---
title: Improper Partition Matrices
url: https://www.emergentmind.com/topics/improper-partition-matrices
type: topic
---

# Improper Partition Matrices

Searching arXiv for the cited papers and related usages of “improper partition matrices” to ground the article in the literature.
“Improper partition matrices” is not a single standardized notion across the arXiv literature. The phrase is used explicitly in enumerative combinatorics for a parity-restricted subclass of partition matrices, while closely related work uses it informally for graph matrix partitions that allow edges inside parts, and for partition-regular matrices that fall outside familiar structural criteria. The common thread is a departure from a canonical “proper” regime, but the underlying objects, equivalence notions, and proof methods differ substantially [2508.21318].

## 1. Terminological scope

The literature separates into several distinct frameworks. In the combinatorics of set-partition matrices, “improper partition matrices” is an explicit term. In graph matrix partitions and Ramsey-theoretic partition regularity, by contrast, the phrase is interpretive rather than standard; the papers define the relevant classes precisely but do not adopt a uniform global terminology for “improperness” [1006.1312].

| Framework | Underlying object | Sense of “improper” |
|---|---|---|
| Partition matrices on $[n]$ | Upper-triangular set-valued matrices | Explicit parity condition on ascents/descents |
| $M$-partitions of graphs | Symmetric $M\in\{0,1,\ast\}^{m\times m}$ | Allowing clique or unconstrained diagonal parts |
| Partition regular matrices | Finite or infinite rational matrices | Interpreted as violating customary structural criteria |

In the 2010 theory of partition matrices, a partition matrix on a finite set $X$ is an upper triangular matrix over the powerset of $X$ such that every row and column contains a non-empty set, the non-empty entries partition $X$, and the column-ordering condition $\operatorname{col}(i)<\operatorname{col}(j)\Rightarrow i<j$ holds. That paper does not define “improper partition matrices,” but it explicitly contrasts partition matrices with composition matrices, which satisfy the first two axioms but not the ordering condition. A plausible interpretation in that framework is therefore: matrices satisfying the base axioms but violating the ordering constraint [1006.1312].

By contrast, the 2025 paper “Signed counting of partition matrices” introduces improper partition matrices as a specific subclass of partition matrices defined by parity restrictions on local ascents and descents, and derives signed-enumerative and lattice-path correspondences from that definition [2508.21318].

## 2. Graph matrix partitions and improper colorings

In graph theory, a symmetric matrix $M\in\{0,1,\ast\}^{m\times m}$ defines an $M$-partition problem: for a graph $G=(V,E)$, one asks whether $V$ can be partitioned as $V=V_1\cup\cdots\cup V_m$ so that for all $i,j$ and all $x\in V_i$, $y\in V_j$, the implications
$$
M_{ij}=1 \Rightarrow xy\in E(G), \qquad
M_{ij}=0 \Rightarrow xy\notin E(G)
$$
hold, while $M_{ij}=\ast$ imposes no restriction. The case $i=j$ is included, so $M_{ii}=1$ forces $V_i$ to induce a clique, $M_{ii}=0$ forces $V_i$ to be an independent set, and $M_{ii}=\ast$ imposes no internal constraint. This framework generalizes proper graph coloring and graph homomorphisms [1403.3548].

Within this setting, “improper” refers to allowing edges inside parts. Choosing $M_{ii}=1$ gives clique parts; choosing $M_{ii}=\ast$ allows arbitrary internal structure. The paper assumes no $\ast$ on the diagonal, since otherwise all graphs have an $M$-partition unless additional constraints such as nonempty parts are imposed. Proper colorings correspond to diagonal zeroes, while improper colorings in the sense of allowing monochromatic edges are represented by taking $M_{ii}=1$ for the relevant colors [1403.3548].

