[0.4.0] - 2026-06-05

Added

• Autodiff (src/autodiff/): forward/reverse mode automatic differentiation with higher-order support (Hessian, Jacobian, hyperdual, Taylor)
• Distributed computing (src/distributed/): data/model parallelism framework, collectives, distributed optimizers, coordinator (~8,000 lines)
• Visualization (src/viz/): plotters-based 2D/3D/matrix/stats/performance plots (pure Rust)
• WebAssembly bindings (src/wasm/): array, linalg, stats, utils bindings; JS/Vite demo app in examples/wasm/
• Reinforcement learning (src/new_modules/rl/): DQN, Actor-Critic, PPO agents, prioritized replay, environment wrappers
• Quantum computing (src/new_modules/quantum/): gates, circuits, state vector simulation, measurements, Deutsch-Jozsa, Grover, QFT
• Model I/O & serving (src/new_modules/model_io/, src/new_modules/serving/): ONNX-compatible serialization, inference engine, model registry, real-time monitoring
• CMA-ES optimizer (src/optimize/cma_es/): IPOP-CMA-ES with step-size adaptation, rank-μ/rank-one covariance updates
• Bayesian optimization (src/optimize/bayesian_opt.rs): GP surrogate, EI/PI/UCB acquisition functions, Matérn/RBF kernels
• Computer vision (src/new_modules/cv/): Gaussian/bilateral/median filters, Canny edges, Harris/FAST corners, morphological ops
• Computational geometry (src/new_modules/geometry/): convex hull (Graham scan), Delaunay triangulation, Voronoi diagrams, polygon ops
• FEM solver (src/new_modules/fem/): 1D/2D FEM with Dirichlet/Neumann BCs, Gaussian quadrature, direct/iterative solvers
• Wavelets (src/new_modules/wavelets/): DWT/CWT, Haar, Daubechies (db2–db10), Symlet, Coiflet families
• Graph algorithms (src/new_modules/graph/): BFS/DFS, Dijkstra/A*/Floyd-Warshall, MST (Kruskal/Prim), max-flow (Dinic)
• Information theory (src/new_modules/information_theory/): Shannon entropy, mutual information, KL/JS divergence, channel capacity
• Control systems (src/new_modules/control/): transfer functions, state space, Routh-Hurwitz, Bode/Nyquist, PID/LQR/pole placement
• Physical constants (src/new_modules/constants/): CODATA 2018/2022 constants with uncertainties, unit conversions
• average_with_weights in src/stats/basic.rs: public function returning (weighted_avg, sum_of_weights) tuple (NumPy returned=True semantics)
• skew() and kurtosis() in src/math/statistics.rs and src/math/mod.rs: compute sample skewness (Fisher-Pearson) and excess kurtosis (fourth standardised moment minus 3); both accept an optional axis parameter and are re-exported from prelude
• f_dist() in src/random/distributions.rs: sample from the F distribution (ratio of two chi-squared variates) given numerator and denominator degrees of freedom
• instance_norm() in src/nn/normalization.rs: instance normalisation for 2-D tensors (normalises each row independently with configurable epsilon)
• Python optimizer methods (src/python/optimize.rs): py_minimize now supports "bfgs" method (numerical gradient via central differences) in addition to the existing "nelder-mead" default
• VECM fitting via Johansen procedure (src/new_modules/timeseries/var.rs): VecmModel::fit now implements the full Johansen (1988) cointegration/error-correction estimation instead of a placeholder
• FEM 2D point evaluation (src/new_modules/fem/solvers.rs): FEM solver can now evaluate solutions at arbitrary 2D points using element shape functions

Changed

• scirs2-core, scirs2-stats, scirs2-linalg, scirs2-ndimage, scirs2-spatial, scirs2-special, scirs2-fft, scirs2-numpy updated from 0.4.2 → 0.5.0
• oxiarc-archive and oxiarc-lz4 updated from 0.2.6 → 0.3.2
• oxicode updated from 0.2 → 0.2.4

Fixed

• Build fix: Updated oxiarc-core lockfile resolution from 0.2.6 to 0.2.8; resolves oxiarc_core::cancel/progress missing-module compile errors introduced by scirs2-core 0.4.4's transitive oxiarc-lz4 0.2.8/oxiarc-zstd 0.2.8 dependencies
• Stable SVD for large matrices (src/linalg_stable.rs): svd_bidiagonal now runs full Golub–Kahan bidiagonalization + Jacobi SVD instead of silently falling back to the n≤3 path
• Stable eigendecomposition for large matrices (src/linalg_stable.rs): symmetric_eigen_tridiagonal now runs a cyclic Jacobi sweeps algorithm instead of silently falling back to the n≤3 path
• Quantum partial trace (src/new_modules/quantum/statevector.rs): correct bit-interleaving index reconstruction; previous code used full_i = i (identity mapping), producing a wrong density matrix for all multi-qubit traces
• Polynomial domain mapping (src/new_modules/polynomial/utils.rs): polyscale now performs real polynomial composition under the affine map; previously returned the input coefficients unchanged
• FEM matrix operations for n>3 (src/new_modules/fem/elements.rs): matrix_determinant and matrix_inverse now handle arbitrary n×n via LU with partial pivoting; previously hard-errored for n>3, blocking 3D/higher-order FEM elements
• Schur decomposition (src/new_modules/matrix_decomp/schur.rs): replaced single Rayleigh-shift with Francis implicit double-shift QR + deflation for correct real Schur form including complex-conjugate eigenpair blocks
• Boltzmann exploration (src/new_modules/rl/utils.rs): BoltzmannExploration::select_action now computes temperature-scaled softmax over Q-values; previously fell back to greedy action selection regardless of temperature
• gamma_ln accuracy (src/new_modules/probabilistic/distributions.rs): replaced custom Stirling approximation with scirs2_special::loggamma (Lanczos g=7, ~15-digit accuracy) with reflection formula for x<0.5
• Unsafe block warnings (src/memory_optimize/cache_layout.rs): removed unnecessary unsafe {} wrappers around std::arch::x86_64::__cpuid (now a safe function in current Rust)
• Clippy compliance (src/simd_optimize/avx2_enhanced/): replaced manual copy loops with copy_from_slice; replaced vec![...] with array literals in tests

Notes

• WASM blocker resolved: scirs2-spatial 0.4.4 feature-gates tokio under async (not enabled by default); numrs2 uses only the parallel feature, so no transitive tokio. Browser WASM (wasm32-unknown-unknown) still requires disabling the gpu and distributed features (which pull tokio directly)
• Known future work: distributed/collective.rs operations (scatter/gather/all-reduce) are stub-implemented returning empty results pending a real network transport layer; distributed/linalg.rs distributed linear algebra returns NotImplemented; GPU NotImplemented branches in gpu/batching.rs (Conv2D batching) and gpu/ops.rs (N-D broadcasting)

Full Changelog: v0.3.3...v0.4.0

Full Changelog: v0.3.2...v0.3.3

Full Changelog: v0.3.1...v0.3.2

Full Changelog: v0.3.0...v0.3.1

Full Changelog: v0.1.1...v0.3.0

Full Changelog: v0.1.1...v0.2.0

Full Changelog: https://github.com/cool-japan/numrs/commits/v0.1.1

Full Changelog: v0.1.0-rc.1...v0.1.0-rc.2

Full Changelog: v0.1.0-beta.3...v0.1.0-rc.1