SciPy 1.16.0 Release Notes

SciPy 1.16.0 is the culmination of 6 months of hard work. It contains many new features, numerous bug-fixes, improved test coverage and better documentation. There have been a number of deprecations and API changes in this release, which are documented below. All users are encouraged to upgrade to this release, as there are a large number of bug-fixes and optimizations. Before upgrading, we recommend that users check that their own code does not use deprecated SciPy functionality (to do so, run your code with python -Wd and check for DeprecationWarning s). Our development attention will now shift to bug-fix releases on the 1.16.x branch, and on adding new features on the main branch.

This release requires Python 3.11-3.13 and NumPy 1.25.2 or greater.

Highlights of this release

• Improved experimental support for the Python array API standard, including new support in scipy.signal, and additional support in scipy.stats and scipy.special. Improved support for JAX and Dask backends has been added, with notable support in scipy.cluster.hierarchy, many functions in scipy.special, and many of the trimmed statistics functions.
• scipy.optimize now uses the new Python implementation from the PRIMA package for COBYLA. The PRIMA implementation fixes many bugs in the old Fortran 77 implementation with a better performance on average.
• scipy.sparse.coo_array now supports n-D arrays with reshaping, arithmetic and reduction operations like sum/mean/min/max. No n-D indexing or scipy.sparse.random_array support yet.
• Updated guide and tools for migration from sparse matrices to sparse arrays.
• Nearly all functions in the scipy.linalg namespace that accept array arguments now support N-dimensional arrays to be processed as a batch.
• Two new scipy.signal functions, scipy.signal.firwin_2d and scipy.signal.closest_STFT_dual_window, for creation of a 2-D FIR filter and scipy.signal.ShortTimeFFT dual window calculation, respectively.
• A new class, scipy.spatial.transform.RigidTransform, provides functionality to convert between different representations of rigid transforms in 3-D space.
• A new function scipy.ndimage.vectorized_filter for generic filters that take advantage of a vectorized Python callable was added.

New features

scipy.io improvements

• scipy.io.savemat now provides informative warnings for invalid field names.
• scipy.io.mmread now provides a clearer error message when provided with a source file path that does not exist.
• scipy.io.wavfile.read can now read non-seekable files.

scipy.integrate improvements

• The error estimate of scipy.integrate.tanhsinh was improved.

scipy.interpolate improvements

• Batch support was added to scipy.interpolate.make_smoothing_spline.

scipy.linalg improvements

• Nearly all functions in the scipy.linalg namespace that accept array arguments now support N-dimensional arrays to be processed as a batch. See linalg_batch for details.
• scipy.linalg.sqrtm is rewritten in C and its performance is improved. It also tries harder to return real-valued results for real-valued inputs if possible. See the function docstring for more details. In this version the input argument disp and the optional output argument errest are deprecated and will be removed four versions later. Similarly, after changing the underlying algorithm to recursion, the blocksize keyword argument has no effect and will be removed two versions later.
• Wrappers for ?stevd, ?langb, ?sytri, ?hetri and ?gbcon were added to scipy.linalg.lapack.
• The default driver of scipy.linalg.eigh_tridiagonal was improved.
• scipy.linalg.solve can now estimate the reciprocal condition number and the matrix norm calculation is more efficient.

scipy.ndimage improvements

• A new function scipy.ndimage.vectorized_filter for generic filters that take advantage of a vectorized Python callable was added.
• scipy.ndimage.rotate has improved performance, especially on ARM platforms.

scipy.optimize improvements

• COBYLA was updated to use the new Python implementation from the PRIMApackage. The PRIMA implementation fixes many bugs in the old Fortran 77 implementation. In addition, it results in fewer function evaluations on average but it depends on the problem and for some problems it can result in more function evaluations or a less optimal result. For those cases the user can try modifying the initial and final trust region radii given by rhobeg and tol respectively. A larger rhobeg can help the algorithm take bigger steps initially, while a smaller tol can help it continue and find a better solution. For more information, see the PRIMA documentation.
• Several of the scipy.optimize.minimize methods, and the scipy.optimize.least_squares function, have been given a workers keyword. This allows parallelization of some calculations via a map-like callable, such as multiprocessing.Pool. These parallelization opportunities typically occur during numerical differentiation. This can greatly speed up minimization when the objective function is expensive to calculate.
• The lm method of scipy.optimize.least_squares can now accept 3-point and cs for the jac keyword.
• The SLSQP Fortran 77 code was ported to C. When this method is used now the constraint multipliers are exposed to the user through the multiplier keyword of the returned scipy.optimize.OptimizeResult object.
• NNLS code has been corrected and rewritten in C to address the performance regression introduced in 1.15.x
• scipy.optimize.root now warns for invalid inner parameters when using the newton_krylov method
• The return value of minimization with method='L-BFGS-B' now has a faster hess_inv.todense() implementation. Time complexity has improved from cubic to quadratic.
• scipy.optimize.least_squares has a new callback argument that is applicable to the trf and dogbox methods. callback may be used to track optimization results at each step or to provide custom conditions for stopping.

