Biclique Partition Number

• Biclique partition number is a graph parameter defined as the minimum number of complete bipartite subgraphs needed to partition the edge set of a graph exactly once.
• It interacts with related parameters such as biclique cover, local variants, and the binary rank of 0–1 matrices, which are crucial for algorithm design and combinatorial optimization.
• Research on biclique partition numbers spans extremal graph theory, NP-hardness of computation, and connections to algebraic topology, highlighting both structural results and open challenges.

Searching arXiv for papers on biclique partition number and related results. The biclique partition number is a graph parameter that measures the minimum number of complete bipartite subgraphs needed to partition the edge set of a graph. For a graph $G$, a biclique is a complete bipartite subgraph of $G$, and $bp(G)$ denotes the least number of bicliques whose edge-sets partition $E(G)$ exactly once (Pinto, 2013, Lyu et al., 2022). This parameter occupies a central position at the interface of extremal graph theory, graph decompositions, matrix factorization, communication-style rectangle partitions, and, in some settings, algebraic and topological invariants of associated constructions (Ghosal et al., 10 Feb 2025, Civan et al., 2024). Classical results such as the Graham–Pollak theorem identify $bp(K_n)=n-1$ for complete graphs, while recent work has developed upper and lower bounds, asymptotics, algorithmic heuristics, structural exact formulas for special graph classes, random-graph thresholds, and links to regularity, independence complexes, and binary rank (Pinto, 2013, Rohatgi et al., 2020, Bohman et al., 2022, Civan et al., 2024).

1. Definitions and basic variants

Let $G=(V,E)$ be a finite simple graph. A biclique is a complete bipartite subgraph $K_{A,B}\subseteq G$ (Pinto, 2013). In the standard edge-partition formulation, a biclique partition of $G$ is a collection of bicliques whose edge-sets partition $E(G)$, and the biclique partition number $bp(G)$ is the minimum size of such a partition (Pinto, 2013, Lyu et al., 2022). Several papers use the synonymous term biclique decomposition for an edge-partition into bicliques (Cardinal et al., 8 Jun 2026).

A distinct notion, emphasized in recent work on subdivision graphs, is the biclique vertex partition number. There, one partitions the vertex set rather than the edge set: if $G$0 can be partitioned into $G$1 bicliques, then the minimum such $G$2 is again denoted $G$3 in that paper’s notation (Civan et al., 2024). Because both the edge-partition and vertex-partition parameters appear in the literature under the same symbol, disambiguation depends on context. In the edge-partition setting, $G$4 interacts naturally with adjacency matrices, covers, and decompositions of $G$5 (Pinto, 2013, Ghosal et al., 10 Feb 2025). In the vertex-partition setting, it appears in inequalities relating graph decompositions to Castelnuovo–Mumford regularity of subdivision graphs (Civan et al., 2024).

Several standard variants refine the edge-partition problem. The biclique cover number $G$6 is the minimum number of bicliques whose union covers all edges, without requiring disjointness (Pinto, 2013). Local versions constrain how many bicliques may contain a given vertex, leading to the local biclique cover number $G$7 and local biclique partition number $G$8 (Pinto, 2013). A multiplicity-$G$9 generalization, denoted $bp(G)$0, asks for the minimum number of bicliques needed so that every edge is covered exactly $bp(G)$1 times (Rohatgi et al., 2020). For complete graphs, an “almost balanced ordered” extension further constrains the orientation balance with which each edge is covered (Babu et al., 7 Jun 2026).

The parameter also has a matrix-theoretic formulation in the bipartite case. If $bp(G)$2 and $bp(G)$3 is the associated bipartite graph, then the binary rank of $bp(G)$4 equals the biclique partition number of $bp(G)$5, written $bp(G)$6 (Ghosal et al., 10 Feb 2025). This identification is one of the principal reasons biclique partitions recur in matrix factorization and combinatorial optimization.

2. Classical results and foundational bounds

The foundational theorem is the Graham–Pollak theorem: for the complete graph $bp(G)$7,

$bp(G)$8

This appears repeatedly as the canonical exact evaluation of biclique partition number (Pinto, 2013, Babu et al., 10 Jul 2025). It shows that even though $bp(G)$9 is highly symmetric, its edge set cannot be partitioned into fewer than $E(G)$0 bicliques. The theorem has standard proofs via linear algebra, rank arguments, or addressing interpretations in $E(G)$1 (Babu et al., 10 Jul 2025, Babu et al., 7 Jun 2026).

