SciPy 1.9.0 Release Notes

Note: SciPy 1.9.0 is not released yet!

SciPy 1.9.0 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
1.9.x branch, and on adding new features on the main branch.

This release requires Python 3.8-3.11 and NumPy 1.18.5 or greater.

For running on PyPy, PyPy3 6.0+ is required.

Highlights of this release

• We have modernized our build system to use meson, substantially improving
  our build performance, and providing better build-time configuration and
  cross-compilation support,
• Added scipy.optimize.milp, new function for mixed-integer linear
  programming,
• Added scipy.stats.fit for fitting discrete and continuous distributions
  to data,
• Tensor-product spline interpolation modes were added to
  scipy.interpolate.RegularGridInterpolator,
• A new global optimizer (DIviding RECTangles algorithm)
  scipy.optimize.direct.

New features

scipy.interpolate improvements

• Speed up the RBFInterpolator evaluation with high dimensional
  interpolants.
• Added new spline based interpolation methods for
  scipy.interpolate.RegularGridInterpolator and its tutorial.
• scipy.interpolate.RegularGridInterpolator and scipy.interpolate.interpn
  now accept descending ordered points.
• RegularGridInterpolator now handles length-1 grid axes.
• The BivariateSpline subclasses have a new method partial_derivative
  which constructs a new spline object representing a derivative of an
  original spline. This mirrors the corresponding functionality for univariate
  splines, splder and BSpline.derivative, and can substantially speed
  up repeated evaluation of derivatives.

scipy.linalg improvements

• scipy.linalg.expm now accepts nD arrays. Its speed is also improved.
• Minimum required LAPACK version is bumped to 3.7.1.

scipy.fft improvements

• Added uarray multimethods for scipy.fft.fht and scipy.fft.ifht
  to allow provision of third party backend implementations such as those
  recently added to CuPy.

scipy.optimize improvements

• A new global optimizer, scipy.optimize.direct (DIviding RECTangles algorithm)
  was added. For problems with inexpensive function evaluations, like the ones
  in the SciPy benchmark suite, direct is competitive with the best other
  solvers in SciPy (dual_annealing and differential_evolution) in terms
  of execution time. See
  gh-14300 <https://github.com/scipy/scipy/pull/14300>__ for more details.

• Add a full_output parameter to scipy.optimize.curve_fit to output
  additional solution information.

• Add a integrality parameter to scipy.optimize.differential_evolution,
  enabling integer constraints on parameters.

• Add a vectorized parameter to call a vectorized objective function only
  once per iteration. This can improve minimization speed by reducing
  interpreter overhead from the multiple objective function calls.

• The default method of scipy.optimize.linprog is now 'highs'.

• Added scipy.optimize.milp, new function for mixed-integer linear
  programming.

• Added Newton-TFQMR method to newton_krylov.

• Added support for the Bounds class in shgo and dual_annealing for
  a more uniform API across scipy.optimize.

• Added the vectorized keyword to differential_evolution.

• approx_fprime now works with vector-valued functions.

scipy.signal improvements

• The new window function scipy.signal.windows.kaiser_bessel_derived was
  added to compute the Kaiser-Bessel derived window.
• Single-precision hilbert operations are now faster as a result of more
  consistent dtype handling.

scipy.sparse improvements

• Add a copy parameter to scipy.sparce.csgraph.laplacian. Using inplace
  computation with copy=False reduces the memory footprint.
• Add a dtype parameter to scipy.sparce.csgraph.laplacian for type casting.
• Add a symmetrized parameter to scipy.sparce.csgraph.laplacian to produce
  symmetric Laplacian for directed graphs.
• Add a form parameter to scipy.sparce.csgraph.laplacian taking one of the
  three values: array, or function, or lo determining the format of
  the output Laplacian:
  • array is a numpy array (backward compatible default);
  • function is a pointer to a lambda-function evaluating the
    Laplacian-vector or Laplacian-matrix product;
  • lo results in the format of the LinearOperator.

scipy.sparse.linalg improvements

• lobpcg performance improvements for small input cases.

scipy.spatial improvements

• Add an order parameter to scipy.spatial.transform.Rotation.from_quat
  and scipy.spatial.transform.Rotation.as_quat to specify quaternion format.

scipy.stats improvements

• scipy.stats.monte_carlo_test performs one-sample Monte Carlo hypothesis
  tests to assess whether a sample was drawn from a given distribution. Besides
  reproducing the results of hypothesis tests like scipy.stats.ks_1samp,
  scipy.stats.normaltest, and scipy.stats.cramervonmises without small sample
  size limitations, it makes it possible to perform similar tests using arbitrary
  statistics and distributions.

