Biclique Partition in Graph Theory

• Biclique partition is the decomposition of a graph's edges into disjoint complete bipartite subgraphs, fundamental for understanding graph structure and binary rank.
• The concept connects various parameters and variants—including ordered, local, and fractional partitions—with implications in communication complexity and succinct representations.
• Recent research has optimized the number and weight of bicliques to enhance algorithmic performance and achieve asymptotically tight bounds in both sparse and dense graphs.

Biclique partition is the decomposition of the edge set of a graph into edge-disjoint complete bipartite subgraphs, also called biclique decompositions. The associated parameter, the biclique partition number $\operatorname{bp}(G)$, measures the minimum number of bicliques needed to partition $E(G)$. In the literature, biclique partition sits beside biclique cover, local, ordered, weighted, and fractional variants, and it also admits a matrix formulation in terms of binary rank. As a result, the topic links extremal graph theory, graph decompositions, $0$–$1$ matrix factorization, communication complexity, and succinct graph representations (Pinto, 2013, Ghosal et al., 10 Feb 2025, Cardinal et al., 8 Jun 2026).

1. Definitions and parameter landscape

A biclique in a graph $G=(V,E)$ is a complete bipartite subgraph $K_{A,B}$ with $A,B\subseteq V$, $A\cap B=\varnothing$, and all edges between $A$ and $B$ present. A biclique partition of $E(G)$0 is a family of bicliques whose edge sets are pairwise disjoint and whose union is $E(G)$1. Formally, a decomposition $E(G)$2 satisfies

$E(G)$3

with each $E(G)$4 complete bipartite. The minimum cardinality of such a family is $E(G)$5 (Cardinal et al., 8 Jun 2026, Pinto, 2013).

The corresponding cover parameter $E(G)$6 only requires the union of biclique edge sets to contain $E(G)$7, so overlaps are allowed. Consequently, $E(G)$8 for every graph. The local variants refine this by controlling per-vertex participation: $E(G)$9 and $0$0 are the smallest $0$1 such that there exists a biclique cover or partition in which each vertex belongs to at most $0$2 bicliques (Pinto, 2013).

Several papers optimize not the number of bicliques but a size functional. If $0$3, one may define

$0$4

equivalently the total number of vertex appearances across all bicliques. In the same vein, the load of a vertex $0$5 is the number of bicliques containing $0$6, and the total weight of a biclique partition is

$0$7

These objectives are central in succinct graph representations and downstream algorithms, but they are distinct from minimizing $0$8 itself (Cardinal et al., 8 Jun 2026, Krapivin et al., 14 Nov 2025).

For bipartite graphs, biclique partition has a direct matrix interpretation. If $0$9 is a $1$0–$1$1 matrix, viewed as the bipartite adjacency matrix of $1$2, then the binary rank $1$3 equals $1$4. The fractional biclique partition number is the LP relaxation obtained by assigning nonnegative weights $1$5 to bicliques and imposing

$1$6

while minimizing $1$7 (Ghosal et al., 10 Feb 2025).

2. Classical complete-graph results and global inequalities

The complete graph $1$8 is the classical benchmark. Hansel proved that the biclique cover number satisfies

$1$9

whereas the Graham–Pollak theorem gives the much larger partition value

$G=(V,E)$0

This already shows that exact partitioning can be exponentially more expensive than covering (Babu et al., 7 Jun 2026, Pinto, 2013).

A sharp general relation between cover and partition was established in the form

$G=(V,E)$1

This bound is tight. The extremal construction uses the universal graph $G=(V,E)$2 with vertex set $G=(V,E)$3, where two vertices are adjacent if and only if some coordinate realizes the pair $G=(V,E)$4. For this family, $G=(V,E)$5 and

$G=(V,E)$6

with tightness certified by an adjacency-matrix rank argument (Pinto, 2013).

Complement structure also constrains biclique partition indirectly through the cover number. If $G=(V,E)$7 denotes the number of maximal cliques of the complement, then for every graph

$G=(V,E)$8

For co-chordal graphs this yields

$G=(V,E)$9

Complete graphs show that this inequality can be tight: $K_{A,B}$0 and $K_{A,B}$1, so $K_{A,B}$2 (Lyu et al., 2023).