A complementary comparison involves the biclique cover number. For complete graphs,

$E(G)$2

so in general the partition number can be exponentially larger than the cover number (Pinto, 2013). More broadly, if $E(G)$3, then

$E(G)$4

and this bound is best possible (Pinto, 2013). The proof proceeds by embedding $E(G)$5 as an induced subgraph of a universal graph on $E(G)$6, then computing the biclique partition number of that universal graph exactly (Pinto, 2013).

Lower bounds often come from linear algebra. If $E(G)$7 is the adjacency matrix of $E(G)$8, then every biclique has rank $E(G)$9, yielding

$bp(K_n)=n-1$0

via a theorem attributed there to Tverberg (Pinto, 2013). For complete graphs this aligns with the Graham–Pollak phenomenon, though the sharp lower bound requires a more refined argument (Pinto, 2013, Babu et al., 7 Jun 2026).

The following table summarizes several basic quantities and exact values recorded in the literature.

Graph or parameter	Value or bound	Source
$bp(K_n)=n-1$1	$bp(K_n)=n-1$2	(Pinto, 2013)
$bp(K_n)=n-1$3	$bp(K_n)=n-1$4	(Pinto, 2013)
If $bp(K_n)=n-1$5	$bp(K_n)=n-1$6	(Pinto, 2013)
$bp(K_n)=n-1$7	$bp(K_n)=n-1$8	(Pinto, 2013)
$bp(K_n)=n-1$9 for even $G=(V,E)$0	$G=(V,E)$1	(Pinto, 2013)

These results establish a recurring theme: edge partitions into bicliques are substantially more rigid than covers, and exact partition counts tend to encode nontrivial structural information.

3. Algorithmic complexity and constructive methods

Computing biclique partition number is generally hard. For the vertex-partition parameter, the decision problem “$G=(V,E)$2?” is NP-complete in general and remains NP-hard even when $G=(V,E)$3 is bipartite, including restricted subclasses such as chordal bipartite graphs (Civan et al., 2024). For the edge-partition setting, papers state that computing $G=(V,E)$4 or $G=(V,E)$5 is NP-hard (Rohatgi et al., 2020). This computational hardness motivates both structural exact formulas on special classes and heuristic or approximation-style constructions.

For co-chordal graphs $G=(V,E)$6, meaning $G=(V,E)$7 is chordal, a constructive upper bound is available: $G=(V,E)$8 where $G=(V,E)$9 denotes the number of maximal cliques of the complement (Lyu et al., 2022). The proof uses a clique tree of $K_{A,B}\subseteq G$0. Recursively deleting a clique-tree edge $K_{A,B}\subseteq G$1 with middle set $K_{A,B}\subseteq G$2 defines vertex sets $K_{A,B}\subseteq G$3 and $K_{A,B}\subseteq G$4 such that $K_{A,B}\subseteq G$5 is a biclique in $K_{A,B}\subseteq G$6, and the recursion yields exactly $K_{A,B}\subseteq G$7 bicliques (Lyu et al., 2022).

The same paper gives two explicit heuristics. A clique-tree divide-and-conquer heuristic runs in $K_{A,B}\subseteq G$8, while a LexBFS-based heuristic using moplexes runs in

$K_{A,B}\subseteq G$9

Both output a biclique partition of size $G$0 (Lyu et al., 2022). Moreover, if $G$1 is chordal and clique vertex irreducible, then this upper bound is exact: $G$2 This exactness is obtained by combining the constructive upper bound with lower-bound arguments tied to maximal cliques and the Graham–Pollak theorem (Lyu et al., 2022).

A different algorithmic direction appears in work on compact graph representations. There, the objective is not minimizing the number of bicliques, but minimizing the total vertex-incidence size of a biclique decomposition: $G$3 If a graph has contiguity $G$4, then it admits a biclique decomposition with at most $G$5 bicliques, each vertex appearing in at most $G$6 bicliques, hence

$G$7

(Cardinal et al., 8 Jun 2026). More generally, if the total contiguity is $G$8, then a decomposition of size $G$9 exists and can be output in $E(G)$0 time (Cardinal et al., 8 Jun 2026). This does not directly determine $E(G)$1, but it shows how ordered neighborhood structure can force compact biclique decompositions.

For bipartite graphs arising from $E(G)$2-$E(G)$3 matrices, exact optimization is often replaced by linear programming. The fractional biclique partition number

$E(G)$4

relaxes the integer program for $E(G)$5, where $E(G)$6 is the edge-biclique incidence matrix (Ghosal et al., 10 Feb 2025). Since the number of bicliques can be doubly exponential, the paper develops a column-generation approach, with a restricted master problem, a pricing subproblem over weighted bicliques, inductive initialization for Kronecker powers, and a pruning rule that removes columns slack for more than $E(G)$7 consecutive iterations (Ghosal et al., 10 Feb 2025). This is an algorithmic contribution to the biclique-partition problem from the matrix-factorization side.