A graph with no $M$-partition but whose every proper induced subgraph does have an $M$-partition is a minimal obstruction. For friendly matrices, defined by the absence of the $2\times2$ principal submatrices
$$
\begin{pmatrix}0&\ast\\ \ast&0\end{pmatrix},
\qquad
\begin{pmatrix}1&\ast\\ \ast&1\end{pmatrix},
$$
Feder, Hell, and Xie had shown that friendly matrices can have finitely many minimal obstructions, infinitely many minimal obstructions, or NP-complete partition problems. The 2014 paper proves that almost all friendly matrices have infinitely many minimal obstructions and an NP-complete partition problem. More precisely, if $M$ is drawn uniformly from friendly $2n\times2n$ matrices with $n$ zeroes and $n$ ones on the diagonal, then
$$
P(M\text{ has infinitely many minimal obstructions})\to1
$$
and
$$
P(M\text{-partition is NP-complete})\to1
$$
as $n\to\infty$; moreover, for the unrestricted random type model, almost all matrices have NP-complete partition problems [1403.3548].

The proof operates through the equivalent language of types: red vertices encode diagonal $0$, blue vertices encode diagonal $1$, and red/blue/green edges encode $0/1/\ast$. Two structural ingredients are central. First, random types satisfy a fixed-point phenomenon for edge-homomorphisms of large subtypes. Second, common-neighborhood sets
$$
N(A)=\{\,v\in V(\tau)\setminus A:\text{ there do not exist }r,b\in A\text{ with }vr\text{ red and }vb\text{ blue}\,\}
$$
admit sharp lower and upper bounds that support obstruction gadgets and NP-hardness reductions. The resulting picture is that, in the friendly regime with both clique and independent-set parts, “almost all” large matrices lie on the intractable and infinite-obstruction side [1403.3548].

A restricted but fully classified example appears for chordal graphs. For
$$
M_1=
\begin{pmatrix}
0&\ast&\ast\\
\ast&0&1\\
\ast&1&0
\end{pmatrix},
$$
an $M_1$-partition requires $V_1,V_2,V_3$ to be independent sets, with $V_2$ and $V_3$ completely joined and no restrictions between $V_1$ and the other parts. The complete family of chordal minimal obstructions consists of seven fixed graphs $F_1,\dots,F_7$ together with an infinite family $\mathcal F_1$ obtained from an odd chordless path of length at least $5$ by adjoining a vertex adjacent to all internal path vertices and non-adjacent to the endpoints. A chordal graph admits an $M_1$-partition if and only if it is $\mathcal F$-free [2004.01229].

## 3. Improper partition matrices in enumerative combinatorics

The explicit modern definition appears in “Signed counting of partition matrices.” A partition matrix $P$ on $[n]$ is an upper-triangular square matrix whose entries are subsets of $[n]$, with each row and column containing at least one nonempty subset, the nonempty subsets partitioning $[n]$, and the column monotonicity condition $\operatorname{col}(i)<\operatorname{col}(j)\Rightarrow i<j$. On such a matrix, for $1<i<n$, a descent at $i$ occurs if $\operatorname{col}(i)=\operatorname{col}(i+1)$ and $\operatorname{row}(i)>\operatorname{row}(i+1)$; an ascent occurs if $\operatorname{col}(i)=\operatorname{col}(i+1)$ and $\operatorname{row}(i)<\operatorname{row}(i+1)$. Writing $c_{\min}$ for the minimal element in the column of $i$, a descent or ascent is proper if $i\equiv c_{\min}\pmod 2$ and improper otherwise. An improper partition matrix is then a partition matrix in which every descent and every ascent, if any, is improper [2508.21318].

This notion is tied to the inversion statistic. A pair $(i,j)$ is an inversion in $P$ if $i>j$, $\operatorname{col}(i)=\operatorname{col}(j)$, and $\operatorname{row}(i)<\operatorname{row}(j)$; let $\operatorname{inv}(P)$ be the number of such pairs. The paper defines a sign-reversing involution $\Theta$ on all partition matrices: if $P$ is not improper, $\Theta$ swaps the labels of the smallest proper ascent or descent $i$ and $i+1$, and one has
$$
|\operatorname{inv}(P)-\operatorname{inv}(\Theta(P))|=1.
$$
The fixed points of $\Theta$ are precisely the improper partition matrices, and every improper partition matrix has even inversion number. Consequently,
$$
S_n(-1)=\sum_{P\in PM_n}(-1)^{\operatorname{inv}(P)}=|IPPM_n|.
$$
The main theorem sharpens this by identifying the signed count with inversion-sequence avoidance:
$$
S_n(-1)=\sum_{P\in PM_n}(-1)^{\operatorname{inv}(P)}=|I_n(-,-,=)|,
$$
where $I_n(-,-,=)$ is the set of inversion sequences of length $n$ with no triple $i<j<k$ satisfying $e_i=e_k$ [2508.21318].

