Statistics

Computes the X^T * X matrix manually to save memory.

Returns the class label with the highest probability.

Computes logits for a single sample against a CxK coefficient matrix. Row 0 of beta is assumed to be the zeroed reference class.

Computes approximate per-class standard errors from the Hessian at convergence. Uses the standard class-vs-rest approximation H_c = Σ p_ic(1-p_ic) x_i x_i^T + λI.

Computes the VIF of a given matrix.

Returns the condition number of a given matrix.

Smooths age cohorts and couples sex ratios simultaneously. Enforces non-increasing monotonicity along the age dimension using weighted PAVA on total cohort population density, then allocates totals between male and female using the model's empirical sex ratio bounded within a biological plausibility envelope (Option 1).

Graduates total age-cohort density and reallocates it between sexes.

Defaults are appropriate for stochastic model-generated compositions: raw sex ratios are preserved, monotonicity is disabled, and the total population of the graduated suffix is preserved by final rescaling.

Density is normalised by the mean density of the graduated suffix. huber_delta is therefore expressed relative to that mean.

Fidelity weights default to band widths. These are integration weights, not estimated inverse-variance weights for a multinomial model.

Sex-ratio allocation applies to every cohort, including the prefix. Density graduation and total rescaling apply only from start_index.

Legacy options.buffers is not used; working calculations use Float64Array.

Evaluates a trained multinomial logit model against a covariates object, returning overall accuracy and a confusion matrix.

Generates multi-class probability rasters from a trained ALR compositional model.

Generates raster predictions from a trained multinomial logit model. Supports argmax class maps, single-class probability surfaces, or one probability raster per class.

Concurrently generates a series of Multinomial Logit predicted rasters across background worker processes.

Generates an OLS raster from a stack of coefficients.

Concurrently generates a series of OLS predicted rasters across background worker processes.

Geometrically average OLS models across a folder path, with base OLS prefixes.

Returns the class label at a given index of Y, unwrapping single-element arrays.

Computes a normalised logistic transition kernel weight w in [0, 1] over a domain [start_val, end_val]. Returns 0 for values <= start_val, 1 for values >= end_val, and an S-curve blend for intermediate values.

Loads a set of covariate rasters, resolving function- or string-valued paths and dropping rasters that fail to load. Shared by all covariate-stacking routines.

Loads a stack of covariates for a categorical class raster for multinomial logit training. Unlike loadOLSCovariates, zero-valued class labels are meaningful and retained.

Loads a stack of covariates for a specific utility file path for OLS training.

Loads a stack of covariates for point-based categorical data (e.g. household survey age/sex buckets) for multinomial logit training for a specific year.

Loads a stack of covariates for point-based data for OLS training for a specific year.

Performs L2-regularised Multinomial Logistic Regression (softmax regression) via batch gradient descent with momentum. The first observed class is the reference class with coefficients fixed at zero for identifiability.

Enforces non-increasing monotonicity on a sequence using the weighted Pool Adjacent Violators Algorithm (PAVA). Values before arg2_start_index are preserved unconstrained.

Computes predicted compositional class probabilities for a single feature object using a trained ALR model.

Predicts class probabilities from an ensemble of models using a convex combination in probability space.

Predicts class probabilities for a single feature object against a trained multinomial logit model. Supports both discrete multinomial_logit models and probability-space multinomial_ensemble mixtures.

Processes and adjusts OLS model coefficients against target and covariate rasters using Multiplicative Update Rules (NMF).

Removes high VIF features for a given matrix.

Performs Ridge Regression on two matrices with RMS scale stabilization and pseudo-inverse fallbacks.

Computes a numerically-stable softmax over an array of logits.

Trains an Additive Log-Ratio (ALR) regularised ridge regression model for continuous compositional targets. Bypasses gradient descent under-convergence by solving normal equations analytically.

Trains a multinomial logit model given a fitted covariates object { X, Y, keys } and saves it to JSON. Coefficients are stored per non-reference class.

Concurrently trains a series of multinomial logit models across background worker processes.

Trains a raster-based OLS model given a fitted covariates object { X, Y, keys }.

Concurrently trains a series of OLS models across background worker processes.

Performs Whittaker-Henderson graduation with weighted Huber fidelity, second-difference roughness, and optional soft non-increasing monotonicity.

Penalties apply only within [start_index, values.length). Values before start_index are copied without modification.

Computes the X^T * X matrix manually to save memory.

Returns the class label with the highest probability.

Computes logits for a single sample against a CxK coefficient matrix. Row 0 of beta is assumed to be the zeroed reference class.

Computes approximate per-class standard errors from the Hessian at convergence. Uses the standard class-vs-rest approximation H_c = Σ p_ic(1-p_ic) x_i x_i^T + λI.

Computes the VIF of a given matrix.

Returns the condition number of a given matrix.

Smooths age cohorts and couples sex ratios simultaneously. Enforces non-increasing monotonicity along the age dimension using weighted PAVA on total cohort population density, then allocates totals between male and female using the model's empirical sex ratio bounded within a biological plausibility envelope (Option 1).

