Citation and evidence

SciPy

17 min full readUpdated 26 references

This article's verification

Report a problem with this article

Edit article
More

Use this article

Raw MarkdownExplore connections

Improve this page

Revision historyDiscuss this article

Topics

Data ScienceMachine LearningMathematics

Cite this article

SciPy is an open-source library of numerical algorithms for the Python programming language, covering optimization, integration, interpolation, eigenvalue problems, algebraic and differential equations, statistics, signal processing and related tasks, together with data structures such as sparse matrices and k-dimensional trees.[1] It is built on NumPy, which supplies the array type and basic array routines, and it is in turn the base on which higher-level libraries such as scikit-learn and scikit-image are built.[1] The project began in 2001, released version 1.0 in October 2017 and is distributed under a BSD license.[1][3][5]

The library's developers describe it as a de facto standard for scientific algorithms in Python. Their 2020 paper in Nature Methods, "SciPy 1.0: fundamental algorithms for scientific computing in Python", reported more than 600 code contributors, over 100,000 dependent repositories and millions of downloads a year, and has become the project's recommended citation.[1][8] As of November 2026 the current release is SciPy 1.18.1, announced on 2026-08-21.[4][5]

DeveloperSciPy community (steering council chaired by Ralf Gommers; BDFL Pauli Virtanen)[6][7]
First releaseAugust 2001[1]
Version 1.02017-10-25[3][4]
Latest release (November 2026)1.18.1, 2026-08-21[4][5]
Languages (1.0 paper)About 50% Python, 25% Fortran, 20% C, 3% Cython, 2% C++[1]
RequiresPython 3.12 to 3.14 and NumPy 2.0.0 or later (SciPy 1.18)[20]
LicenseModified (3-clause) BSD[4][6]
Websitescipy.org[5]

Expanded article table

SciPy logo: a white S-shaped curve ending in an arrow on a dark blue circle
Logo of SciPy, the open-source Python library for scientific and technical computing. Not sure, via Wikimedia Commons, Public domain.

History

Background: Python and Numeric

The history in the SciPy 1.0 paper notes that Python, as a general-purpose language, had no special support for scientific data structures or algorithms, but scientists valued its ability to wrap existing C and Fortran libraries and drive them interactively, without managing memory by hand.[1] In 1995 Jim Hugunin, then a graduate student at MIT, wrote the first message on a new Python Matrix Special Interest Group (Matrix-SIG) mailing list. Discussions there among Jim Fulton, Hugunin, Paul Dubois, Konrad Hinsen and Guido van Rossum led to Numeric, a package with a multidimensional array object, and the availability of a standard array type set off rapid growth in scientific packages for Python.[1]

Origins (1998 to 2001)

By the late 1990s, Matrix-SIG members were asking for a complete scientific data analysis environment.[1] Travis Oliphant, a PhD student at the Mayo Clinic, released several packages built on Numeric that provided signal processing, special functions, sparse matrices, quadrature, optimization and fast Fourier transforms. One of them, Multipack, wrapped Fortran and C libraries to solve nonlinear equations and least-squares problems, integrate differential equations and fit splines.[1] Around the same time Pearu Peterson, a PhD student from Estonia, released F2PY, a tool for binding Python to Fortran code, and wrote linear algebra and interpolation modules. Eric Jones, a graduate student at Duke University, wrote a parallel job scheduler and a genetic optimizer for his dissertation, and Gary Strangman, a postdoctoral fellow at Harvard Medical School, published statistical routines.[1] In his note for the 1.0 release, Oliphant wrote that Robert Kern, Peterson and Jones joined their efforts with his in 2000.[3]

In 2001 Eric Jones and Travis Vaught founded Enthought Scientific Computing Solutions (now Enthought, Inc.) in Austin, Texas, and created the SciPy project around a library meant to absorb these separate packages.[1] A website and code repository appeared in February 2001, a mailing list was announced in June, the first release was announced in August, and the first documentation followed in September. The first SciPy workshop, a two-day single-track event with 50 participants, was held at Caltech in September 2002.[1] The SciPy 0.1 announcement listed special functions, signal and image processing, 2D plotting, integration, ODE solvers, optimization, genetic algorithms, a Numeric-to-C++ expression compiler, parallel programming tools, and splines and interpolation.[2]