scipy.signal improvements

• A new function scipy.signal.firwin_2d for the creation of a 2-D FIR Filter using the 1-D window method was added.
• scipy.signal.cspline1d_eval and scipy.signal.qspline1d_eval now provide an informative error on empty input rather than hitting the recursion limit.
• A new function scipy.signal.closest_STFT_dual_window to calculate the scipy.signal.ShortTimeFFT dual window of a given window closest to a desired dual window.
• A new classmethod scipy.signal.ShortTimeFFT.from_win_equals_dual to create a scipy.signal.ShortTimeFFT instance where the window and its dual are equal up to a scaling factor. It allows to create short-time Fourier transforms which are unitary mappings.
• The performance of scipy.signal.convolve2d was improved.

scipy.sparse improvements

• scipy.sparse.coo_array now supports n-D arrays using binary and reduction operations.
• Faster operations between two DIA arrays/matrices for: add, sub, multiply, matmul.
• scipy.sparse.csgraph.dijkstra shortest_path is more efficient.
• scipy.sparse.csgraph.yen has performance improvements.
• Support for lazy loading of sparse.csgraph and sparse.linalg was added.

scipy.spatial improvements

• A new class, scipy.spatial.transform.RigidTransform, provides functionality to convert between different representations of rigid transforms in 3-D space, its application to vectors and transform composition. It follows the same design approach as scipy.spatial.transform.Rotation.
• scipy.spatial.transform.Rotation now has an appropriate __repr__ method, and improved performance for its scipy.spatial.transform.Rotation.apply method.

scipy.stats improvements

• A new function scipy.stats.quantile, an array API compatible function for quantile estimation, was added.
• scipy.stats.make_distribution was extended to work with existing discrete distributions and to facilitate the creation of custom distributions in the new random variable infrastructure.
• A new distribution, scipy.stats.Binomial, was added.
• An equal_var keyword was added to scipy.stats.tukey_hsd (enables the Games-Howell test) and scipy.stats.f_oneway (enables Welch ANOVA).
• The moment calculation for scipy.stats.gennorm was improved.
• The scipy.stats.mode implementation was vectorized, for faster batch calculation.
• Support for axis, nan_policy, and keepdims keywords was added to scipy.stats.power_divergence, scipy.stats.chisquare, scipy.stats.pointbiserialr, scipy.stats.kendalltau, scipy.stats.weightedtau, scipy.stats.theilslopes, scipy.stats.siegelslopes, scipy.stats.boxcox_llf, and scipy.stats.linregress.
• Support for keepdims and nan_policy keywords was added to scipy.stats.gstd.
• The performance of scipy.stats.special_ortho_group and scipy.stats.pearsonr was improved.
• Support for an rng keyword argument was added to the logcdf and cdf methods of multivariate_normal_gen and multivariate_normal_frozen.

Array API Standard Support

Experimental support for array libraries other than NumPy has been added to multiple submodules in recent versions of SciPy. Please consider testing these features by setting the environment variable SCIPY_ARRAY_API=1 and providing PyTorch, JAX, CuPy or Dask arrays as array arguments.

Many functions in scipy.stats, scipy.special, scipy.optimize, and scipy.constants now provide tables documenting compatible array and device types as well as support for lazy arrays and JIT compilation. New features with support and old features with support added for SciPy 1.16.0 include:

• Most of the scipy.signal functionality
• scipy.ndimage.vectorized_filter
• scipy.special.stdtrit
• scipy.special.softmax
• scipy.special.log_softmax
• scipy.stats.quantile
• scipy.stats.gstd
• scipy.stats.rankdata

Features with extended array API support (generally, improved support for JAX and Dask) in SciPy 1.16.0 include:

• many of the scipy.cluster.hierarchy functions
• many functions in scipy.special
• many of the trimmed statistics functions in scipy.stats

SciPy now has a CI job that exercises GPU (CUDA) support, and as a result using PyTorch, CuPy or JAX arrays on GPU with SciPy is now more reliable.

