/**
 * Optimized hierarchical agglomerative clustering using the nearest-neighbor
 * chain algorithm with Lance–Williams distance updates.
 *
 * NN-chain follows reciprocal nearest neighbors until it finds a reducible
 * merge pair, then updates distances in place. For single, complete, average,
 * and Ward linkage this gives guaranteed O(n²) time and O(n²) memory.
 *
 * The distance matrix is stored as a flat Float64Array of size n×n with
 * index-based active tracking (Uint8Array flags). This avoids the
 * O(n²)-per-merge cost of Array.splice and is cache-friendly.
 *
 * Lance–Williams recurrence for the updated distance D(t,k) when merging
 * clusters i and j into t:
 *
 *   Single   : min(D(i,k), D(j,k))
 *   Complete : max(D(i,k), D(j,k))
 *   Average  : (n_i·D(i,k) + n_j·D(j,k)) / (n_i + n_j)
 *   Ward     : sqrt(max(((n_i+n_k)·D(i,k)² + (n_j+n_k)·D(j,k)² - n_k·D(i,j)²)
 *                       / (n_i + n_j + n_k), 0))
 */
export type LinkageCriterion = 'single' | 'complete' | 'average' | 'ward';
export interface MergeRecord {
    /** Lower active slot index of the two merged clusters. */
    cluster_a: number;
    /** Higher active slot index of the two merged clusters. */
    cluster_b: number;
    distance: number;
    new_size: number;
}
/**
 * Emitted merges are sorted by distance before returning, matching the
 * scipy/fastcluster convention for NN-chain. The raw discovery order is not a
 * valid dendrogram order for cutting.
 *
 * @param D Flat n×n distance matrix (Float64Array). Mutated in place.
 */
export declare function nn_chain_cluster(D: Float64Array, n: number, linkage: LinkageCriterion): MergeRecord[];