Scipy 1 17 1

scipy 1.17.1

SciPy is the core scientific computing library for Python, built on NumPy arrays. It provides efficient numerical routines across mathematics, science, and engineering domains.

Overview

SciPy organizes its functionality into submodules, each accessible as scipy.<module>. The main namespace exports submodule names (lazy-loaded), plus __version__, LowLevelCallable, show_config, and test.

Submodule Map

Deprecated / Legacy Modules

• scipy.fftpack — legacy FFT; use scipy.fft instead
• scipy.misc — deprecated, removed in 2.0.0
• scipy.odr — deprecated since 1.17.0, removed in 1.19.0; migrate to odrpack on PyPI
• odeint — old ODE API; prefer solve_ivp for new code

Usage

Import Patterns

# Submodule-level import (recommended)
from scipy import optimize, integrate, stats, linalg, signal, fft, sparse, spatial, special, ndimage, io, cluster, constants, differentiate, interpolate

# Function-level import for frequently used functions
from scipy.optimize import minimize, curve_fit, differential_evolution
from scipy.integrate import quad, solve_ivp
from scipy.stats import norm, ttest_ind, pearsonr
from scipy.linalg import eig, svd, cholesky, expm
from scipy.signal import find_peaks, welch, firwin, lfilter
from scipy.fft import fft, ifft, rfft, next_fast_len
from scipy.interpolate import interp1d, CubicSpline, RegularGridInterpolator
from scipy.sparse import csr_array, diags_array, eye_array
from scipy.sparse.linalg import spsolve, gmres, eigs
from scipy.spatial import KDTree, ConvexHull, distance
from scipy.special import gamma, erf, jv, lambertw
from scipy.ndimage import gaussian_filter, label, binary_erosion
from scipy.io import loadmat, savemat
from scipy.cluster.vq import kmeans
from scipy.cluster.hierarchy import linkage, dendrogram
from scipy.constants import pi, c, h, value

Quick Reference by Task

Optimization: minimize(func, x0, method='LBFGS') for unconstrained; add bounds or constraints for constrained. Use differential_evolution() for global optimization. Use curve_fit(model, xdata, ydata) for fitting data to a model.

Root-finding: root_scalar(func, bracket=[a, b], method='brentq') for scalar; root(func, x0) for multivariate. Bracketing methods (brentq, bisect) are guaranteed to converge.

Integration: quad(func, a, b) for definite integrals. solve_ivp(fun, t_span, y0, method='RK45') for ODEs. trapezoid(y, x) or simpson(y, x) for numerical integration from samples.

Linear Algebra: Prefer scipy.linalg over numpy.linalg — it offers more methods and consistent behavior. Use solve(A, b) instead of inv(A) @ b. Use eig(), eigh() (Hermitian), svd(), cholesky().

Statistics: Distributions are objects: norm.pdf(x, loc=0, scale=1), norm.cdf(x), norm.rvs(size=1000). Hypothesis tests return (statistic, pvalue): ttest_ind(a, b), pearsonr(x, y).

Signal Processing: find_peaks(signal, height=threshold, distance=min_distance). welch(x, fs=sampling_rate) for power spectral density. firwin(numtaps, cutoff, fs=fs) for FIR filter design. lfilter(b, a, x) to apply a filter.

FFT: fft(x) and ifft(X). Use rfft(x) for real-valued input (faster, half-size output). next_fast_len(n) finds optimal zero-pad length. Prefer scipy.fft over numpy.fft for multi-threading and backend control.

Sparse Matrices: Use _array classes (csr_array, not csr_matrix) for new code. Matrix multiplication uses @ operator, not *. Use spsolve(A, b) for direct solve, gmres(A, b) for iterative.

Spatial: KDTree(points).query(query_points, k=5) for nearest neighbors. distance.pdist(X, metric='euclidean') for pairwise distances. ConvexHull(points) and Delaunay(points) for geometry.

Special Functions: Vectorized over NumPy arrays. gamma(x), erf(x), jv(n, x) (Bessel J), lambertw(x). Error handling via seterr(), errstate().

Version Compatibility

SciPy 1.17.1 requires NumPy ≥ 1.26.4 and < 2.7.0. A warning is emitted if the installed NumPy version is outside this range.

Gotchas

• scipy.sparse arrays use @ for matrix multiplication — the * operator does element-wise multiplication (like NumPy). This differs from the old _matrix classes where * was matrix multiply. Always use _array classes, not _matrix.
• scipy.odr is deprecated since 1.17.0 and will be removed in 1.19.0. Migrate to the standalone odrpack package on PyPI.
• scipy.fftpack is legacy — use scipy.fft which supports multi-threading, backends, and real-valued optimizations (rfft).
• odeint vs solve_ivp — odeint is the old Fortran-based API. solve_ivp is the modern Python API with event detection, dense output, and multiple solver methods. Prefer solve_ivp.
• Distribution parameters use loc and scale — not mu/sigma. E.g., norm.pdf(x, loc=5, scale=2) for N(5, 4). The *args positional form still works but keyword form is clearer.
• Hypothesis tests return (statistic, pvalue) tuples — not just p-values. Unpack both: stat, p = ttest_ind(a, b).
• scipy.linalg extends numpy.linalg — same function names but with more methods and consistent behavior across dtypes. Prefer scipy.linalg when you need features beyond what NumPy provides.
• find_peaks returns indices, not values — use signal[peaks] to get peak values. Parameters like height, distance, prominence, and width control filtering.
• Sparse solver choice matters — spsolve uses SuperLU (direct, good for small/medium). Iterative solvers (gmres, cg) scale better for large systems but need preconditioners. Use LinearOperator when A is implicit (e.g., defined by a function).
• quad expects a callable — for data-driven integration, use trapezoid() or simpson() instead. quad adapts its sampling internally.
• minimize methods have different capabilities — Nelder-Mead needs no gradient but is slow. L-BFGS-B supports bounds. trust-constr supports general constraints. Choose method based on problem structure.

References

Detailed function listings and usage patterns for each submodule:

• 01-optimize.md — Optimization, root-finding, curve fitting, LP/MILP
• 02-integrate.md — Numerical integration, ODE/BVP solvers
• 03-stats.md — Probability distributions, hypothesis tests, summary statistics
• 04-linalg.md — Dense linear algebra (decompositions, eigenvalues, matrix functions)
• 05-signal.md — Signal processing, filter design, spectral analysis, LTI systems
• 06-fft.md — Discrete Fourier transforms, DCT/DST, Hankel transforms
• 07-interpolate.md — Interpolation and spline fitting
• 08-sparse.md — Sparse matrices and sparse linear algebra
• 09-spatial.md — Spatial algorithms, distance metrics, geometry, transforms
• 10-special.md — Special mathematical functions (Bessel, gamma, erf, orthogonal polynomials)
• 11-ndimage.md — N-dimensional image processing (filters, morphology, measurements)
• 12-io.md — File I/O (MATLAB, Matrix Market, NetCDF, Fortran, WAV)
• 13-cluster.md — Clustering algorithms (k-means, hierarchical)
• 14-constants.md — Physical/mathematical constants, units, SI prefixes