Graduates total age-cohort density and reallocates it between sexes.

Defaults are appropriate for stochastic model-generated compositions: raw sex ratios are preserved, monotonicity is disabled, and the total population of the graduated suffix is preserved by final rescaling.

Density is normalised by the mean density of the graduated suffix. huber_delta is therefore expressed relative to that mean.

Fidelity weights default to band widths. These are integration weights, not estimated inverse-variance weights for a multinomial model.

Sex-ratio allocation applies to every cohort, including the prefix. Density graduation and total rescaling apply only from start_index.

Legacy options.buffers is not used; working calculations use Float64Array.

Evaluates a trained multinomial logit model against a covariates object, returning overall accuracy and a confusion matrix.

Generates multi-class probability rasters from a trained ALR compositional model.

Generates raster predictions from a trained multinomial logit model. Supports argmax class maps, single-class probability surfaces, or one probability raster per class.

Concurrently generates a series of Multinomial Logit predicted rasters across background worker processes.

Generates an OLS raster from a stack of coefficients.

Concurrently generates a series of OLS predicted rasters across background worker processes.

Geometrically average OLS models across a folder path, with base OLS prefixes.

Returns the class label at a given index of Y, unwrapping single-element arrays.

Computes a normalised logistic transition kernel weight w in [0, 1] over a domain [start_val, end_val]. Returns 0 for values <= start_val, 1 for values >= end_val, and an S-curve blend for intermediate values.

Loads a set of covariate rasters, resolving function- or string-valued paths and dropping rasters that fail to load. Shared by all covariate-stacking routines.

Loads a stack of covariates for a categorical class raster for multinomial logit training. Unlike loadOLSCovariates, zero-valued class labels are meaningful and retained.

Loads a stack of covariates for a specific utility file path for OLS training.

Loads a stack of covariates for point-based categorical data (e.g. household survey age/sex buckets) for multinomial logit training for a specific year.

Loads a stack of covariates for point-based data for OLS training for a specific year.

Performs L2-regularised Multinomial Logistic Regression (softmax regression) via batch gradient descent with momentum. The first observed class is the reference class with coefficients fixed at zero for identifiability.

Enforces non-increasing monotonicity on a sequence using the weighted Pool Adjacent Violators Algorithm (PAVA). Values before arg2_start_index are preserved unconstrained.

Computes predicted compositional class probabilities for a single feature object using a trained ALR model.

Predicts class probabilities from an ensemble of models using a convex combination in probability space.

Predicts class probabilities for a single feature object against a trained multinomial logit model. Supports both discrete multinomial_logit models and probability-space multinomial_ensemble mixtures.

Processes and adjusts OLS model coefficients against target and covariate rasters using Multiplicative Update Rules (NMF).

Removes high VIF features for a given matrix.

Performs Ridge Regression on two matrices with RMS scale stabilization and pseudo-inverse fallbacks.

Computes a numerically-stable softmax over an array of logits.

Trains an Additive Log-Ratio (ALR) regularised ridge regression model for continuous compositional targets. Bypasses gradient descent under-convergence by solving normal equations analytically.

Trains a multinomial logit model given a fitted covariates object { X, Y, keys } and saves it to JSON. Coefficients are stored per non-reference class.

Concurrently trains a series of multinomial logit models across background worker processes.

Trains a raster-based OLS model given a fitted covariates object { X, Y, keys }.

Concurrently trains a series of OLS models across background worker processes.

Performs Whittaker-Henderson graduation with weighted Huber fidelity, second-difference roughness, and optional soft non-increasing monotonicity.

Penalties apply only within [start_index, values.length). Values before start_index are copied without modification.

Computes the X^T * X matrix manually to save memory.

Returns the class label with the highest probability.

Computes logits for a single sample against a CxK coefficient matrix. Row 0 of beta is assumed to be the zeroed reference class.

Computes approximate per-class standard errors from the Hessian at convergence. Uses the standard class-vs-rest approximation H_c = Σ p_ic(1-p_ic) x_i x_i^T + λI.

Computes the VIF of a given matrix.

Returns the condition number of a given matrix.

Smooths age cohorts and couples sex ratios simultaneously. Enforces non-increasing monotonicity along the age dimension using weighted PAVA on total cohort population density, then allocates totals between male and female using the model's empirical sex ratio bounded within a biological plausibility envelope (Option 1).

Graduates total age-cohort density and reallocates it between sexes.

Defaults are appropriate for stochastic model-generated compositions: raw sex ratios are preserved, monotonicity is disabled, and the total population of the graduated suffix is preserved by final rescaling.

Density is normalised by the mean density of the graduated suffix. huber_delta is therefore expressed relative to that mean.

Fidelity weights default to band widths. These are integration weights, not estimated inverse-variance weights for a multinomial model.

Sex-ratio allocation applies to every cohort, including the prefix. Density graduation and total rescaling apply only from start_index.

Legacy options.buffers is not used; working calculations use Float64Array.