Bipartiteness Ratio: Spectral & Algorithmic Analysis

• Bipartiteness Ratio is a graph metric that quantifies how far a graph deviates from a perfect bipartite structure by penalizing internal and boundary edges normalized by vertex volume.
• It connects directly to spectral theory since its value is controlled by the top eigenvalue of the normalized Laplacian, serving as a dual Cheeger quantity to standard conductance measures.
• The measure underlies practical algorithms for local subgraph extraction, flow-based approximations, and even extends to graphon models and cut-matching game frameworks.

Searching arXiv for relevant papers on bipartiteness ratio and closely related formulations. arXiv search query: "bipartiteness ratio Trevisan graphons cut matching games bipartiteness". Bipartiteness ratio is a cut-based quantitative measure of how closely a graph, or a selected region of a graph, behaves like a bipartite object. In its standard graph-theoretic form, it penalizes edges internal to the two sides of a proposed bipartition and edges leaving the active region, and normalizes these penalties by incident volume. It is the dual counterpart of conductance: conductance is controlled by the bottom of the normalized Laplacian spectrum, whereas bipartiteness ratio is controlled by the top of that spectrum. The notion has been developed in finite graphs, extended to graphons, and used algorithmically for local subgraph extraction, spectral characterization, and flow-based approximation (Li et al., 2012, Soma et al., 17 Jul 2025, Pokharanakar, 21 Feb 2025).

1. Definitions and combinatorial meaning

In the formulation for an undirected weighted graph $G=(V,E)$ with weighted degree $d(v)$, volume

$vol(S):=\sum_{v\in S} d(v),$

and disjoint sets $L,R\subseteq V$, the bipartiteness ratio is

$$\beta(L,R)=\frac{2e(L)+2e(R)+e(U,\bar{U})}{vol(U)}, \qquad U:=L\cup R.$$

The induced set version is

$$\beta(S):=\min_{(L,R)\,\textrm{partition of }S}\beta(L,R),$$

the whole-graph version is

$\beta(G):=\min_{S\subseteq V}\beta(S),$

and the dense bipartite profile is

$$\beta(k):=\min_{\substack{L,R: L\cap R=\emptyset\ vol(L\cup R)\leq k}}\beta(L,R).$$

These are the definitions used to detect “small dense bipartite-like subgraphs” (Li et al., 2012).

A closely related weighted-vector formulation uses a positive vertex-weight function $b:V\to \mathbb{Z}_{++}$ and a nonzero vector $\mathbf{x}\in \{0,\pm 1\}^V$: $d(v)$0 with

$d(v)$1

Trevisan’s original bipartiteness ratio is recovered by taking

$d(v)$2

so that

$d(v)$3

If $d(v)$4 induces the tripartition

$d(v)$5

then

$d(v)$6

The same paper also records

$d(v)$7

so low bipartiteness ratio means that, on the active induced subgraph $d(v)$8, there is a cut $d(v)$9 that cuts almost all incident edge volume (Soma et al., 17 Jul 2025).

Setting	Quantity	Normalization
Finite graph, pair form	$vol(S):=\sum_{v\in S} d(v),$0	$vol(S):=\sum_{v\in S} d(v),$1
Finite graph, vector form	$vol(S):=\sum_{v\in S} d(v),$2	$vol(S):=\sum_{v\in S} d(v),$3
Graphon	$vol(S):=\sum_{v\in S} d(v),$4	$vol(S):=\sum_{v\in S} d(v),$5

The numerator has a direct combinatorial interpretation. The terms $vol(S):=\sum_{v\in S} d(v),$6 and $vol(S):=\sum_{v\in S} d(v),$7 count edge mass that stays inside the two sides and therefore fails to be cut by the candidate bipartition, while $vol(S):=\sum_{v\in S} d(v),$8 counts edge mass that leaves the active region. Since

$vol(S):=\sum_{v\in S} d(v),$9

the only term not penalized is $L,R\subseteq V$0, the cross-edge mass. This makes bipartiteness ratio the fraction of incident edge volume that does not support a bipartite-like structure (Li et al., 2012).

2. Spectral theory and dual Cheeger structure

Bipartiteness ratio is the “dual Cheeger” quantity associated with the top of the normalized Laplacian spectrum. In Trevisan’s inequality, as used for small dense bipartite-like subgraphs,

$L,R\subseteq V$1

where $L,R\subseteq V$2 is the largest normalized Laplacian eigenvalue under the paper’s indexing. The same work gives a profile statement: $L,R\subseteq V$3 Accordingly, a large $L,R\subseteq V$4-th largest Laplacian eigenvalue forces existence of a low-volume set with good bipartite structure (Li et al., 2012).

