betaFeedback

WorksEngineering

On Real Eventual Exponential Positive property of complex-valued Laplacians and their pseudoinverses

Abstract

The property of real eventual exponential positivity guarantees the state trajectories of a linear time-invariant system to evolve asymptotically to right half of the complex-plane [1]. It has applications in the characterization of herdability and reachability of states in networked systems. Networks with complex-valued edge-weights arise in AC power systems and opinion dynamics. Recent work has introduced real eventual exponential positive (rEEP) Laplacians, for which the trajectories of the associated network flow eventually converge to consensus. We establish further properties of rEEP Laplacians and their pseudoinverses. Firstly, we show that the rEEP property for complex-valued matrices exhibits invariance under specific rotational transformations. Secondly, we establish that the property of rEEP is preserved under the MoorePenrose pseudoinverse operation, highlighting its robustness under algebraic manipulations. These results apply to effective impedance computation in AC networks and to consensus in multi-agent systems. Finally, we provide numerical examples to validate our theoretical results.

INTRODUCTION

Complex-valued networks, whose edge weights are complex numbers, arise in quantum physics, electromagnetics, power and societal systems [2]. Although complex-valued Laplacians have appeared in formation control and spectral graph theory [3], [4], dynamics of such networks remain far less understood than their real-valued counterparts [5], [6]. The main obstacle is largely structural. For Hermitian weights, the familiar tools of real symmetric matrices largely survive. For complex symmetric or edge asymmetric weights, which arise in AC power networks and recently in opinion dynamics [7], these tools no longer apply: eigenvalues are complex, positive semi-definiteness is undefined, and the Perron–Frobenius (PF hereafter) theorem no longer applies.

Eventual positivity [8] has proven useful for signed real networks, where it recovers PF-type conclusions without nonnegative weights [9], [10]. Recent literature has pushed this idea to networks with complex-valued edge weights [1], [11]–[13]. Building on these PF-type arguments [14], these papers introduced rEEP Laplacians, for which the real parts of steady-state trajectories eventually become positive. Interestingly, [12] studied the dynamics and applications of the associated pseudoinverse flows, and analyzed them on undirected and weight-balanced digraphs.

We compute the Moore-Penrose pseudoinverse of the Laplacian matrix to obtain the pseudoinverse Laplacian. Building upon our previous work [1], we extend the notion of EEP from real-valued systems to complex-valued systems. To the best of our knowledge, we are the first to answer the following question for complex-valued networks: When does the trajectory from an initial point x0x_0 become (entrywise) real positive and remain real positive for all time thereafter in a linear system? Below, we summarize our key contributions:

  1. We establish that the property of rEEP is invariant under certain global rotations of Laplacian matrices under some mathematical conditions. Equivalently, if all the complex-valued edge-weights are rotated by a given angle, then the resulting Laplacian matrix still satisfies the property of rEEP. Further, we show that rEEP can be characterized through real dominant eigenvectors only.

  2. We derive the necessary and sufficient conditions to prove that the property of rEEP is closed under the operation of pseudoinverse in complex-valued digraphs. We use complex PF theory to establish the conditions for the rEEP property of the matrices.

Related Literature: In [9], [10], the authors show the inter-relation between the property of EEP and the spectrum of Laplacian matrices for signed networks with real edge-weights. Recently, researchers have started exploring complex-valued Laplacians; the authors provide necessary and sufficient conditions for the Laplacian matrix to be singular, based on the topology of graphs in [3]. Further, we introduce the rEEP property of complex-valued Laplacians in [15] and show that this property ensures consensus in complex-valued digraphs [1]. The pseudoinverse Laplacian flow systems for complex-valued networks are introduced in [12] for undirected and weight-balanced digraphs. Moreover, the notions of structural balance and anti-balance for Hermitian matrices are proposed in [16].

Our existing works introduced the rEEP property and established its role in studying consensus in Laplacian flows. In contrast, the present work investigates the implications of rEEP property in the pseudoinverse Laplacian flows beyond the class of undirected and weight-balanced digraphs. Additionally, the invariance of this property under rotations of complex-valued Laplacian matrix and the possible rotations in eigenvectors are discussed here; unlike the previous works.

