Parallel Linkage Construction Techniques

• Parallel linkage construction is a cross-disciplinary method that concurrently builds linkage objects while preserving domain-specific invariants such as reducibility, rank-gap, and acyclicity.
• It is applied in hierarchical clustering (e.g., ParChain), constant-dimension coding, graph-theoretic 2-linkages, and mechanical as well as genetic linkage mapping to enhance performance.
• This approach enables substantial speedups and reduced memory usage by decomposing complex problems into invariant-preserving substructures suitable for parallel execution.

Parallel linkage construction is a cross-disciplinary term used for several structurally related tasks in which a linkage object is assembled, extended, or reconstructed by simultaneous composition of compatible substructures rather than by a purely sequential procedure. In hierarchical clustering, it denotes parallel construction of a dendrogram or linkage tree; in constant-dimension coding, it denotes blockwise linkage families combined in parallel; in graph theory, it denotes spaces generated by pairs of vertex-disjoint paths; in mechanism theory and knot theory, it denotes parallelized realizations of moving graphs or welded link diagrams (Yu et al., 2021, Heinlein, 2019, Chen et al., 2017, Qi, 2019, Kamada et al., 22 Dec 2025).

1. Scope and common structural themes

Across the cited literature, the phrase does not refer to a single algorithmic template. What remains common is the requirement that parallel composition preserve an exact invariant of the underlying linkage object. In parallel hierarchical agglomerative clustering (HAC), the invariant is a valid exact or certified approximate dendrogram (Yu et al., 2021, Bateni et al., 26 Jul 2025). In constant-dimension coding, it is minimum subspace distance under blockwise code composition (Gluesing-Luerssen et al., 2015, Heinlein, 2019). In the graph-theoretic theory of 2-linkages, it is realizability of endpoint tensors by vertex-disjoint path pairs (Chen et al., 2017). In collision-free L-linkages, it is avoidance of vertex-edge interference under a height assignment (Qi, 2019). In welded-link parallelization, it is welded-link equivalence under doubling of a diagram (Kamada et al., 22 Dec 2025).

The technical consequence is that parallelism is never introduced as mere concurrent execution. Each domain isolates a structural condition that makes simultaneous construction sound. ParChain relies on reducibility of the linkage criterion in HAC (Yu et al., 2021). Parallel linkage constructions for constant-dimension codes rely on rank restrictions, pivot separation, and cross-distance bounds (Heinlein, 2019, He, 2019). The space of 2-linkages is controlled by nearly planar cyclic-order obstructions (Chen et al., 2017). Collision-free L-linkages use the partition condition on a collision graph (Qi, 2019). This suggests a unifying viewpoint: parallel linkage construction is fundamentally an invariant-preserving decomposition problem rather than a generic scheduling problem.

2. Hierarchical clustering and parallel dendrogram construction

In exact HAC, parallel linkage construction means constructing the full dendrogram without reducing the task to independent distance evaluations. The most explicit formulation is ParChain, a shared-memory framework for exact HAC under reducible linkage criteria such as complete linkage, average linkage, and Ward’s linkage (Yu et al., 2021). The framework parallelizes the nearest-neighbor chain algorithm by maintaining many chains at once and merging all current reciprocal nearest-neighbor pairs in the same round. Its correctness hinge is the reducibility condition

$$(A, B) < (A \cup B, C) \quad \text{when } (A,B)<(A,C) \text{ or } (A,B)<(B,C),$$

which implies that a cluster’s nearest-neighbor distance never decreases unless its nearest neighbor itself participates in a merge. Operationally, this allows nonterminal chain segments to remain valid across unrelated merges. The framework stores only active clusters, chain pointers, a dendrogram tree, geometric search structures, and an optional $O(ns)$ cache, avoiding the $\Theta(n^2)$ distance matrix used by most prior exact parallel methods. The implementation combines parallel chain extension, priority concurrent writes $WriteMin/WriteMax$, $k$d-tree range queries, and bounded distance caching, and reports $5.8$–$110.1\times$ speedup over prior parallel HAC implementations, $13.75$–$54.23\times$ self-relative speedup, and up to $237.3\times$ less memory on 48 cores with hyper-threading (Yu et al., 2021).

