---
title: Complex-Valued Laplacian Flows
url: https://www.emergentmind.com/topics/complex-valued-laplacian-flows
type: topic
---

# Complex-Valued Laplacian Flows

Complex-valued Laplacian flows are network dynamical systems in which the coupling operator is a Laplacian with complex entries and the state evolves in a complex vector space. In the most direct formulation, the node-state vector \(x(t)\in\mathbb C^n\) satisfies
\[
\dot{x}(t)=-Lx(t),
\]
where \(L\) is a complex-valued Laplacian associated with a graph whose adjacency matrix has complex weights. This setting is motivated by applications such as AC power networks, impedance networks, electrodynamics, planar coordination, and directed graph signal propagation. Recent work has also introduced the pseudoinverse Laplacian flow
\[
\dot{x}(t)=-L^\dagger x(t),
\]
and has shown that, for connected undirected graphs and unsigned strongly connected weight-balanced digraphs, Laplacian and pseudoinverse Laplacian flows have the same consensus behavior [2411.09254].

## 1. Definitions, operator conventions, and model scope

A standard modern convention defines a complex-valued network by
\[
\mathcal G(A)=(\mathcal V,\mathcal E,A), \qquad A=[a_{ij}]\in \mathbb C^{n\times n},
\]
with
\[
D_{\rm out}=\operatorname{diag}(A\mathbb 1_n), \qquad L=D_{\rm out}-A.
\]
In this convention, \(L\mathbb 1_n=0\) for unsigned graphs, and for undirected graphs \(L\) is complex symmetric when \(A\) is complex symmetric. This is the convention used in recent consensus analyses on complex-valued Laplacian flows and in the corresponding pseudoinverse theory [2410.13700].

An earlier but still influential convention defines the complex Laplacian by
\[
L=D-A, \qquad d_i=\sum_{j\in\mathcal N_i}|a_{ij}|,
\]
so that the diagonal degree term is built from \(|a_{ij}|\) rather than from \(a_{ij}\) itself. In that framework, the flow
\[
\dot z(t)=-\kappa L z(t)
\]
acts on planar agent states \(z_i=x_i+{\rm j}y_i\in\mathbb C\), and the nullspace need not be \(\operatorname{span}\{\mathbb 1\}\); instead it may be \(\operatorname{span}\{\zeta\}\) for some \(\zeta\in\mathbb T^n\) [1406.1862].

These two definitions are not interchangeable. One of the most persistent misconceptions in the area is that “complex Laplacian” denotes a single canonical operator. The literature instead contains at least two distinct Laplacian constructions with different diagonal conventions, different nullspace geometry, and different consensus notions. In the row-sum-zero setting of \(L=D_{\rm out}-A\), consensus means convergence to \(\operatorname{span}\{\mathbb 1_n\}\). In the modulus-degree setting, the limiting one-dimensional subspace is generally \(\operatorname{span}\{\zeta\}\), and the induced coordination pattern can encode fixed relative phases rather than identical complex states [2411.09254].

The same nonuniqueness appears in directed-graph signal processing. There, the magnetic Laplacian is defined by
\[
\mathbf{L}_{q} = \mathbf{D}_{s} - \mathbf{A}_{s}\odot{\mathbf{T}_{q},
\qquad
\mathbf{T}_{q}{(u,v)} = {\exp}{\left(i2{\pi}q{\left(\mathbf{A}(u,v) - \mathbf{A}(v,u)\right)}\right)},
\]
and preserves directionality through complex phases while remaining Hermitian. This is a different operator family again, designed for filtering and propagation on directed graphs rather than for the consensus theory of row-sum-zero complex Laplacians [2110.07570].

## 2. Consensus dynamics and asymptotic subspaces

For the row-sum-zero complex Laplacian flow,
\[
\dot{x}(t)=-Lx(t), \qquad x(t)=e^{-Lt}x_0,
\]
consensus means asymptotic agreement of all complex states:
\[
x_i(t)-x_j(t)\to 0 \quad \text{for all } i,j,
\]
equivalently,
\[
\lim_{t\to\infty}x(t)=\mathbb 1_n c
\]
for some complex scalar \(c\). When \(0\) is a simple eigenvalue and the nonzero eigenvalues of \(L\) lie in the open right half-plane, the semigroup has the asymptotic rank-one form
\[
\lim_{t\to\infty} e^{-Lt}=\alpha \mathbb 1_n z^H,
\]
hence
\[
\lim_{t\to\infty}x(t)=\mathbb 1_n z^H x_0.
\]
The real and imaginary parts may converge to different common values, because the consensus value is complex [2410.13700].

