a clear ConversionError instead of leaking an AttributeError when a source distribution lacks mean()/covariance() support. (#2535)
• Resolve issue where the velocity-locked MEM-QKF tracker now correctly returns early without modifying state when given an empty measurement set. (#2555)
• Resolve issue where the UKF prediction now correctly regenerates cached propagated sigma points after folding additive process noise into the prediction covariance. (#2548)
• Resolve issue where PyTorch dot now correctly handles 0-D scalar operands by using multiplication instead of einsum/dot paths. (#2521)
• Resolve issue where backend size() now correctly infers shapes for Python array-like inputs, and comb(n, k) now returns 0 for out-of-range selections (e.g., comb(3, 4) == 0). (#2520)
• Resolve issue where PyTorch random.randint now correctly handles array-valued and broadcasted low/high bounds with NumPy-compatible semantics. (#2519)
• Resolve issue where JAX and PyTorch random.rand now correctly accept NumPy-style positional dimension arguments (e.g., rand(2, 3)). (#2518)
• Resolve issue where PyTorch random.multinomial now correctly supports NumPy-style size= parameter with proper trial count and probability validation. (#2517)
• Resolve issue where JAX random.multinomial now correctly supports NumPy-style size= parameter and validates trial counts and probability vector shape before sampling. (#2515)
• Resolve issue where the NumPy backend random.rand now correctly accepts the size= keyword contract alongside legacy positional dimension calls. (#2513)
• Resolve issue where the PyTorch random backend now validates size arguments, rejecting boolean, non-integral, string, and negative dimensions before passing them to PyTorch. (#2511)
• Resolve issue where PyTorch random.uniform now samples independent values over the broadcasted low/high shape when size is omitted, instead of correlating all output elements. (#2510)
• Resolve issue where JAX random.normal now samples independent values over the broadcasted loc/scale shape when size is omitted, instead of producing perfectly correlated outputs. (#2504)
• Fix JAX autodiff backend behavior and tighten PyTorch backend contract handling for linspace and apply_along_axis. (#2539)

⚡ Improvements

• Improve PyTorch backend NumPy compatibility for atleast_* helpers, meshgrid, expand_dims, and cross-product with NumPy axis arguments and mixed dimensions. (#2543)
• Add NumPy-compatible PyTorch reduction handling for mean/std edge cases, including improved argument normalization and dtype behavior. (#2544)
• Align NumPy and JAX backend matmul behavior with NumPy-style vector operands for vector-vector, matrix-vector, and vector-matrix products. (#2545)
• Accept array-like inputs in JAX and PyTorch backend helper wrappers, and add NumPy-style axis/keepdims support in PyTorch reduction helpers. (#2561)
• Handle scalar operands in shared backend dot helpers using multiplication instead of vector einsum paths, and coerce array-like inputs for outer and matvec helpers across backends. (#2550)
• Handle scalar and array-like products consistently across NumPy/autograd, JAX, and PyTorch backends. (#2554)
• Expand backend product helpers to support scalar and array-like inputs consistently across NumPy/autograd, JAX, and PyTorch. (#2551)
• Accept array-like inputs in backend assignment, diagonal/triangular matrix reconstruction, and dot helpers, and add stacked cart-product distribution metadata. (#2549)
• Tighten PyTorch backend contract handling for array helpers and matmul edge cases. (#2541)
• Add autograd to SUPPORTED_BACKENDS in pyrecest.backends, making get_backend("autograd") an officially recognized entry point. (#2505)

📦 Dependencies

• Bump astropy-iers-data to 0.2026.6.1.17.39.59 with updated IERS Earth rotation and leap second tables. (#2525)
• Bump filelock from 3.29.0 to 3.29.1. (#2524)
• Bump idna from 3.17 to 3.18. (#2526)
• Bump platformdirs from 4.9.6 to 4.10.0. (#2509)
• Bump idna from 3.16 to 3.17 with 75% memory usage reduction and input length cap for validation entry points. (#2508)
• Bump setuptools from 81.0.0 to 82.0.1. (#2506)
• Bump astropy-iers-data from 0.2026.5.18.1.11.28 to 0.2026.5.25.1.14.13 with updated IERS Earth rotation and leap second tables. (#2507)

