Skip to content

curvefit_stats

Statistical inference and estimation for SciPy-style nonlinear models.

All ten features are re-exported here without wrappers. Direct imports from :mod:curvefit_stats.inference and :mod:curvefit_stats.estimators retain the same objects and signatures. These are the only functional submodules.

Covariance conventions

A covariance array alone cannot encode its estimator, noise model, degrees of freedom, or boundary status. Callers must preserve that context:

  • SciPy curve_fit with an unknown common Gaussian scale supports approximate Student-t/Wald inference using the original fitting N-p. Known absolute scales use generate_error_bands(..., reference='normal') instead.
  • GlobalFitGroup uses one global degree of freedom per estimated shared parameter. For per-dataset bands, project the global covariance using the group's integer mapping and pass the global dof explicitly, without supplying per-dataset y to infer a new local dof. Use explicit prediction_variance from the global noise model rather than reestimating it from a single dataset.
  • EIV covariance comes from ODR, not vertical-response least squares. It can support Delta Method mean bands at exact evaluation coordinates and Gaussian parameter propagation. audit_fit, profile_likelihood_errors, fit_summary, residual bootstrap, and robust_bounded_pcov use response-only least-squares assumptions and do not automatically become ODR methods when passed ODR estimates. Future uncertain predictors require their own observation model.
  • Bootstrap covariance is an empirical parameter covariance, not a Student-t pivot. Gaussian propagation from it is a further approximation; bootstrap percentile intervals should be calculated directly from bootstrap samples.
  • Bound-active covariance has NaN rows and columns. Full-matrix propagation is rejected. Extracting a finite free block only permits inference conditional on the active parameters, never unrestricted marginal uncertainty for them.

See docs/integration.md for complete examples and limitations. No estimator provenance is inferred from plain parameter/covariance arrays.