The weighted-vector formulation makes the dual-Cheeger connection explicit. For the largest eigenvalue $L,R\subseteq V$5 of the normalized Laplacian, the recalled inequality is

$L,R\subseteq V$6

This is the finite-graph analogue of ordinary Cheeger inequalities, but with the top rather than the bottom of the spectrum (Soma et al., 17 Jul 2025).

A complementary adjacency-matrix statement is classical for $L,R\subseteq V$7-regular graphs: if $L,R\subseteq V$8 is the normalized adjacency matrix, then

$L,R\subseteq V$9

if and only if $$\beta(L,R)=\frac{2e(L)+2e(R)+e(U,\bar{U})}{vol(U)}, \qquad U:=L\cup R.$$0 is bipartite. In the Cayley-graph setting, non-bipartite expansion forces quantitative separation from $$\beta(L,R)=\frac{2e(L)+2e(R)+e(U,\bar{U})}{vol(U)}, \qquad U:=L\cup R.$$1: if $$\beta(L,R)=\frac{2e(L)+2e(R)+e(U,\bar{U})}{vol(U)}, \qquad U:=L\cup R.$$2 is a non-bipartite Cayley graph constructed using a group and a symmetric generating set of size $$\beta(L,R)=\frac{2e(L)+2e(R)+e(U,\bar{U})}{vol(U)}, \qquad U:=L\cup R.$$3, then

$$\beta(L,R)=\frac{2e(L)+2e(R)+e(U,\bar{U})}{vol(U)}, \qquad U:=L\cup R.$$4

for $$\beta(L,R)=\frac{2e(L)+2e(R)+e(U,\bar{U})}{vol(U)}, \qquad U:=L\cup R.$$5 an absolute constant. The same abstract emphasizes that such a result could not be true for general non-bipartite graphs (Moorman et al., 2020).

This spectral viewpoint explains why bipartiteness ratio is a dual notion to conductance. Conductance is organized by eigenvalues near $$\beta(L,R)=\frac{2e(L)+2e(R)+e(U,\bar{U})}{vol(U)}, \qquad U:=L\cup R.$$6, sparse cuts, and low-expansion communities. Bipartiteness ratio is organized by eigenvalues near $$\beta(L,R)=\frac{2e(L)+2e(R)+e(U,\bar{U})}{vol(U)}, \qquad U:=L\cup R.$$7, signed structure, and regions whose incident volume is concentrated on cross edges rather than internal edges (Li et al., 2012, Pokharanakar, 21 Feb 2025).

3. Small dense bipartite-like subgraphs and local extraction

One of the main algorithmic uses of bipartiteness ratio is the detection of small dense bipartite-like subgraphs. The central global guarantee is bicriteria. If there exists a subset $$\beta(L,R)=\frac{2e(L)+2e(R)+e(U,\bar{U})}{vol(U)}, \qquad U:=L\cup R.$$8 with

$$\beta(L,R)=\frac{2e(L)+2e(R)+e(U,\bar{U})}{vol(U)}, \qquad U:=L\cup R.$$9

where $$\beta(S):=\min_{(L,R)\,\textrm{partition of }S}\beta(L,R),$$0 and $$\beta(S):=\min_{(L,R)\,\textrm{partition of }S}\beta(L,R),$$1, then for any $$\beta(S):=\min_{(L,R)\,\textrm{partition of }S}\beta(L,R),$$2, algorithm $$\beta(S):=\min_{(L,R)\,\textrm{partition of }S}\beta(L,R),$$3 finds $$\beta(S):=\min_{(L,R)\,\textrm{partition of }S}\beta(L,R),$$4 with

$$\beta(S):=\min_{(L,R)\,\textrm{partition of }S}\beta(L,R),$$5

and

$$\beta(S):=\min_{(L,R)\,\textrm{partition of }S}\beta(L,R),$$6

The corresponding local guarantee states that if

$$\beta(S):=\min_{(L,R)\,\textrm{partition of }S}\beta(L,R),$$7

where $$\beta(S):=\min_{(L,R)\,\textrm{partition of }S}\beta(L,R),$$8 and $$\beta(S):=\min_{(L,R)\,\textrm{partition of }S}\beta(L,R),$$9, then there exists $\beta(G):=\min_{S\subseteq V}\beta(S),$0 such that

$\beta(G):=\min_{S\subseteq V}\beta(S),$1

and for any $\beta(G):=\min_{S\subseteq V}\beta(S),$2, $\beta(G):=\min_{S\subseteq V}\beta(S),$3 finds a subgraph $\beta(G):=\min_{S\subseteq V}\beta(S),$4 of

$\beta(G):=\min_{S\subseteq V}\beta(S),$5

with

$\beta(G):=\min_{S\subseteq V}\beta(S),$6

in time

$\beta(G):=\min_{S\subseteq V}\beta(S),$7

independent of the size of the graph (Li et al., 2012).