Deprecated features

• The unused atol argument of scipy.optimize.nnls is deprecated and will be removed in SciPy 1.18.0.
• The disp argument of scipy.linalg.signm, scipy.linalg.logm, and scipy.linalg.sqrtm will be removed in SciPy 1.18.0.
• scipy.stats.multinomial now emits a FutureWarning if the rows of p do not sum to 1.0. This condition will produce NaNs beginning in SciPy 1.18.0.
• The disp and iprint arguments of the l-bfgs-b solver of scipy.optimize have been deprecated, and will be removed in SciPy 1.18.0.

Expired Deprecations

• scipy.sparse.conjtransp has been removed. Use .T.conj() instead.
• The quadrature='trapz' option has been removed from scipy.integrate.quad_vec, and scipy.stats.trapz has been removed. Use trapezoid in both instances instead.
• scipy.special.comb and scipy.special.perm now raise when exact=True and arguments are non-integral.
• Support for inference of the two sets of measurements from the single argument x has been removed from scipy.stats.linregress. The data must be specified separately as x and y.
• Support for NumPy masked arrays has been removed from scipy.stats.power_divergence and scipy.stats.chisquare.
• A significant number of functions from non-public namespaces (e.g., scipy.sparse.base, scipy.interpolate.dfitpack) were cleaned up. They were previously already emitting deprecation warnings.

Backwards incompatible changes

• Several of the scipy.linalg functions for solving a linear system (e.g. scipy.linalg.solve) documented that the RHS argument must be either 1-D or 2-D but did not always raise an error when the RHS argument had more the two dimensions. Now, many-dimensional right hand sides are treated according to the rules specified in linalg_batch.
• scipy.stats.bootstrap now explicitly broadcasts elements of data to the same shape (ignoring axis) before performing the calculation.
• Several submodule names are no longer available via from scipy.signal import *, but may still be imported directly, as detailed at https://github.com/scipy/scipy-stubs/pull/549.

Build and packaging related changes

• The minimum supported version of Clang was bumped from 12.0 to 15.0.
• The lowest supported macOS version for wheels on PyPI is now 10.14 instead of 10.13.
• The sdist contents were optimized, resulting in a size reduction of about 50%, from 60 MB to 30 MB.
• For Cython>=3.1.0, SciPy now uses the new cython --generate-shared functionality, which reduces the total size of SciPy's wheels and on-disk installations significantly.
• SciPy no longer contains an internal shared library that requires RPATH support, after sf_error_state was removed from scipy.special.
• A new build option -Duse-system-libraries has been added. It allows opting in to using system libraries instead of using vendored sources. Currently Boost.Math and Qhull are supported as system build dependencies.

Other changes

• A new accompanying release of scipy-stubs (v1.16.0.0) is available at https://github.com/scipy/scipy-stubs/releases/tag/v1.16.0.0
• The internal dependency of scipy._lib on scipy.sparse was removed, which reduces the import time of a number of other SciPy submodules.
• Support for free-threaded CPython was improved: the last known thread-safety issues in scipy.special were fixed, and pytest-run-parallel is now used in a CI job to guard against regressions.
• Support for spin as a developer CLI was added, including support for editable installs. The SciPy-specific python dev.py CLI will be removed in the next release cycle in favor of spin.
• The vendored Qhull library was upgraded from version 2019.1 to 2020.2.
• A large amount of the C++ code in scipy.special was moved to the new header-only xsf library. That library was included back in the SciPy source tree as a git submodule.
• The namedtuple-like bunch objects returned by some SciPy functions now have improved compatibility with the polars library.
• The output of the rvs method of scipy.stats.wrapcauchy is now mapped to the unit circle between 0 and 2 * pi.
• The lm method of scipy.optimize.least_squares now has a different behavior for the maximum number of function evaluations, max_nfev. The default for the lm method is changed to 100 * n, for both a callable and a numerically estimated jacobian. This limit on function evaluations excludes those used for any numerical estimation of the Jacobian. Previously the default when using an estimated jacobian was 100 * n * (n + 1), because the method included evaluations used in the estimation. In addition, for the lm method the number of function calls used in Jacobian approximation is no longer included in OptimizeResult.nfev. This brings the behavior of lm, trf, and dogbox into line.