For undirected graphs and weight-balanced digraphs, the consensus subspace is \(\operatorname{span}\{\mathbb 1_n\}\). In that case the left eigenvector associated with the zero eigenvalue determines the limiting projection, and in weight-balanced graphs the limiting value corresponds to average consensus in the sense emphasized by the recent consensus papers [2411.09254].

A different asymptotic picture appears in the modulus-degree framework developed for planar coordination. There the continuous-time flow
\[
\dot z(t)=-\kappa L z(t)
\]
reaches **complex consensus** iff the adjacency matrix is essentially nonnegative and the graph has a spanning tree, and then
\[
\lim_{t\to\infty} z(t)= (\eta^T z(0))\,\zeta.
\]
This notion is weaker than ordinary equality of complex node states: all agents converge to values with equal modulus and generally different phases. In geometric terms, the limiting pattern is
\[
z_i(\infty)=c\,\zeta_i, \qquad |\zeta_i|=1,
\]
so the agents lie on a common circle centered at the origin, with angular offsets determined by \(\zeta\) [1406.1862].

This divergence of consensus notions is substantive rather than terminological. In one branch of the literature, complex-valued Laplacian flows generalize average consensus to \(\mathbb C^n\); in another, they provide distributed laws for planar formations, set surrounding, and phase-coherent patterns. A plausible implication is that any encyclopedic treatment of complex-valued Laplacian flows must separate “consensus to \(\mathbb 1_n\)” from “coordination on \(\zeta\)-shaped manifolds” rather than subsuming both under a single asymptotic template.

## 3. Real eventual exponential positivity and spectral criteria

A key technical notion in recent theory is **real eventual exponential positivity**. A matrix \(M\in\mathbb C^{n\times n}\) is rEEP if there exists \(t_0\) such that
\[
\Re(e^{Mt})>0,\qquad \forall t\ge t_0.
\]
This extends eventual positivity from real matrices to complex matrices by requiring eventual positivity only of the real part of the semigroup. In complex-valued Laplacian flows, rEEP replaces the classical positivity machinery that is unavailable once \(e^{-Lt}\) is complex-valued [2410.13700].

The basic mechanism is spectral. One introduces the shifted matrix
\[
B=dI_n-L,
\]
choosing \(d\) so that the eigenvalue corresponding to consensus becomes a positive dominant eigenvalue. Under strong complex Perron–Frobenius assumptions, real eventual positivity of powers of \(B\) is equivalent to rEEP of \(-L\). The dominant eigendirection then yields positivity of the real part of the asymptotic rank-one projection, which in turn supports consensus analysis [2410.13700].

For unsigned undirected graphs with complex symmetric Laplacian, the recent theory proves
\[
-L \text{ is real EEP} \iff 0 \text{ is a simple eigenvalue of }L.
\]
For unsigned strongly connected weight-balanced digraphs, the same equivalence holds:
\[
-L \text{ is real EEP} \iff 0 \text{ is a simple eigenvalue of }L.
\]
In both cases, rEEP implies convergence of the Laplacian flow to the consensus subspace. The role of weight balance is to ensure that both the right and left eigenvectors associated with the dominant shifted eigenvalue lie in \(\operatorname{span}\{\mathbb 1_n\}\), which makes the Perron–Frobenius argument tractable in the nonsymmetric complex setting [2410.13700].

The theory was subsequently extended to general complex-valued directed networks by introducing the additional constraint of **real dominance**. A vector \(z\in \mathbb{C}^n\) is real dominant if
\[
\Re({z})\geq\left | \Im({z}) \right|,
\]
equivalently, each entry has phase angle in \([ -45^{\circ},45^{\circ}]\). In strongly connected digraphs, strict real dominance of the zero-eigenvectors leads to
\[
\text{\(-L\) is rEEP} \quad \Longleftrightarrow \quad \text{\(-L\) is marginally stable of corank \(1\)},
\]
and consensus is guaranteed iff the reduced Jordan block has all eigenvalues in the open right half-plane [2509.04196].

The status of rEEP is therefore nuanced. In symmetric and balanced classes it yields necessary-and-sufficient criteria. In more general complex digraphs, later work adds real-dominance hypotheses, angle constraints, and separate treatment of weakly connected graphs [2509.04196].