The same paper isolates a nondecreasing subclass. A partition matrix is nondecreasing if $i<j$ implies $\operatorname{row}_P(i)\le \operatorname{row}_P(j)$ and $\operatorname{col}_P(i)\le \operatorname{col}_P(j)$. The subset $NDIPPM_n$ of nondecreasing improper partition matrices is equinumerous with Motzkin paths of length $n$, and the pair of statistics $(\operatorname{blk},\operatorname{odd})$ on $NDIPPM_n$ is equidistributed with $(\operatorname{comp},\operatorname{level})$ on Motzkin paths. In particular,
$$
|NDIPPM_n|=|M_n|,
$$
the $n$th Motzkin number [2508.21318].

This usage should be distinguished from the earlier 2010 partition-matrix theory. There, partition matrices are in bijection with inversion tables, row-ordered partition matrices correspond to non-decreasing inversion tables and are counted by Catalan and Narayana numbers, and composition matrices correspond to $(2+2)$-free posets. Since that paper does not use the adjective “improper,” any such terminology in its context is necessarily interpretive rather than canonical [1006.1312].

## 4. Partition regular matrices: standard terminology and informal “improper” readings

In Ramsey theory, the established terms are image partition regular (IPR) and kernel partition regular (KPR) matrices, not “improper partition matrices.” A finite or infinite matrix $A$ is image partition regular over $\mathbb N$ if whenever $\mathbb N$ is finitely colored, one can find $\vec x$ such that all entries of $A\vec x$ lie in a single color class. Centrally image partition regular (CIPR) matrices strengthen this by demanding monochromatic images inside every central set in $(\mathbb N,+)$, with the algebra of $\beta\mathbb N$ providing the basic machinery [1703.05531].

A principal closure result concerns diagonal sums. If $A$ is a subtracted centrally image partition regular matrix and $B$ is a Milliken–Taylor matrix, then
$$
A\oplus B=
\begin{pmatrix}
A&0\\
0&B
\end{pmatrix}
$$
is image partition regular over $\mathbb N$; the same conclusion holds when $A$ is subtracted segmented image partition regular. The proof uses the sets
$$
I(A)=\{\,p\in\beta\mathbb N:\forall P\in p\ \exists \vec x\text{ with }A\vec x\in P^u\,\},
$$
together with additive and multiplicative ideal properties in $\beta\mathbb N$, and combines these with Milliken–Taylor structure. The paper also proves that if $A$ is finite IPR and $B$ is any infinite IPR matrix, then $A\oplus B$ is IPR [1703.05531].

The same framework is extended from “near zero” to “near an idempotent” in an arbitrary Hausdorff semitopological semigroup. For an idempotent $e$, a matrix is image partition regular over $S$ near $e$ if monochromatic images can be forced inside arbitrary neighborhoods of $e$. For finite matrices, three conditions are equivalent: classical IPR over $S$, IPR over $S$ near $e$, and the property that every central set near $e$ contains an image of the matrix. For infinite systems, insertion matrices built from finite IPR blocks and Milliken–Taylor matrices are IPR near $e$, and segmented image partition regular matrices are centrally IPR near $e$ under the stated IP$^\ast$ hypothesis on scalar multiples $cS$ [1409.5945].

A different dual perspective appears over $\mathbb Q$. For any rational matrix $B$, there exists a matrix $C$ such that
$$
Bx=0 \iff Cx=x,
$$
so $C$ is an idempotent projection with range equal to the kernel of $B$. This yields an exact duality: $B$ is KPR over a nontrivial subsemigroup $S\le \mathbb Q$ if and only if $C$ is IPR over $S$, equivalently weakly IPR over $S$. Conversely, for any rational matrix $A$, one can build a matrix $B$ whose monochromatic kernel vectors are exactly the monochromatic images of $A$ in $(\mathbb Q\setminus\{0\})^u$ [1609.03225].