The exact parallelizability of HAC is, however, linkage-dependent. For average linkage in the sparse-graph model, the picture is sharply negative: exact average-linkage HAC is CC-hard even on trees of diameter $O(ns)$0 (Bateni et al., 2024). The same paper proves positive results only on special structures, notably paths, where average linkage is in NC with

$O(ns)$1

assuming polynomial aspect ratio, and shows that a nearest-neighbor-chain implementation runs in $O(ns)$2 time when the output dendrogram has height $O(ns)$3 (Bateni et al., 2024). The resulting controversy is therefore not whether average linkage is useful, but whether exact average-linkage linkage construction admits broad parallelization at all; the evidence in the graph setting is strongly negative outside special cases.

For single linkage, the dominant reduction is to minimum spanning trees. cuSLINK reformulates exact single-linkage HAC on the GPU as sparse graph construction, MST/MSF computation, and post hoc dendrogram relabeling, using $O(ns)$4 space and $O(ns)$5 time (Nolet et al., 2023). PANDORA goes further at the dendrogram-construction stage: starting from an MST, it recursively contracts non-branching edges, constructs a dendrogram on the contracted tree, and reconstructs full chains, obtaining $O(ns)$6 work independent of dendrogram skewness (Sao et al., 2024). The GPU implementation reports $O(ns)$7 speed-up on AMD GPUs and $O(ns)$8 on NVIDIA GPUs over multithreaded PANDORA, and yields up to a 6-fold speedup for HDBSCAN when dendrogram construction had previously remained on the CPU (Sao et al., 2024). In the fully dynamic setting, explicit single-linkage dendrogram maintenance is possible with update costs asymptotically smaller than recomputation: insertions in $O(ns)$9, deletions in $\Theta(n^2)$0, and parallel versions with $\Theta(n^2)$1 work and $\Theta(n^2)$2 depth, where $\Theta(n^2)$3 is dendrogram height (Man et al., 23 Jun 2025).

Approximate HAC introduces a different notion of parallel linkage construction. For non-monotone geometric linkages, including centroid and Ward’s linkage, the central structural result is that any constant-approximate HAC in sufficiently low dimension has shallow height (Bateni et al., 26 Jul 2025). The paper formalizes a class of “well-behaved” linkages through packability, an approximate triangle inequality, weight-stability, average-reducibility, and poly-bounded diameter, and proves that centroid and Ward’s linkage satisfy these conditions. This yields NC-style $\Theta(n^2)$4-approximate algorithms in low-dimensional Euclidean space, including $\Theta(n^2)$5 expected work and $\Theta(n^2)$6 depth w.h.p. for centroid linkage when $\Theta(n^2)$7, and the analogous bound for Ward’s linkage when

$\Theta(n^2)$8

under polynomial aspect ratio (Bateni et al., 26 Jul 2025). The same paper shows that $\Theta(n^2)$9-approximate promise decision HAC with centroid linkage is CC-hard in $WriteMin/WriteMax$0, so low dimension is not an incidental assumption but part of the parallelizability boundary (Bateni et al., 26 Jul 2025).

3. Parallel linkage constructions for constant-dimension codes

In coding theory, linkage construction is a method for building longer constant-dimension codes from shorter constituent codes and rank-metric codes. The basic linkage theorem constructs

$WriteMin/WriteMax$1

from two constituent constant-dimension codes and a rank-metric code, with resulting cardinality

$WriteMin/WriteMax$2

and minimum subspace distance

$WriteMin/WriteMax$3

(Gluesing-Luerssen et al., 2015). In a special three-block form,

$WriteMin/WriteMax$4

the associated decoding algorithm reduces to decoding smaller constituent codes, and in the hybrid case the last two constituent decodings can be performed in parallel (Gluesing-Luerssen et al., 2015). Here, then, “parallel linkage” first appears as parallel decodability induced by blockwise construction.

The term becomes explicit in the constant-dimension-code literature with the parallel linkage construction and its generalizations. The original parallel construction, as summarized in Theorem 20 of the 2019 generalization paper, combines two families in the same ambient space: $WriteMin/WriteMax$5 where the second family replaces the zero block of ordinary linkage by a low-rank block $WriteMin/WriteMax$6 coming from a rank-restricted rank-metric code (Heinlein, 2019). The strengthened version, Theorem 21, yields

$WriteMin/WriteMax$7

and the generalized linkage construction of Theorem 25 subsumes both improved linkage and parallel linkage as special cases (Heinlein, 2019). The proof mechanism is a pivot-separation argument formalized in Lemma 23: a rank gap of at least $WriteMin/WriteMax$8 in a block forces subspace distance at least $WriteMin/WriteMax$9. The same paper proves that generalized linkage is strictly stronger than its strengthened parallel linkage on $k$0, giving

$k$1

compared with

$k$2

from the parallel linkage construction (Heinlein, 2019).