Growth and maturation (2002 to 2016)

Numeric worked well for small arrays but not for the large images processed by the Space Telescope Science Institute, which wrote a replacement called NumArray; NumArray in turn was slow for small arrays. In 2005 Oliphant combined the two designs into NumPy, and NumPy 1.0 was released in October 2006.[1] SciPy moved onto NumPy in 2005.[3]

From the mid-2000s the library's scope narrowed while the surrounding ecosystem grew through "scikits", separate add-on packages.[1] Documentation improved after the 2006 Guide to NumPy: the SciPy Documentation Project, which used Sphinx and the wiki-like pydocweb editor and offered contributors T-shirts in exchange for text, gathered contributions from more than 75 people and produced an 884-page manual.[1][2] The SciPy workshop grew into an international conference; by 2010 the US conference had several tracks, and volunteers ran satellite events such as EuroSciPy (from 2008) and SciPy India (from 2009).[1] In 2012 Perry Greenfield, John Hunter, Jarrod Millman, Oliphant and Fernando Perez founded NumFOCUS, a 501(c)(3) charity whose sponsored projects include SciPy.[2]

The release notes for SciPy 1.0 summarise the project's milestones up to that point:[3]

YearMilestone
2001First SciPy release
2005Transition to NumPy
2007Creation of scikits
2008scipy.spatial module and first Cython code added
2010Move to a six-monthly release cycle
2011Development moves to GitHub; Python 3 support
2012Sparse graph module and unified optimization interface added; scipy.maxentropy removed
2013Continuous integration with Travis CI
2015Cython interface for BLAS/LAPACK and a benchmark suite added
2017Unified C API with scipy.LowLevelCallable; scipy.weave removed
2017SciPy 1.0 release

Expanded article table

SciPy 1.0 (2017)

SciPy 1.0.0 was published on 2017-10-25, about 16 years after version 0.1.[3][4] The release notes argued that the number was "long overdue" because the library had been used in production for years, and that recent goals, both technical (Windows wheels, continuous integration) and organisational (a governance structure, a code of conduct and a roadmap), had now been met.[3] The project had adopted a formal governance document on August 3, 2017, and approved its code of conduct on October 24, 2017.[1]

Highlights of 1.0 included the first Windows wheels on PyPI, continuous integration on Windows and macOS as well as Linux, a set of new ordinary differential equation solvers behind a unified interface (scipy.integrate.solve_ivp), two new trust-region optimizers, a new linear programming method, and complete BLAS wrappers. It required Python 2.7 or 3.4 and later, and NumPy 1.8.2 or later.[3]

After 1.0

Feature releases have continued at roughly six-month intervals; most release notes describe the release as the result of six months of work (1.14.0 says three).[10][16][20] The table lists releases with notable changes; dates are the first upload of each final release to PyPI.[4]