Release 2.2.4

✨ Features

• Add delayed-decision MHT utilities for deferred track management decisions in multi-hypothesis tracking. (#2169)
• Add SO3 product sequence filtering utilities for multi-step orientation filtering over particle sets. (#2154)
• Add RaFT-UAV helper utilities for UAV tracking with range and angle measurements. (#2153)
• Add SO3 tangent Savitzky-Golay smoother for polynomial smoothing of orientation sequences. (#2152)
• Add replay grid likelihood helpers for offline likelihood evaluation over pre-recorded grids. (#2151)
• Add experimental DVS (Dynamic Vision Sensor) core processing utilities. (#2150)
• Add max-cardinality assignment helper for optimal assignment problems with cardinality constraints. (#2149)
• Add fusion utility candidates including multi-sensor combination helpers. (#2148)
• Add discrete state space utilities for Markov chain and hidden Markov model filtering. (#2146)
• Add SCGP filter support for measurement reliability weights, enabling downweighting of uncertain observations. (#2145)
• Add JAX backend helpers for triangular vector pack/unpack operations. (#2159)
• Implement JAX positive-definite matrix check (is_positive_definite) in the JAX backend. (#2190)
• Add PyTorch backend wrappers for reduction operations (sum, prod, any, all) with correct axis/keepdims semantics. (#2216)
• Add PyTorch backend wrapper for argument reduction operations (argmin/argmax). (#2239)