## 4. Pseudoinverse Laplacian flows

The pseudoinverse Laplacian flow is defined by
\[
\dot{x}(t)=-L^\dagger x(t),
\]
where \(L^\dagger\) is the Moore–Penrose pseudoinverse. Structurally, the pseudoinverse preserves the zero eigenvalue and inverts all nonzero spectral modes:
\[
\operatorname{spec}(L^\dagger)=\{0,1/\lambda_2,\dots,1/\lambda_n\},
\]
assuming \(0\) is simple. Thus the pseudoinverse changes non-consensus modes from \(e^{-\lambda_j t}\) to \(e^{-t/\lambda_j}\), so asymptotic consensus may be identical while transient time scales can be very different [2411.09254].

For undirected graphs and weight-balanced digraphs, \(L\) and \(L^\dagger\) have the same zero-eigenspace. Using \(L\mathbb 1_n=0\) and \(\mathbb 1_n^TL=0\), one obtains
\[
L^\dagger \mathbb 1_n=0,\qquad \mathbb 1_n^T L^\dagger=0.
\]
Hence
\[
\ker L=\ker L^\dagger=\operatorname{span}\{\mathbb 1_n\}
\]
under the standing assumptions. This shared nullspace is the basis for the main equivalence theorem:
\[
\textit{In undirected graphs and weight-balanced digraphs, the Laplacian flow achieves consensus if and only if the pseudoinverse Laplacian flow achieves consensus.}
\]
The two flows therefore have the same consensus/non-consensus dichotomy and the same asymptotic consensus form on those graph classes [2411.09254].

The rEEP characterization also transfers. For connected undirected signed graphs with complex symmetric \(L\) and \(L^\dagger\), the following are equivalent:
\[
\begin{aligned}
&(i)\quad L \text{ is psd of corank }1,\\
&(ii)\quad -L \text{ is rEEP},\\
&(iii)\quad L^\dagger \text{ is psd of corank }1,\\
&(iv)\quad -L^\dagger \text{ is rEEP}.
\end{aligned}
\]
For connected undirected unsigned graphs, a corollary states the equivalence of
\[
(i)\ -L \text{ is rEEP}, \qquad (ii)\ -L^\dagger \text{ is psd of corank }1,\qquad (iii)\ -L^\dagger \text{ is rEEP}.
\]
For unsigned strongly connected weight-balanced digraphs, the paper proves the equivalence of
\[
\begin{aligned}
&(i)\quad -L \text{ is rEEP},\\
&(ii)\quad -L^\dagger \text{ has a simple }0\text{ eigenvalue},\\
&(iii)\quad -L^\dagger \text{ is rEEP},\\
&(iv)\quad (L^\dagger)_s \text{ is psd of corank }1.
\end{aligned}
\]
These results show that, in the symmetric and balanced cases studied, the pseudoinverse preserves exactly the semistability structure needed for consensus [2411.09254].

The model is also physically motivated. In an impedance network with complex Laplacian \(L\), Kirchhoff’s law gives
\[
I(t)=Lv(t),
\]
and the shunt inductor relation gives
\[
v(t)=-\mathcal L_{\rm ind}\,\frac{dI}{dt}.
\]
Combining them yields a current dynamics model of pseudoinverse-Laplacian type. This establishes that \(L^\dagger\)-based dynamics are not merely algebraic variations of standard Laplacian flows but arise directly in electrical-network modeling [2411.09254].

## 5. Related operator families and application domains

One major application domain is power systems. In that setting, the bus-admittance matrix is exactly a complex Laplacian:
\[
[Y]_{m,k}= \begin{cases} \sum_{n\in\mathcal N_m} y_{m,n}, & m=k,\\[2mm] -\,y_{m,k}, & m\neq k, \end{cases}
\]
with complex line admittances \(y_{m,k}\in\mathbb C\). The resulting matrix \(Y\) is symmetric but generally non-Hermitian, and decomposes as
\[
Y = G + j\tilde B,
\]
where both \(G\) and \(\tilde B\) are real Laplacians. Although that literature studies estimation rather than autonomous consensus dynamics, it formalizes a class of complex Laplacians that is directly relevant to AC and DLPF network models, and it matches the power-network motivation used in the consensus and pseudoinverse-flow papers [2308.03392].