A distinct 2019 development studies two parallel versions of linkage construction directly as unions of the two block orientations

$k$3

with a rank restriction

$k$4

controlling cross-distance (He, 2019). The two resulting lower bounds improve earlier linkage-type bounds for several parameters, including $k$5, $k$6, and $k$7 (He, 2019). In these constructions, “parallel” means simultaneous use of the two natural linkage orientations rather than repeated application of one orientation.

More recent work places a multilevel layer on top of a parallel mixed-dimension construction. The generalized bilateral multilevel construction from parallel mixed dimension begins with a code $k$8 produced by a prior parallel mixed-dimension theorem and appends a bilateral family

$k$9

indexed by bilateral identifying vectors (Li et al., 10 Jul 2025). Compatibility with the parallel mixed-dimension families is ensured by the constraints

$5.8$0

and

$5.8$1

The resulting union

$5.8$2

is again an $5.8$3 CDC, and the paper states that its corollaries provide at least 49 new lower bounds (Li et al., 10 Jul 2025). A recurrent misconception in this literature is that “parallel linkage” names one theorem; in fact it denotes a family of blockwise constructions whose common feature is additive enlargement by mutually compatible components.

4. Graph-theoretic 2-linkages and algebraic realizability

In graph theory, a parallel linkage is a pair of vertex-disjoint paths connecting two terminal pairs. For $5.8$4, the paper “On the Space of 2-Linkages” encodes such objects algebraically through

$5.8$5

where $5.8$6 are disjoint oriented paths and $5.8$7 records the ordered endpoints (Chen et al., 2017). The ambient module $5.8$8 consists of tensors satisfying zero row-sum, zero column-sum, and zero diagonal conditions, and the central problem is to determine when

$5.8$9

Under the paper’s $110.1\times$0-connectedness hypothesis, failure of equality has an exact topological characterization: $110.1\times$1 if and only if the graph is nearly planar with a cyclic ordering of $110.1\times$2 in which $110.1\times$3 and $110.1\times$4 cross (Chen et al., 2017). In the obstructed case,

$110.1\times$5

where $110.1\times$6 is the cyclic linking functional associated with the boundary order (Chen et al., 2017). The obstruction is therefore planar and cyclic-order based, not an arbitrary global pathology. This gives “parallel linkage construction” a precise algebraic meaning: the realizable space of two disjoint path linkages is everything allowed by the balancing constraints unless a nearly planar crossing obstruction intervenes.

5. Mechanical and topological parallelizations

In mechanism theory, the relevant object is the L-linkage associated with a moving graph. A moving graph $110.1\times$7 is lifted into a spatial mechanism by representing edges as horizontal bars at distinct heights and vertices as vertical sticks; all revolute axes are therefore parallel (Qi, 2019). An L-model is a pair $110.1\times$8, where $110.1\times$9 is injective. Collision-freeness is defined by the condition that for every collision pair $13.75$0,

$13.75$1

The paper introduces the collision graph $13.75$2, whose vertices are edges of $13.75$3, and proves a sufficient condition for collision-free realization: if the collision graph satisfies the partition condition—its vertex set can be split into two parts whose induced directed subgraphs are acyclic—then the moving graph has a collision-free L-model (Qi, 2019). The theorem is constructive: once such a partition is known, heights are assigned by repeatedly removing minimal vertices in the induced acyclic graphs. The criterion is explicitly sufficient but not necessary, and the paper classifies two classical families sharply: every Dixon-1 moving graph has a collision-free L-model, whereas no Dixon-2 moving graph does (Qi, 2019).