VersionPyPI dateSelected changes
1.2.02018-12-18Last release to support Python 2.7, made a long-term support series with bug-fix backports until 1 January 2020; new root_scalar interface, dual_annealing and shgo global optimizers, quaternion-based scipy.spatial.transform[10]
1.4.02019-12-16New scipy.fft module supersedes scipy.fftpack, adding long double transforms, release of the global interpreter lock and threading control; Edmonds-Karp maximum flow in scipy.sparse.csgraph[11]
1.8.02022-02-05Sparse array API added for early testing; PROPACK vendored for svds; new scipy.stats.sampling built on the UNU.RAN C library[12]
1.9.02022-07-29Build system moved to Meson; scipy.optimize.milp for mixed-integer linear programming; scipy.stats.fit; DIRECT global optimizer[13]
1.10.02023-01-03New scipy.datasets submodule; scipy.misc deprecated; make_smoothing_spline[14]
1.11.02023-06-25New sparse.sparray base class; survival analysis, multiple hypothesis testing and censored-data tools in scipy.stats[15]
1.14.02024-06-24Wheels built against Apple's Accelerate library for macOS 14 and later; COBYQA derivative-free solver in minimize[16]
1.15.02025-01-03Sparse arrays fully functional for 1-D and 2-D; preliminary support for free-threaded Python 3.13; new scipy.differentiate submodule; scipy.optimize.elementwise[17]
1.16.02025-06-22Wider experimental array API support with JAX and Dask backends; COBYLA replaced by the PRIMA implementation; batched input across nearly all of scipy.linalg[18]
1.17.02026-01-10Native batching of N-dimensional input in many functions; ARPACK and PROPACK rewritten from Fortran 77 to C; initial ILP64 BLAS/LAPACK support; scipy.odr deprecated for removal in 1.19.0[19]
1.18.02026-06-19Remaining Fortran-to-C translations completed and an experimental Fortran-free build offered; FFT backend switched from pocketfft to ducc0.fft; many scipy.stats functions support lazy arrays and JAX JIT[20]

Expanded article table

The supported Python versions have moved with the language. SciPy 1.4 required Python 3.5 or later, 1.15 supported Python 3.10 to 3.13, and 1.18 requires Python 3.12 to 3.14 and NumPy 2.0.0 or later.[11][17][20]

"Keynote: SciPy 1.0 and Beyond: A Story of Code and Community | SciPy 2018 | Ralf Gommers", video from Enthought, published July 13, 2018. Watch on YouTube.

Subpackages

SciPy is organised as a set of subpackages, which the documentation calls submodules. The 1.0 paper counted 16 as of February 2019.[1] The API reference for SciPy 1.18 lists the following public submodules, some of which contain further public modules such as scipy.sparse.linalg, scipy.sparse.csgraph, scipy.spatial.distance and scipy.stats.qmc:[9]

SubmoduleScope
clusterVector quantization and k-means (cluster.vq); hierarchical and agglomerative clustering (cluster.hierarchy)[1]
constantsPhysical and mathematical constants, including CODATA values of the fundamental constants[1]
datasetsExample datasets, added in 1.10 to replace dataset access through scipy.misc[14]
differentiateDerivative estimation for black-box functions, added in 1.15[17]
fftDiscrete Fourier transforms, added in 1.4; supersedes fftpack[11]
fftpackThe older Fourier transform module, with discrete sine and cosine transforms[1][11]
integrateDefinite integrals and ordinary differential equation solvers, including initial value and two-point boundary value problems[1]
interpolateSplines, univariate and multivariate interpolation, and wrappers for the FITPACK library[1]
ioReaders and writers for MATLAB, IDL, Matrix Market, Fortran, NetCDF, Harwell-Boeing, WAV and ARFF files[1]
linalgMatrix functions, solvers for $Ax = b$ and structured matrices, least squares, eigenvalue problems and decompositions (Cholesky, LU, QR, singular value and others)[1]
ndimageMultidimensional image processing: filters, interpolation, region labeling and morphology[1]
odrOrthogonal distance regression wrapping ODRPACK; deprecated in 1.17[1][19]
optimizeLocal and global minimization, linear and mixed-integer programming, least-squares curve fitting and root finding[1][13]
signalFiltering and filter design, convolution, linear systems, window functions, peak finding and spectral analysis[1]
sparseSparse matrix and sparse array formats, sparse linear algebra (sparse.linalg) and graph algorithms (sparse.csgraph)[1][17]
spatialk-d trees, Delaunay triangulation, convex hulls, Voronoi diagrams, distance functions and rotations[1][10]
specialSpecial functions such as Airy, Bessel and gamma, orthogonal polynomials and related functions[1]
statsProbability distributions, statistical tests, correlation measures, kernel density estimation and non-uniform random variate sampling (stats.sampling)[1][12]

