SciPy 1.18.0 Release Notes

note: SciPy 1.18.0 is not released yet!

SciPy 1.18.0 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 1.18.x branch, and on adding new features on the main branch.

This release requires Python 3.12-3.14 and NumPy 2.0.0 or greater.

Highlights of this release

• SciPy now supports three different build modes for BLAS and LAPACK LP64/ILP64 support, and machinery is provided for downstream cython_lapack consumers to gracefully handle LP64/ILP64 backend builds. ILP64 support has been substantially improved across the SciPy library.
• Remaining Fortran to C translations have been completed---an experimental Fortran-free build option is now available to developers for testing purposes. Developer feedback is welcome on Fortran-free builds.
• scipy.signal.whittaker_henderson now provides access to Whittaker-Henderson smoothing of a discrete signal.
• A large number of scipy.stats functions now support lazy arrays and JAX JIT. Array API support has been improved substantially in SciPy, with at least 21 functions gaining new support in this release. 16 scipy.stats functions have also gained support for MArray input.

New features

scipy.fft improvements

• SciPy's internal FFT backend has switched from pocketfft to its successor package ducc0.fft, which features several incremental improvements. The most significant of those from SciPy's perspective is probably that storage requirements for internally cached plans have been significantly reduced for most long 1D transforms. Plans that require more storage than 1MB will no longer be cached; this mainly affects huge 1D transforms of prime and near-prime sizes.

scipy.interpolate improvements

• Users may now increase the QHull simplex assignment tolerance via the new simplex_tolerance argument to the _call__ methods of LinearNDInterpolator and CloughTocher2dInterpolator. This can help users avoid holes in certain interpolation problems.
• The FITPACK Fortran code has been ported to C.

scipy.differentiate improvements

• scipy.differentiate.derivative now supports passing kwargs to the function whose derivative is desired.

scipy.linalg improvements

• We now support three different build modes for BLAS and LAPACK: LP64-only, ILP64-only, and ILP64 for everyting except cython_blas/cython_lapack/ linalg.blas/linalg.lapack (support for Accelerate and MKL).
• Machinery is now provided for downstream cython_lapack users to gracefully handle LP64/ILP64 backend builds. Worked examples, including build system details, have been included in this release.
• An overwrite_b keyword argument was added to eigvals, for consistency with other similar linalg functions.
• linalg.cholesky now leverages symmetry properties for performance improvements, especially for real matrices. The batching loop of cholesky has now also been moved to a C implementation.
• scipy.linalg.lu and scipy.linalg.det have been rewritten in C++ with batching support in the compiled code.
• Added ILP64 support to scipy.linalg.expm and scipy.linalg.sqrtm.
• The batching loops of scipy.linalg.qr, scipy.linalg.eig, scipy.linalg.lstsq, and scipy.linalg.svd have been moved to C, providing a substantial speedup for batched input.
• The performance of scipy.linalg.expm has been improved.
• The performance for scipy.linalg.solve has improved for batched inputs.

scipy.optimize improvements

• The trust_constr method for minimize was adjusted so that if the x array would result in infeasible constraints, and those constraints were marked as keep_feasible, then the objective function is not called with that x array.
• The COBYQA method for minimize now supports being called concurrently by multiple threads. Previously, multiple threads calling this function would only run one at a time.
• scipy.optimize.nnls, and minimize methods SLSQP and L-BFGS-B now have support for ILP64 LAPACK, when available.
• Functions in scipy.optimize.elementwise now support passing kwargs to the callable function.

scipy.signal improvements

• The new ~scipy.signal.whittaker_henderson implements Whittaker-Henderson smoothing of a discrete signal. It offers different penalties to control the smoothness as well as automatic selection of the penalty strength via optimization of the restricted maximum likelihood (REML) criterion. It is a valuable alternative for the Savitzky-Golay filter ~scipy.signal.savgol_filter. In econometrics, Whittaker-Henderson graduation of penalty order 2 is also known as Hodrick-Prescott filter.
• lfilter_zi was refactored for improved numerical stability and efficiency. It now raises a ValueError if parameter a has leading zeros, i.e., a[0] == 0, since lfilter and filtfilt do not support that as well. Furthermore, a ValueError instead of a LinAlgError is raised if the filter is unstable due to having a pole at z = 1.