• Several scipy.stats functions support new axis (integer or tuple of
  integers) and nan_policy ('raise', 'omit', or 'propagate'), and
  keepdims arguments.
  These functions also support masked arrays as inputs, even if they do not have
  a scipy.stats.mstats counterpart. Edge cases for multidimensional arrays,
  such as when axis-slices have no unmasked elements or entire inputs are of
  size zero, are handled consistently.

• Add a weight parameter to scipy.stats.hmean.

• Several improvements have been made to scipy.stats.levy_stable. Substantial
  improvement has been made for numerical evaluation of the pdf and cdf,
  resolving #12658 and
  #14944. The improvement is
  particularly dramatic for stability parameter alpha close to or equal to 1
  and for alpha below but approaching its maximum value of 2. The alternative
  fast Fourier transform based method for pdf calculation has also been updated
  to use the approach of Wang and Zhang from their 2008 conference paper
  Simpson’s rule based FFT method to compute densities of stable distribution,
  making this method more competitive with the default method. In addition,
  users now have the option to change the parametrization of the Levy Stable
  distribution to Nolan's "S0" parametrization which is used internally by
  SciPy's pdf and cdf implementations. The "S0" parametrization is described in
  Nolan's paper Numerical calculation of stable densities and distribution
  functions upon which SciPy's
  implementation is based. "S0" has the advantage that delta and gamma
  are proper location and scale parameters. With delta and gamma fixed,
  the location and scale of the resulting distribution remain unchanged as
  alpha and beta change. This is not the case for the default "S1"
  parametrization. Finally, more options have been exposed to allow users to
  trade off between runtime and accuracy for both the default and FFT methods of
  pdf and cdf calculation. More information can be found in the documentation
  here (to be linked).

• Added scipy.stats.fit for fitting discrete and continuous distributions to
  data.

• The methods "pearson" and "tippet" from scipy.stats.combine_pvalues
  have been fixed to return the correct p-values, resolving
  #15373. In addition, the
  documentation for scipy.stats.combine_pvalues has been expanded and improved.

• Unlike other reduction functions, stats.mode didn't consume the axis
  being operated on and failed for negative axis inputs. Both the bugs have been
  fixed. Note that stats.mode will now consume the input axis and return an
  ndarray with the axis dimension removed.

• Replaced implementation of scipy.stats.ncf with the implementation from
  Boost for improved reliability.

• Add a bits parameter to scipy.stats.qmc.Sobol. It allows to use from 0
  to 64 bits to compute the sequence. Default is None which corresponds to
  30 for backward compatibility. Using a higher value allow to sample more
  points. Note: bits does not affect the output dtype.

• Add a integers method to scipy.stats.qmc.QMCEngine. It allows sampling
  integers using any QMC sampler.

• Improved the fit speed and accuracy of stats.pareto.

• Added qrvs method to NumericalInversePolynomial to match the
  situation for NumericalInverseHermite.

• Faster random variate generation for gennorm and nakagami.

• lloyd_centroidal_voronoi_tessellation has been added to allow improved
  sample distributions via iterative application of Voronoi diagrams and
  centering operations

• Add scipy.stats.qmc.PoissonDisk to sample using the Poisson disk sampling
  method. It guarantees that samples are separated from each other by a
  given radius.

• Add scipy.stats.pmean to calculate the weighted power mean also called
  generalized mean.