PRELIMINARIES

Let Cn×n\mathbb{C}^{n \times n} denote the space of n×nn \times n complex-valued matrices. For n=1n = 1, the space is C\mathbb{C}. Let R(m)\mathfrak{R}(m) and I(m)\mathfrak{I}(m) denote the real and imaginary parts of m∈Cm \in\mathbb{C}, respectively. The imaginary unit ι=−1\iota= \sqrt{-1}. The right half of the complex- and real-planes are C+={R(m)≥0, for all m∈C}\mathbb{C}_{+} = \{\mathfrak{R}(m) \ge0,\ \text{for all }m \in\mathbb{C}\} and R+={m≥0, for all m∈R}\mathbb{R}_{+} = \{m \ge0,\ \text{for all }m \in\mathbb{R}\}, respectively, where the open right half of complex plane is denoted as ORHP. 1n\mathbf{1}_{n} and 0n\mathbf{0}_{n} are all ones and zeros vectors.

Matrix Algebra

A matrix M∈Cn×nM \in\mathbb{C}^{n \times n} is real positive if R(mij)>0\mathfrak{R}(m_{ij}) > 0 for all the entries (hereafter R(M)>0\mathfrak{R}(M) > 0). A matrix is real eventually positive (rEP hereafter) if R(Mk)>0\mathfrak{R}(M^{k}) > 0 for some k≥k0k \ge k_{0}. The complex conjugate transpose and transpose of matrix MM are MHM^{H} and MTM^{T}, resp. The pseudoinverse of MM is denoted by M†M^{\dagger}. The spectrum of MM is spec⁡(M)={λ1,λ2,…,λn}\operatorname{spec}(M) = \{\lambda_{1}, \lambda_{2}, \ldots, \lambda_{n}\}, where λi\lambda_{i} is the ithi^{\text{th}} eigenvalue of MM. An eigenvalue having the largest magnitude is the dominant eigenvalue and is denoted by λ1\lambda_{1}. The corresponding eigenvectors are referred as dominant eigenvectors [14]. The right and left dominant eigenvectors are denoted by vv and wHw^{H}, respectively. The possible angle of rotation for the dominant eigenvector is represented as ψ\psi.

Now, we cull some important results, which play a crucial role in our analysis. More details are provided in [1], [14].

Definition 1. (Strong Complex PF Property [14]) A matrix M∈Cn×nM \in\mathbb{C}^{n \times n} has the strong complex Perron-Frobenius property if its dominant eigenvalue λ1\lambda_{1} is positive and simple with λ1>∣λi∣\lambda_{1} > |\lambda_{i}| for all i∈{2,…,n}i \in\{2, \ldots, n\}. The right eigenvector v∈Cnv \in\mathbb{C}^{n} associated with λ1\lambda_{1} is such that R(v)>0n\mathfrak{R}(v) > \mathbf{0}_{n}.

Definition 2. (**Real Eventually Exponentially Positive matrix (rEEP) [15]) A matrix M∈Cn×nM \in\mathbb{C}^{n \times n} is rEEP if there exists an exponential index τ0\tau_{0} such that R(eMτ)>0\mathfrak{R}(e^{M\tau}) > 0 for all τ≥τ0\tau\ge\tau_{0}.

Definition 3. (**Moore-Penrose pseudoinverse [17]) The Moore-Penrose pseudoinverse of M∈Cn×mM \in\mathbb{C}^{n \times m} satisfies [17]: (i) MM†M=MMM^{\dagger}M = M, (ii) M†MM†=M†M^{\dagger}MM^{\dagger} = M^{\dagger}, (iii) MM†∈Cm×mMM^{\dagger} \in\mathbb{C}^{m \times m} and M†M∈Cn×nM^{\dagger}M \in\mathbb{C}^{n \times n} are self-adjoint.

