---
title: Triangle-Free Operator Systems
url: https://www.emergentmind.com/topics/triangle-free-operator-systems
type: topic
---

# Triangle-Free Operator Systems

Searching arXiv for the specified papers to ground the article and citations.
arXiv search query: id:2010.11891 OR id:2509.17782
Triangle-free operator systems arise in two related but distinct senses in recent research. In bipartite quantum information, a triangle-free graph \(G\) on \([d]\) determines an operator system \(S_G \subseteq M_d \otimes M_d\) through a prescribed zero pattern on diagonal entries in the product basis, and this structure supports an entanglement-detection method based on local averaging and comparison matrices [2010.11891]. In noncommutative graph theory, an operator system \(\mathcal{V} \subseteq M_n(\mathbb{C})\) is treated as a quantum graph; in that setting, “triangle-free” means the absence of a \(3\)-clique, while “strongly triangle-free” requires that every rank-\(3\) compression contain a rank-\(2\) anticlique [2509.17782]. The subject therefore connects graph-theoretic sparsity, operator-system structure, and separability or clique-anticlique phenomena in matrix algebras.

## 1. Graph-supported operator systems from triangle-free graphs

Let \(G=(V,E)\) be a simple undirected graph on the vertex set \(V=[d]\). It is \(\Delta\)-free, or triangle-free, if it contains no \(3\)-cycle. The associated operator system is
\[
S_G
=
\Bigl\{
X\in M_d\otimes M_d
:
\langle i\,j\,,\,i\,j\rvert\,X\,\lvert i\,j\rangle = 0
\ \text{whenever }i\neq j\text{ and }\{i,j\}\notin E
\Bigr\}.
\]
Equivalently, in the product basis \(\{\lvert i\,j\rangle\}_{i,j=1}^d\), all diagonal matrix entries \(\langle ij|X|ij\rangle\) vanish away from positions \((i,j)\) corresponding either to an edge of \(G\) or to \(i=j\) [2010.11891].

This construction isolates a graph-controlled support condition inside \(M_d\otimes M_d\). In practice, one often works with the larger subspace of local diagonal orthogonal-invariant matrices, but the diagonal-entry pattern remains dictated by \(G\). The triangle-free hypothesis is not merely combinatorial: it is the condition under which the separability test developed for these systems collapses to a tractable positivity criterion on auxiliary matrices.

## 2. Local averaging and the LDOI normal form

Let \(\mathrm{DO}_d \subset M_d\) be the group of diagonal orthogonal matrices, that is, matrices of the form \(\mathrm{diag}(\pm 1,\dots,\pm 1)\). The local-averaging map is
\[
\Phi:M_d\otimes M_d\longrightarrow M_d\otimes M_d,
\qquad
\Phi(\rho)=
\mathbb{E}_{O\in\mathrm{DO}_d}\bigl[(O\otimes O)\rho(O\otimes O)\bigr].
\]
This map is the orthogonal projection onto the subspace of local diagonal orthogonal-invariant (LDOI) matrices. For arbitrary \(\rho\), the averaged state has the form
\[
\Phi(\rho)=\rho_{A,B,C}
=
\sum_{i,j=1}^d A_{ij}\,\lvert i\,j\rangle\langle i\,j\rvert
+
\sum_{i\neq j}B_{ij}\,\lvert i\,i\rangle\langle j\,j\rvert
+
\sum_{i\neq j}C_{ij}\,\lvert i\,j\rangle\langle j\,i\rvert,
\]
with
\[
A_{ij}=\langle ij|\rho|ij\rangle,\qquad
B_{ij}=\langle ii|\rho|jj\rangle,\qquad
C_{ij}=\langle ij|\rho|ji\rangle.
\]
Because \(\Phi\) is a separable, in fact LOCC, operation, it preserves separability:
\[
\rho \text{ separable}
\Longrightarrow
\Phi(\rho)=\rho_{A,B,C}\text{ is separable in the LDOI class}
\]
[2010.11891].