Expanded article table

Examples

A few widely used functions show the style of the library:

  • scipy.optimize.minimize gives one interface to many local minimizers. When no method is given, it picks BFGS, L-BFGS-B or SLSQP depending on whether the problem has bounds or constraints.[22]
  • scipy.optimize.linprog solves linear programs of the form

min⁡x  cTxsubject toAubx≤bub,Aeqx=beq,l≤x≤u\min_x \; c^T x \quad \text{subject to} \quad A_{ub} x \le b_{ub}, \quad A_{eq} x = b_{eq}, \quad l \le x \le u

where $c$ is the cost vector, $A_{ub}$ and $b_{ub}$ define the inequality constraints, $A_{eq}$ and $b_{eq}$ the equality constraints, and $l$ and $u$ the bounds on $x$. Its default method is 'highs', which uses the HiGHS solvers.[21]

  • scipy.integrate.solve_ivp integrates systems of ordinary differential equations $dy/dt = f(t, y)$ from an initial value, using the explicit Runge-Kutta method of order 5(4), 'RK45', by default.[23]

Design and implementation

Scope

The 1.0 paper explains that SciPy's breadth was derived from the Guide to Available Mathematical Software (GAMS) classification. In slowly changing areas such as linear algebra it aims for complete coverage; elsewhere it supplies building blocks and leaves specialised work to other packages. In statistics, for example, SciPy covers what a textbook would (distributions, hypothesis tests, correlation), while statsmodels covers more advanced estimators and scikit-learn covers machine learning.[1] New algorithms are accepted if they are relevant to several fields of science and demonstrably important, for example classic enough to appear in textbooks or based on a well-cited peer-reviewed paper.[2]

At the 1.0 release SciPy's architectural scope matched NumPy's: in-memory algorithms on a single machine. Distributed computing and GPU support were explicitly out of scope, though the published paper notes that the roadmap had since been revised and that the new scipy.fft allowed experimental use of multiple backends.[1] Later releases moved further in that direction through support for input that follows the Python array API standard; SciPy 1.16 added improved support for JAX and Dask backends, and 1.18 added JAX JIT support to many scipy.stats functions.[18][20]

Languages and wrapped libraries

The SciPy 1.0 paper, using GitHub's linguist tool, put SciPy at about 50% Python, 25% Fortran, 20% C, 3% Cython and 2% C++.[1] Much of the compiled code wraps long-established numerical libraries: the Fortran packages FFTPACK, ODEPACK, QUADPACK, FITPACK, ODRPACK, MINPACK, ARPACK, ALGORITHM 644 and CDFLIB, and the C libraries trlib, SuperLU, Qhull and Cephes.[2] Cython serves as glue between these libraries and the Python interface and is used to speed up heavily used inner loops. For new code the project's stated preference is Python first, then Cython, then C, C++ or Fortran, chosen for maintainability.[1]

That balance has shifted since. The 1.17 release rewrote ARPACK and PROPACK from Fortran 77 in C, and the 1.18 release notes report that the remaining Fortran-to-C translations are complete, with an experimental Fortran-free build available to developers.[19][20]

API stability and testing

The SciPy 1.0 paper put the API at about 1,500 functions and classes and described a policy that existing functionality is removed or changed only when the benefit outweighs the cost to users, after deprecation warnings lasting at least one year.[1] The current API reference states that public submodules are unlikely to be renamed or changed incompatibly and that a deprecation warning is raised for one release before such a change.[9] Besides the Python API, SciPy offers C and Cython interfaces, including Cython wrappers for BLAS, LAPACK and scipy.special, and scipy.LowLevelCallable for passing compiled callback functions with low overhead.[1][2]

At the 1.0 release the test suite had more than 13,000 unit tests run with pytest, covering 87% of Python code and 45% of compiled code; the authors note that the compiled figure includes vendored, generated and deprecated code.[1] Performance has been tracked with the Airspeed Velocity benchmarking tool since February 2015.[2]