🐛 Fixes

• Fix underscored symmetric hypersphere mean dispatch so mean correctly delegates to the internal _mean computation. (#2138)
• Fix backend contract edge cases including indexing and shape operations across all backends. (#2160)
• Fix various runtime edge cases in PyRecEst distributions and filters. (#2161)
• Fix backend contract shape handling edge cases across array operations. (#2163)
• Fix runtime edge cases in various distribution and filter classes. (#2164)
• Complete MHT tracker fixes including track pruning and merging edge cases. (#2166)
• Harden multi-Bernoulli tracker and backend edge cases for robustness. (#2167)
• Fix backend and MHT edge cases including empty track set handling. (#2168)
• Fix multi-session assignment indexing so tracks are correctly mapped across sessions. (#2170)
• Fix time-offset calibration aggregation to correctly combine offset estimates from multiple sensors. (#2171)
• Fix hypertoroidal wrapped-normal distribution to correctly handle edge case inputs such as zero sigma and boundary points. (#2173)
• Fix backend meshgrid to correctly coerce array-like inputs before grid construction. (#2174)
• Fix wrapped-normal distribution to correctly handle edge case inputs. (#2175)
• Fix hypertoroidal distributions to correctly accept scalar and list inputs across methods. (#2176)
• Fix PartiallyWrappedNormalDistribution to correctly handle edge case inputs. (#2177)
• Fix JAX multinomial sampling to correctly manage the random number generator state. (#2178)
• Fix backend meshgrid and JPDAF warning handling edge cases. (#2179)
• Fix time-offset utility to validate input shapes before processing. (#2181)
• Fix JAX backend assignment indexing to correctly handle all index types. (#2182)
• Complete autograd backend contract, adding missing operations needed for gradient-based filtering. (#2183)
• Fix HistoryRecorder to correctly handle array-like inputs with padding. (#2184)
• Fix PyTorch tuple reduction operations (min/max returning value and index) to work correctly. (#2187)
• Fix HypercylindricalDiracDistribution conditioning to correctly handle edge case observations. (#2188)
• Fix tracklet Viterbi multi-session assignment to handle edge cases including empty sessions and boundary tracks. (#2189)
• Fix S2 coordinate conversion to validate inputs before conversion, preventing silent NaN propagation. (#2195)
• Fix positive-definite matrix shape validation to work correctly for NumPy and PyTorch backends. (#2196)
• Add support for scalar likelihood callbacks in the wrapped-normal filter update step. (#2194)
• Fix WrappedExponentialDistribution to correctly handle scalar parameter inputs. (#2198)
• Fix backend and wrapped distribution input edge cases across multiple distributions. (#2200)
• Fix track metric min_length denominator so track quality scores are computed correctly. (#2201)
• Fix circular distribution constructors to correctly handle scalar parameters. (#2202)
• Fix backend and scalar parameter handling edge cases across distributions. (#2203)
• Fix association model to fall back correctly to an alternative solver when the primary solver fails. (#2205)
• Fix PyTorch random.choice to correctly handle the population array and sampling semantics. (#2208)
• Fix PyTorch random.choice to correctly sample with and without replacement. (#2209)
• Fix scalar process noise handling in the random matrix tracker. (#2210)
• Fix JAX random.randint to follow the NumPy contract for shape and dtype. (#2211)
• Fix PyTorch mean to correctly handle the axis argument. (#2212)
• Fix PyTorch backend quantile computation to correctly handle the axis argument. (#2214)
• Fix dtype promotion and bias validation edge cases in the shared backend. (#2215)
• Fix PyTorch random.normal to correctly handle array parameters for loc and scale. (#2217)
• Fix JAX meshgrid to correctly coerce array-like inputs before constructing the coordinate grid. (#2218)
• Fix PyTorch backend compatibility gaps including correct handling of zeros_like, ones_like, and type promotion edge cases. (#2219)
• Fix PyTorch split and hsplit to match NumPy semantics for section-count and size-list arguments. (#2223)
• Fix PyTorch random.multinomial to correctly accept sequence inputs. (#2225)
• Fix MEM-RBPF FFBSi backward smoothing to correctly handle tuple record types. (#2226)
• Fix backend assignment to correctly handle empty index arrays as no-ops across all backends. (#2227)
• Fix GNN tracker pairwise-cost validation to raise clear errors for malformed cost matrices. (#2232)
• Fix time-offset summary reporting to correctly identify the best-row offset estimate. (#2233)
• Fix PyTorch sum to correctly implement the keepdims contract. (#2234)
• Fix PyTorch reduction operations to correctly implement the keepdims contract. (#2237)
• Fix positive-definite matrix checks to correctly handle non-symmetric matrices without false positives. (#2238)
• Fix scalar non-assignment costs in Murty k-best assignment to be handled correctly. (#2244)
• Fix PyTorch multivariate_normal to correctly accept array-like inputs. (#2245)
• Fix PyTorch std to correctly accept array-like inputs. (#2246)
• Validate negative time-offset maximum delta, raising a clear error instead of silently producing incorrect offsets. (#2247)
• Validate Fibonacci Gaussian sampler inputs for finiteness and consistency. (#2248)
• Fix PyTorch linalg functions to correctly accept array-like inputs. (#2249)
• Fix association gate monotonicity so the gate function correctly enforces its ordering constraint. (#2254)
• Fix PyTorch mean to correctly handle integer array inputs. (#2257)
• Fix PyTorch take and pad backend functions to correctly follow the NumPy contract. (#2258)
• Filter non-finite reference values from calibration reference sets before processing. (#2262)
• Validate Euclidean sampler integer arguments before sampling to prevent cryptic failures. (#2263)
• Fix PyTorch backend copy to correctly handle scalar inputs without shape errors. (#2265)
• Validate numerical repair helper parameters with explicit error checks. (#2266)
• Fix backend matmul to correctly accept array-like inputs including Python lists. (#2269)
• Validate numerical tolerance parameters for finiteness and positivity. (#2270)
• Fix JAX UKF sigma-point updates to correctly propagate state covariance. (#2271)
• Fix non-vectorized Dirac transforms to work correctly across NumPy, PyTorch, and JAX backends. (#2272)
• Fix infinite log-weight normalization to clamp weights to finite values instead of producing NaN. (#2273)
• Validate sensor bias feature row counts before matrix construction. (#2274)
• Validate bias application feature row counts before applying sensor bias corrections. (#2275)
• Validate bias prediction row counts before matrix construction. (#2277)
• Validate finite sequence association costs, rejecting non-finite values before assignment. (#2281)
• Handle non-finite track observation IDs gracefully rather than propagating them into track data structures. (#2282)
• Reject NaN time delta cutoff parameters with a clear error. (#2283)
• Reject non-finite association weight arrays before running data association. (#2285)
• Validate ROI assignment threshold parameters with explicit error checks. (#2286)
• Validate non-assignment cost parameters for finiteness. (#2287)
• Validate chi-square gating state dimension parameter. (#2288)
• Validate track metric min_length parameter to be a positive integer. (#2290)
• Validate multi-session tracker max gap parameter explicitly. (#2291)
• Validate multi-session tracker integer inputs explicitly. (#2292)
• Validate candidate pruning top-k parameter to be a positive integer. (#2294)
• Validate adaptive process noise configuration parameters explicitly. (#2295)
• Validate linear update planning scalar parameters explicitly. (#2296)
• Validate bias calibration fitting parameters with explicit error checks. (#2297)
• Validate track manager integer inputs (e.g., max track count) explicitly. (#2301)
• Validate association gate parameters with explicit error checks. (#2302)
• Validate tracklet Viterbi tracker configuration parameters explicitly. (#2303)
• Validate diagnostic summary configuration parameters explicitly. (#2304)
• Validate dynamic model parameters with explicit error checks. (#2305)
• Validate sensor model parameters with explicit error checks. (#2306)
• Reject boolean values as model dimension parameters. (#2307)
• Validate numerical utility sc...

