SciPy 1.4.0

SciPy 1.4.0 Release Notes

SciPy 1.4.0 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
1.4.x branch, and on adding new features on the master branch.

This release requires Python 3.5+ and NumPy >=1.13.3 (for Python 3.5, 3.6),
>=1.14.5 (for Python 3.7), >= 1.17.3 (for Python 3.8)

For running on PyPy, PyPy3 6.0+ and NumPy 1.15.0 are required.

Highlights of this release

• a new submodule, scipy.fft, now supersedes scipy.fftpack; this
  means support for long double transforms, faster multi-dimensional
  transforms, improved algorithm time complexity, release of the global
  intepreter lock, and control over threading behavior
• support for pydata/sparse arrays in scipy.sparse.linalg
• substantial improvement to the documentation and functionality of
  several scipy.special functions, and some new additions
• the generalized inverse Gaussian distribution has been added to
  scipy.stats
• an implementation of the Edmonds-Karp algorithm in
  scipy.sparse.csgraph.maximum_flow
• scipy.spatial.SphericalVoronoi now supports n-dimensional input,
  has linear memory complexity, improved performance, and
  supports single-hemisphere generators

New features

Infrastructure

Documentation can now be built with runtests.py --doc

A Dockerfile is now available in the scipy/scipy-dev repository to
facilitate getting started with SciPy development.

scipy.constants improvements

scipy.constants has been updated with the CODATA 2018 constants.

scipy.fft added

scipy.fft is a new submodule that supersedes the scipy.fftpack submodule.
For the most part, this is a drop-in replacement for numpy.fft and
scipy.fftpack alike. With some important differences, scipy.fft:

• uses NumPy's conventions for real transforms (rfft). This means the
  return value is a complex array, half the size of the full fft output.
  This is different from the output of fftpack which returned a real array
  representing complex components packed together.
• the inverse real to real transforms (idct and idst) are normalized
  for norm=None in thesame way as ifft. This means the identity
  idct(dct(x)) == x is now True for all norm modes.
• does not include the convolutions or pseudo-differential operators
  from fftpack.

This submodule is based on the pypocketfft library, developed by the
author of pocketfft which was recently adopted by NumPy as well.
pypocketfft offers a number of advantages over fortran FFTPACK:

• support for long double (np.longfloat) precision transforms.
• faster multi-dimensional transforms using vectorisation
• Bluestein’s algorithm removes the worst-case O(n^2) complexity of
  FFTPACK
• the global interpreter lock (GIL) is released during transforms
• optional multithreading of multi-dimensional transforms via the workers
  argument

Note that scipy.fftpack has not been deprecated and will continue to be
maintained but is now considered legacy. New code is recommended to use
scipy.fft instead, where possible.

scipy.fftpack improvements

scipy.fftpack now uses pypocketfft to perform its FFTs, offering the same
speed and accuracy benefits listed for scipy.fft above but without the
improved API.

scipy.integrate improvements

The function scipy.integrate.solve_ivp now has an args argument.
This allows the user-defined functions passed to the function to have
additional parameters without having to create wrapper functions or
lambda expressions for them.

scipy.integrate.solve_ivp can now return a y_events attribute
representing the solution of the ODE at event times

New OdeSolver is implemented --- DOP853. This is a high-order explicit
Runge-Kutta method originally implemented in Fortran. Now we provide a pure
Python implementation usable through solve_ivp with all its features.

scipy.integrate.quad provides better user feedback when break points are
specified with a weighted integrand.

scipy.integrate.quad_vec is now available for general purpose integration
of vector-valued functions

scipy.interpolate improvements

scipy.interpolate.pade now handles complex input data gracefully

scipy.interpolate.Rbf can now interpolate multi-dimensional functions

scipy.io improvements

scipy.io.wavfile.read can now read data from a WAV file that has a
malformed header, similar to other modern WAV file parsers

scipy.io.FortranFile now has an expanded set of available Exception
classes for handling poorly-formatted files

scipy.linalg improvements

The function scipy.linalg.subspace_angles(A, B) now gives correct
results for complex-valued matrices. Before this, the function only returned
correct values for real-valued matrices.

