SciPy 1.7.0 Release Notes

Note: Scipy 1.7.0 is not released yet!

SciPy 1.7.0 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
1.7.x branch, and on adding new features on the master branch.

This release requires Python 3.7+ and NumPy 1.16.5 or greater.

For running on PyPy, PyPy3 6.0+ is required.

Highlights of this release

• A new submodule for quasi-Monte Carlo, scipy.stats.qmc, was added
• The documentation design was updated to use the same PyData-Sphinx theme as
  other NumFOCUS packages like NumPy.
• We now vendor and leverage the Boost C++ library to enable numerous
  improvements for long-standing weaknesses in scipy.stats
• scipy.stats has six new distributions, eight new (or overhauled)
  hypothesis tests, a new function for bootstrapping, a class that enables
  fast random variate sampling and percentile point function evaluation,
  and many other enhancements.
• cdist and pdist distance calculations are faster for several metrics,
  especially weighted cases, thanks to a rewrite to a new C++ backend framework
• A new class for radial basis function interpolation, RBFInterpolator, was
  added to address issues with the Rbf class.

We gratefully acknowledge the Chan-Zuckerberg Initiative Essential Open Source
Software for Science program for supporting many of the improvements to
scipy.stats.

New features

scipy.cluster improvements

An optional argument, seed, has been added to kmeans and kmeans2 to
set the random generator and random state.

scipy.interpolate improvements

Improved input validation and error messages for fitpack.bispev and
fitpack.parder for scenarios that previously caused substantial confusion
for users.

The class RBFInterpolator was added to supersede the Rbf class. The new
class has usage that more closely follows other interpolator classes, corrects
sign errors that caused unexpected smoothing behavior, includes polynomial
terms in the interpolant (which are necessary for some RBF choices), and
supports interpolation using only the k-nearest neighbors for memory
efficiency.

scipy.linalg improvements

An LAPACK wrapper was added for access to the tgexc subroutine.

scipy.ndimage improvements

scipy.ndimage.affine_transform is now able to infer the output_shape from
the out array.

scipy.optimize improvements

The optional parameter bounds was added to
_minimize_neldermead to support bounds constraints
for the Nelder-Mead solver.

trustregion methods trust-krylov, dogleg and trust-ncg can now
estimate hess by finite difference using one of
["2-point", "3-point", "cs"].

halton was added as a sampling_method in scipy.optimize.shgo.
sobol was fixed and is now using scipy.stats.qmc.Sobol.

halton and sobol were added as init methods in
scipy.optimize.differential_evolution.

differential_evolution now accepts an x0 parameter to provide an
initial guess for the minimization.

least_squares has a modest performance improvement when SciPy is built
with Pythran transpiler enabled.

When linprog is used with method 'highs', 'highs-ipm', or
'highs-ds', the result object now reports the marginals (AKA shadow
prices, dual values) and residuals associated with each constraint.

scipy.signal improvements

get_window supports general_cosine and general_hamming window
functions.

scipy.signal.medfilt2d now releases the GIL where appropriate to enable
performance gains via multithreaded calculations.

scipy.sparse improvements

Addition of dia_matrix sparse matrices is now faster.

scipy.spatial improvements

distance.cdist and distance.pdist performance has greatly improved for
certain weighted metrics. Namely: minkowski, euclidean, chebyshev,
canberra, and cityblock.

Modest performance improvements for many of the unweighted cdist and
pdist metrics noted above.

The parameter seed was added to scipy.spatial.vq.kmeans and
scipy.spatial.vq.kmeans2.

The parameters axis and keepdims where added to
scipy.spatial.distance.jensenshannon.

The rotation methods from_rotvec and as_rotvec now accept a
degrees argument to specify usage of degrees instead of radians.

scipy.special improvements

Wright's generalized Bessel function for positive arguments was added as
scipy.special.wright_bessel.

An implementation of the inverse of the Log CDF of the Normal Distribution is
now available via scipy.special.ndtri_exp.

scipy.stats improvements

Hypothesis Tests

The Mann-Whitney-Wilcoxon test, mannwhitneyu, has been rewritten. It now
supports n-dimensional input, an exact test method when there are no ties,
and improved documentation. Please see "Other changes" for adjustments to
default behavior.

The new function scipy.stats.binomtest replaces scipy.stats.binom_test. The
new function returns an object that calculates a confidence intervals of the
proportion parameter. Also, performance was improved from O(n) to O(log(n)) by
using binary search.

The two-sample version of the Cramer-von Mises test is implemented in
scipy.stats.cramervonmises_2samp.

