API stability and support#
SplineOps 2.3 provides stable CPU APIs for all seven numerical modules,
including TensorSpline and its reusable geometry plans. Stability is a
compatibility commitment for the documented shapes, precisions, boundaries,
and operators below. It does not imply that different spline methods are
interchangeable or that a benchmark speedup applies to every machine.
Supported numerical APIs#
Module |
Supported contract |
Numerical validation |
|---|---|---|
N-D interpolation and projection antialiasing; explicit axes, output geometry, buffers, reusable plans, and native/Python execution. |
Independent projection quadrature, native/reference parity, cache and concurrency tests, and installed-wheel checks. |
|
Uniform N-D construction grids; arbitrary point and tensor-grid queries; all named bases and zero, mirror, and periodic modes; float32/float64 and complex64/complex128; reusable geometry and refitting. |
SciPy references, independent B-spline/derivative definitions for every O-MOMS degree, dense cardinal solves for all boundaries and short signals, mixed-axis complex references, and threaded geometry reuse. |
|
2-D/3-D pull-back transforms, degrees 0–7, float32/float64, explicit batch/channel axes, output buffers, reusable coefficients and archives. SciPy and NumPy evaluators share the coefficient contract. |
Analytical geometry, matched SciPy/reference comparisons, immutable archive fixtures, corruption checks, and repeated process/thread reuse. |
|
Cubic whole-sample mirror derivatives in 2-D/3-D, physical spacing, selected gradient/Hessian/Laplacian families, explicit axes and exact structured buffers. Legacy angular maps remain 2-D. |
Polynomial and trigonometric invariants, independent clamped cubic piecewise polynomials, scalar/batched parity, and repeated buffer reuse. |
|
Finite real inputs; periodic 1-D fractional interpolation, periodic N-D radial smoothing, and first-order recursive smoothing with steady-state ends. Each formulation keeps its own operator. |
Dense DFT and linear-system references, analytical frequency responses, constant preservation, periodic autocorrelation, and bounded batch FFT workspace. |
|
Finite real, strictly increasing 1-D sample locations; TV denoising, reusable factorizations, lambda paths, diagnostics, and sparse linear interpolation. Check convergence before using an iterative fit. |
Independent primal and dual optimizers, known sparse hinge models, coordinate-translation invariance, and integer-promotion regressions. |
|
Mirror pyramids with explicit odd/singleton behavior; 2-D Haar and
spline wavelets with explicit batch axes and lengths divisible by
|
Independent Haar coefficients and energy, randomized reconstruction, legacy numerical fixtures, and scalar/batch comparisons. |
Use float32 or float64 for the supported real-valued numerical paths.
TensorSpline also supports the matching complex precisions; its
construction coordinates must have the corresponding real precision.
Integer promotion and result precision are specified in each module guide.
Half and extended precision are not covered by the stability promise.
Compatibility policy#
Public imports, argument meanings, shapes, dtype rules, boundary conventions, and documented return types are maintained within the 2.x series.
Bug fixes may correct numerical results or reject invalid inputs. Such changes are described in the changelog with regression tests.
An intentional incompatible API change needs a deprecation path and migration guidance. Internal names beginning with an underscore are private.
Cached plans describe fixed operators. Construct a new affine, smoothing, or denoising plan when its geometry or fixed parameters change. Reassigning settings cannot silently leave old caches active.
Existing affine schema-1 and schema-2 coefficient archives remain readable. Archive fixtures are retained when a future format is introduced.
TensorSplineQueryPlanremains an alias ofTensorSplineGeometryPlan.Differentialsremains available alongsideDifferentialPlan; neither legacy name is deprecated in 2.3.Wavelet coefficient arrays preserve axis order and dtype; batched results may use non-contiguous views over contiguous spatial planes. C-contiguity is not part of the coefficient-array contract.
Scope that remains experimental#
CuPy interoperability is limited to parts of TensorSpline and lacks a
dedicated GPU CI runner. It remains experimental independently of the stable
NumPy implementation. Custom basis/mode extensions are the caller’s numerical
responsibility. See Backend support contract for details.
The order-5 spline wavelet is a stable API for the existing approximate tap table, not a perfect-reconstruction guarantee. Reported reconstruction error depends on signal magnitude, shape, and scale count; the fixture bounds in the guide are measured cases rather than universal absolute bounds.
How stability is checked#
The test matrix covers Python 3.11–3.13 on Linux, macOS, and Windows, plus the declared minimum NumPy/SciPy versions. Release validation also installs the 12 platform wheels, runs reference comparisons, repeats public-API workflows, and builds the documentation. Performance checks compare numerical outputs before considering timings. See 2.3 numerical API review for the 2.3 review and Release validation for the exact release gates.