New boolean keyword argument check_finite for scipy.linalg.norm; whether
to check that the input matrix contains only finite numbers. Disabling may
give a performance gain, but may result in problems (crashes, non-termination)
if the inputs do contain infinities or NaNs.

scipy.linalg.solve_triangular has improved performance for a C-ordered
triangular matrix

LAPACK wrappers have been added for ?geequ, ?geequb, ?syequb,
and ?heequb

Some performance improvements may be observed due to an internal optimization
in operations involving LAPACK routines via _compute_lwork. This is
particularly true for operations on small arrays.

Block QR wrappers are now available in scipy.linalg.lapack

scipy.ndimage improvements

scipy.optimize improvements

It is now possible to use linear and non-linear constraints with
scipy.optimize.differential_evolution.

scipy.optimize.linear_sum_assignment has been re-written in C++ to improve
performance, and now allows input costs to be infinite.

A ScalarFunction.fun_and_grad method was added for convenient simultaneous
retrieval of a function and gradient evaluation

scipy.optimize.minimize BFGS method has improved performance by avoiding
duplicate evaluations in some cases

Better user feedback is provided when an objective function returns an array
instead of a scalar.

scipy.signal improvements

Added a new function to calculate convolution using the overlap-add method,
named scipy.signal.oaconvolve. Like scipy.signal.fftconvolve, this
function supports specifying dimensions along which to do the convolution.

scipy.signal.cwt now supports complex wavelets.

The implementation of choose_conv_method has been updated to reflect the
new FFT implementation. In addition, the performance has been significantly
improved (with rather drastic improvements in edge cases).

The function upfirdn now has a mode keyword argument that can be used
to select the signal extension mode used at the signal boundaries. These modes
are also available for use in resample_poly via a newly added padtype
argument.

scipy.signal.sosfilt now benefits from Cython code for improved performance

scipy.signal.resample should be more efficient by leveraging rfft when
possible

scipy.sparse improvements

It is now possible to use the LOBPCG method in scipy.sparse.linalg.svds.

scipy.sparse.linalg.LinearOperator now supports the operation rmatmat
for adjoint matrix-matrix multiplication, in addition to rmatvec.

Multiple stability updates enable float32 support in the LOBPCG eigenvalue
solver for symmetric and Hermitian eigenvalues problems in
scipy.sparse.linalg.lobpcg.

A solver for the maximum flow problem has been added as
scipy.sparse.csgraph.maximum_flow.

scipy.sparse.csgraph.maximum_bipartite_matching now allows non-square inputs,
no longer requires a perfect matching to exist, and has improved performance.

scipy.sparse.lil_matrix conversions now perform better in some scenarios

Basic support is available for pydata/sparse arrays in
scipy.sparse.linalg

scipy.sparse.linalg.spsolve_triangular now supports the unit_diagonal
argument to improve call signature similarity with its dense counterpart,
scipy.linalg.solve_triangular

assertAlmostEqual may now be used with sparse matrices, which have added
support for __round__

scipy.spatial improvements

The bundled Qhull library was upgraded to version 2019.1, fixing several
issues. Scipy-specific patches are no longer applied to it.

scipy.spatial.SphericalVoronoi now has linear memory complexity, improved
performance, and supports single-hemisphere generators. Support has also been
added for handling generators that lie on a great circle arc (geodesic input)
and for generators in n-dimensions.

scipy.spatial.transform.Rotation now includes functions for calculation of a
mean rotation, generation of the 3D rotation groups, and reduction of rotations
with rotational symmetries.

scipy.spatial.transform.Slerp is now callable with a scalar argument

scipy.spatial.voronoi_plot_2d now supports furthest site Voronoi diagrams

scipy.spatial.Delaunay and scipy.spatial.Voronoi now have attributes
for tracking whether they are furthest site diagrams

scipy.special improvements

The Voigt profile has been added as scipy.special.voigt_profile.

A real dispatch has been added for the Wright Omega function
(scipy.special.wrightomega).