scipy.sparse improvements

• In scipy.sparse.csgraph the computation of strongly connected components for directed graphs is now 2x faster with better cache locality, using algorithmic improvements described in the recent survey by Tarjan and Zwick.
• Added ILP64 BLAS/LAPACK support to SuperLU and PROPACK extensions.
• All sparse array/matrix formats now support matrix_transpose/.mT.
• Support for n-dimensional linear operators has been added to scipy.sparse.linalg.LinearOperator.
• scipy.sparse.linalg.minres now supports complex hermitian matrices.

scipy.integrate improvements

• ILP64 support was added for ODEPACK
• scipy.integrate.tanhsinh and scipy.integrate.nsum now support passing kwargs to the function to be integrated.

scipy.spatial improvements

• 3D area calculations are now faster in scipy.spatial.SphericalVoronoi.
• N-dimensional input is now supported for scipy.spatial.distance.minkowski, scipy.spatial.distance.euclidean, and scipy.spatial.distance.seuclidean.
• It is now possible to return sparse arrays rather than matrices from KDTree.sparse_distance_matrix.
• It is now possible to compose Rotation and RigidTransform directly, by automatically promoting Rotation when the two are composed via a multiplication operator.

scipy.special improvements

• The accuracy of the following functions was improved: scipy.special.bdtrik, scipy.special.bdtrin, scipy.special.nbdtrik, scipy.special.nbdtrin.
• The numerical behavior for scipy.special.eval_jacobi has been improved for several parameter combinations.
• The Bessel functions scipy.special.j0 and scipy.special.y0 have improved accuracy for large arguments.

scipy.stats improvements

• The accuracy of scipy.stats.pmean with tiny, nonzero p has been improved.
• The performance of scipy.stats.halfgennorm has been improved.
• zstatistic has been added to the result object of scipy.stats.mannwhitneyu.
• A large number of stats functions now support lazy arrays and JAX JIT (see Python Array API support section below).
• Support for the nan_policy keyword argument has been added to: scipy.stats.obrientransform, scipy.stats.boxcox, scipy.stats.boxcox_normmax, scipy.stats.yeojohnson, scipy.stats.yeojohnson_normmax, and scipy.stats.sigmaclip.
• scipy.stats.ContinuousDistribution.lmoment has been added for computing population L-moments.

Python Array API Standard Support

• Support has been added for CuPy delegation for: interpolate.PPoly, interpolate.BPoly, and interpolate.BSpline.
• CuPy support has been added for scipy.stats.rankdata.
• Array API support has been added for method and trim usage in scipy.stats.ttest_ind.
• Support for MArrays has been added to: scipy.stats.cramervonmises, scipy.stats.ks_1samp, scipy.stats.ks_2samp, scipy.stats.mode, scipy.stats.rankdata, scipy.stats.kruskal, scipy.stats.brunnermunzel, scipy.stats.spearmanrho, scipy.stats.friedmanchisquare, scipy.stats.cramervonmises_2samp, scipy.stats.mannwhitneyu, scipy.stats.wilcoxon, scipy.stats.fligner, scipy.stats.linregress, scipy.stats.alexandergovern, and scipy.stats.levene.
• Array API support has been added to: scipy.stats.quantile_test, scipy.stats.kendalltau (via NumPy conversion), scipy.stats.kstest, scipy.sparse.linalg.LinearOperator, scipy.stats.cumfreq, scipy.stats.relfreq, scipy.stats.ks_2samp, scipy.stats.theilslopes, scipy.stats.siegelslopes, scipy.stats.obrientransform (including marray), scipy.stats.binomtest, scipy.integrate.fixed_quad, scipy.signal.square, scipy.stats.expectile, scipy.stats.shapiro, scipy.stats.pointbiserialr, scipy.stats.bws_test, scipy.stats.estimated_cdf (new function), scipy.stats.linregress, scipy.integrate.simpson, and scipy.signal.sawtooth.
• The torch support for scipy.signal.fftconvolve now correctly handles the float32 dtype.
• JAX JIT support has been added for: scipy.stats.binomtest (except for method='two-sided'), scipy.stats.mannwhitneyu (except for method='auto'), scipy.stats.lmoment, scipy.stats.moment, scipy.stats.ansari (related to new method argument), scipy.stats.yeojohnson_llf, scipy.stats.epps_singleton_2samp, scipy.stats.wilcoxon (except for method='exact' and method='auto'), scipy.stats.rankdata (via delegation), scipy.signal.oaconvolve, scipy.signal.hilbert, and scipy.signal.hilbert2.