The significance of this reduction is structural. It replaces an arbitrary bipartite state by a canonical LDOI representative whose relevant data are carried by the three matrices \(A,B,C\). In the triangle-free regime, the resulting separability question is governed by comparison-matrix positivity rather than by a full direct analysis of the original state.

## 3. Comparison matrices and the graph-theoretic detection criterion

For any matrix \(M\in M_d\), the comparison matrix \(M(M)\) is defined by
\[
\bigl[M(M)\bigr]_{ii}=|M_{ii}|,\qquad
\bigl[M(M)\bigr]_{ij}=-|M_{ij}|
\quad(i\neq j).
\]
Two graph-theoretic propositions drive the entanglement test. First, if \(G\) is acyclic, that is, a forest, and \(B\ge 0\) is any matrix-realization of \(G\) with \(B_{ij}\neq 0\) exactly on the edges of \(G\), then
\[
B\ge 0,\; G(B)=G
\quad\Longrightarrow\quad
M(B)\succeq 0.
\]
Second, if \(G\) is triangle-free but contains at least one cycle of length \(\ge 4\), then there exists a positive semidefinite \(B\ge 0\) with \(G(B)=G\) such that
\[
M(B)\not\succeq 0.
\]
From these facts one obtains the main graph test:
\[
M(B)\not\succeq 0
\quad\Longrightarrow\quad
\rho_{A,B,C}\text{ is entangled},
\]
and, more sharply, for a triangle-free graph \(G\), there exist entangled PPT-LDOI states detectable by this criterion if and only if \(G\) contains an induced cycle of length \(\ge 4\). Equivalently, nontrivial detection on \(S_G\) occurs exactly when \(G\) is not a forest [2010.11891].

The criterion therefore separates triangle-free graphs into two classes. Forests never yield a nontrivial comparison-matrix obstruction, whereas triangle-free cyclic graphs do. In that sense, induced cycles of length at least \(4\) are the precise graph-theoretic source of detectable PPT entanglement within this operator-system framework.

## 4. Construction of PPT entangled families on triangle-free supports

Fix a triangle-free cyclic graph \(G\) on \(d\ge 4\) vertices. There exists a positive matrix
\[
B=X X^T \succeq 0,\qquad G(B)=G,\qquad M(B)\not\succeq 0.
\]
Choose matrices \(A\succeq 0\) and \(C\succeq 0\) such that

- \(\operatorname{diag}A=\operatorname{diag}B=\operatorname{diag}C\),
- \(G(A)=G(B)=G(C)=G\),
- for all \(i\neq j\),
  \[
  A_{ij}A_{ji}\ge \max\bigl\{|B_{ij}|^2,\;|C_{ij}|^2\bigr\}.
  \]

Then
\[
\rho_{A,B,C}
=
\sum_{i,j}A_{ij}\,\lvert i\,j\rangle\langle i\,j\rvert
+
\sum_{i\neq j}B_{ij}\,\lvert i\,i\rangle\langle j\,j\rvert
+
\sum_{i\neq j}C_{ij}\,\lvert i\,j\rangle\langle j\,i\rvert
\]
is a PPT LDOI state, and because \(M(B)\not\succeq 0\), it fails the comparison-matrix separability test. This yields a continuous family of PPT entangled states supported on \(G\). By varying over all triangle-free cyclic graphs and all admissible choices of \((A,C)\), one obtains infinitely many inequivalent PPT entangled families in every dimension [2010.11891].

A concrete example is the \(4\otimes 4\) \(C_4\) family. For the \(4\)-cycle \(G=C_4\), one may take
\[
X=
\begin{pmatrix}
1&0&0\\
1&1&0\\
0&1&1\\
1&-1&0
\end{pmatrix},
\qquad
B=XX^T,
\qquad
M(B)\not\succeq 0,
\]
and then choose any \(A\succeq 0\), \(C\succeq 0\) with
\[
G(A)=G(B)=G(C)=C_4,\qquad
A_{ij}A_{ji}\ge \max\{|B_{ij}|^2,|C_{ij}|^2\}.
\]
Each resulting \(\rho_{A,B,C}\) is PPT but entangled. The construction is notable because it does not merely prove existence; it gives a recipe for generating large families parameterized by graph support and admissible positive data.