A second important branch concerns planar multi-agent coordination. There, complex edge weights encode rotations and scalings between neighboring planar positions. The central structural condition is **essential nonnegativity**, meaning that there exists a diagonal unit-modulus matrix \(D_\zeta\) such that
\[
D_\zeta^{-1} A D_\zeta \ge 0.
\]
Under this condition and a spanning tree, the transformed Laplacian becomes an ordinary consensus Laplacian, and the original complex flow converges to
\[
(\eta^T z(0))\,\zeta.
\]
The same theory recovers bipartite consensus on signed graphs as the real special case \(\zeta_i\in\{\pm1\}\) [1406.1862].

A third branch replaces row-sum-zero Laplacians by direction-sensitive Hermitian deformations. The magnetic Laplacian
\[
\mathcal L_q = \mathbf I_n - \mathbf D_s^{-1/2}\mathbf A_s\mathbf D_s^{-1/2}\odot \mathbf T_q
\]
is a complex Hermitian PSD operator, interpreted as a discrete Schrödinger operator with magnetic field. The corresponding graph filters use repeated powers of a renormalized magnetic adjacency, and the paper derives denoisers such as
\[
\bar{\mathbf x}=\left(\mathbf I_n+\frac1\mu\mathcal L_q\right)^{-1}\mathbf x.
\]
The paper describes diffusion, denoising, and filtering rather than a continuous-time consensus flow; this suggests a discrete-time, phase-aware propagation paradigm closely related to complex-valued Laplacian evolution on directed graphs [2110.07570].

These application domains share a common algebraic theme—complex couplings encode amplitude and phase simultaneously—but they do not share a single dynamical objective. Power-system models emphasize admittance structure and physical flow laws; planar coordination emphasizes switching equivalence and one-dimensional phase-shaped limit sets; magnetic-Laplacian models emphasize direction-preserving propagation on digraphs.

## 6. Limitations, controversies, and open directions

The strongest exact consensus theorems currently apply only to restricted graph classes. For the row-sum-zero model, the equivalence between Laplacian and pseudoinverse Laplacian flows is proved for connected undirected graphs and unsigned strongly connected weight-balanced digraphs, and not for arbitrary complex-weighted digraphs. The role of weight balance is essential because it fixes the left zero eigenvector in a tractable way and supports average-consensus-type behavior [2411.09254].

Outside those classes, rEEP is not exhaustive. The pseudoinverse-flow paper explicitly states that for general directed graphs the results are only sufficient, not necessary-and-sufficient, and gives a counterexample: a weakly connected digraph for which \(-L\) is not rEEP, yet the Laplacian flow still achieves consensus. This makes clear that eventual positivity of the real part of the semigroup is a powerful criterion, but not a complete characterization of complex-digraph consensus [2411.09254].

A second recurring misconception is that connectivity alone should guarantee consensus, as in classical real-valued theory. Later work on complex-valued directed networks shows that this is false in general: connectivity must be supplemented by spectral placement and, in the developed theory, by real-dominance conditions on the zero-eigenvectors. Even strongly connected complex digraphs can have nonzero eigenvalues in the wrong half-plane, and then
\[
\dot{x}(t)=-Lx(t)
\]
diverges. That paper therefore proposes modified flows
\[
\dot{x}(t)=-L_m x(t)
\]
obtained by eigenspectrum assignment, with the goal of moving the nonzero eigenvalues to the open right half-plane while preserving the zero eigenspace [2509.04196].

A third limitation concerns the operator family itself. The complex magnetic Laplacian in the PDE setting is genuinely non-selfadjoint when the magnetic potential is complex-valued. Under a relative form-boundedness assumption on the imaginary part of the magnetic potential, the operator is \(m\)-sectorial, which provides the operator-theoretic basis for parabolic evolution. In two dimensions, the same paper proves sufficient conditions for compact resolvent and constructs semiclassical pseudomodes for non-critical complex magnetic fields—pseudomodes that do not exist when the magnetic field is real-valued. This suggests that non-normal transient behavior is intrinsic to some complex-valued Laplacian flows and cannot be read off from eigenvalues alone [2410.01377].

Open problems follow directly from these limitations. The recent pseudoinverse theory leaves non-weight-balanced and signed digraphs for future work [2411.09254]. The later general-digraph theory identifies real dominance as sufficient, not necessary, and leaves finer edge-modification designs open [2509.04196]. More broadly, the coexistence of row-sum-zero, modulus-degree, magnetic, and admittance-based Laplacians indicates that “complex-valued Laplacian flow” is best understood as a family of related but nonidentical theories rather than as a single closed formalism.

Source: https://www.emergentmind.com/topics/complex-valued-laplacian-flows