---
title: Directed Quantum Graphs
url: https://www.emergentmind.com/topics/directed-quantum-graphs
type: topic
---

# Directed Quantum Graphs

Directed quantum graphs are a family of non-equivalent constructions in which orientation is incorporated into a quantum, operator-algebraic, geometric, or algorithmic graph formalism. In one line of work, a directed quantum graph is a quantum graph on a quantum set with no imposed symmetry condition, so that “undirected” becomes an additional constraint rather than the default. In other lines, directedness is carried by control–target assignments in graph-state constructions, by non-symmetric transport operators in quantum walks on digraphs, by arrow-dependent metric weights in discrete quantum geometry, or by directed access and orientation constraints in quantum graph algorithms. This breadth of usage reflects the wider development of quantum graph theory, which “has been introduced by Duan, Severini, and Winter to describe the zero-error behaviour of quantum channels,” and later expanded into several distinct research programs [2603.23651][2507.15534][2509.05214][1606.00992][2002.11851].

## 1. Historical scope and semantic range

In the operator-algebraic literature, the modern directed theory is explicitly presented as an extension of a formerly undirected framework. “Later, Wasilweski provided a solid theory of directed quantum graphs which was formerly only established for undirected quantum graphs,” and this extension already becomes nontrivial on the smallest noncommutative algebra \(M_2\) [2507.15534]. In that setting, directedness is not a perturbation of classical orientation but a relaxation of self-adjointness-type constraints on the quantum edge structure.

A different usage appears in quantum-information constructions indexed by classical digraphs. There the starting object is an ordinary directed simple graph \(G(V,L)\), and the graph is made quantum by assigning qubits to vertices and controlled unitaries to oriented edges. The orientation is operational: the arrow \(a\to b\) means that qubit \(a\) is the control and qubit \(b\) is the target [2509.05214].

A third usage concerns quantum transport on digraphs. For continuous-time or stochastic quantum walks, directionality conflicts with Hermiticity if one uses the raw directed adjacency or Laplacian as Hamiltonian data. Several papers therefore replace the naive generator by a Hermitian adjacency with phases, a pseudo-Hermitian similarity transform, or a GKSL generator on a digraph [1606.00992][1607.02673][2109.12978]. This suggests that “directed quantum graph” is best treated as a family of related notions rather than a single stabilized definition.

## 2. Operator-algebraic directed quantum graphs

For a finite-dimensional quantum set \((B,\psi)\), a quantum graph \(\mathcal G\) may be presented in three equivalent ways: as a \(B'-B'\)-bimodule \(S\subset B(\mathcal H)\), as a quantum adjacency matrix \(A:B\to B\) satisfying
\[
A(x^*)=A(x)^* \quad \text{for all }x\in B, \qquad m(A\otimes A)m^*=A,
\]
or as a projection \(P\in B\otimes B^{\mathrm{op}}\) [2507.15534]. In this framework, directedness is the generic case. Undirectedness requires extra symmetry: on a nontracial quantum set, a quantum graph is GNS-undirected iff
\[
S=S^* \quad\text{and}\quad \rho S\rho^{-1}=S,
\]
while it is KMS-undirected iff
\[
S=S^*.
\]
In the tracial case these coincide; in the nontracial case they do not [2507.15534].

The complete classification on \(M_2\) makes this distinction concrete. In the tracial case \((M_2,\tau)\), if \(\mathcal G\) is loopfree with exactly one quantum edge, then it is isomorphic to a unique \(\mathcal G_{0,\beta}^{(1B)}\), \(\beta\in[0,1]\), with adjacency matrix
\[
A_{0,\beta}^{(1B)}= \frac{1}{1+\beta^2} \begin{pmatrix} 0 & 0 & 0 & (1+\beta)^2 \\
0 & 0 & 1-\beta^2 & 0 \\
0 & 1-\beta^2 & 0 & 0 \\
(1-\beta)^2 & 0 & 0 & 0 \end{pmatrix},
\]
and it is undirected iff \(\beta=0\) [2507.15534]. On \((M_2,\psi_q)\), \(q\in(0,1)\), there are infinitely many loopfree KMS-undirected quantum graphs that are not GNS-undirected. The paper exhibits the family \(\mathcal G_{0,0,\gamma}^{(q,1C)}\), \(\gamma\ge 0\), as explicit examples of this separation [2507.15534].