The Alexander-Govern test is implemented in the new function
scipy.stats.alexandergovern.

The new functions scipy.stats.barnard_exact and scipy.stats. boschloo_exact
respectively perform Barnard's exact test and Boschloo's exact test
for 2x2 contingency tables.

The new function scipy.stats.page_trend_test performs Page's test for ordered
alternatives.

The new function scipy.stats.somersd performs Somers' D test for ordinal
association between two variables.

An option, permutations, has been added in scipy.stats.ttest_ind to
perform permutation t-tests. A trim option was also added to perform
a trimmed (Yuen's) t-test.

The alternative parameter was added to the skewtest, kurtosistest,
ranksums, mood, ansari, linregress, and spearmanr functions
to allow one-sided hypothesis testing.

Sample statistics

The new function scipy.stats.differential_entropy estimates the differential
entropy of a continuous distribution from a sample.

The boxcox and boxcox_normmax now allow the user to control the
optimizer used to minimize the negative log-likelihood function.

A new function scipy.stats.contingency.relative_risk calculates the
relative risk, or risk ratio, of a 2x2 contingency table. The object
returned has a method to compute the confidence interval of the relative risk.

Performance improvements in the skew and kurtosis functions achieved
by removal of repeated/redundant calculations.

Substantial performance improvements in scipy.stats.mstats.hdquantiles_sd.

The new function scipy.stats.contingency.association computes several
measures of association for a contingency table: Pearsons contingency
coefficient, Cramer's V, and Tschuprow's T.

The parameter nan_policy was added to scipy.stats.zmap to provide options
for handling the occurrence of nan in the input data.

The parameter ddof was added to scipy.stats.variation and
scipy.stats.mstats.variation.

The parameter weights was added to scipy.stats.gmean.

Statistical Distributions

We now vendor and leverage the Boost C++ library to address a number of
previously reported issues in stats. Notably, beta, binom,
nbinom now have Boost backends, and it is straightforward to leverage
the backend for additional functions.

The skew Cauchy probability distribution has been implemented as
scipy.stats.skewcauchy.

The Zipfian probability distribution has been implemented as
scipy.stats.zipfian.

The new distributions nchypergeom_fisher and nchypergeom_wallenius
implement the Fisher and Wallenius versions of the noncentral hypergeometric
distribution, respectively.

The generalized hyperbolic distribution was added in
scipy.stats.genhyperbolic.

The studentized range distribution was added in scipy.stats.studentized_range.

scipy.stats.argus now has improved handling for small parameter values.

Better argument handling/preparation has resulted in performance improvements
for many distributions.

The cosine distribution has added ufuncs for ppf, cdf, sf, and
isf methods including numerical precision improvements at the edges of the
support of the distribution.

An option to fit the distribution to data by the method of moments has been
added to the fit method of the univariate continuous distributions.

Other

scipy.stats.bootstrap has been added to allow estimation of the confidence
interval and standard error of a statistic.

The new function scipy.stats.contingency.crosstab computes a contingency
table (i.e. a table of counts of unique entries) for the given data.

scipy.stats.NumericalInverseHermite enables fast random variate sampling
and percentile point function evaluation of an arbitrary univariate statistical
distribution.

New scipy.stats.qmc module

This new module provides Quasi-Monte Carlo (QMC) generators and associated
helper functions.

It provides a generic class scipy.stats.qmc.QMCEngine which defines a QMC
engine/sampler. An engine is state aware: it can be continued, advanced and
reset. 3 base samplers are available:

• scipy.stats.qmc.Sobol the well known Sobol low discrepancy sequence.
  Several warnings have been added to guide the user into properly using this
  sampler. The sequence is scrambled by default.
• scipy.stats.qmc.Halton: Halton low discrepancy sequence. The sequence is
  scrambled by default.
• scipy.stats.qmc.LatinHypercube: plain LHS design.

And 2 special samplers are available:

• scipy.stats.qmc.MultinomialQMC: sampling from a multinomial distribution
  using any of the base scipy.stats.qmc.QMCEngine.
• scipy.stats.qmc.MultivariateNormalQMC: sampling from a multivariate Normal
  using any of the base scipy.stats.qmc.QMCEngine.

The module also provide the following helpers:

• scipy.stats.qmc.discrepancy: assess the quality of a set of points in terms
  of space coverage.
• scipy.stats.qmc.update_discrepancy: can be used in an optimization loop to
  construct a good set of points.
• scipy.stats.qmc.scale: easily scale a set of points from (to) the unit
  interval to (from) a given range.

