Chromatic Discrepancy in Graph Theory

• Chromatic discrepancy is a graph parameter that measures the excess number of colors used in any induced subgraph compared to its chromatic number.
• It captures how a proper coloring may fail to be locally optimal by quantifying the gap between used colors and minimal necessary colors in both general and connected induced subgraphs.
• Important bounds, probabilistic behaviors, and NP-hard computational challenges underscore its practical relevance in graph coloring and optimization.

Searching arXiv for recent and foundational papers on chromatic discrepancy and closely related terminology. Chromatic discrepancy is a graph-coloring parameter that measures how far a proper coloring of a graph is from remaining color-efficient on every induced subgraph. For a proper coloring, one examines each induced subgraph $H$, compares the number of colors actually used on $H$ with its intrinsic chromatic number $\chi(H)$, and records the largest excess. Minimizing this excess over all proper colorings yields the chromatic discrepancy $\phi(G)$; restricting to connected induced subgraphs yields the connected variant $\hat\phi(G)$. These parameters were introduced to study colorings that use as few colors as possible not only on the ambient graph but also on all its induced subgraphs (Aravind et al., 2014). Subsequent work connected $\phi(G)$ to local colorability, forbidden subgraph conditions, exact extremal constructions, random graphs, and several nearby notions that use similar terminology but are formally distinct (Corsini et al., 5 Aug 2025).

1. Definition and basic variants

Let $G=(V,E)$ be a finite simple graph, and let $c$ be a proper coloring of $G$. For any induced subgraph $H$ of $H$0, write $H$1 for the chromatic number of $H$2, and let $H$3 denote the set of colors used by $H$4 on $H$5. The discrepancy of $H$6 is

$H$7

and the chromatic discrepancy of $H$8 is

$H$9

If the maximum is restricted to connected induced subgraphs, one obtains

$\chi(H)$0

By construction, $\chi(H)$1, and both parameters are monotone under taking induced subgraphs (Aravind et al., 2014).

A later formulation rewrites the parameter through an auxiliary extremal function. If $\chi(H)$2 denotes the set of proper $\chi(H)$3-colorings of $\chi(H)$4, define

$\chi(H)$5

Then

$\chi(H)$6

This formulation is useful for lower bounds: an upper bound on $\chi(H)$7 immediately yields a lower bound on $\chi(H)$8 (Corsini et al., 5 Aug 2025).

Conceptually, $\chi(H)$9 means that some proper coloring uses exactly $\phi(G)$0 colors on every induced subgraph $\phi(G)$1. Positive discrepancy measures unavoidable “over-use” of colors somewhere inside the induced-subgraph lattice.

2. General bounds and exact values

Several basic inequalities are known. For any graph $\phi(G)$2 with at least one edge,

$\phi(G)$3

If $\phi(G)$4, $\phi(G)$5, and $\phi(G)$6, then

$\phi(G)$7

The lower-bound side is governed by the gap between chromatic and clique structure: $\phi(G)$8 In particular, if $\phi(G)$9 is triangle-free then

$\hat\phi(G)$0

For triangle-free graphs one also has

$\hat\phi(G)$1

where $\hat\phi(G)$2 is the local chromatic number, and since $\hat\phi(G)$3, this yields $\hat\phi(G)$4 as well (Aravind et al., 2014).

A number of special classes admit exact values. Complete graphs satisfy

$\hat\phi(G)$5

Odd cycles show a small but nontrivial dichotomy: if $\hat\phi(G)$6 is odd and $\hat\phi(G)$7, then

$\hat\phi(G)$8

whereas for odd $\hat\phi(G)$9,

$\phi(G)$0

The Mycielski graphs $\phi(G)$1, with $\phi(G)$2 and $\phi(G)$3, satisfy

$\phi(G)$4

(Aravind et al., 2014).

Later work substantially sharpened the triangle-free lower bound: every triangle-free graph $\phi(G)$5 satisfies

$\phi(G)$6

and this is best possible because the Mycielski graphs already realize equality (Corsini et al., 5 Aug 2025). This replaces the earlier $\phi(G)$7 lower bound by an essentially optimal linear statement.

3. Zero discrepancy and structural characterization

The zero sets of $\phi(G)$8 and $\phi(G)$9 admit exact structural descriptions. One has

$G=(V,E)$0

and

$G=(V,E)$1

The connected variant is thus governed by a perfect-graph condition plus exclusion of the paw, whereas the unrestricted variant is rigid enough to force complete multipartite structure (Aravind et al., 2014).

The proof sketch recorded for $G=(V,E)$2 uses the notion of a perfect coloring, namely a coloring that uses exactly $G=(V,E)$3 colors on every connected induced subgraph $G=(V,E)$4. Such a perfect coloring exists exactly for paw-free perfect graphs. For $G=(V,E)$5, the characterization is equivalent to the statement that complete multipartite graphs are exactly the graphs without an induced $G=(V,E)$6 (Aravind et al., 2014).

The two parameters can differ. A basic example is $G=(V,E)$7: here $G=(V,E)$8, any $G=(V,E)$9-coloring yields $c$0, while the largest connected induced subgraphs always match their chromatic numbers, so $c$1 (Aravind et al., 2014). More generally, if $c$2 is connected, then

$c$3

This shows that for graphs with small independence number the unrestricted and connected discrepancies remain close (Aravind et al., 2014).