Authors

• Name (commits)
• h-vetinari (4)
• aiudirog (1) +
• Anton Akhmerov (2)
• Thorsten Alteholz (1) +
• Gabriel Augusto (1) +
• Backfisch263 (1) +
• Nickolai Belakovski (5)
• Peter Bell (1)
• Benoît W. (1) +
• Evandro Bernardes (1)
• Gauthier Berthomieu (1) +
• Maxwell Bileschi (1) +
• Sam Birch (1) +
• Florian Bourgey (3) +
• Charles Bousseau (2) +
• Richard Strong Bowen (2) +
• Jake Bowhay (127)
• Matthew Brett (1)
• Dietrich Brunn (53)
• Evgeni Burovski (254)
• Christine P. Chai (12) +
• Gayatri Chakkithara (1) +
• Saransh Chopra (2) +
• Omer Cohen (1) +
• Lucas Colley (91)
• Yahya Darman (3) +
• Benjamin Eisele (1) +
• Donnie Erb (1)
• Sagi Ezri (58) +
• Alexander Fabisch (2) +
• Matthew H Flamm (1)
• Karthik Viswanath Ganti (1) +
• Neil Girdhar (1)
• Ralf Gommers (162)
• Rohit Goswami (4)
• Saarthak Gupta (4) +
• Matt Haberland (326)
• Sasha Hafner (1) +
• Joren Hammudoglu (11)
• Chengyu Han (1) +
• Charles Harris (1)
• Kim Hsieh (4) +
• Yongcai Huang (2) +
• Lukas Huber (1) +
• Yuji Ikeda (2) +
• Guido Imperiale (105) +
• Robert Kern (2)
• Harin Khakhi (2) +
• Agriya Khetarpal (4)
• Daniil Kiktenko (1) +
• Kirill R. (2) +
• Tetsuo Koyama (1)
• Jigyasu Krishnan (1) +
• Abhishek Kumar (2) +
• Pratham Kumar (3) +
• David Kun (1) +
• Eric Larson (3)
• lciti (1)
• Antony Lee (1)
• Kieran Leschinski (1) +
• Thomas Li (2) +
• Yuxi Long (2) +
• Christian Lorentzen (2)
• Loïc Estève (4)
• Panos Mavrogiorgos (1) +
• Nikolay Mayorov (2)
• Melissa Weber Mendonça (10)
• Michał Górny (1)
• Miguel Cárdenas (2) +
• Swastik Mishra (1) +
• Sturla Molden (2)
• Andreas Nazlidis (1) +
• Andrew Nelson (209)
• Parth Nobel (1) +
• Nick ODell (9)
• Giacomo Petrillo (1)
• Victor PM (10) +
• pmav99 (1) +
• Ilhan Polat (74)
• Tyler Reddy (128)
• Érico Nogueira Rolim (1) +
• Pamphile Roy (10)
• Mikhail Ryazanov (6)
• Atsushi Sakai (9)
• Marco Salathe (1) +
• sanvi (1) +
• Neil Schemenauer (2) +
• Daniel Schmitz (20)
• Martin Schuck (1) +
• Dan Schult (33)
• Tomer Sery (19)
• Adrian Seyboldt (1) +
• Scott Shambaugh (4)
• ShannonS00 (1) +
• sildater (3) +
• Param Singh (1) +
• G Sreeja (7) +
• Albert Steppi (133)
• Kai Striega (3)
• Anushka Suyal (2)
• Julia Tatz (1) +
• Tearyt (1) +
• Elia Tomasi (1) +
• Jamie Townsend (2) +
• Edgar Andrés Margffoy Tuay (4)
• Matthias Urlichs (1) +
• Mark van Rossum (1) +
• Jacob Vanderplas (2)
• David Varela (2) +
• Christian Veenhuis (3)
• vfdev (1)
• Stefan van der Walt (2)
• Warren Weckesser (5)
• Jason N. White (1) +
• windows-server-2003 (5)
• Zhiqing Xiao (1)
• Pavadol Yamsiri (1)
• Rory Yorke (3)
• Irwin Zaid (4)
• Austin Zhang (1) +
• William Zijie Zhang (1) +
• Zaikun Zhang (1) +
• Zhenyu Zhu (1) +
• Eric Zitong Zhou (11) +
• Case Zumbrum (2) +
• ਗਗਨਦੀਪ ਸਿੰਘ (Gagandeep Singh) (45)

A total of 126 people contributed to this release. People with a "+" by their names contributed a patch for the first time. This list of names is automatically generated, and may not be fully complete.