Release 2.2.3

📚 Docs

• Clarify the InformationFormDistributedKalmanNode/IDKFNode prediction-input conventions in the API docs so node-local inputs are interpreted correctly. (#2088)
• Expand the documentation with API-selection, distribution-taxonomy, error-handling, and install-footprint guidance for choosing and deploying PyRecEst. (#2102)
• Document the public API registry and backend capability matrix so supported modules and examples are easier to discover. (#2105)
• Add scientific validation documentation, a backend-portable workflows tutorial, and backend API matrix checks for cross-backend usage. (#2109)
• Document backend contracts and advanced tracking behavior while tightening distribution API and scientific-invariant coverage. (#2111)

📦 Dependencies

• Update generated requirements to use NumPy 2.4.6. (#2089)
• Refresh the locked pymdown-extensions version for documentation builds. (#2093)
• Refresh the locked idna version in Poetry installs to 3.15. (#2094)
• Raise the runtime and development idna pins to 3.15 and add a temporary OSV override so dependency scans report the fixed version. (#2095)
• Regenerate the dependency files, including new requirements-all.txt and requirements-min.txt install variants. (#2097)
• Refresh generated requirement files after the dependency set changes. (#2099)
• Update generated requirements to use idna 3.15. (#2100)
• Refresh the Poetry lockfile to keep packaged dependencies in sync. (#2115)
• Refresh the generated requirements and lockfile entries after dependency changes. (#2122)
• Update generated requirements to use the latest astropy-iers-data tables package. (#2125)
• Update generated requirements to use requests 2.34.2. (#2126)
• Update generated requirements to use networkx 3.6.1. (#2127)
• Update generated requirements to use boule 0.6.0. (#2128)
• Update generated requirements to use packaging 26.2. (#2129)
• Refresh the generated requirements for the Healpy, JAX, and PyTorch install variants. (#2131)
• Update generated requirements to use requests 2.34.2. (#2133)
• Update generated requirements to use pandas 3.0.3. (#2134)
• Update generated requirements to use setuptools 82.0.1. (#2135)
• Update generated requirements to use packaging 26.2. (#2136)
• Update generated requirements to use idna 3.16. (#2137)
• Refresh the generated requirements for the Healpy, JAX, and PyTorch install variants after the latest dependency bumps. (#2140)

🐛 Fixes

• Make the JAX array_from_sparse conversion place multidimensional entries correctly when rebuilding dense arrays. (#2091)
• Make HypersphericalGridDistribution.to_hemisphere() reject invalid even-sized grids with a clear ValueError. (#2092)
• Make NumPy and JAX backend contract operations such as reductions, take, cross, scatter_add, and slicing behave more consistently. (#2104)
• Resolve hyperrectangular distribution edge cases so NumPy, JAX, and PyTorch backends handle them consistently. (#2108)
• Restore custom distribution behavior and filter lazy exports so distributions and filter modules load consistently again. (#2110)
• Make Fibonacci, spherical, circular, and Driscoll-Healy samplers return valid grids for singleton and low-density edge cases. (#2117)
• Make evaluation runs pass Kalman process noise correctly, honor likelihood-based updates, and accept simulator-produced MTT ground-truth arrays. (#2119)
• Resolve evaluation distance selection and object-array estimate extraction so filter evaluations use the intended metrics and mean extractors. (#2121)