3. Ordered, local, and fractional variants

One major extension replaces exact one-time coverage by constrained multicovers with orientation conditions. In the ordered biclique partition number of $K_{A,B}$3, denoted $K_{A,B}$4, every edge is covered at least once and at most twice, and if an edge is covered twice then its endpoints appear with opposite orientations across the two bicliques. An explicit construction gives

$K_{A,B}$5

improving a previous $K_{A,B}$6 bound. The construction is explicit and is based on the families $K_{A,B}$7, $K_{A,B}$8, and $K_{A,B}$9 on the vertex set $A,B\subseteq V$0, together with carefully ordered star partitions to enforce the orientation rule (Shigeta et al., 2013).

The 2026 notion of an almost balanced ordered biclique cover of order $A,B\subseteq V$1 further generalizes this. For a family $A,B\subseteq V$2 on $A,B\subseteq V$3, let

$A,B\subseteq V$4

$A,B\subseteq V$5

The conditions are $A,B\subseteq V$6 and $A,B\subseteq V$7 for every edge $A,B\subseteq V$8. The minimum $A,B\subseteq V$9 is denoted $A\cap B=\varnothing$0. Here $A\cap B=\varnothing$1, while $A\cap B=\varnothing$2 is precisely the ordered biclique partition number. For all $A\cap B=\varnothing$3,

$A\cap B=\varnothing$4

and

$A\cap B=\varnothing$5

For even $A\cap B=\varnothing$6, the exponents match asymptotically at $A\cap B=\varnothing$7; for odd $A\cap B=\varnothing$8, the lower and upper exponents are $A\cap B=\varnothing$9 and $A$0, respectively (Babu et al., 7 Jun 2026).

Local parameters behave differently from global ones. Although $A$1 is bounded in terms of $A$2, no analogous statement holds for $A$3 in terms of $A$4. In particular, $A$5 can be arbitrarily large even for graphs with $A$6. More quantitatively, there are graphs with a $A$7-local $A$8-cover for which

$A$9

while any graph with a $B$0-local $B$1-cover satisfies

$B$2

Crown graphs $B$3 are the key comparison family in these bounds (Pinto, 2013).

The fractional viewpoint replaces discrete partitions by LP weights. For the Domino matrix

$B$4

the fractional biclique partition number is

$B$5

Thus the fractional binary rank is sub-multiplicative but not multiplicative under Kronecker product. The same work computes

$B$6

showing that the asymptotic fractional binary rank of the Domino is strictly smaller than its one-shot fractional value (Ghosal et al., 10 Feb 2025).

4. Co-chordal and split graphs

For co-chordal graphs, the complement $B$7 is chordal and carries clique-tree structure. A 2022 construction shows that any co-chordal graph admits a biclique partition of size

$B$8

where $B$9 is the number of maximal cliques of the complement. The proof recursively cuts a clique tree of $E(G)$00, and each cut yields a partitioning biclique. A second heuristic uses LexBFS and moplexes. Their complexities are $E(G)$01 and $E(G)$02, respectively. If $E(G)$03 is chordal and clique vertex irreducible, then the upper bound is exact:

$E(G)$04

For split graphs, the same paper proved

$E(G)$05

(Lyu et al., 2022).

A subsequent note claimed the exact identity

$E(G)$06

for every split graph (Babu et al., 10 Jul 2025). This claim was later disproved. A 2026 counterexample constructs a split graph on $E(G)$07 vertices with clique part $E(G)$08 and independent part $E(G)$09 such that

$E(G)$10

so

$E(G)$11

The same paper gives an infinite family of split graphs with

$E(G)$12

again forcing a gap of $E(G)$13 (Babu et al., 7 Apr 2026).

The negative result does not eliminate exact formulas on subclasses. For unbalanced split graphs, the 2026 paper proves

$E(G)$14

so the identity $E(G)$15 does hold there. For balanced split graphs, the same paper proves structural restrictions on any partition with at most $E(G)$16 bicliques, including the fact that no biclique can be a star centered at a vertex of the independent side and no biclique can have an entire part contained in that side (Babu et al., 7 Apr 2026).

These results make split graphs a useful cautionary case. Clique-tree constructions still yield the universal upper bound $E(G)$17, but exactness depends on finer structure than chordality alone.