4. Structural formulas on special graph classes

One major line of work seeks exact formulas for $E(G)$8 on restricted graph classes. For co-chordal graphs, the general upper bound $E(G)$9 is always available (Lyu et al., 2022). For split graphs, the same paper proved

$bp(G)$0

(Lyu et al., 2022). This bound already indicated that split graphs are close to the Graham–Pollak paradigm but may exhibit a one-unit gap.

A 2025 note claimed the exact formula

$bp(G)$1

for every split graph, presenting it as an extension of the Graham–Pollak theorem (Babu et al., 10 Jul 2025). The exposition there divides split graphs into balanced and unbalanced cases and argues that the obstruction is controlled by the number of maximal cliques in the complement (Babu et al., 10 Jul 2025).

However, a subsequent 2026 paper disproved this conjectured exact formula by constructing a split-graph counterexample with

$bp(G)$2

so $bp(G)$3 (Babu et al., 7 Apr 2026). It further constructed an infinite family of balanced split graphs $bp(G)$4 on $bp(G)$5 vertices such that

$bp(G)$6

hence

$bp(G)$7

for all $bp(G)$8 (Babu et al., 7 Apr 2026). In this sense, the earlier split-graph identity does not hold universally.

At the same time, the 2026 paper reports that unbalanced split graphs do satisfy

$bp(G)$9

with $G$00 and $G$01 in that case (Babu et al., 7 Apr 2026). The paper also states that balanced split graphs obey

$G$02

hence $G$03, while $G$04, so the conjectured formula fails by exactly $G$05 in the balanced case (Babu et al., 7 Apr 2026). This suggests that the split-graph behavior is subtler than a single complement-clique-count identity.

Another structural theorem concerns chordal bipartite graphs, but here for the vertex-partition parameter. If $G$06 is chordal bipartite, then for the $G$07-subdivision graph $G$08,

$G$09

where $G$10 denotes the biclique vertex partition number (Civan et al., 2024). The proof uses induction, bisimplicial edges, and a case analysis tracking simultaneous changes in regularity and partition number (Civan et al., 2024). Although this is not the edge-partition parameter, it shows that biclique partition numbers can admit exact formulas tightly connected to graph structure.

5. Generalizations: local, multiple, ordered, and random settings

The local parameters $G$11 and $G$12 reveal that global and local biclique complexity can diverge sharply. In particular, there exist graphs $G$13 with

$G$14

for every $G$15 (Pinto, 2013). Thus no analogue of the exponential upper bound $G$16 holds for local measures. The construction uses graphs $G$17 on vectors in $G$18 with exactly $G$19 asterisks, together with induced crown subgraphs that force large local partition number (Pinto, 2013).

For complete graphs, the multiplicity-$G$20 variant $G$21 was studied in connection with a conjecture of de Caen, Gregory, and Pritikin. For every fixed $G$22,

$G$23

and more precisely, for all sufficiently large $G$24,

$G$25

(Rohatgi et al., 2020). This asymptotically determines the number of bicliques needed to cover each edge of $G$26 exactly $G$27 times. The lower bound generalizes Graham–Pollak by a rank argument, and the upper bound uses combinatorial designs together with auxiliary coverings of small cliques and a final star-padding step (Rohatgi et al., 2020).

An ordered and “almost balanced” extension is given by the parameter $G$28, where each edge of $G$29 is covered between $G$30 and $G$31 times, and the two orientations of each covered edge must differ in multiplicity by at most $G$32 (Babu et al., 7 Jun 2026). For $G$33, this reduces to the usual biclique partition number of $G$34, giving $G$35 (Babu et al., 7 Jun 2026). For $G$36, it recovers the ordered biclique partition number and satisfies

$G$37

for positive constants $G$38 (Babu et al., 7 Jun 2026). More generally, for fixed $G$39,

$G$40

and

$G$41

so for even $G$42 and odd $G$43, the exponent becomes $G$44 in both cases (Babu et al., 7 Jun 2026).

Random graphs exhibit a threshold phenomenon. For $G$45, there is always the trivial upper bound

$G$46

obtained by partitioning edges into stars centered outside a maximum independent set (Bohman et al., 2022). A 2022 paper identifies a critical probability $G$47, the unique root in $G$48 of

$G$49

and proves the following. If $G$50, then with high probability

$G$51

verifying a conjecture of Chung and Peng in this range (Bohman et al., 2022). If $G$52, then there exists $G$53 such that with high probability

$G$54

and in particular

$G$55

(Bohman et al., 2022). The proof is organized around “special subgraphs” whose existence is equivalent to improving upon the $G$56 star-based upper bound (Bohman et al., 2022).