Deprecated features and future changes

• passing lwork parameter to scipy.linalg.qr has been deprecated. The functionality was rarely used; the function computes the optimal size of the work arrays automatically, therefore users should simply remove their uses of the lwork parameter.
• The sparse construction functions kron, kronsum and build_diag choose return type sparray or spmatrix depending on the type of the sparse input arrays. When no inputs are sparse, the output is chosen to be spmatrix. That has been deprecated. The return type when no inputs are sparse will be changing to sparray. You can control the output type by ensuring that at least one input array is sparse. If any are sparray, the output will be sparray. If all sparse inputs are spmatrix, the output will be spmatrix.
• A FutureWarning is now issued for calling {r}matvec on column vectors with LinearOperator. Identical behavior can be achieved (and extended to batch dimensions) via {r}matmat.
• scipy.linalg functions are now stricter--using non-LAPACK dtypes is deprecated. When the deprecations expire, this will effectively limit the dtypes allowed in linear algebra functions to: integers (upcast to float), and single/double precision float/complex dtypes.
• scipy.spatial.minkowsi_distance, scipy.spatial.minkowsi_distance_p, and scipy.spatial.distance_matrix have been deprecated in favor of other superior functions.
• scipy.spatial.tsearch has been deprecated because it duplicates functionality more conveniently provided within the Delaunay class proper.
• The following functions have been deprecated because they were deemed not practically useful: scipy.interpolate.pade, scipy.interpolate.lagrange, and scipy.interpolate.approximate_taylor_polynomial.
• Setting spmatrix=True for the scipy.io readers mmio, FFM, hb, and matlab/_mio is now deprecated, including when set as the default value.
• The unintentionally public scipy.cluster.vq.py_vq has been deprecated.

Backwards incompatible changes

• The output of scipy.stats.rankdata is now always of a floating point dtype -- the result dtype of the input and a Python float.
• The behavior of the residuals returned by scipy.linalg.lstsq has been changed. For lapack_driver == "gelsy" or the system being either underdetermined or square, empty residuals are still returned. For lapack_driver == "gesld"/"gelss" in combination with an overdetermined system a non-empty residual is always returned. However, in the case where a slice is not full column rank, the corresponding residual is set to NaN.
• The 2nd output object of scipy.stats.contingency.crosstab when kwarg sparse=True is now a sparse array holding the counts instead of a sparse matrix. This allows it to be nD, so can accept more than 2 sequences as inputs, but it is a different class. Most operations work the same for sparse arrays and matrixes with notable differences for matrix: * means matmul and always-2D. For more info see migration_to_sparray.
• scipy.stats.obrientransform now returns a tuple of arrays instead of a single ndarray.
• scipy.stats.multinomial now returns NaNs when the category probability (p) rows/arrays do not sum to unity. This is an expiration of the deprecated behavior of adjusting the final element in the p array to compensate. Note that multinomial.rvs will now raise an error in such cases, since it has an integral return type.
• The iprint and disp parameters of scipy.optimize.fmin_l_bfgs_b have been removed, following the expiry of their deprecation.
• For scipy.linalg.{sqrtm, logm, signm}, disp (and sqrtm blocksize) parameters were removed (expired deprecations).
• The deprecated atol argument of scipy.optimize.nnls has been removed.