Governance and funding

SciPy's governance document, adopted on August 3, 2017, gives the project a Benevolent Dictator for Life (BDFL) and a Steering Council. As of November 2026 the BDFL is Pauli Virtanen; the document says the BDFL normally defers to community consensus and the council and uses a final "overriding" vote only rarely, such as in a deadlock.[1][7] Council members are contributors with substantial work sustained over at least a year, and the council has a chair and appoints a release manager.[7] The project's About page lists 17 council members, with Ralf Gommers as chair and Tyler Reddy as release manager, and names co-creators Eric Jones, Pearu Peterson and Travis Oliphant among emeritus members.[6] The governance document states that the project itself is not a legal entity.[7]

The 1.0 paper described the project as largely unfunded, developed mostly by graduate students, faculty and industry staff in their free time, and cited an Open Hub estimate that its development cost exceeded $10 million.[1] Donations are handled by NumFOCUS, which provides administrative and legal services. The About page lists Quansight and Los Alamos National Laboratory as institutional partners that employ SciPy contributors, and thanks Enthought for hosting and for holding the SciPy trademark, along with companies that donated continuous integration credit, software licences and Google Summer of Code support.[6]

The SciPy 1.0 paper

"SciPy 1.0: fundamental algorithms for scientific computing in Python" was posted to arXiv on 2019-07-23 and published in Nature Methods (volume 17, issue 3, pages 261-272), online on 2020-02-03, with an author correction on 2020-02-24.[1][2][24] Its 34 named authors, beginning with Pauli Virtanen, Ralf Gommers, Travis E. Oliphant, Matt Haberland and Tyler Reddy, are followed by the group author "SciPy 1.0 Contributors".[8] The paper covers the project's history, the scope of each subpackage, implementation choices, technical improvements in the three years before 1.0, testing, and governance.[1]

Before the paper, the SciPy website was the default citation and had been cited more than 3,000 times.[1] The project now asks researchers to cite the paper, and to cite the original authors of specific algorithms where relevant.[8] Crossref's record for the article listed more than 40,000 citing works as of November 2026.[24]

The paper also collected usage figures: as of February 2019 the library had nearly 600,000 lines of code, and more than 110,000 GitHub repositories and 6,500 packages depended on it. It recorded 13,096,468 downloads from PyPI and 5,776,017 from the default conda channel in 2017, and 610 source code contributors with more than 19,000 commits as of May 2018.[1]

Use in science and machine learning

The 1.0 paper names the scripts used to analyse gravitational waves at LIGO and the M87 black hole imaging project among SciPy's prominent uses, and notes that SciPy shipped with macOS and in the Intel Distribution for Python.[1] Citing a January 2019 GitHub blog post on machine learning projects, it states that SciPy was used by 47% of all machine learning projects on GitHub.[1]

Much of that use is indirect: scikit-learn and scikit-image are built on top of SciPy.[1] Individual SciPy functions also appear directly in data analysis and machine learning code. scipy.stats.wasserstein_distance computes the Wasserstein-1 distance, also called the Earth mover's distance, between two one-dimensional discrete distributions.[25] scipy.stats.zscore standardises data for z-score normalization; its ddof argument defaults to 0, which gives the population standard deviation.[26]