The analytic continuation of the Riemann zeta function has been added. (The
Riemann zeta function is the one-argument variant of scipy.special.zeta.)

The complete elliptic integral of the first kind (scipy.special.ellipk) is
now available in scipy.special.cython_special.

The accuracy of scipy.special.hyp1f1 for real arguments has been improved.

The documentation of many functions has been improved.

scipy.stats improvements

scipy.stats.multiscale_graphcorr added as an independence test that
operates on high dimensional and nonlinear data sets. It has higher statistical
power than other scipy.stats tests while being the only one that operates on
multivariate data.

The generalized inverse Gaussian distribution (scipy.stats.geninvgauss) has
been added.

It is now possible to efficiently reuse scipy.stats.binned_statistic_dd
with new values by providing the result of a previous call to the function.

scipy.stats.hmean now handles input with zeros more gracefully.

The beta-binomial distribution is now available in scipy.stats.betabinom.

scipy.stats.zscore, scipy.stats.circmean, scipy.stats.circstd, and
scipy.stats.circvar now support the nan_policy argument for enhanced
handling of NaN values

scipy.stats.entropy now accepts an axis argument

scipy.stats.gaussian_kde.resample now accepts a seed argument to empower
reproducibility

scipy.stats.kendalltau performance has improved, especially for large inputs,
due to improved cache usage

scipy.stats.truncnorm distribution has been rewritten to support much wider
tails

Deprecated features

scipy deprecations

Support for NumPy functions exposed via the root SciPy namespace is deprecated
and will be removed in 2.0.0. For example, if you use scipy.rand or
scipy.diag, you should change your code to directly use
numpy.random.default_rng or numpy.diag, respectively.
They remain available in the currently continuing Scipy 1.x release series.

The exception to this rule is using scipy.fft as a function --
:mod:scipy.fft is now meant to be used only as a module, so the ability to
call scipy.fft(...) will be removed in SciPy 1.5.0.

In scipy.spatial.Rotation methods from_dcm, as_dcm were renamed to
from_matrix, as_matrix respectively. The old names will be removed in
SciPy 1.6.0.

Method Rotation.match_vectors was deprecated in favor of
Rotation.align_vectors, which provides a more logical and
general API to the same functionality. The old method
will be removed in SciPy 1.6.0.

Backwards incompatible changes

scipy.special changes

The deprecated functions hyp2f0, hyp1f2, and hyp3f0 have been
removed.

The deprecated function bessel_diff_formula has been removed.

The function i0 is no longer registered with numpy.dual, so that
numpy.dual.i0 will unconditionally refer to the NumPy version regardless
of whether scipy.special is imported.

The function expn has been changed to return nan outside of its
domain of definition (x, n < 0) instead of inf.

scipy.sparse changes

Sparse matrix reshape now raises an error if shape is not two-dimensional,
rather than guessing what was meant. The behavior is now the same as before
SciPy 1.1.0.

CSR and CSC sparse matrix classes should now return empty matrices
of the same type when indexed out of bounds. Previously, for some versions
of SciPy, this would raise an IndexError. The change is largely motivated
by greater consistency with ndarray and numpy.matrix semantics.

scipy.signal changes

scipy.signal.resample behavior for length-1 signal inputs has been
fixed to output a constant (DC) value rather than an impulse, consistent with
the assumption of signal periodicity in the FFT method.

scipy.signal.cwt now performs complex conjugation and time-reversal of
wavelet data, which is a backwards-incompatible bugfix for
time-asymmetric wavelets.

scipy.stats changes

scipy.stats.loguniform added with better documentation as (an alias for
scipy.stats.reciprocal). loguniform generates random variables
that are equally likely in the log space; e.g., 1, 10 and 100
are all equally likely if loguniform(10 ** 0, 10 ** 2).rvs() is used.

Other changes

The LSODA method of scipy.integrate.solve_ivp now correctly detects stiff
problems.

scipy.spatial.cKDTree now accepts and correctly handles empty input data

scipy.stats.binned_statistic_dd now calculates the standard deviation
statistic in a numerically stable way.

scipy.stats.binned_statistic_dd now throws an error if the input data
contains either np.nan or np.inf. Similarly, in scipy.stats now all
continuous distributions' .fit() methods throw an error if the input data
contain any instance of either np.nan or np.inf.