The infinite-dimensional extension replaces bounded operator bimodules by Hilbert–Schmidt bimodules. An HS quantum relation is a closed subspace \(V\subseteq HS(H)\) such that \(M'VM'\subseteq V\). Such \(V\) are in bijection with projections
\[
e\in M\bar\otimes M^{\mathrm{op}},
\]
and, when \(e\) is integrable for the slice-map \(id\otimes\varphi^{\mathrm{op}}\), there is a normal CP map \(A\colon M\to M\) called a quantum adjacency operator [2511.23121]. Its Kraus form is
\[
A(x)=\sum_i \alpha_i x \alpha_i^*.
\]
In this formulation, a general projection \(e\) or CP map \(A\) is the directed object; symmetry requires extra conditions involving the tensor swap \(\tau(e)\) and, in the nontracial case, KMS-adjoint relations [2511.23121].

## 3. Directed graph states and orientation-dependent entanglement

In the graph-state construction based on a directed simple graph \(G(V,L)\), with \(V=\{1,\ldots,M\}\), the oriented adjacency matrix is
\[
\Gamma_{ab}= \begin{cases} 1 & \text{if } (a,b)\in L,\\ 0 & \text{otherwise}. \end{cases}
\]
Each vertex carries a qubit, so
\[
\mathcal{H}=(\mathbb{C}^2)^{\otimes M}.
\]
An oriented edge \((a,b)\) is encoded by
\[
U_{ab}=\Pi_0^{(a)}\mathbb{I}^{(b)}+\Pi_1^{(a)}\bar U^{(b)},
\]
and the graph state is
\[
\ket{G}=\prod_{(a,b)\in L} U_{ab}\ket{\phi}^{\otimes M}.
\]
The paper emphasizes that the arrow \(a\to b\) has a direct operational meaning: qubit \(a\) is the control and qubit \(b\) is the target [2509.05214].

To obtain a graph-defined state independent of any arbitrary ordering of edges, the construction imposes
\[
[U_{ab},U_{bc}]=0,\qquad \forall a,b,c\in V,
\]
and
\[
U_{ab}=U,\qquad \forall (a,b)\in L.
\]
These assumptions force
\[
\bar U=e^{-i\psi} \begin{pmatrix} e^{i\theta} & 0\\ 0 & e^{-i\theta} \end{pmatrix},
\qquad \psi,\theta\in\mathbb{R},
\]
so the interaction becomes controlled phase-type [2509.05214]. For the maximally entangling choice, the input is
\[
\ket{\phi}=\frac{1}{\sqrt2}(\ket0+\ket1).
\]

The entanglement observable studied there is the Entanglement Distance,
\[
E(\ket{G})=1-\frac{1}{M}\sum_{i=1}^M \left\|\bra{G}\boldsymbol{\sigma}^{(i)}\ket{G}\right\|^2.
\]
Its main formula is
\[
E(\theta;\{d(i)\})=1-\frac{1}{M}\sum_{i\in V}[\cos(\theta)]^{2d(i)},
\]
where
\[
d(i)=d_\rightarrow(i)+d_\leftarrow(i)
\]
is the total degree [2509.05214]. The directional information enters the phase
\[
\Phi=\delta_0-\delta_1-d(i)\delta+d_\rightarrow(i)\psi+d_\leftarrow(i)\theta,
\]
but the squared Bloch-vector norm is phase-blind, so the final ED depends only on \(d(i)\), not on the split between in-degree and out-degree. A common misconception is therefore avoided only by distinguishing the state from the observable: the construction is genuinely directed at the kinematic level, but the chosen entanglement functional is insensitive to orientation [2509.05214].

## 4. Quantum walks, stochastic walks, and transport on digraphs

For continuous-time quantum walks on a directed graph with adjacency matrix \(A\neq A^T\), one Hermitian formulation is
\[
\mathcal A(\alpha) = e^{i\alpha}A+e^{-i\alpha}A^T = (A+A^T)\cos\alpha+i(A-A^T)\sin\alpha.
\]
At \(\alpha=0\), directionality disappears; at \(\alpha=\pi/2\), only the antisymmetric directed part survives. For directed bipartite graphs, transport between the two partitions is completely suppressed at
\[
\alpha=\frac{\pi}{2},
\]
because the resulting Hamiltonian becomes block diagonal [1606.00992]. The same model proves two mirror symmetries around \(\alpha=0\) and \(\pi/2\), yielding \(\pi\)-periodicity in \(\alpha\) [1606.00992].