Deprecated features

• Due to collision with the shape parameter n of several distributions,
  use of the distribution moment method with keyword argument n is
  deprecated. Keyword n is replaced with keyword order.
• Similarly, use of the distribution interval method with keyword arguments
  alpha is deprecated. Keyword alpha is replaced with keyword
  confidence.
• The 'simplex', 'revised simplex', and 'interior-point' methods
  of scipy.optimize.linprog are deprecated. Methods highs, highs-ds,
  or highs-ipm should be used in new code.
• Support for non-numeric arrays has been deprecated from stats.mode.
  pandas.DataFrame.mode can be used instead.
• The function spatial.distance.kulsinski has been deprecated in favor
  of spatial.distance.kulczynski1.
• The maxiter keyword of the truncated Newton (TNC) algorithm has been
  deprecated in favour of maxfun.
• The vertices keyword of Delauney.qhull now raises a
  DeprecationWarning, after having been deprecated in documentation only
  for a long time.
• The extradoc keyword of rv_continuous, rv_discrete and
  rv_sample now raises a DeprecationWarning, after having been deprecated in
  documentation only for a long time.

Expired Deprecations

There is an ongoing effort to follow through on long-standing deprecations.
The following previously deprecated features are affected:

• Object arrays in sparse matrices now raise an error.
• Inexact indices into sparse matrices now raise an error.
• Passing radius=None to scipy.spatial.SphericalVoronoi now raises an
  error (not adding radius defaults to 1, as before).
• Several BSpline methods now raise an error if inputs have ndim > 1.
• The _rvs method of statistical distributions now requires a size
  parameter.
• Passing a fillvalue that cannot be cast to the output type in
  scipy.signal.convolve2d now raises an error.
• scipy.spatial.distance now enforces that the input vectors are
  one-dimensional.
• Removed stats.itemfreq.
• Removed stats.median_absolute_deviation.
• Removed n_jobs keyword argument and use of k=None from
  kdtree.query.
• Removed right keyword from interpolate.PPoly.extend.
• Removed debug keyword from scipy.linalg.solve_*.
• Removed class _ppform scipy.interpolate.
• Removed BSR methods matvec and matmat.
• Removed mlab truncation mode from cluster.dendrogram.
• Removed cluster.vq.py_vq2.
• Removed keyword arguments ftol and xtol from
  optimize.minimize(method='Nelder-Mead').
• Removed signal.windows.hanning.
• Removed LAPACK gegv functions from linalg; this raises the minimally
  required LAPACK version to 3.7.1.
• Removed spatial.distance.matching.
• Removed the alias scipy.random for numpy.random.
• Removed docstring related functions from scipy.misc (docformat,
  inherit_docstring_from, extend_notes_in_docstring,
  replace_notes_in_docstring, indentcount_lines, filldoc,
  unindent_dict, unindent_string).
• Removed linalg.pinv2.

Backwards incompatible changes

• Several scipy.stats functions now convert np.matrix to np.ndarrays
  before the calculation is performed. In this case, the output will be a scalar
  or np.ndarray of appropriate shape rather than a 2D np.matrix.
  Similarly, while masked elements of masked arrays are still ignored, the
  output will be a scalar or np.ndarray rather than a masked array with
  mask=False.
• The default method of scipy.optimize.linprog is now 'highs', not
  'interior-point' (which is now deprecated), so callback functions and
  some options are no longer supported with the default method. With the
  default method, the x attribute of the returned OptimizeResult is
  now None (instead of a non-optimal array) when an optimal solution
  cannot be found (e.g. infeasible problem).
• For scipy.stats.combine_pvalues, the sign of the test statistic returned
  for the method "pearson" has been flipped so that higher values of the
  statistic now correspond to lower p-values, making the statistic more
  consistent with those of the other methods and with the majority of the
  literature.
• scipy.linalg.expm due to historical reasons was using the sparse
  implementation and thus was accepting sparse arrays. Now it only works with
  nDarrays. For sparse usage, scipy.sparse.linalg.expm needs to be used
  explicitly.
• The definition of scipy.stats.circvar has reverted to the one that is
  standard in the literature; note that this is not the same as the square of
  scipy.stats.circstd.
• Remove inheritance to QMCEngine in MultinomialQMC and
  MultivariateNormalQMC. It removes the methods fast_forward and reset.
• Init of MultinomialQMC now require the number of trials with n_trials.
  Hence, MultinomialQMC.random output has now the correct shape (n, pvals).
• Several function-specific warnings (F_onewayConstantInputWarning,
  F_onewayBadInputSizesWarning, PearsonRConstantInputWarning,
  PearsonRNearConstantInputWarning, SpearmanRConstantInputWarning, and
  BootstrapDegenerateDistributionWarning) have been replaced with more
  general warnings.

