import { RegressorBase } from '../base';
import { Params } from '../base/estimator';
export type IsotonicOutOfBounds = 'clip' | 'nan' | 'raise';
export interface IsotonicRegressionProps {
    /** Lower bound on the fitted values (sklearn `y_min`). */
    yMin?: number;
    /** Upper bound on the fitted values (sklearn `y_max`). */
    yMax?: number;
    /**
     * Whether the fitted function is non-decreasing (`true`, default),
     * non-increasing (`false`) or inferred from the sign of the Spearman rank
     * correlation between x and y (`'auto'`, sklearn `check_increasing`).
     */
    increasing?: boolean | 'auto';
    /**
     * How `predict`/`transform` handle inputs outside the training range:
     * 'clip' (default) evaluates at the nearest boundary, 'nan' returns NaN,
     * 'raise' throws.
     */
    outOfBounds?: IsotonicOutOfBounds;
}
/**
 * Pool-adjacent-violators algorithm producing the weighted least-squares
 * non-decreasing fit. Direct port of scikit-learn's
 * `_inplace_contiguous_isotonic_regression` (Best & Chakravarti 1990 active
 * set, O(n)).
 */
export declare function pavaNonDecreasing(y: number[], w: number[]): number[];
/**
 * Isotonic (monotonic) regression, mirroring `sklearn.isotonic.IsotonicRegression`.
 *
 * Fits a free-form, non-decreasing (or non-increasing) step-linear function to
 * 1-D data by weighted pool-adjacent-violators (PAVA), then predicts by linear
 * interpolation between the fitted thresholds.
 *
 * Input shape: the estimator is inherently 1-D, so `fit`, `predict` and
 * `transform` accept either a plain `number[]` or a single-column
 * `number[][]` matrix (so it composes with matrix-based meta-estimators).
 *
 * Matches sklearn:
 *  - ties in x are lexicographically sorted by (x, y) and merged into a single
 *    point whose y is the sample-weight-weighted mean (sklearn `_make_unique`);
 *  - `increasing='auto'` picks the direction from the sign of the Spearman
 *    rank correlation (sklearn `check_increasing`, `rho >= 0` → increasing);
 *  - `yMin`/`yMax` clip the fitted values after PAVA;
 *  - `outOfBounds` = 'clip' | 'nan' | 'raise' as in sklearn.
 *
 * Omitted vs sklearn: the redundant-threshold trimming sklearn applies before
 * building its interpolator (`_build_y` keeps only points where the fit
 * changes slope). We keep every unique-x point; the interpolated predictions
 * are identical. The `increasing='auto'` confidence-interval warning is also
 * omitted.
 */
export declare class IsotonicRegression extends RegressorBase {
    private yMin?;
    private yMax?;
    private increasing;
    private outOfBounds;
    /** Unique ascending x thresholds of the fitted function. */
    private xThresholds;
    /** Fitted values aligned with `xThresholds`. */
    private yThresholds;
    /** Direction actually used at fit time (resolves 'auto'). */
    increasingFitted: boolean | null;
    private fitted;
    constructor(props?: IsotonicRegressionProps);
    getParams(): Params;
    fit(X: number[][] | number[], y: number[], sampleWeight?: number[]): void;
    /** Piecewise-linear interpolation over the fitted thresholds. */
    private interpolate;
    predict(X: number[][] | number[]): number[];
    /** Alias of `predict` (sklearn exposes both). */
    transform(X: number[][] | number[]): number[];
}