The sweep procedure is a signed analogue of threshold sweeps for conductance. Given a vector $\beta(G):=\min_{S\subseteq V}\beta(S),$8, vertices are ordered by

$\beta(G):=\min_{S\subseteq V}\beta(S),$9

and the sweep sets are

$$\beta(k):=\min_{\substack{L,R: L\cap R=\emptyset\ vol(L\cup R)\leq k}}\beta(L,R).$$0

$$\beta(k):=\min_{\substack{L,R: L\cap R=\emptyset\ vol(L\cup R)\leq k}}\beta(L,R).$$1

$$\beta(k):=\min_{\substack{L,R: L\cap R=\emptyset\ vol(L\cup R)\leq k}}\beta(L,R).$$2

The algorithm sweeps vectors of the form

$$\beta(k):=\min_{\substack{L,R: L\cap R=\emptyset\ vol(L\cup R)\leq k}}\beta(L,R).$$3

where

$$\beta(k):=\min_{\substack{L,R: L\cap R=\emptyset\ vol(L\cup R)\leq k}}\beta(L,R).$$4

is the quasi-Laplacian. This is coupled with the potential

$$\beta(k):=\min_{\substack{L,R: L\cap R=\emptyset\ vol(L\cup R)\leq k}}\beta(L,R).$$5

which plays the role of a Lovász–Simonovits curve. The convergence lemma states that if $$\beta(k):=\min_{\substack{L,R: L\cap R=\emptyset\ vol(L\cup R)\leq k}}\beta(L,R).$$6, then for $$\beta(k):=\min_{\substack{L,R: L\cap R=\emptyset\ vol(L\cup R)\leq k}}\beta(L,R).$$7,

$$\beta(k):=\min_{\substack{L,R: L\cap R=\emptyset\ vol(L\cup R)\leq k}}\beta(L,R).$$8

This yields upper bounds on propagated indicator vectors unless a good low-volume sweep set appears, while a matching lower bound is produced by any planted low-$$\beta(k):=\min_{\substack{L,R: L\cap R=\emptyset\ vol(L\cup R)\leq k}}\beta(L,R).$$9 region (Li et al., 2012).

The same paper also gives a spectral characterization of the dense bipartite profile by the $b:V\to \mathbb{Z}_{++}$0-th largest eigenvalue. If

$b:V\to \mathbb{Z}_{++}$1

then there is a polynomial-time algorithm that, for any $b:V\to \mathbb{Z}_{++}$2, finds $b:V\to \mathbb{Z}_{++}$3 of volume at most

$b:V\to \mathbb{Z}_{++}$4

and bipartiteness ratio

$b:V\to \mathbb{Z}_{++}$5

This places small-set bipartite structure in direct analogy with higher-order small-set expansion, but at the top of the spectrum rather than the bottom (Li et al., 2012).

4. Flow-based approximation and cut-matching games

A later development gives a flow-based $b:V\to \mathbb{Z}_{++}$6-approximation algorithm for bipartiteness ratio in undirected graphs. The key idea is to pass to a doubled auxiliary graph

$b:V\to \mathbb{Z}_{++}$7

where each original edge $b:V\to \mathbb{Z}_{++}$8 becomes the two cross-edges

$b:V\to \mathbb{Z}_{++}$9

In this graph, if $\mathbf{x}\in \{0,\pm 1\}^V$0 induces $\mathbf{x}\in \{0,\pm 1\}^V$1 and one defines

$\mathbf{x}\in \{0,\pm 1\}^V$2

then

$\mathbf{x}\in \{0,\pm 1\}^V$3

Hence

$\mathbf{x}\in \{0,\pm 1\}^V$4

The optimization becomes a constrained cut problem over cuts with skew-symmetric structure (Soma et al., 17 Jul 2025).

The same work introduces a corresponding notion of well-linkedness for skew-symmetric graphs. For a symmetric pair

$\mathbf{x}\in \{0,\pm 1\}^V$5

one builds a network $\mathbf{x}\in \{0,\pm 1\}^V$6 by attaching a super-source and super-sink and giving each edge in $\mathbf{x}\in \{0,\pm 1\}^V$7 capacity $\mathbf{x}\in \{0,\pm 1\}^V$8. The central flow-cut theorem is

$\mathbf{x}\in \{0,\pm 1\}^V$9

This is the structural basis for the cut-matching game.