A different repair of non-Hermiticity uses pseudo-Hermiticity. Starting from the directed-graph Laplacian
\[
H_{ij}=\left(\sum_k A_{ik}\right)\delta_{ij}-A_{ij},
\]
one assumes \(H\) is diagonalizable with real spectrum, finds \(\eta\) such that
\[
\tilde H=\eta H\eta^{-1}
\]
is Hermitian, and evolves with
\[
\tilde U(t)=e^{-i\tilde H t}.
\]
This produces a unitary walk in the ordinary inner product and is equivalent to mapping the directed graph to an undirected weighted complete graph with self-loops on the same vertex set [1607.02673]. The associated directed-graph quantum centrality is defined by the long-time average
\[
v_j= \lim_{T\to\infty} \frac1T\int_0^T \left| \bra{j} e^{-i\eta H\eta^{-1}t} \left(\frac1{\sqrt N}\sum_k \ket{k}\right) \right|^2 dt,
\]
and the statistical analysis shows strong agreement with PageRank on directed acyclic graphs [1607.02673].

For arbitrary digraphs, the dissertation-length open-system route replaces unitary CTQW by a GKSL evolution
\[
\frac{\dd \varrho}{\dd t} = - [H, \varrho] + \sum_{L\in \mathbb L} \left(  L \varrho L^\dagger - \frac12 \{L ^\dagger L, \varrho\} \right),
\]
and distinguishes local, global, and nonmoralizing global interaction models [2109.12978]. The local model uses
\[
\mathbb L= \{c_{ij}\ketbra{j}{i}: (i,j)\in\vec E,\ c_{ij}\neq 0 \},
\]
and the limiting-properties analysis proves that, for strongly connected digraphs and for weakly connected digraphs whose condensation graph has exactly one sink, the local-environment-interaction evolution is relaxing [1703.01792]. By contrast, the standard global model moralizes directed graphs, and there exists a directed graph and an initial state for which the global-interaction evolution is non-convergent for every \(\omega\in(0,1]\) [1703.01792]. The nonmoralizing global interaction QSW enlarges the Hilbert space by splitting each vertex into \(\operatorname{indeg}(v)\) copies, constructs Lindblad blocks with orthogonal columns, and adds a local rotating Hamiltonian to avoid premature localization; numerically it restores digraph observance much better than standard global interaction [2109.12978][1703.01792].

In the metric-graph branch, a star graph with arm Hamiltonians
\[
H_j=-\frac{\partial^2}{\partial x^2}+\left[V_0\cos\left(\frac{2\pi}{d}x\right)+F_j(t)x\right]
\]
exhibits controlled directed transport under arm-dependent periodic fields
\[
F_j(t)=f_j\sin(\omega t+\phi_j).
\]
For the three-arm example, a suitable choice of \(\phi_2=\pi/2\) almost suppresses transmission into one arm and channels nearly \(70\%\) of the wave-packet norm into the desired arm while approximately preserving the packet width [1503.02253].

A discrete-time scattering theory on balanced directed graphs adds a resonance perspective. For a family \(U(\varepsilon)\) on a finite directed core plus semi-infinite directed tails, the scattering matrix \(\Sigma(\varepsilon,z)\) generally fails to converge to \(\Sigma(0,z)\) at those \(z\in\mathbb S^1\) approached by resonances \(\lambda_\varepsilon\). The discrepancy is described by rank-one resonant terms
\[
M_{\lambda,\varepsilon}(z)\alpha^-= \frac{\lambda_\varepsilon^2}{z-\lambda_\varepsilon} (\alpha^-,\chi(\Omega^-)\psi^\circledast_{\lambda,\varepsilon})\,\chi(\Omega^+)\psi_{\lambda,\varepsilon},
\]
and under a channel-balance condition on the resonant state this yields a directed-graph analog of resonant tunneling, including asymptotically anti-diagonal scattering for two channels [2504.15584].

## 5. Directed quantum geometry and graph-algebraic morphisms

In the discrete quantum-geometric approach, quantum geometry on a discrete set means “a directed graph with a weight associated to each arrow defining the quantum metric,” and the weights need not be independent of direction [2002.11851]. For \(A=k(X)\), the differential calculus is determined by arrows \(x\to y\), with basis \( \omega_{x\to y}\), and the canonical Laplacian satisfies
\[
-\Delta_\theta f=(\mathrm d f,\theta)=\sum_{x\to y}(f(y)-f(x))\,\delta_x\,p_{y\to x}.
\]
A discrete-time Markov process then becomes
\[
\partial_+ f = (-\Delta_\theta + q - p)f,
\]
with
\[
p(x)=\sum_{y:x\to y} p_{x\to y}, \qquad q(x)=\sum_{y:x\to y} p_{y\to x}.
\]
In this sense, directed transition probabilities are reinterpreted as metric data [2002.11851].

The same paper introduces a discrete Schrödinger process
\[
\partial_+\psi=\imath(-\Delta+V)\psi,
\]
where \(\Delta\) comes from a bimodule connection rather than the canonical \(\nabla_\theta\). For the two-state graph, the one-step evolution matrix is solved explicitly, yielding a \(1\)-parameter family of unitary connections and an induced generalized Markov process for \(f=|\psi|^2\) with an additional source current built from \(\psi\) [2002.11851].

