scipy-best-practices

SciPy Best Practices

Expert guidelines for SciPy development, focusing on scientific computing, optimization, signal processing, and statistical analysis.

Code Style and Structure

• Write concise, technical Python code with accurate SciPy examples
• Prioritize numerical accuracy and computational efficiency
• Use functional programming patterns for mathematical operations
• Prefer vectorized operations over explicit loops
• Use descriptive variable names reflecting scientific context
• Follow PEP 8 style guidelines

scipy.optimize - Optimization

• Use scipy.optimize.minimize() for general-purpose optimization
• Choose appropriate method based on problem characteristics:
  • 'BFGS' for smooth, unconstrained problems
  • 'L-BFGS-B' for bounded problems
  • 'SLSQP' for constrained optimization
  • 'Nelder-Mead' for non-differentiable functions
• Provide gradients when available for faster convergence
• Use scipy.optimize.curve_fit() for nonlinear least squares fitting
• Use scipy.optimize.root() for finding roots of equations

scipy.linalg - Linear Algebra

• Prefer scipy.linalg over numpy.linalg for additional functionality
• Use scipy.linalg.solve() instead of computing matrix inverse
• Leverage specialized solvers for structured matrices (banded, triangular)
• Use scipy.linalg.lu_factor() and lu_solve() for multiple right-hand sides
• Use sparse matrix solvers from scipy.sparse.linalg for large sparse systems

scipy.stats - Statistics

• Use distribution objects for probability calculations
• Leverage scipy.stats.describe() for summary statistics
• Use hypothesis testing functions: ttest_ind(), chi2_contingency(), mannwhitneyu()
• Generate random samples with .rvs() method on distributions
• Use .fit() for parameter estimation from data

scipy.interpolate - Interpolation

• Use scipy.interpolate.interp1d() for 1D interpolation
• Use scipy.interpolate.griddata() for scattered data interpolation
• Choose appropriate interpolation method: 'linear', 'cubic', 'nearest'
• Use spline functions for smooth interpolation: UnivariateSpline, BSpline
• Consider RegularGridInterpolator for regular grid data

scipy.integrate - Integration

• Use scipy.integrate.quad() for single integrals
• Use scipy.integrate.dblquad(), tplquad() for multiple integrals
• Use scipy.integrate.solve_ivp() for ordinary differential equations
• Choose appropriate ODE method: 'RK45', 'BDF', 'LSODA'
• Provide Jacobian for stiff systems to improve performance

scipy.signal - Signal Processing

• Use scipy.signal.butter(), cheby1(), ellip() for filter design
• Apply filters with scipy.signal.filtfilt() for zero-phase filtering
• Use scipy.signal.welch() for power spectral density estimation
• Use scipy.signal.find_peaks() for peak detection
• Leverage scipy.signal.convolve() and correlate() for convolution

scipy.sparse - Sparse Matrices

• Use appropriate sparse format for your use case:
  • csr_matrix for efficient row slicing and matrix-vector products
  • csc_matrix for efficient column slicing
  • coo_matrix for constructing sparse matrices
  • lil_matrix for incremental construction
• Convert to optimal format before operations
• Use scipy.sparse.linalg solvers for sparse linear systems

Performance Optimization

• Use appropriate data types (float64 for precision, float32 for memory)
• Leverage BLAS/LAPACK through SciPy for optimized linear algebra
• Pre-allocate arrays when possible
• Use in-place operations when available

Error Handling and Validation

• Check convergence status of optimization routines
• Validate numerical results for reasonableness
• Handle ill-conditioned problems gracefully
• Use appropriate tolerances for convergence criteria

Testing Scientific Code

• Test against known analytical solutions
• Use np.testing.assert_allclose() for numerical comparisons
• Test edge cases and boundary conditions
• Verify conservation laws and invariants

Key Conventions

• Import specific submodules: from scipy import optimize, stats, linalg
• Use snake_case for variables and functions
• Document algorithm choices and parameters
• Include convergence diagnostics in output