5. Size-sensitive decompositions and algorithmic constructions

A different line of work studies biclique decompositions under the size functional

$E(G)$18

rather than the cardinality $E(G)$19. The central structural notion is contiguity with respect to a vertex order $E(G)$20: if every neighborhood is a union of few intervals in $E(G)$21, then the graph admits compact biclique decompositions. Specifically, any $E(G)$22-vertex graph of contiguity $E(G)$23 has a biclique decomposition with at most $E(G)$24 bicliques, and every vertex belongs to at most $E(G)$25 bicliques. If the total contiguity is $E(G)$26, then there exists a decomposition with

$E(G)$27

(Cardinal et al., 8 Jun 2026).

Welzl-type orderings convert neighborhood complexity into contiguity bounds. If the shatter function satisfies $E(G)$28 with $E(G)$29, then the graph admits an order with contiguity $E(G)$30; if $E(G)$31, then the contiguity is $E(G)$32. Combined with the dyadic-interval decomposition, this yields a randomized $E(G)$33-time algorithm computing a biclique decomposition $E(G)$34 with

$E(G)$35

under $E(G)$36 and

$E(G)$37

under $E(G)$38 (Cardinal et al., 8 Jun 2026).

The same paper derives algorithmic consequences from these compact decompositions. If $E(G)$39 has linear neighborhood complexity and $E(G)$40 is its adjacency matrix, then for any real matrix $E(G)$41, the product $E(G)$42 can be computed in time $E(G)$43. If $E(G)$44 has a biclique decomposition of size $E(G)$45 in the $E(G)$46 sense, then the graph state $E(G)$47 can be prepared by a stabilizer circuit of size $E(G)$48. On graphs of linear neighborhood complexity, APSP can be solved in randomized time $E(G)$49 (Cardinal et al., 8 Jun 2026).

The weight/load perspective of biclique partition gives an asymptotically optimal worst-case representation theorem. Every graph admits a biclique partition with

$E(G)$50

equivalently

$E(G)$51

For fixed edge density $E(G)$52, there is a density-aware refinement

$E(G)$53

and both bounds are asymptotically tight. The constructions run in $E(G)$54 time in the worst case and in $E(G)$55 time in the density-aware setting (Krapivin et al., 14 Nov 2025).

These results do not minimize $E(G)$56, but they show that biclique partitions can be information-theoretically optimal representations when the objective is total weight rather than biclique count.

6. Random graphs and broader connections

For Erdős–Rényi random graphs, the asymptotic behavior of $E(G)$57 exhibits a threshold phenomenon. Let $E(G)$58 be the unique real root of

$E(G)$59

numerically $E(G)$60. If $E(G)$61 is constant, or if $E(G)$62 with $E(G)$63, then

$E(G)$64

If $E(G)$65 is constant, then

$E(G)$66

The proof uses special subgraphs: $E(G)$67 if and only if $E(G)$68 contains a special subgraph of order $E(G)$69 (Bohman et al., 2022).

Ordered biclique partitions have direct consequences in communication complexity. From an ordered biclique partition of $E(G)$70 of size $E(G)$71, one obtains an $E(G)$72 $E(G)$73–$E(G)$74 matrix $E(G)$75 with $E(G)$76 and $E(G)$77, hence

$E(G)$78

The same construction yields an infinite family of graphs on $E(G)$79 vertices for which the nondeterministic communication complexity of Clique vs. Independent Set is at least

$E(G)$80

This matches the best known lower bound on the deterministic version quoted in that work (Shigeta et al., 2013).

The matrix viewpoint extends further. Biclique partition equals binary rank, biclique cover equals Boolean rank, and fractional biclique partition admits dual certificates through edge weights. This connects graph decomposition to factorization, fooling-set bounds, and Kronecker-product behavior (Ghosal et al., 10 Feb 2025). On the algorithmic side, compact biclique decompositions support matrix multiplication, quantum circuit synthesis, and shortest path algorithms in classes of low neighborhood complexity (Cardinal et al., 8 Jun 2026).

A broader extension replaces graphs by $E(G)$81-uniform hypergraphs and bicliques by complete $E(G)$82-partite hypergraphs. In the $E(G)$83 case, the resulting per-vertex load bound specializes to the graph statement

$E(G)$84

which is presented there as answering a question of Chung, Erdős, and Spencer (Krapivin et al., 14 Nov 2025).

Taken together, these developments place biclique partition at the intersection of exact edge decomposition, structural graph theory, randomized graph asymptotics, and low-rank representations. The complete graph remains the canonical extremal case, but recent work shows that once one varies the notion of partition—ordered, local, fractional, weighted, or structured by vertex orderings—the parameter landscape becomes substantially richer.