## 5. Extension beyond the triangle-free case and complexity

For an arbitrary \(\rho\in M_d\otimes M_d\), not necessarily itself \(\Delta\)-free, define
\[
A_{ij}=\langle ij|\rho|ij\rangle,\qquad
B_{ij}=\langle ii|\rho|jj\rangle,\qquad
C_{ij}=\langle ij|\rho|ji\rangle.
\]
If some principal submatrix \(\widetilde A\) of \(A\), of size \(\widetilde d<d\), has graph \(G(\widetilde A)\) that is triangle-free, then one may project to the corresponding \(\widetilde d\otimes \widetilde d\) subsystem,
\[
\widetilde\rho=(I\otimes P)\rho(I\otimes P),
\]
and apply the same test to \(\widetilde B\) and \(\widetilde C\). The resulting theorem states that if there exists such a principal submatrix \(\widetilde A\) with
\[
M(\widetilde B)\not\succeq 0
\quad\text{or}\quad
M(\widetilde C)\not\succeq 0,
\]
then \(\rho\) is entangled [2010.11891].

The search problem underlying this extension is computationally hard. Finding a triangle-free induced subgraph of maximal size is a known NP-complete problem, since it generalizes the independent-set problem. The worst-case complexity of the associated detection method is therefore NP-hard. This places the method in a common pattern of entanglement theory: the underlying certificate is explicit once the right subsystem is found, but locating that subsystem may itself dominate the computational cost.

## 6. Triangle-free and strongly triangle-free operator systems as quantum graphs

In the noncommutative setting, a quantum graph on an \(n\)-dimensional Hilbert space is an operator system
\[
\mathcal{V}=S\subseteq M_n,
\]
that is, a linear subspace containing the identity \(I_n\) and closed under the adjoint operation. The ambient space \(M_n\) is equipped with the Hilbert–Schmidt inner product
\[
\langle X,Y\rangle = \operatorname{Tr}(X^*Y).
\]
The orthogonal projection onto \(\mathcal{V}\) is denoted
\[
P_{\mathcal{V}}:M_n\longrightarrow \mathcal{V},
\]
characterized by
\[
P_{\mathcal{V}}(X)\in\mathcal{V},
\qquad
\langle Y,\;X-P_{\mathcal{V}}(X)\rangle = 0
\quad \forall\,Y\in\mathcal{V}.
\]
The noncommutative adjacency relation is defined by declaring two unit vectors \(v,w\in\mathbb{C}^n\) adjacent if there exists \(A\in\mathcal{V}\) such that
\[
\langle v,\;A\,w\rangle \neq 0.
\]
For a rank-\(k\) projection \(P\in M_n\), the compression is
\[
P\mathcal{V}P=\{PAP:A\in\mathcal{V}\}\subseteq M_k.
\]
A \(k\)-clique is a rank-\(k\) projection \(P\) with
\[
P\mathcal{V}P \cong M_k
\quad\Longleftrightarrow\quad
\dim(P\mathcal{V}P)=k^2,
\]
while a \(k\)-anticlique is a rank-\(k\) projection \(P\) with
\[
P\mathcal{V}P=\mathbb{C}P
\quad\Longleftrightarrow\quad
\dim(P\mathcal{V}P)=1.
\]
An operator system \(\mathcal{V}\) is triangle-free if it admits no \(3\)-clique, equivalently if every rank-\(3\) projection \(P\) satisfies
\[
\dim(P\mathcal{V}P)<9.
\]
It is strongly triangle-free if every rank-\(3\) projection \(Q\) dominates a \(2\)-anticlique: there exists a rank-\(2\) projection \(P\le Q\) such that
\[
P\mathcal{V}P=\mathbb{C}P
\]
[2509.17782].

