/**
 * Variational Bayesian estimation of a Gaussian mixture, following the update
 * equations of sklearn.mixture.BayesianGaussianMixture (Bishop, PRML ch. 10.2):
 *
 *  - weights carry a Dirichlet prior ('dirichletDistribution') or a
 *    stick-breaking Dirichlet-process prior ('dirichletProcess'), updated with
 *    digamma expectations;
 *  - each component's mean/precision carries a Gaussian-Wishart prior updated
 *    from weighted sufficient statistics;
 *  - the E-step uses the expected log-likelihood (digamma-corrected Gaussian
 *    densities), and convergence is declared when the change in the variational
 *    lower bound falls below `tol`.
 *
 * covarianceType 'full' and 'diag' are implemented; 'tied' and 'spherical'
 * throw a clear not-implemented error.
 */
import { ClusterBase } from '../base/cluster';
import { Params } from '../base/estimator';
export type BayesianCovarianceType = 'full' | 'diag';
export type WeightConcentrationPriorType = 'dirichletProcess' | 'dirichletDistribution';
export interface BayesianGaussianMixtureProps {
    /** number of mixture components; with a DP prior extra ones shrink to ~0 weight (default 1) */
    nComponents?: number;
    /** covariance parametrization: 'full' | 'diag' (default 'full') */
    covarianceType?: BayesianCovarianceType;
    /** convergence threshold on the lower-bound change (default 1e-3) */
    tol?: number;
    /** non-negative regularization added to covariance diagonals (default 1e-6) */
    regCovar?: number;
    /** maximum number of variational iterations (default 100) */
    maxIter?: number;
    /** number of restarts; the run with the best lower bound wins (default 1) */
    nInit?: number;
    /** responsibility initialization: 'kmeans' | 'random' (default 'kmeans') */
    initParams?: 'kmeans' | 'random';
    /** seed for reproducible initialization */
    randomState?: number;
    /** prior family on the weights (default 'dirichletProcess') */
    weightConcentrationPriorType?: WeightConcentrationPriorType;
    /** Dirichlet/stick-breaking concentration; default 1 / nComponents */
    weightConcentrationPrior?: number;
    /** precision prior on the mean distributions (default 1) */
    meanPrecisionPrior?: number;
    /** prior on the mean distributions (default: column means of X) */
    meanPrior?: number[];
    /** Wishart degrees-of-freedom prior; must exceed nFeatures − 1 (default nFeatures) */
    degreesOfFreedomPrior?: number;
    /** covariance prior: d×d matrix for 'full', length-d variance vector for 'diag' (default: empirical) */
    covariancePrior?: number[][] | number[];
}
type BayesianCovariances = number[][][] | number[][];
export declare class BayesianGaussianMixture extends ClusterBase {
    private nComponents;
    private covarianceType;
    private tol;
    private regCovar;
    private maxIter;
    private nInit;
    private initParams;
    private randomState?;
    private weightConcentrationPriorType;
    private weightConcentrationPrior?;
    private meanPrecisionPrior?;
    private meanPrior;
    private degreesOfFreedomPrior?;
    private covariancePrior;
    private weights_;
    private weightConcentrationA_;
    private weightConcentrationB_;
    private meanPrecision_;
    private means_;
    private degreesOfFreedom_;
    private covariances_;
    /** whether the best run reached the `tol` criterion */
    converged: boolean;
    /** number of variational iterations performed by the best run */
    nIter: number;
    private lowerBound_;
    private lowerBoundHistory_;
    constructor(props?: BayesianGaussianMixtureProps);
    getParams(): Params;
    private resolvePriors;
    private mStep;
    /** E[log πₖ] under the Dirichlet / stick-breaking posterior */
    private static expectedLogWeights;
    /** n × k matrix E[log πₖ] + E[log N(xᵢ | μₖ, Λₖ⁻¹)] */
    private weightedLogProbFromState;
    private computeLowerBound;
    private runVB;
    fit(X: number[][]): this;
    /** posterior expected mixture weights (normalized) */
    private static expectedWeights;
    private checkFitted;
    private fittedState;
    /** hard assignment: argmax of the responsibilities */
    predict(X: number[][]): number[];
    /** posterior probability (responsibility) of each component for each sample */
    predictProba(X: number[][]): number[][];
    fitPredict(X: number[][]): number[];
    /** per-sample expected log-likelihood under the variational posterior */
    scoreSamples(X: number[][]): number[];
    /** average per-sample expected log-likelihood */
    score(X: number[][]): number;
    getWeights(): number[] | null;
    getMeans(): number[][] | null;
    getCovariances(): BayesianCovariances | null;
    /** final variational lower bound (sklearn's simplified ELBO) of the best run */
    getLowerBound(): number;
    /** lower-bound trajectory of the best run (one entry per iteration, non-decreasing) */
    getLowerBoundHistory(): number[];
}
export {};