In welded-link theory, the relevant operation is called a parallelization rather than a linkage construction, but it fits the same pattern of duplicating a structure while preserving an equivalence relation. The map

$13.75$4

replaces each strand of a welded-link diagram by a parallel copy, using prescribed local replacements at positive, negative, and virtual crossings (Kamada et al., 22 Dec 2025). The central theorem states that if two diagrams are equivalent as welded links, then their parallel diagrams are also equivalent; thus $13.75$5 and $13.75$6 descend to well-defined maps on welded links (Kamada et al., 22 Dec 2025). The right-hand subdiagram is a copy of the original diagram, while the left-hand subdiagram has only virtual crossings and therefore represents a trivial welded link. Orientation matters: $13.75$7 admits a mod-$13.75$8 Alexander numbering and is checkerboard colorable, whereas $13.75$9 admits an Alexander numbering and is almost classical (Kamada et al., 22 Dec 2025). The paper also derives

$54.23\times$0

from a presentation of the fundamental quandle of $54.23\times$1, showing that the parallelization can be tracked algebraically (Kamada et al., 22 Dec 2025).

6. Linkage map construction in genetics and scalable decomposition

In statistical genetics, linkage construction usually means construction of a genetic linkage map rather than a dendrogram or path linkage. The ASMap package wraps MSTmap to cluster markers into linkage groups and order them efficiently inside R (Taylor et al., 2017). MSTmap builds an edge-weighted complete graph on markers using Hamming distances $54.23\times$2, estimates recombination probabilities by

$54.23\times$3

and uses a threshold derived from Hoeffding’s inequality to split the graph into linkage groups (Taylor et al., 2017). Within a linkage group it uses an MST/TSP-style ordering strategy together with EM-style missing-data imputation and bad-data detection. The paper is explicit that ASMap itself does not provide built-in parallel execution, but its workflow naturally decomposes by linkage group, and the bychr = TRUE mode reconstructs each group independently (Taylor et al., 2017). In this sense, large-scale genetic linkage construction is compatible with external parallelization even when the package is not itself a native parallel implementation.

A different high-dimensional formulation is given by the graphical-model approach implemented in netgwas. Here linkage groups are inferred from a sparse conditional-independence graph estimated through a sparse Gaussian copula or a nonparanormal skeptic model, and ordering is recovered by one-dimensional MDS in inbred populations or reverse Cuthill–McKee bandwidth reduction in outbred populations (Behrouzi et al., 2017). The key object is the precision matrix $54.23\times$4, where zeros encode conditional independence. Tuning is selected by

$54.23\times$5

and the package is designed for high-density diploid and polyploid dosage data (Behrouzi et al., 2017). The paper states that, in simulations, the map construction functions in netgwas were run in parallel (Behrouzi et al., 2017). This is a different sense of parallel linkage construction from the HAC and coding literatures: the central object is a genomic linkage map, and parallelism arises from scalable sparse estimation and decomposition rather than simultaneous blockwise code assembly or reciprocal nearest-neighbor merging.

A recurrent misunderstanding across these genetic papers is to treat linkage-map construction as synonymous with pairwise recombination-threshold methods. Both ASMap and netgwas show broader formulations: MST-based ordering with iterative repair in one case, and multivariate conditional-independence inference in the other (Taylor et al., 2017, Behrouzi et al., 2017).

7. General perspective

The literature does not support a single universal definition of parallel linkage construction. Instead, it supports a family resemblance among methods that build linkage objects through concurrency-compatible decompositions. In clustering, the dominant issues are reducibility, MST equivalence, and the complexity-theoretic limits of exact parallelization (Yu et al., 2021, Bateni et al., 2024, Bateni et al., 26 Jul 2025). In constant-dimension coding, the dominant issues are block placement, pivot-vector separation, and low-rank restrictions that bound cross-intersection (Heinlein, 2019, He, 2019, Li et al., 10 Jul 2025). In graph theory, the dominant issue is whether algebraically admissible endpoint pairings are realizable by disjoint paths, with nearly planar crossing as the exact obstruction (Chen et al., 2017). In mechanism design and welded-link theory, the focus shifts to height assignments and doubled diagrams that preserve collision-freeness or welded equivalence (Qi, 2019, Kamada et al., 22 Dec 2025).

The phrase therefore names a methodological pattern rather than a single formal object. Parallel linkage construction is the simultaneous synthesis of a linkage while retaining the invariant that defines correctness in the relevant domain. Where that invariant is sufficiently rigid—reducibility for HAC, rank-gap for CDCs, acyclicity of collision constraints, or topological compatibility of path pairings—parallel construction is exact and constructive. Where it is not, the literature records sharp boundaries: average linkage is CC-hard even on diameter-4 trees, and approximate centroid HAC is CC-hard in arbitrary dimension (Bateni et al., 2024, Bateni et al., 26 Jul 2025).