✨ Features

• Add CLI scenario tooling and backend-support utilities while resolving several tracking, distribution, and Gaussian numerics edge cases across PyRecEst. (#2096)

⚡ Improvements

• Add backend portability, optional-dependency, and numerical-stability utilities plus compatibility and release-note tooling for easier setup across environments. (#2103)
• Add backend capability and public API contract checks plus benchmark baselines to make supported features and regressions easier to track. (#2107)
• Add pyrecest.api_registry and install, benchmark, and public-API checks so supported imports and dependency footprints are easier to audit. (#2112)
• Add reproducibility and scenario-registry helpers and improve diagnostics, CLI discovery, and evaluation/filter backend boundaries. (#2114)

Release 2.2.2

What's Changed

• Fix MegaLinter lint errors across three files by @Copilot in #2067
• [MegaLinter] Apply linters automatic fixes by @github-actions[bot] in #2069
• Add fixed-lag and fixed-interval smoothers for MEM-QKF trackers by @FlorianPfaff in #2068
• [MegaLinter] Apply linters automatic fixes by @github-actions[bot] in #2070
• Add RB-FFBSi smoother for MEM-RBPF trackers by @FlorianPfaff in #2071
• [MegaLinter] Apply linters automatic fixes by @github-actions[bot] in #2072
• Align LOMEM update with paper reduction by @FlorianPfaff in #2074
• Fix MegaLinter E402 and invalid parameterized test case causing CI failures by @Copilot in #2078
• Ensure hypertoroidal wrapped-normal mean is 1-D by @FlorianPfaff in #2082
• Fix von Mises moment inversion by @FlorianPfaff in #2081
• Fix JAX wrapped-normal PDF convergence check by @FlorianPfaff in #2079
• Bump numpy from 2.4.4 to 2.4.5 by @dependabot[bot] in #2073
• Handle undefined hyperspherical mean directions by @FlorianPfaff in #2080
• Guard undefined hyperspherical mean directions by @FlorianPfaff in #2077
• Fix von Mises moment inversion edge cases by @FlorianPfaff in #2075
• Adapt RM smoother to Granstrom extent recursion by @FlorianPfaff in #2076
• [MegaLinter] Apply linters automatic fixes by @github-actions[bot] in #2085
• Bump version from 2.2.1 to 2.2.2 by @FlorianPfaff in #2086
• Update requirements by @github-actions[bot] in #2087