The algorithmic theorem is a decision-plus-certificate statement: given $d(v)$00 with $d(v)$01 an integer, Algorithm 1 finds either $d(v)$02 with $d(v)$03, or a certificate proving that

$d(v)$04

with probability at least $d(v)$05. By binary search, this yields an $d(v)$06-approximation randomized algorithm for $d(v)$07. The procedure uses

$d(v)$08

single-commodity max-flow computations in the decision phase, and only polylogarithmically many such max-flow computations overall; with current fastest undirected max-flow algorithms, it runs in nearly linear time (Soma et al., 17 Jul 2025).

The same paper gives a max-cut application. Given a graph whose maximum cut deletes a $d(v)$09 fraction of edges, it finds a cut that deletes a

$d(v)$10

fraction of edges in running time $d(v)$11. Conceptually, this uses the identity

$d(v)$12

to turn approximate minimization of bipartiteness ratio into recursive extraction of nearly bipartite active regions (Soma et al., 17 Jul 2025).

5. Graphon generalization

The graphon analogue replaces finite subsets by measurable sets. For a connected graphon $d(v)$13, with

$d(v)$14

and measurable disjoint sets $d(v)$15 with $d(v)$16, the graphon bipartiteness ratio is

$d(v)$17

and

$d(v)$18

This is presented as the exact graphon analogue of Trevisan’s graph bipartiteness ratio (Pokharanakar, 21 Feb 2025).

The graphon paper also gives a variational characterization through $d(v)$19-valued functions: $d(v)$20 For the graphon Laplacian $d(v)$21, the top of the spectrum is

$d(v)$22

and one has the identity

$d(v)$23

From these formulas comes the dual Cheeger–Buser inequality for graphons: $d(v)$24 Thus $d(v)$25 and $d(v)$26 are quantitatively equivalent up to constants (Pokharanakar, 21 Feb 2025).

The extremal theory matches the finite-graph picture. The paper proves the easy bound

$d(v)$27

and shows the bound is sharp by the example of a nonzero constant graphon, for which

$d(v)$28

Under the additional assumption

$d(v)$29

it proves

$d(v)$30

are equivalent. This transfers the usual correspondence “zero bipartiteness ratio $d(v)$31 exact bipartiteness $d(v)$32 top eigenvalue $d(v)$33” to the graphon setting (Pokharanakar, 21 Feb 2025).

6. Alternative notions and neighboring frameworks

A recurring source of confusion is that not every quantitative notion of non-bipartiteness is a bipartiteness ratio. Several recent lines of work use different normalizations and different objectives.

In bounded-degree property testing, one paper states plainly that it does not explicitly use the term bipartiteness ratio. Its primary global quantity is instead

$d(v)$34

where

$d(v)$35

Its most local surrogate is the parity-separation quantity

$d(v)$36

which measures overlap between even- and odd-parity lazy-walk endpoint distributions. The analysis converts small average parity collapse into a high-value Goemans–Williamson SDP certificate and then into Max-Cut value close to $d(v)$37 (Fei et al., 11 Jun 2026).

In the dense-graph model, the dominant quantity is often normalized edit distance to bipartiteness rather than a conductance-style ratio. One definition is

$d(v)$38

with

$d(v)$39

and

$d(v)$40

This is a normalized edge-deletion or cut-deficiency measure, not the standard sparse-graph bipartiteness ratio (Ghosh et al., 2022). An earlier dense-graph testing framework likewise measures $d(v)$41-farness by the minimum number of edge deletions needed to make the graph bipartite, normalized by $d(v)$42, and studies preservation of that quantity under random induced subgraphs (Bogdanov et al., 2010).

In planar and minor-free property testing, $d(v)$43-farness is again deletion-based: $d(v)$44 That theory proceeds through odd-cycle detection by constant-length random walks and does not invoke the standard spectral bipartiteness ratio (Czumaj et al., 2014).

Beyond graphs, one higher-dimensional analogue does not define a formal bipartiteness ratio at all. For simplicial complexes, the relevant spectral-combinatorial notion is disorientability. Under the no-twisting assumption, the closest analogues of a bipartiteness defect are the gap

$d(v)$45

the number of simplex splittings needed to make the complex disorientable, and the number of non-branching simple odd cycles in the down dual graph; the paper is explicit that it does not provide a classical higher-order bipartiteness ratio (Eidi et al., 2024).

These neighboring frameworks are mathematically close to bipartiteness ratio because they all quantify failure of a perfect two-coloring, but they differ in normalization, in whether inactive vertices are allowed, and in whether the objective is local, global, spectral, or edit-distance based. A precise use of the term “bipartiteness ratio” therefore refers most naturally to the Trevisan-style quantity and its graphon analogue, together with the dual-Cheeger spectral theory and the algorithmic frameworks built around them (Li et al., 2012, Pokharanakar, 21 Feb 2025, Soma et al., 17 Jul 2025).