Continue reading

More to read

Topics:Data ScienceMachine LearningMathematics

References

  1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39 ^40 ^41 ^42 ^43 ^44 ^45 ^46 ^47 ^48 ^49 ^50Virtanen, P., Gommers, R., Oliphant, T. E., et al. "SciPy 1.0: fundamental algorithms for scientific computing in Python". Nature Methods 17, 261-272, 2020. nature.com/...s41592-019-0686-2
  2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Virtanen, P., Gommers, R., Oliphant, T. E., et al. "SciPy 1.0--Fundamental Algorithms for Scientific Computing in Python". arXiv:1907.10121, 2019-07-23. arxiv.org/...1907.10121
  3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8SciPy developers. "SciPy 1.0.0 Release Notes". SciPy documentation. docs.scipy.org/...1.0.0-notes
  4. ^1 ^2 ^3 ^4 ^5 ^6Python Package Index. "scipy" (release history). PyPI. pypi.org/...scipy
  5. ^1 ^2 ^3 ^4SciPy developers. "SciPy". scipy.org. scipy.org
  6. ^1 ^2 ^3 ^4SciPy developers. "About Us". scipy.org. scipy.org/about
  7. ^1 ^2 ^3 ^4SciPy developers. "SciPy Project Governance". SciPy documentation. docs.scipy.org/...governance
  8. ^1 ^2 ^3SciPy developers. "Citing SciPy". scipy.org. scipy.org/citing-scipy
  9. ^1 ^2SciPy developers. "SciPy API". SciPy v1.18.0 Manual. docs.scipy.org/...reference
  10. ^1 ^2 ^3SciPy developers. "SciPy 1.2.0 Release Notes". SciPy documentation. docs.scipy.org/...1.2.0-notes
  11. ^1 ^2 ^3 ^4SciPy developers. "SciPy 1.4.0 Release Notes". SciPy documentation. docs.scipy.org/...1.4.0-notes
  12. ^1 ^2SciPy developers. "SciPy 1.8.0 Release Notes". SciPy documentation. docs.scipy.org/...1.8.0-notes
  13. ^1 ^2SciPy developers. "SciPy 1.9.0 Release Notes". SciPy documentation. docs.scipy.org/...1.9.0-notes
  14. ^1 ^2SciPy developers. "SciPy 1.10.0 Release Notes". SciPy documentation. docs.scipy.org/...1.10.0-notes
  15. ^SciPy developers. "SciPy 1.11.0 Release Notes". SciPy documentation. docs.scipy.org/...1.11.0-notes
  16. ^1 ^2SciPy developers. "SciPy 1.14.0 Release Notes". SciPy documentation. docs.scipy.org/...1.14.0-notes
  17. ^1 ^2 ^3 ^4SciPy developers. "SciPy 1.15.0 Release Notes". SciPy documentation. docs.scipy.org/...1.15.0-notes
  18. ^1 ^2SciPy developers. "SciPy 1.16.0 Release Notes". SciPy documentation. docs.scipy.org/...1.16.0-notes
  19. ^1 ^2 ^3SciPy developers. "SciPy 1.17.0 Release Notes". SciPy documentation. docs.scipy.org/...1.17.0-notes
  20. ^1 ^2 ^3 ^4 ^5 ^6SciPy developers. "SciPy 1.18.0 Release Notes". SciPy documentation. docs.scipy.org/...1.18.0-notes
  21. ^SciPy developers. "scipy.optimize.linprog". SciPy v1.18.0 Manual. docs.scipy.org/...scipy.optimize.linprog
  22. ^SciPy developers. "scipy.optimize.minimize". SciPy v1.18.0 Manual. docs.scipy.org/...scipy.optimize.minimize
  23. ^SciPy developers. "scipy.integrate.solve_ivp". SciPy v1.18.0 Manual. docs.scipy.org/...scipy.integrate.solve_ivp
  24. ^1 ^2Crossref. "SciPy 1.0: fundamental algorithms for scientific computing in Python" (metadata record, DOI 10.1038/s41592-019-0686-2). Crossref REST API. api.crossref.org/...s41592-019-0686-2
  25. ^SciPy developers. "scipy.stats.wasserstein_distance". SciPy v1.18.0 Manual. docs.scipy.org/...scipy.stats.wasserstein_distance
  26. ^SciPy developers. "scipy.stats.zscore". SciPy v1.18.0 Manual. docs.scipy.org/...scipy.stats.zscore

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

2 revisions · v2 · 3,343 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Independent fact-check before publication on November 18, 2026: 142 claims checked against 28 cited and independent sources; 0 material and 4 minor corrections applied before publication.

Cite this page: AI Wiki. "SciPy." aiwiki.ai, updated 18 Nov 2026, fact-checked 18 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/scipy

Edit article