Taken together, these results suggest that in Ramsey theory “improperness” is best understood as an informal label for matrices that lie outside the cleanest finite templates—such as diagonal-sum closure for arbitrary infinite IPR matrices, exact finite columns criteria, or strict image-versus-kernel separations—rather than as a standard technical term.

## 5. Infinite partition regularity beyond the columns property and maximality phenomena

A stronger informal sense of “improper” appears in the study of infinite kernel partition regular matrices that fail the columns property. For a matrix $A$ with rational entries, partition regularity in the kernel sense means that every finite coloring of $\mathbb N$ admits a monochromatic vector $x$ with $Ax=0$. Rado’s theorem characterizes finite partition regular matrices by the columns property, but the infinite case is more subtle. The paper “Partition regularity without the columns property” constructs an infinite system
$$
2x_n+x_{2^n}+x_{2^n+1}+\cdots+x_{2^{n+1}-1}=y_n \qquad (n\in\omega)
$$
whose coefficient matrix is $[A\mid -I]$, with entries in $\{-1,0,1,2\}$ and bounded column support, and proves that it is partition regular over $\mathbb N$ even though no non-empty subset of columns sums to the zero vector. Thus the matrix fails even the first requirement of the columns property [1401.1377].

The proof replaces columns-based arguments with ultrafilter algebra and central-set largeness. A key lemma states that if $C\subseteq\mathbb N$ is central, then there exist $m,K\in\mathbb N$ such that for all $k\ge K$,
$$
m\mathbb N\subseteq C-kC.
$$
This absorption property allows an inductive construction of monochromatic solutions inside a central set. The same paper conjectures that bounded row sums may restore a columns-style necessity: if an $\omega\times\omega$ rational matrix is partition regular and has bounded row sums, then it should satisfy an infinite version of the columns property. The proved first step is that bounded row sums already force the existence of a non-empty set of columns summing to zero [1401.1377].

A related extension problem arises on the image side. In “Maximality of Infinite Partition Regular Matrices,” a matrix is maximal IPR if no new finitely supported row can be added while preserving IPR. The paper also defines image domination and image maximality: $A$ image dominates $B$ if every image set of $B$ is contained in some image set of $A$. Within this theory, the Finite Sums matrix $F$ is not image maximal, but it is maximal among rapidly IPR integer matrices; the DH-matrix $D$ is finitely image maximal; and no universally image maximal matrix is known. Throughout, “improper” solutions are understood in the standard IPR sense that variables need not be distinct unless a stronger notion such as strongly centrally IPR is imposed [1408.2429].

## 6. Comparative perspective and open directions

Across these literatures, “improperness” consistently marks the failure of a preferred normal form. In graph matrix partitions, it is the failure of strictly independent color classes, replaced by clique or unrestricted diagonal parts. In partition-matrix enumeration, it is the failure of parity-compatible local ascents and descents. In Ramsey theory, it is the failure of standard finite criteria such as the columns property, or the need to move from classical partition regularity to central, near-idempotent, or translated infinite systems. This suggests a unifying editorial description: improper partition matrices are matrices whose partition-theoretic behavior is controlled by a relaxation of the regularity constraints that define the canonical model in their respective frameworks.

The open problems are correspondingly framework-specific. For graph $M$-partitions, it remains unknown whether every fixed $M$-partition problem is either polynomial-time solvable or NP-complete, and a full characterization of which friendly matrices have finitely many minimal obstructions or polynomial-time algorithms is still open [1403.3548]. For enumerative improper partition matrices, open directions include a direct bijective proof of the equidistribution with inversion sequences and further analysis of the polynomials $S_n(q)$ [2508.21318]. In the diagonal-sum and central-set theory of IPR matrices, the paper poses questions on simultaneous left/right images, common-vector products of images, and avoidance partitions for product-of-sums configurations [1703.05531]. For infinite kernel partition regularity, the bounded-row-sums conjecture remains a central unresolved problem [1401.1377].

The resulting encyclopedia picture is therefore plural rather than singular. “Improper partition matrices” denotes a precise parity-defined class in one active combinatorial line of work, while in graph partition theory and Ramsey theory it functions more as a descriptive label for matrices whose partition behavior departs from the standard proper regime. The term is best understood through its local framework, not as a universally fixed definition.

Source: https://www.emergentmind.com/topics/improper-partition-matrices