/** Options for `ridge`. */
export interface RidgeOptions {
    /** Center X/y and fit an unpenalized intercept (default true). */
    intercept?: boolean;
}
/**
 * Options for `lasso` and `elasticNet`, which add an iteration limit and a tolerance
 * to `RidgeOptions`.
 */
export interface CoordinateDescentOptions extends RidgeOptions {
    /** Maximum coordinate-descent sweeps (default 1000). */
    maxIter?: number;
    /** Convergence tolerance on the max coefficient change per sweep (default 1e-7). */
    tol?: number;
}
/** Result of a regularized regression: the coefficients and the intercept. */
export interface RegularizedRegressionResult {
    /** Fitted coefficients, one per predictor column (intercept excluded). */
    coefficients: number[];
    /** Fitted intercept (0 if `opts.intercept === false`). */
    intercept: number;
}
/**
 * Ridge regression (L2-penalized least squares) with a closed-form solution on
 * centered data: `β = (XᵀX + αI)⁻¹Xᵀy`. The intercept is never penalized.
 *
 * @param X - Design matrix (rows = observations, cols = predictors)
 * @param y - Response vector (length = number of observations)
 * @param alpha - L2 penalty strength (`alpha >= 0`; `alpha = 0` is OLS)
 * @param opts - `intercept` (default true)
 *
 * @example
 * ridge([[1], [2], [3], [4]], [2, 4, 6, 8], 0)
 * // => { coefficients: [2], intercept: 0 } (recovers the exact OLS fit)
 */
export declare function ridge(X: number[][], y: number[], alpha: number, opts?: RidgeOptions): RegularizedRegressionResult;
/**
 * Lasso regression (L1-penalized least squares) via cyclic coordinate descent
 * with soft-thresholding on standardized columns. Unlike ridge, large enough
 * penalties drive coefficients to exactly 0 (sparse solutions). The intercept
 * is never penalized.
 *
 * @param X - Design matrix (rows = observations, cols = predictors)
 * @param y - Response vector (length = number of observations)
 * @param alpha - L1 penalty strength (`alpha >= 0`; `alpha = 0` ~ OLS)
 * @param opts - `intercept` (default true), `maxIter` (default 1000), `tol` (default 1e-7)
 *
 * @example
 * lasso([[1], [2], [3], [4]], [2, 4, 6, 8], 100)
 * // => coefficients[0] === 0 (penalty overwhelms the signal)
 */
export declare function lasso(X: number[][], y: number[], alpha: number, opts?: CoordinateDescentOptions): RegularizedRegressionResult;
/**
 * Elastic-net regression combining an L1 penalty (`alpha * l1Ratio`, soft-
 * thresholded) and an L2 penalty (`alpha * (1 - l1Ratio)`, added to the
 * coordinate-descent denominator) via the same cyclic coordinate descent as
 * `lasso`. `l1Ratio = 1` is pure lasso; `l1Ratio = 0` is (coordinate-descent)
 * ridge. The intercept is never penalized.
 *
 * @param X - Design matrix (rows = observations, cols = predictors)
 * @param y - Response vector (length = number of observations)
 * @param alpha - Overall penalty strength (`alpha >= 0`)
 * @param l1Ratio - Mixing parameter in `[0, 1]` between L1 and L2
 * @param opts - `intercept` (default true), `maxIter` (default 1000), `tol` (default 1e-7)
 *
 * @example
 * elasticNet([[1], [2], [3], [4]], [2, 4, 6, 8], 0.1, 0.5)
 * // => finite coefficients blending ridge shrinkage and lasso sparsity
 */
export declare function elasticNet(X: number[][], y: number[], alpha: number, l1Ratio: number, opts?: CoordinateDescentOptions): RegularizedRegressionResult;
//# sourceMappingURL=regularized-regression.d.ts.map