◉ COOLJAPAN
← All posts
March 6, 2026 • 8 min read

NumRS2 0.3.0 Released — Periodic Splines in O(n), Compressed Pure-Rust .npz, and the SciRS2 0.3.0 Lift

NumRS2 0.3.0 replaces the O(n²) cyclic spline solve with a Sherman-Morrison O(n) cyclic Thomas algorithm, enables DEFLATE compression for .npz via OxiARC (no zlib, no C), and lifts the entire scirs2-* stack to v0.3.0 — faster KD-trees, SIMD, and special functions.

release numrs2 numpy numerical-computing scirs2 pure-rust linear-algebra oxiblas

Periodic cubic splines just went from O(n²) to O(n) — and your .npz files now write compressed, in pure Rust.

Today we released NumRS2 0.3.0 — a release that pairs a genuine algorithmic speedup with compressed pure-Rust storage and a full lift onto the SciRS2 0.3.0 stack.

This one earns its minor-version bump. The headline is a real interpolation win: periodic cubic-spline solves are now linear-time. Around it, .npz archives finally write DEFLATE-compressed — through the COOLJAPAN pure-Rust archive layer, with no zlib and no C — and every SciRS2 dependency moves up to v0.3.0, pulling in faster spatial structures, sharper SIMD, and updated special functions.

No C. No Fortran. No system BLAS/LAPACK. No zlib for compression. No Python interpreter overhead. No FFI. Just clean, blazing-fast N-dimensional arrays — broadcasting, fancy indexing, SVD, FFT, autodiff and all — that compile to a single static binary (or WASM) and run everywhere: from laptops to browsers to edge devices to cloud clusters.

Why NumRS2 0.3.0 matters

NumPy is the bedrock of scientific Python, but its foundation shows its age — built on C and Fortran, dependent on a system BLAS/LAPACK you have to link at install time, throttled by the GIL, and stuck reaching for native zlib the moment you want compressed .npz on disk. Periodic interpolation, in particular, is the kind of thing that quietly costs you: a naive cyclic spline solve scales quadratically, and on long periodic signals that is real wall-clock time.

NumRS2 0.3.0 answers with concrete wins:

Technical Deep Dive: The O(n) Cyclic Thomas Solver, Compressed .npz, and the 0.3.0 Stack

Layer 1 — From O(n²) to O(n) periodic splines. A natural cubic spline reduces to a tridiagonal system, which the Thomas algorithm solves in linear time. A periodic (cyclic) cubic spline breaks that nice structure: the periodicity constraint adds corner entries to the matrix, turning it from tridiagonal into “tridiagonal plus a rank-one perturbation.” The old solver fell back to general Gaussian elimination on that system — O(n²). NumRS2 0.3.0 instead applies the Sherman-Morrison identity to handle the rank-one corner correction, then runs the standard tridiagonal Thomas solve twice. The result is mathematically exact and runs in O(n). For periodic data — closed curves, cyclic time series, anything that wraps around — this is the difference between a quadratic blowup and a flat, linear cost.

Layer 2 — DEFLATE .npz through OxiARC. A .npz is, structurally, a multi-file archive of .npy members. Writing it compressed means running each member through DEFLATE inside an archive container. This had been waiting on OxiARC to fix a multi-file bug in its v0.3.0+ line; with that resolved, NumRS2 enables DEFLATE compression for .npz writes. The payoff is twofold: smaller files, and a storage path that is pure Rust top to bottom — no zlib, no native codec linked into your binary. Compressed array I/O is now as sovereign as the math that produced the arrays.

Layer 3 — The SciRS2 0.3.0 lift, across the whole stack. Every scirs2-* dependency moves to v0.3.0 in one coordinated step: scirs2-core 0.3.0 (enhanced SIMD, parallel, and random ops), scirs2-linalg 0.3.0 (linear-algebra improvements on OxiBLAS), scirs2-stats 0.3.0 (statistical-function enhancements), scirs2-fft 0.3.0 (FFT improvements), scirs2-ndimage 0.3.0 (N-dimensional image processing), scirs2-spatial 0.3.0 (spatial algorithms with improved KD-trees), scirs2-special 0.3.0 (special-function updates), and scirs2-numpy 0.3.0 (Python-bindings compatibility). Because the stack moves together, NumRS2 inherits the gains without version-skew between the layers it sits on.

Layer 4 — WASM stays honest. The WebAssembly build is held to the same standard as everything else: the version-assertion tests were updated to expect “0.3”, so the same numerical code that runs on a laptop or a cluster keeps running bit-for-bit in the browser, now on the 0.3.0 stack.

Getting Started

cargo add numrs2

use numrs2::prelude::*;

fn main() -> Result<()> {
    // N-dimensional arrays with NumPy-style broadcasting
    let a = Array::from_vec(vec![1.0, 2.0, 3.0, 4.0]).reshape(&[2, 2]);
    let b = Array::from_vec(vec![5.0, 6.0, 7.0, 8.0]).reshape(&[2, 2]);
    let e = a.matmul(&b)?;             // matrix multiply on OxiBLAS
    println!("a @ b = {}", e);

    // Descriptive statistics
    let data = Array::from_vec(vec![1.0, 2.0, 3.0, 4.0, 5.0]);
    println!("mean = {}, std = {}", data.mean()?, data.std()?);

    // FFT on a real signal
    let signal = Array::from_vec(vec![1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]);
    let spectrum = signal.fft()?;
    println!("spectrum = {}", spectrum);

    Ok(())
}

For periodic data, reach for cubic-spline interpolation: as of 0.3.0 the periodic (cyclic) solve runs in O(n), so fitting a closed curve or a wrap-around time series is linear-time. And when you persist the result, saving to .npz now writes a DEFLATE-compressed, pure-Rust archive — smaller on disk, with no zlib in your binary.

What’s New in 0.3.0

Tips

This is the foundation

NumRS2 is the NumPy-class N-dimensional array core of the COOLJAPAN scientific stack, and 0.3.0 lifts it onto the SciRS2 0.3.0 generation. It sits directly beneath SciRS2, depending on the full scirs2-* family at v0.3.0 — core, linalg, stats, fft, ndimage, spatial, and special — for unified SIMD, statistics, transforms, spatial structures, and special functions. Linear algebra rides on OxiBLAS, serialization on OxiCode, and compressed .npz now on OxiARC — all pure-Rust COOLJAPAN crates.

By March 2026 the ecosystem around it is broad and deepening: OptiRS and PandRS alongside NumRS2; the OxiFFT, OxiARC, OxiZ, OxiGDAL, and OxiLean infrastructure crates; ML and application work spanning ToRSh, VoiRS, TenRSo, TensorLogic, legalis, mecrab, chie, mielin, and oxirag. NumRS2 is the array layer all of them can build on — now faster on periodic data, smaller on disk, and aligned to the 0.3.0 stack.

Repository: https://github.com/cool-japan/numrs

Star the repo if you want a NumPy-class numerical foundation with linear-time periodic splines and compressed pure-Rust storage — and not a line of C or Fortran in the build.

Pure Rust numerical computing is here — fast, safe, and sovereign.

— KitaSan at COOLJAPAN OÜ March 6, 2026

↑ Back to all posts