A different, graph-algebraic notion of directionality is categorical. The relation category of graphs \(\mathsf{RG}\) replaces graph maps by relations
\[
R\subseteq FP(E)\times FP(F)
\]
on path sets, and admissible relation morphisms form a category \(\mathsf{ARG}\) admitting a contravariant functor to \(k\)-Alg and to graph \(C^*\)-algebras [2503.23343]. The induced maps have the uniform form
\[
R^*(S_y)=\sum_{x\in R^{-1}(y)}S_x.
\]
This does not define a quantum graph in the operator-system sense, but it gives a graph-side formalism for morphisms between directed graph-based noncommutative spaces, including Leavitt path algebras, graph \(C^*\)-algebras, Cuntz algebras, quantum spheres, and quantum balls [2503.23343].

## 6. Quantum algorithms on directed graphs

Algorithmic work uses directed graphs not as static quantum objects but as structures to be queried or tested quantumly. One example arises in causal Feynman loop configurations, where the task is to query which edge orientations of an underlying undirected graph define valid directed acyclic graph configurations. The oracle is built only from multicontrolled Toffoli and \(X\) gates, with edge orientations encoded in a register \(\ket e\), clause ancillas \(\ket a\), and a phase-marking output qubit. The basic Boolean forms are
\[
s_j \equiv \underset{i\in s_j}{\bigwedge} e_i,\qquad \bar s_j \equiv \underset{i\in s_j}{\bigwedge} \bar e_i,
\]
and the oracle marks assignments that avoid all forbidden directed cycles [2404.03544].

In sparse-graph property testing, the unidirectional model allows only outgoing-edge queries. For \(k\)-source-subgraph-freeness in this model, the quantum query complexity is
\[
O\!\left( N^{\frac12\left(1-\frac{1}{2^k-1}\right)} \right),
\]
while the lower bound obtained through a property-testing version of the \(k\)-collision problem is
\[
\tilde{\Omega}\!\left( N^{\frac12\left(1-\frac1k\right)} \right).
\]
The same work also proves that testing \(3\)-colorability in bounded-degree undirected graphs still requires \(\Omega(N)\) quantum queries [2410.05001]. A later general theorem shows that any property of bounded-degree directed graphs that is testable with \(O_{\varepsilon,d}(1)\) classical queries in the bidirectional model is testable in the quantum unidirectional model with
\[
n^{1/2-\Omega_{\varepsilon,d}(1)}
\]
queries, and that the number of occurrences of any constant-size subgraph \(H\) can be approximated up to additive error \(\delta n\) using \(o(\sqrt n)\) quantum queries [2604.07954]. These results belong to quantum algorithms on directed graphs rather than to quantum graph objects themselves, but they form an important computational branch of the subject.

## 7. Conceptual distinctions and recurrent misconceptions

A recurrent misconception is that directed quantum graphs should have a single formal definition. The literature instead divides sharply by what is being quantized. In operator-algebraic quantum graph theory, directedness is the absence of symmetry constraints, and undirectedness is extra structure such as \(S=S^*\), \(\rho S\rho^{-1}=S\), or \(e=\tau(e)\) [2507.15534][2511.23121]. In graph-state constructions, directedness is encoded operationally by control–target asymmetry, yet the entanglement distance studied in one prominent model depends only on the total degree sequence, not on orientation [2509.05214]. In walk-based theories, the basic obstruction is different: naive directed adjacency destroys Hermiticity, so one must choose between Hermitian phase adjacencies, pseudo-Hermitian similarity transforms, or open-system generators, and these choices do not preserve directed structure in the same way [1606.00992][1607.02673][2109.12978].

Another misconception is that “directed” automatically implies nonreciprocal observables. The directed graph-state model shows the opposite possibility: the preparation rule is directed, but the chosen multipartite entanglement functional is orientation-blind [2509.05214]. Conversely, the global stochastic-walk literature shows that a formally unitary or dissipative evolution may fail to respect the intended digraph because of moralization, so directional data can be present in the definition yet degraded dynamically [1703.01792].

These divergences suggest a useful editorial principle: the phrase “directed quantum graphs” names a research area rather than a single object class. Its stable core is the study of how orientation interacts with quantum adjacency, quantum transport, noncommutative edge spaces, or quantum access to graph structure. Its unstable boundary lies in the fact that the meaning of “graph,” “adjacency,” and even “undirected” changes from one framework to another.

Source: https://www.emergentmind.com/topics/directed-quantum-graphs