Other changes

• A draft developer CLI is available for SciPy, leveraging the doit,
  click and rich-click tools. For more details, see
  gh-15959.

• The SciPy contributor guide has been reorganized and updated
  (see #15947 for details).

• QUADPACK Fortran routines in scipy.integrate, which power
  scipy.integrate.quad, have been marked as recursive. This should fix rare
  issues in multivariate integration (nquad and friends) and obviate the need
  for compiler-specific compile flags (/recursive for ifort etc). Please file
  an issue if this change turns out problematic for you. This is also true for
  FITPACK routines in scipy.interpolate, which power splrep,
  splev etc., and *UnivariateSpline and *BivariateSpline classes.

• the USE_PROPACK environment variable has been renamed to
  SCIPY_USE_PROPACK; setting to a non-zero value will enable
  the usage of the PROPACK library as before

• Building SciPy on windows with MSVC now requires at least the vc142
  toolset (available in Visual Studio 2019 and higher).

Lazy access to subpackages

Before this release, all subpackages of SciPy (cluster, fft, ndimage,
etc.) had to be explicitly imported. Now, these subpackages are lazily loaded
as soon as they are accessed, so that the following is possible (if desired
for interactive use, it's not actually recommended for code,
see :ref:scipy-api):
import scipy as sp; sp.fft.dct([1, 2, 3]). Advantages include: making it
easier to navigate SciPy in interactive terminals, reducing subpackage import
conflicts (which before required
import networkx.linalg as nla; import scipy.linalg as sla),
and avoiding repeatedly having to update imports during teaching &
experimentation. Also see
the related community specification document.

SciPy switched to Meson as its build system

This is the first release that ships with Meson as
the build system. When installing with pip or pypa/build, Meson will be
used (invoked via the meson-python build hook). This change brings
significant benefits - most importantly much faster build times, but also
better support for cross-compilation and cleaner build logs.

Note:

This release still ships with support for numpy.distutils-based builds
as well. Those can be invoked through the setup.py command-line
interface (e.g., python setup.py install). It is planned to remove
numpy.distutils support before the 1.10.0 release.

When building from source, a number of things have changed compared to building
with numpy.distutils:

• New build dependencies: meson, ninja, and pkg-config.
  setuptools and wheel are no longer needed.
• BLAS and LAPACK libraries that are supported haven't changed, however the
  discovery mechanism has: that is now using pkg-config instead of hardcoded
  paths or a site.cfg file.
• The build defaults to using OpenBLAS. See :ref:blas-lapack-selection for
  details.

The two CLIs that can be used to build wheels are pip and build. In
addition, the SciPy repo contains a python dev.py CLI for any kind of
development task (see its --help for details). For a comparison between old
(distutils) and new (meson) build commands, see :ref:meson-faq.

For more information on the introduction of Meson support in SciPy, see
gh-13615 <https://github.com/scipy/scipy/issues/13615>__ and
this blog post <https://labs.quansight.org/blog/2021/07/moving-scipy-to-meson/>__.