Authors

• @endolith
• @wenhui-prudencemed +
• Abhinav +
• Anne Archibald
• ashwinpathak20nov1996 +
• Danilo Augusto +
• Nelson Auner +
• aypiggott +
• Christoph Baumgarten
• Peter Bell
• Sebastian Berg
• Arman Bilge +
• Benedikt Boecking +
• Christoph Boeddeker +
• Daniel Bunting
• Evgeni Burovski
• Angeline Burrell +
• Angeline G. Burrell +
• CJ Carey
• Carlos Ramos Carreño +
• Mak Sze Chun +
• Malayaja Chutani +
• Christian Clauss +
• Jonathan Conroy +
• Stephen P Cook +
• Dylan Cutler +
• Anirudh Dagar +
• Aidan Dang +
• dankleeman +
• Brandon David +
• Tyler Dawson +
• Dieter Werthmüller
• Joe Driscoll +
• Jakub Dyczek +
• Dávid Bodnár
• Fletcher Easton +
• Stefan Endres
• etienne +
• Johann Faouzi
• Yu Feng
• Isuru Fernando +
• Matthew H Flamm
• Martin Gauch +
• Gabriel Gerlero +
• Ralf Gommers
• Chris Gorgolewski +
• Domen Gorjup +
• Edouard Goudenhoofdt +
• Jan Gwinner +
• Maja Gwozdz +
• Matt Haberland
• hadshirt +
• Pierre Haessig +
• David Hagen
• Charles Harris
• Gina Helfrich +
• Alex Henrie +
• Francisco J. Hernandez Heras +
• Andreas Hilboll
• Lindsey Hiltner
• Thomas Hisch
• Min ho Kim +
• Gert-Ludwig Ingold
• jakobjakobson13 +
• Todd Jennings
• He Jia
• Muhammad Firmansyah Kasim +
• Andrew Knyazev +
• Holger Kohr +
• Mateusz Konieczny +
• Krzysztof Pióro +
• Philipp Lang +
• Peter Mahler Larsen +
• Eric Larson
• Antony Lee
• Gregory R. Lee
• Chelsea Liu +
• Jesse Livezey
• Peter Lysakovski +
• Jason Manley +
• Michael Marien +
• Nikolay Mayorov
• G. D. McBain +
• Sam McCormack +
• Melissa Weber Mendonça +
• Kevin Michel +
• mikeWShef +
• Sturla Molden
• Eric Moore
• Peyton Murray +
• Andrew Nelson
• Clement Ng +
• Juan Nunez-Iglesias
• Renee Otten +
• Kellie Ottoboni +
• Ayappan P
• Sambit Panda +
• Tapasweni Pathak +
• Oleksandr Pavlyk
• Fabian Pedregosa
• Petar Mlinarić
• Matti Picus
• Marcel Plch +
• Christoph Pohl +
• Ilhan Polat
• Siddhesh Poyarekar +
• Ioannis Prapas +
• James Alan Preiss +
• Yisheng Qiu +
• Eric Quintero
• Bharat Raghunathan +
• Tyler Reddy
• Joscha Reimer
• Antonio Horta Ribeiro
• Lucas Roberts
• rtshort +
• Josua Sassen
• Kevin Sheppard
• Scott Sievert
• Leo Singer
• Kai Striega
• Søren Fuglede Jørgensen
• tborisow +
• Étienne Tremblay +
• tuxcell +
• Miguel de Val-Borro
• Andrew Valentine +
• Hugo van Kemenade
• Paul van Mulbregt
• Sebastiano Vigna
• Pauli Virtanen
• Dany Vohl +
• Ben Walsh +
• Huize Wang +
• Warren Weckesser
• Anreas Weh +
• Joseph Weston +
• Adrian Wijaya +
• Timothy Willard +
• Josh Wilson
• Kentaro Yamamoto +
• Dave Zbarsky +

A total of 142 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.