import type { CoreClusteringParams, DataMatrix } from '../clustering/types';
/**
 * Principal Component Analysis matching scikit-learn's `PCA(svd_solver='full')`
 * numerically (up to per-component sign).
 *
 * The TensorFlow.js backend has no eigendecomposition, so the leading
 * components are found by power iteration with deflation on the (mean-centered)
 * covariance matrix. The same power-iteration core seeds SOM `'linear'`/`'pca'`
 * weight initialization, keeping map seeding and public reduction numerically
 * identical.
 */
export interface PCAParams extends CoreClusteringParams {
    /** Must be ≤ n_features. */
    n_components: number;
}
export interface PCAJSON {
    params: PCAParams;
    components_: number[][];
    explained_variance_: number[];
    mean_: number[];
}
export interface EigResult {
    /** Eigenvectors as rows, most-significant first. */
    components: number[][];
    /** Corresponding eigenvalues, descending. */
    eigenvalues: number[];
}
/**
 * Unit-norm power-iteration start vector for component `c`: the candidate
 * normalized, or the `c % d` standard basis vector when the candidate has
 * zero norm (so iteration always starts from a valid direction).
 */
export declare function unit_init_vector(candidate: number[], c: number): number[];
/**
 * Top-`k` eigenvectors/eigenvalues of a symmetric matrix via power iteration
 * with deflation. Deterministic for a fixed `random_state`. Shared by
 * {@link PCA} and SOM principal-component initialization.
 *
 */
export declare function power_iteration_eig(matrix: number[][], k: number, random_state?: number): EigResult;
export declare class PCA {
    readonly params: PCAParams;
    /** Principal axes as rows (`n_components × n_features`). Null until `fit`. */
    components_: number[][] | null;
    explained_variance_: number[] | null;
    /** Per-feature empirical mean used for centering. Null until `fit`. */
    mean_: number[] | null;
    constructor(params: PCAParams);
    fit(X: DataMatrix): this;
    transform(X: DataMatrix): number[][];
    fit_transform(X: DataMatrix): number[][];
    inverse_transform(Z: DataMatrix): number[][];
    /** PCA is a dimensionality reducer, not a clusterer. */
    fit_predict(_X: DataMatrix): never;
    to_json(): PCAJSON;
    static from_json(json: PCAJSON): PCA;
}
//# sourceMappingURL=pca.d.ts.map