6. Relations to algebra, topology, and matrix rank

Biclique partitions connect naturally to binary rank. For a $G$57-$G$58 matrix $G$59, the identity

$G$60

translates biclique partition questions into exact factorizations $G$61 over $G$62-$G$63 matrices (Ghosal et al., 10 Feb 2025). The integer-program formulation

$G$64

and its fractional relaxation organize the problem as a column-generation LP over bicliques (Ghosal et al., 10 Feb 2025). For the Domino graph

$G$65

the paper computes

$G$66

with corresponding $G$67th roots decreasing to $G$68 at $G$69 (Ghosal et al., 10 Feb 2025). From submultiplicativity and a lower bound via fractional biclique cover number, the asymptotic fractional binary rank satisfies

$G$70

The same work shows that the fractional binary rank is not multiplicative under the Kronecker product: $G$71 but $G$72 (Ghosal et al., 10 Feb 2025).

The biclique partition number also arises indirectly in algebraic topology through graph subdivision. For any graph $G$73, the $G$74-subdivision $G$75 satisfies

$G$76

where $G$77 is the biclique vertex partition number and $G$78 is the Castelnuovo–Mumford regularity (Civan et al., 2024). In fact, for chordal bipartite $G$79,

$G$80

(Civan et al., 2024). The proof uses the fact that if $G$81 is partitioned into bicliques $G$82, then each $G$83 induces a copy of $G$84 inside $G$85, these copies are vertex-disjoint, and their regularities add to yield the lower bound (Civan et al., 2024).

The same paper classifies the topology of the independence complex $G$86 for chordal bipartite $G$87. If $G$88 admits a complete simple bisimplicial elimination sequence of length $G$89, then $G$90 is homotopy equivalent to a sphere of dimension

$G$91

otherwise it is contractible (Civan et al., 2024). Moreover, one can decide which case occurs in polynomial time $G$92 and compute the sphere dimension when appropriate (Civan et al., 2024). This is a notably different role for biclique partitioning: not as an end in itself, but as a parameter controlling regularity and homotopy type.

7. Examples, interpretations, and open directions

A few benchmark examples recur across the literature. For complete bipartite graphs,

$G$93

in the vertex-partition setting, and for chordal bipartite graphs this matches the formula $G$94 since $G$95 (Civan et al., 2024). For paths $G$96, the vertex-partition number satisfies $G$97 by covering with stars of size two, and the regularity of the $G$98-subdivision matches $G$99 (Civan et al., 2024). For the $bp(G)$00-cube $bp(G)$01, the edge-partition and cover numbers coincide: $bp(G)$02 and when $bp(G)$03 is even,

$bp(G)$04

(Pinto, 2013).

In extremal and representational contexts, biclique partitions are often interpreted as addressings or subcube representations. The exponential inequality $bp(G)$05 is proved by encoding vertices as strings in $bp(G)$06, where adjacency corresponds to differing in some coordinate by $bp(G)$07 (Pinto, 2013). The same paper relates $bp(G)$08 to subcube intersection representations, showing that the least dimension $bp(G)$09 in which $bp(G)$10 is an intersection graph of subcubes equals $bp(G)$11 (Pinto, 2013). This suggests that biclique partitions and covers are part of a broader geometry of combinatorial encodings.

Several open problems remain explicit in the cited work. For local measures, one question is whether $bp(G)$12 can be bounded in terms of $bp(G)$13 under the additional condition $bp(G)$14 (Pinto, 2013). For multiplicity covers, it remains open for fixed $bp(G)$15 whether $bp(G)$16 eventually equals $bp(G)$17 exactly, rather than merely $bp(G)$18 (Rohatgi et al., 2020). In the matrix setting, the exact multiplicativity of binary rank and biclique partition number under Kronecker product remains unresolved; the fractional analogue is already known to fail (Ghosal et al., 10 Feb 2025). In random graphs, the behavior near and beyond $bp(G)$19 remains delicate, and the case $bp(G)$20 is not settled by the threshold theorem at $bp(G)$21 (Bohman et al., 2022).

Taken together, these results show that the biclique partition number is not a single narrowly extremal parameter, but a nexus linking decomposition theory, complement structure, randomness, local-versus-global complexity, matrix rank, and even topological invariants of graph-derived complexes. The parameter is easy to define but often difficult to compute, and its exact behavior appears to depend sensitively on whether one partitions edges or vertices, whether multiplicities or locality are allowed, and which structural graph class is under consideration (Civan et al., 2024, Pinto, 2013, Lyu et al., 2022).