Other changes

• The vendored Boost.Math was updated from 1.89.0 to 1.91.0.
• SciPy now has a Pixi package definition, allowing developers to easily build SciPy from source inside Pixi workspaces.
• Developers may be interested in the private build option _without-fortran, which allows building SciPy from source in the absence of a Fortran compiler. This is an early prototype of the planned capability of a Fortran-free SciPy.
• The private scipy.interpolate._regrid function may be of experimental interest. It provides an interface for 2-D smoothing B-spline fitting via separable 1-D FITPACK kernels. It is under consideraton for public exposure in some form in the future.

Authors

• Name (commits)
• Joseph Adams (1) +
• Adrián Raso González (1) +
• Virgile Andreani (1)
• AshwathElang0 (1) +
• Mart-Mihkel Aun (1) +
• BarnikRB (2) +
• Richie Bendall (1) +
• J Berg (7) +
• Florian Bourgey (50)
• Jake Bowhay (98)
• Jonathan Brodrick (1) +
• Dietrich Brunn (36)
• Evgeni Burovski (200)
• Matthias Bussonnier (6)
• CJ Carey (9)
• Christine P. Chai (2)
• Lucas Colley (89)
• Dan (3) +
• devdanzin (2) +
• Martin Diehl (4)
• Sam Dolat (2) +
• dphipps-qnx (1) +
• DWesl (2)
• efrat99 (3) +
• fbrandt (1) +
• August Femtehjell (2) +
• Matthew H Flamm (1)
• Juan Flores (1) +
• foreverallama (1) +
• fumoboy007 (4) +
• John Patrick Gallagher (1) +
• Wei Bo Gao (1) +
• Christoph Gohlke (1)
• Nathan Goldbaum (20)
• Ludmila Golomozin (11) +
• Ralf Gommers (172)
• Mathieu Guay-Paquet (1) +
• Matt Haberland (147)
• Joren Hammudoglu (24)
• Jacob Hass (4)
• Maya Horii (1) +
• Guido Imperiale (1)
• Jan Möseritz-Schmidt (2)
• Leo Ji (4) +
• JOD (2) +
• Aditya Kamath (2) +
• Mukunda Rao Katta (1) +
• Robert Kern (1)
• Ria Khatoniar (1) +
• Matthias Koeppe (1)
• krishneetRAJ (1) +
• Iason Krommydas (1) +
• Eric Larson (1)
• Basil Liekens (31) +
• lnzwz (2) +
• Christian Lorentzen (15)
• Alex Lubbock (1) +
• Echedey Luis (2) +
• Lunyxis (1) +
• Zhang Maiyun (1) +
• Diego Medina Medina (1) +
• Elle Musoke (11) +
• Andrew Nelson (102)
• Nick ODell (28)
• Dimitri Papadopoulos Orfanos (1)
• partev (1)
• Matti Picus (7)
• Ilhan Polat (190)
• Adrian Raso (3)
• Aditya Rawat (1) +
• Tyler Reddy (94)
• Martin Reinecke (1)
• Lucas Roberts (6)
• Pamphile Roy (1)
• Daniel Schmitz (26)
• Martin Schuck (4)
• Dan Schult (47)
• Scott Shambaugh (16)
• Sabaa Siddique (1) +
• Nicholas Smith (1) +
• Johannes F. Sommerfeldt (1) +
• SpookyYomo (2) +
• Albert Steppi (80)
• Charalampos Stratakis (16) +
• Taylor (1) +
• thecaptain789 (1) +
• Adam Turner (1)
• Christian Veenhuis (2)
• Sebastiano Vigna (1)
• Rivka Walles (14) +
• Warren Weckesser (10)
• Soeren Wolfers (1) +
• wongaokay (1) +
• Xuefeng Xu (1)
• Aniket Singh Yadav (2) +
• yaochengchen (2) +
• Fadi Younes (2) +
• Isaiah Zimmerman (1) +
• Simon Zwieback (1) +
• ਗਗਨਦੀਪ ਸਿੰਘ (Gagandeep Singh) (19)

A total of 100 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.

Note that the source and binary assets associated with this release candidate were published to PyPI using trusted publishing, and so the trusted assets and their hashes are made available more securely at https://pypi.org/project/scipy/1.18.0rc1/ rather than providing them here in a less secure manner.