Deprecated features

scipy.linalg deprecations

• scipy.linalg.pinv2 is deprecated and its functionality is completely
  subsumed into scipy.linalg.pinv
• Both rcond, cond keywords of scipy.linalg.pinv and
  scipy.linalg.pinvh were not working and now are deprecated. They are now
  replaced with functioning atol and rtol keywords with clear usage.

scipy.spatial deprecations

• scipy.spatial.distance metrics expect 1d input vectors but will call
  np.squeeze on their inputs to accept any extra length-1 dimensions. That
  behaviour is now deprecated.

Backwards incompatible changes

Other changes

We now accept and leverage performance improvements from the ahead-of-time
Python-to-C++ transpiler, Pythran, which can be optionally disabled (via
export SCIPY_USE_PYTHRAN=0) but is enabled by default at build time.

There are two changes to the default behavior of scipy.stats.mannwhitenyu:

• For years, use of the default alternative=None was deprecated; explicit
  alternative specification was required. Use of the new default value of
  alternative, "two-sided", is now permitted.
• Previously, all p-values were based on an asymptotic approximation. Now, for
  small samples without ties, the p-values returned are exact by default.

Support has been added for PEP 621 (project metadata in pyproject.toml)

We now support a Gitpod environment to reduce the barrier to entry for SciPy
development; for more details see quickstart-gitpod.

Authors

• @endolith
• Jelle Aalbers +
• Adam +
• Tania Allard +
• Sven Baars +
• Max Balandat +
• baumgarc +
• Christoph Baumgarten
• Peter Bell
• Lilian Besson
• Robinson Besson +
• Max Bolingbroke
• Blair Bonnett +
• Jordão Bragantini
• Harm Buisman +
• Evgeni Burovski
• Matthias Bussonnier
• Dominic C
• CJ Carey
• Ramón Casero +
• Chachay +
• charlotte12l +
• Benjamin Curtice Corbett +
• Falcon Dai +
• Ian Dall +
• Terry Davis
• droussea2001 +
• DWesl +
• dwight200 +
• Thomas J. Fan +
• Joseph Fox-Rabinovitz
• Max Frei +
• Laura Gutierrez Funderburk +
• gbonomib +
• Matthias Geier +
• Pradipta Ghosh +
• Ralf Gommers
• Evan H +
• h-vetinari
• Matt Haberland
• Anselm Hahn +
• Alex Henrie
• Piet Hessenius +
• Trever Hines +
• Elisha Hollander +
• Stephan Hoyer
• Tom Hu +
• Kei Ishikawa +
• Julien Jerphanion
• Robert Kern
• Shashank KS +
• Peter Mahler Larsen
• Eric Larson
• Cheng H. Lee +
• Gregory R. Lee
• Jean-Benoist Leger +
• lgfunderburk +
• liam-o-marsh +
• Xingyu Liu +
• Alex Loftus +
• Christian Lorentzen +
• Cong Ma
• Marc +
• MarkPundurs +
• Markus Löning +
• Liam Marsh +
• Nicholas McKibben
• melissawm +
• Jamie Morton
• Andrew Nelson
• Nikola Forró
• Tor Nordam +
• Olivier Gauthé +
• Rohit Pandey +
• Avanindra Kumar Pandeya +
• Tirth Patel
• paugier +
• Alex H. Wagner, PhD +
• Jeff Plourde +
• Ilhan Polat
• pranavrajpal +
• Vladyslav Rachek
• Bharat Raghunathan
• Recursing +
• Tyler Reddy
• Lucas Roberts
• Gregor Robinson +
• Pamphile Roy +
• Atsushi Sakai
• Benjamin Santos
• Martin K. Scherer +
• Thomas Schmelzer +
• Daniel Scott +
• Sebastian Wallkötter +
• serge-sans-paille +
• Namami Shanker +
• Masashi Shibata +
• Alexandre de Siqueira +
• Albert Steppi +
• Adam J. Stewart +
• Kai Striega
• Diana Sukhoverkhova
• Søren Fuglede Jørgensen
• Mike Taves
• Dan Temkin +
• Nicolas Tessore +
• tsubota20 +
• Robert Uhl
• christos val +
• Bas van Beek +
• Ashutosh Varma +
• Jose Vazquez +
• Sebastiano Vigna
• Aditya Vijaykumar
• VNMabus
• Arthur Volant +
• Samuel Wallan
• Stefan van der Walt
• Warren Weckesser
• Anreas Weh
• Josh Wilson
• Rory Yorke
• Egor Zemlyanoy
• Marc Zoeller +
• zoj613 +
• 秋纫 +

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.