These definitions separate two notions that coincide in the classical graph setting but not in the quantum one. Triangle-freeness forbids full \(3\times 3\) matrix behavior on a \(3\)-dimensional subspace; strong triangle-freeness requires, in addition, that every \(3\)-dimensional compression contain a \(2\)-dimensional scalar compression.

## 7. Classification, examples, and the classical–quantum divide

A complete characterization is known for strongly triangle-free operator systems. Weaver’s theorem states that an operator system \(\mathcal{V}\subseteq M_n\) is strongly triangle-free if and only if one of the following mutually exclusive cases holds:

1. \(n=2\).
2. \(\dim(\mathcal{V})=2\).
3. \(n=4\) and, up to unitary equivalence,
   \[
   \mathcal{V}
   =
   \left\{
   \begin{pmatrix}
   aI_2 & bI_2\\[6pt]
   bI_2 & cI_2
   \end{pmatrix}
   : a,b,c\in\mathbb{C}
   \right\}
   \subseteq M_4.
   \]
4. \(\mathcal{V}\subseteq \mathcal{K}_{1,n-1}\) for some orthonormal basis, where
   \[
   \mathcal{K}_{1,n-1}
   =
   \left\{
   \begin{pmatrix}
   a & v^*\\
   w & bI_{n-1}
   \end{pmatrix}
   : a,b\in\mathbb{C},\; v,w\in\mathbb{C}^{n-1}
   \right\}.
   \]

Several examples and non-examples delimit the scope of these notions. If \(\dim(\mathcal{V})\le 8\), then \(\mathcal{V}\) is triangle-free. If \(\dim(\mathcal{V})\le 2\), then \(\mathcal{V}\) is strongly triangle-free. The diagonal system \(\mathcal{D}_n\) has no \(2\)-anticlique for \(n\ge 3\), so it fails strong triangle-freeness for \(n\ge 3\), and it is triangle-free exactly when \(n\le 8\). For complete bipartite operator systems \(\mathcal{K}_{m,n}\), \(\mathcal{K}_{1,n}\) is strongly triangle-free for every \(n\), whereas \(\mathcal{K}_{m,n}\) with \(m,n\ge 2\) has a \(3\)-clique and hence is not triangle-free. There are also nine-dimensional triangle-free systems in \(M_4\): the block-diagonal eight-dimensional system \(M_2\oplus M_2\) admits a one-dimensional extension that remains triangle-free, and the skew version
\[
\left\{
\begin{pmatrix}
0&A\\
B&0
\end{pmatrix}
\right\}\oplus\mathbb{C}I
\]
likewise. The exceptional three-dimensional \(M_4\) system displayed above is strongly triangle-free; every rank-\(3\) projection in \(M_4\) compresses it to a system admitting a rank-\(2\) anticlique [2509.17782].

The proof strategy for the classification proceeds through compressions to \(M_3\), eigenvalue constraints, and structural reductions. Rank-\(2\) anticliques for Hermitian matrices in \(M_3\) are classified first; subsequent lemmas derive balance conditions from \(3\)-dimensional compressions, show that the presence of a rank-\(1\) projection forces \(\mathcal{V}\subseteq\mathcal{K}_{1,n-1}\), isolate the exceptional two-eigenvalue case with \(n=4\), and then exclude more complicated spectral behavior. In the comparison with classical theory, a graph \(G\) may be encoded by an operator system of diagonal matrices plus edge matrix units \(E_{ij}\). Classically, “no triangle” means no induced \(K_3\), equivalently every \(3\)-subset has a missing edge, hence a \(2\)-anticlique on that \(3\)-set. In the quantum case, this equivalence splits: triangle-free means no \(3\)-dimensional quantum clique, while strongly triangle-free means every \(3\)-dimensional subspace contains a \(2\)-dimensional anticlique. There is no simple duality between cliques and anticliques in the quantum setting, and most quantum bipartite operator systems \(\mathcal{K}_{m,n}\) with \(m,n\ge 2\) fail triangle-freeness by admitting \(3\)-cliques.

Source: https://www.emergentmind.com/topics/triangle-free-operator-systems