These structural results clarify that chromatic discrepancy is not merely a numerical refinement of $c$4: it detects whether a single global coloring can remain locally optimal across all induced subgraphs.

4. Local colorability and forbidden subgraphs

A major recent direction studies chromatic discrepancy under local colorability assumptions. A graph $c$5 is called locally $c$6-colourable if the closed neighbourhood $c$7 of every vertex $c$8 is properly $c$9-colourable. In particular, every triangle-free graph is locally $G$0-colourable (Corsini et al., 5 Aug 2025).

For locally $G$1-colourable graphs, the key bound is

$G$2

Combined with

$G$3

this yields the triangle-free theorem

$G$4

The proof is based on a rainbow closed-neighbourhood lemma: for any proper $G$5-colouring $G$6 and for each colour class $G$7, there exists a set $G$8 with $G$9, meeting $H$0, such that $H$1 contains all $H$2 colours (Corsini et al., 5 Aug 2025).

The same framework leads to a broader conjecture: $H$3 for every locally $H$4-colourable graph $H$5. This conjecture is proved when

$H$6

in the form stated in the abstract, and the paper also proves the partial general bound

$H$7

(Corsini et al., 5 Aug 2025).

Forbidden cycles provide a large class of examples. If $H$8 is $H$9-free, then each $H$00 is $H$01-colourable, so $H$02 is locally $H$03-colourable. Consequently,

$H$04

For $H$05-free graphs there is a stronger statement: $H$06 For $H$07, if $H$08 is $H$09-free, $H$10, and

$H$11

then

$H$12

In the general $H$13-free case, the paper proves that if $H$14, then every ball $H$15 satisfies

$H$16

and from a $H$17-local argument deduces that for $H$18,

$H$19

for some constant $H$20 (Corsini et al., 5 Aug 2025).

This body of results suggests that chromatic discrepancy is especially sensitive to local obstructions: excluding short cycles or forcing low chromaticity in closed neighbourhoods makes large discrepancy unavoidable.

5. Probabilistic behavior, complexity, and open problems

The parameter also has a probabilistic profile. If

$H$21

then there exists a constant $H$22 such that asymptotically almost surely

$H$23

Thus for random graphs in this range, both discrepancy parameters lie very close to the chromatic number itself (Aravind et al., 2014).

From the algorithmic perspective, computing $H$24 and $H$25 is NP-hard. The reduction recorded in the source starts from a connected graph $H$26 on $H$27 vertices, forms $H$28 by adjoining an independent copy of $H$29, and joins each original vertex to two distinct new vertices; one then shows

$H$30

(Aravind et al., 2014). This places exact computation of chromatic discrepancy among the difficult optimization problems of graph coloring.

Several open questions remain explicit in the literature. One asks whether there is a universal lower bound on $H$31 in terms of $H$32; a conjectured form is

$H$33

Other questions ask whether the decision problems $H$34 and $H$35 lie in NP, for which graph classes $H$36 and $H$37 are polynomial-time computable, and whether

$H$38

always holds (Aravind et al., 2014). The later locally $H$39-colourable program adds the unresolved conjecture $H$40 for all $H$41, together with the problem of improving the logarithmic losses in the general bounds (Corsini et al., 5 Aug 2025).

6. Terminological scope and related parameters

The phrase “chromatic discrepancy” is not uniform across the literature. In the graph-coloring sense discussed above, it denotes the induced-subgraph over-use parameter $H$42 and its connected variant $H$43 (Aravind et al., 2014). In discrepancy theory, however, the same phrase is sometimes used for two-color discrepancy of a set system: if $H$44 is a finite set system and $H$45, then

$H$46

(Csikós et al., 2022).

Related geometric usages also appear. For a checkerboard coloring $H$47 constant on unit squares and a rectifiable curve $H$48,

$H$49

measures black-white length imbalance along $H$50 (Kolountzakis et al., 2012). In edge-colored graph settings, if $H$51 is a $H$52-edge-coloring and $H$53 is a spanning tree, then

$H$54

defines the color discrepancy of $H$55 (Chen et al., 7 Nov 2025). For subgraphs of a $H$56-colored complete graph, one likewise writes

$H$57

(Christoph et al., 3 Feb 2026).

These notions are all discrepancy measures, but they are distinct from graph chromatic discrepancy $H$58, which depends on proper vertex colorings and induced subgraphs rather than $H$59-colorings of edges or set-system elements.

There are also nearby graph invariants that should not be conflated with $H$60. The chromatic gap is

$H$61

or equivalently $H$62 on the complement side (Gyárfás et al., 2011). The list-chromatic gap is

$H$63

which measures failure of chromatic-choosability (Zhu et al., 2022). Decomposition parameters such as the tree-chromatic number $H$64 and path-chromatic number $H$65 satisfy

$H$66

and compare chromatic behavior across tree- and path-decompositions rather than across induced subgraphs (Barrera-Cruz et al., 2017). Taken together, these distinctions show that “chromatic discrepancy” sits inside a wider family of chromatic and discrepancy parameters, but its defining feature is the demand that one global proper coloring behave near-optimally on every induced subgraph.