Full Changelog: 2.2.1...2.2.2

Release 2.2.1

What's Changed

• Add tracking performance metrics by @FlorianPfaff in #2011
• Add fixed-lag out-of-sequence measurement utilities by @FlorianPfaff in #2010
• Add ready-made dynamic and sensor model catalog by @FlorianPfaff in #2014
• Add association hypothesis and gating helpers by @FlorianPfaff in #2016
• Use matvec for camera projection helpers by @FlorianPfaff in #2017
• Add association hypothesis smoke workflow by @FlorianPfaff in #2019
• [MegaLinter] Apply linters automatic fixes by @github-actions[bot] in #2021
• Bump urllib3 from 2.6.3 to 2.7.0 by @dependabot[bot] in #2022
• Add Euclidean boxed particle filter by @FlorianPfaff in #2023
• Fix block particle filter lint issues by @FlorianPfaff in #2025
• [MegaLinter] Apply linters automatic fixes by @github-actions[bot] in #2026
• Fix MEM-RBPF tracker lint issue by @FlorianPfaff in #2027
• Bump urllib3 requirements to 2.7.0 by @FlorianPfaff in #2028
• Add VBRM extended object tracker by @FlorianPfaff in #2031
• Add MEM-QKF tracker by @FlorianPfaff in #2033
• [MegaLinter] Apply linters automatic fixes by @github-actions[bot] in #2034
• Update MEM-QKF robustness by @FlorianPfaff in #2035
• Add LOMEM extended object tracker by @FlorianPfaff in #2036
• Reuse MegaLinter autofix PR branch by @FlorianPfaff in #2041
• Add MEM-QKF batch update mode by @FlorianPfaff in #2037
• [MegaLinter] Apply linters automatic fixes by @github-actions[bot] in #2043
• Use backend in MEM-RBPF tracker by @FlorianPfaff in #2044
• [MegaLinter] Apply linters automatic fixes by @github-actions[bot] in #2045
• Restore velocity locked MEM-QKF aliases by @FlorianPfaff in #2047
• Bump torch from 2.11.0 to 2.12.0 by @dependabot[bot] in #2049
• [MegaLinter] Apply linters automatic fixes by @github-actions[bot] in #2048
• Add tests and exports for VL-MEM-QKF smoother by @FlorianPfaff in #2053
• [MegaLinter] Apply linters automatic fixes by @github-actions[bot] in #2050
• Add velocity-aided MEM-QKF tracker by @FlorianPfaff in #2051
• Add mode RBPF manifold UKF tracker by @FlorianPfaff in #2052
• [MegaLinter] Apply linters automatic fixes by @github-actions[bot] in #2054
• Fix vMF deterministic sigma-point sampler by @FlorianPfaff in #2055
• [MegaLinter] Apply linters automatic fixes by @github-actions[bot] in #2056
• Fix hypertoroidal Dirac marginalization typing by @FlorianPfaff in #2058
• [MegaLinter] Apply linters automatic fixes by @github-actions[bot] in #2059
• [MegaLinter] Apply linters automatic fixes by @github-actions[bot] in #2060
• Fix JAX backend failures in mixture pruning and vMF deterministic sampling by @Copilot in #2061
• [MegaLinter] Apply linters automatic fixes by @github-actions[bot] in #2062
• Bump version from 2.2.0 to 2.2.1 by @FlorianPfaff in #2063
• Update requirements by @github-actions[bot] in #2064
• [MegaLinter] Apply linters automatic fixes by @github-actions[bot] in #2065

Full Changelog: 2.2.0...2.2.1

Release 2.2.0

⚡ Improvements

• Expose linear Gaussian innovation diagnostics through KalmanFilter.innovation_linear(...) and KalmanFilter.normalized_innovation_squared_linear(...) for backend-native Kalman consistency checks. (#1999)
• Add KalmanFilter.update_model_robust(...) so model-based Kalman filters can use robust linear-Gaussian updates with gating, Huber, and Student-t scaling. (#2002)
• Add optional diagnostics to linear Gaussian and Kalman updates, including residuals, normalized innovation squared, measurement-noise scale, and action labels. (#1987)
• Add inverse and composition helpers to AffineTransform for point-set registration workflows. (#1991)
• Add student_t_covariance_scale(...) to compute robust measurement-covariance inflation for Student-t Kalman updates. (#1988)

✨ Features

• Add NamedPairwiseFeatureSchema and CalibratedPairwiseAssociationModel to define named pairwise association features and turn them into calibrated match probabilities or costs. (#2004)
• Add pairwise_mahalanobis_distances and pairwise_covariance_shape_components in pyrecest.utils for generic covariance-based association features. (#2003)
• Add linear_gaussian_update_robust(...) and KalmanFilter.update_linear_robust(...) for gated, Student-t, Huber, and NIS-inflated linear-Gaussian measurement updates. (#1992)
• Add pyrecest.filters.relaxed_s3f_so3 with relaxed S3F prediction helpers for SO(3) and S3+ x R3 state spaces. (#1998)
• Add generic track-evaluation utilities in pyrecest.utils for scoring links, complete tracks, fragmentation, and aggregate track matrices across sessions. (#1997)
• Add confidence-aware, heteroskedastic geodesic log-likelihood updates for SO3ProductParticleFilter and PartitionedSO3ProductParticleFilter. (#1996)
• Expose a public pyrecest.distributions.so3_helpers module for quaternion normalization, exp/log maps, rotation conversion, geodesic distance, and vector rotation on SO(3). (#1990)
• Add so3_right_multiplication_grid_transition(...) and quaternion_grid_transition_density(...) for quaternion-grid prediction on SO(3). (#1995)

🐛 Fixes

• Resolve scalar SO(3) geodesic-distance handling and keep quaternion-grid transition densities normalized for SO(3) particle-filter workflows. (#2000)

