/**
 * Shared core for the plain (non-averaged) stochastic gradient descent
 * estimators (SGDClassifier, SGDRegressor, Perceptron).
 *
 * Faithful port of scikit-learn's `_plain_sgd` (sklearn/linear_model/_sgd_fast.pyx):
 *  - per-sample updates over seeded-shuffled epochs;
 *  - learning-rate schedules `constant`, `optimal` (Léon Bottou's t0
 *    heuristic, exactly as in sklearn's `_init_t`), `invscaling`, `adaptive`;
 *  - L2 regularization by per-step weight scaling (intercept unregularized);
 *  - L1 / elastic-net via the cumulative-penalty trick (Tsuruoka et al. 2009);
 *  - convergence: stop when epoch loss > best - tol for nIterNoChange
 *    consecutive epochs (for `adaptive`, divide eta by 5 instead until
 *    eta <= 1e-6).
 */
export type SGDPenalty = 'l2' | 'l1' | 'elasticnet' | null;
export type SGDLearningRate = 'constant' | 'optimal' | 'invscaling' | 'adaptive';
/**
 * A pointwise loss. `p` is the model output (decision value / prediction),
 * `y` the target (±1 for classification losses, real for regression losses).
 * `dloss` is the (sub)gradient of the loss with respect to `p`.
 */
export interface SGDLoss {
    loss(p: number, y: number): number;
    dloss(p: number, y: number): number;
}
export type ClassificationLossName = 'hinge' | 'logLoss' | 'modifiedHuber' | 'squaredHinge' | 'perceptron';
export declare function getClassificationLoss(name: ClassificationLossName): SGDLoss;
export type RegressionLossName = 'squaredError' | 'huber' | 'epsilonInsensitive' | 'squaredEpsilonInsensitive';
export declare function getRegressionLoss(name: RegressionLossName, epsilon: number): SGDLoss;
export interface PlainSGDConfig {
    X: number[][];
    /** Targets: ±1 for classification losses, real values for regression. */
    y: number[];
    loss: SGDLoss;
    penalty: SGDPenalty;
    alpha: number;
    l1Ratio: number;
    fitIntercept: boolean;
    maxIter: number;
    /** Pass null to disable the convergence check (always run maxIter epochs). */
    tol: number | null;
    shuffle: boolean;
    /** Fit-local RNG in [0, 1). */
    rand: () => number;
    learningRate: SGDLearningRate;
    eta0: number;
    powerT: number;
    nIterNoChange: number;
}
export interface PlainSGDResult {
    weights: number[];
    intercept: number;
    /** Number of epochs actually run. */
    nIter: number;
}
/** In-place Fisher-Yates shuffle driven by the supplied RNG. */
export declare function shuffleIndices(indices: number[], rand: () => number): void;
export declare function plainSGD(config: PlainSGDConfig): PlainSGDResult;
export declare function validateSGDData(X: number[][], y: number[]): void;
export declare function validatePenalty(penalty: unknown): SGDPenalty;
export declare function validateSchedule(learningRate: SGDLearningRate, eta0: number, alpha: number): void;