Authors

• endolith (12)
• Caio Agiani (2) +
• Emmy Albert (1) +
• Joseph Albert (1)
• Tania Allard (3)
• Carsten Allefeld (1) +
• Kartik Anand (1) +
• Virgile Andreani (2) +
• Weh Andreas (1) +
• Francesco Andreuzzi (5) +
• Kian-Meng Ang (2) +
• Gerrit Ansmann (1)
• Ar-Kareem (1) +
• Shehan Atukorala (1) +
• avishai231 (1) +
• Blair Azzopardi (1)
• Sayantika Banik (2) +
• Ross Barnowski (9)
• Christoph Baumgarten (3)
• Nickolai Belakovski (1)
• Peter Bell (9)
• Sebastian Berg (3)
• Bharath (1) +
• bobcatCA (2) +
• boussoffara (2) +
• Islem BOUZENIA (1) +
• Jake Bowhay (41) +
• Matthew Brett (11)
• Dietrich Brunn (2) +
• Michael Burkhart (2) +
• Evgeni Burovski (96)
• Matthias Bussonnier (20)
• Dominic C (1)
• Cameron (1) +
• CJ Carey (3)
• Thomas A Caswell (2)
• Ali Cetin (2) +
• Hood Chatham (5) +
• Klesk Chonkin (1)
• Craig Citro (1) +
• Dan Cogswell (1) +
• Luigi Cruz (1) +
• Anirudh Dagar (5)
• Brandon David (1)
• deepakdinesh1123 (1) +
• Denton DeLoss (1) +
• derbuihan (2) +
• Sameer Deshmukh (13) +
• Niels Doucet (1) +
• DWesl (8)
• eytanadler (30) +
• Thomas J. Fan (5)
• Isuru Fernando (3)
• Joseph Fox-Rabinovitz (1)
• Ryan Gibson (4) +
• Ralf Gommers (323)
• Srinivas Gorur-Shandilya (1) +
• Alex Griffing (2)
• h-vetinari (5)
• Matt Haberland (459)
• Tristan Hearn (1) +
• Jonathan Helgert (1) +
• Samuel Hinton (1) +
• Jake (1) +
• Stewart Jamieson (1) +
• Jan-Hendrik Müller (1)
• Yikun Jiang (1) +
• JuliaMelle01 (1) +
• jyuv (12) +
• Toshiki Kataoka (1)
• Chris Keefe (1) +
• Robert Kern (4)
• Andrew Knyazev (11)
• Matthias Koeppe (4) +
• Sergey Koposov (1)
• Volodymyr Kozachynskyi (1) +
• Yotaro Kubo (2) +
• Jacob Lapenna (1) +
• Peter Mahler Larsen (8)
• Eric Larson (4)
• Laurynas Mikšys (1) +
• Antony Lee (1)
• Gregory R. Lee (2)
• lerichi (1) +
• Tim Leslie (2)
• P. L. Lim (1)
• Smit Lunagariya (43)
• lutefiskhotdish (1) +
• Cong Ma (12)
• Syrtis Major (1)
• Nicholas McKibben (18)
• Melissa Weber Mendonça (10)
• Mark Mikofski (1)
• Jarrod Millman (13)
• Harsh Mishra (6)
• ML-Nielsen (3) +
• Matthew Murray (1) +
• Andrew Nelson (50)
• Dimitri Papadopoulos Orfanos (1) +
• Evgueni Ovtchinnikov (2) +
• Sambit Panda (1)
• Nick Papior (2)
• Tirth Patel (43)
• Petar Mlinarić (1)
• petroselo (1) +
• Ilhan Polat (64)
• Anthony Polloreno (1)
• Amit Portnoy (1) +
• Quentin Barthélemy (9)
• Patrick N. Raanes (1) +
• Tyler Reddy (174)
• Pamphile Roy (197)
• Vivek Roy (2) +
• sabonerune (1) +
• Niyas Sait (2) +
• Atsushi Sakai (25)
• Mazen Sayed (1) +
• Eduardo Schettino (5) +
• Daniel Schmitz (6) +
• Eli Schwartz (4) +
• SELEE (2) +
• Namami Shanker (4)
• siddhantwahal (1) +
• Gagandeep Singh (8)
• Soph (1) +
• Shivnaren Srinivasan (1) +
• Scott Staniewicz (1) +
• Leo C. Stein (4)
• Albert Steppi (7)
• Christopher Strickland (1) +
• Kai Striega (4)
• Søren Fuglede Jørgensen (1)
• Aleksandr Tagilov (1) +
• Masayuki Takagi (1) +
• Sai Teja (1) +
• Ewout ter Hoeven (2) +
• Will Tirone (2)
• Bas van Beek (7)
• Dhruv Vats (1)
• H. Vetinari (6)
• Arthur Volant (1)
• Samuel Wallan (5)
• Stefan van der Walt (8)
• Warren Weckesser (83)
• Anreas Weh (1)
• Nils Werner (1)
• Aviv Yaish (1) +
• Dowon Yi (1)
• Rory Yorke (1)
• Yosshi999 (1) +
• yuanx749 (2) +
• Gang Zhao (23)
• ZhihuiChen0903 (1)
• Pavel Zun (1) +
• David Zwicker (1) +

A total of 155 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.