Release 2.1.0

✨ Features

• Add PartitionedSO3ProductParticleFilter for SO(3)^K states with configurable partitions, per-block weights, and block-wise resampling. (#1980)
• Add configurable post-update resampling to particle filters so users can keep weighted particles, trigger resampling manually, or apply a custom resampling rule. (#1978)
• Let FullSCGPTracker.update() weight or mask individual measurements so unreliable detections can be down-weighted or ignored during shape and kinematic updates. (#1976)
• Add SphericalHarmonicsEOTTracker for 3-D star-convex extended-object tracking with spherical-harmonic extent coefficients. (#1973)

⚡ Improvements

• Speed up BinghamDistribution normalization and moment fitting, especially for 2D and 4D cases, to make Bingham-based estimation much faster. (#1972)
• Make HyperhemisphericalGridFilter.get_point_estimate() return the dominant scatter-matrix eigenvector directly for much faster point estimates. (#1970)

Release 2.0.0

✨ Features

• Introduce AdditiveNoiseTransitionModel and AdditiveNoiseMeasurementModel so nonlinear additive-noise dynamics and sensor models can be defined once and reused across compatible filters. (#1950)
• Add SO(3) and SO(3)^K representation conversion support, including aliases and tangent-Gaussian and Dirac approximations for rotation distributions. (#1939)
• Add reusable model adapters for grid filters so AbstractGridFilter can update from likelihood models and predict from transition-density models. (#1948)
• Expand representation conversion aliases with domain-specific targets, custom alias registration, and a runnable conversion example for more distribution families. (#1964)
• Add predict_model(...) and update_model(...) support to particle filters for reusable transition-sampling and likelihood model objects. (#1947)
• Add predict_model(...) and update_model(...) support to UnscentedKalmanFilter for reusable additive-noise transition and measurement models. (#1944)
• Add KalmanFilter.predict_model(...) and update_model(...) so linear Gaussian model objects can be passed directly to the filter. (#1945)
• Add reusable likelihood, transition-sampling, and transition-density model objects in pyrecest.models for particle, grid, and related estimators. (#1949)
• Add reusable linear Gaussian transition and measurement model classes in pyrecest.models with prediction and noise helper methods. (#1946)
• Add string-based representation conversion aliases such as particles, gaussian, grid, and fourier, plus convert_to(...) convenience methods on distributions. (#1931)
• Introduce a distribution representation-conversion layer with convert_distribution(...), conversion registration helpers, and method-form convert_to(...) access on distributions. (#1929)
• Add proposal-based rejuvenation for goal-conditioned replay particle filters so particles can be pulled back onto high-likelihood position hypotheses after an update. (#1928)
• Add GoalConditionedReplayParticleIMMFilter with per-particle motion modes for stationary, diffusion, momentum, goal-directed, and jump replay dynamics. (#1927)

⚡ Improvements

• Make the representation conversion API available directly from pyrecest.distributions, including package-level exports for convert_distribution(...) and related helpers. (#1940)
• Add pyrecest.protocols.testing helpers that make it easier to check whether custom distributions, filters, and models satisfy PyRecEst capability protocols. (#1953)
• Add public model capability protocols and adapter helpers that let existing Kalman-style APIs accept structurally compatible model objects. (#1962)
• Add a protocol capability matrix for representative distributions and filters, and ensure batched particle-model predictions keep the expected particle layout. (#1965)
• Ship a py.typed marker so PyRecEst exposes its type information to external type checkers and typed downstream projects. (#1954)
• Introduce the public pyrecest.protocols package with common array and dimension contracts plus filter capability protocols for extension code. (#1952)
• Add backend-aware validation utilities for model vectors, matrices, covariances, and inferred state dimensions to make reusable model objects safer to construct. (#1942)
• Speed up StateSpaceSubdivisionFilter linear updates and relaxed S3F covariance calculations with vectorized implementations. (#1926)

📚 Docs

• Document how to extend pyrecest.protocols and add runnable custom distribution and custom filter examples. (#1963)
• Add executable Kalman, Unscented Kalman, and particle filter examples that show how to use reusable model objects across filters. (#1943)