Lemma 1. (**Marginal Stability of L∈Cn×nL \in\mathbb{C}^{n \times n} [1]) Consider a strongly connected digraph G(A)\mathcal{G}(A) and for 1≤i,j≤n1 \le i,j \le n. Let

v=α1nv = \alpha\mathbf{1}_{n}, α∈C\alpha\in\mathbb{C} and w∈Cnw \in\mathbb{C}^{n} be the right and left eigenvectors associated with the zero eigenvalue of LL, respectively. Suppose the phase angles, ϕvi=∠(vi)\phi_{v_i} = \angle(v_i) and ϕwi=∠(wi)\phi_{w_i} = \angle(w_i) lie in [−45∘,45∘][-45^{\circ},45^{\circ}]. Then, the following statements are equivalent:

(i) −L-L is rEEP.

(ii) −L-L is marginally stable of corank 1.

Lemma 1 relates the rEEP property with the marginal stability of the complex-valued Laplacian matrix. This lemma is instrumental to the analysis of pseudoinverse flows.

Graph Theory

Let G(A)=(V,E,A)\mathcal{G}(A) = (\mathcal{V}, \mathcal{E}, A), where V\mathcal{V} be the network or graph of interest and E\mathcal{E} are the sets of vertices and edges respectively. The phase angles θij∈C\theta_{ij} \in\mathbb{C} and ϕk∈C\phi_{k} \in\mathbb{C} are associated to the directed edges from nodes ‘ii’ to ‘jj’ and ‘kkth’ components of the eigenvector. The adjacency matrix is A∈Cn×nA \in\mathbb{C}^{n \times n}. In a strongly connected directed graph (digraph) there exists a directed path between any pair of vertices. In digraphs, out-degree of a vertex is the sum of edge-weights of all the outgoing edges, whereas in-degree is the sum of edge-weights of incoming edges. Collectively for all nodes, the degree matrices are given by Dout=diag⁡(A1n)D_{\mathrm{out}} = \operatorname{diag}(A\mathbf{1}_{n}) and Din=diag⁡(AT1n)D_{\mathrm{in}} = \operatorname{diag}(A^{T}\mathbf{1}_{n}) [18]. The Laplacian matrix is defined as L=Dout−AL = D_{\mathrm{out}} - A.

We make use of polar as well as rectangular form representation of the complex number. Thus for A∈Cn×nA \in\mathbb{C}^{n \times n}, with non-zero aija_{ij}, we let aij=rij∠θija_{ij} = r_{ij}\angle\theta_{ij}, where r>0r > 0 is the magnitude and θij∈[0∘,360∘)\theta_{ij} \in[0^{\circ}, 360^{\circ}) (measured in degrees) is the phase angle. The angles of edge-weights may represent social interactions among agents [7].

Problem Formulation

Consider a G(A)\mathcal{G}(A) with nn nodes with complex-valued edge-weights. Let x(τ)∈Cnx(\tau) \in\mathbb{C}^{n} be the state vector and consider the (generalized) Laplacian flow system

x˙(τ)=−Lˉx(τ),τ≥0,(1)\dot{x}(\tau) = -\bar{L}x(\tau), \quad\tau\ge0, \tag*{(1)}

where Lˉ:Cn×n→Cn×n\bar{L} : \mathbb{C}^{n \times n} \to\mathbb{C}^{n \times n}. We consider the cases where Lˉ=L\bar{L} = L and Lˉ=L†\bar{L} = L^{\dagger}.

We are interested in studying the relation between the spectrum of the Laplacians and the pseudoinverse Laplacians. Specifically, we establish that rEEP Laplacians imply rEEP pseudoinverse Laplacians. Also, this property remains unchanged when all the edge-weights are rotated by an angle under some mathematical conditions. These results hold for a class of complex-valued networks, which are characterized by real dominant eigenvectors. Real dominant complex-valued networks arise in settings like phase-dependent interactions, distributed coordination, power system synchronization, and wireless communication. We use complex PF theory (see Def.1) to derive our results.

Main Results

We begin by establishing relationships between the Perron-Frobenius property in Def.1 and the proposed notion of rEEP in Def.2. Our approach relies on the spectral decomposition of complex matrices and is inspired by [14]. These relationships serve as a starting point for analyzing reachability in complex-valued networks. The rEEP matrices have significant applications in control theory to study reachability1 and holdability2 of C+n\mathbb{C}_{+}^{n}.

Let the eigenvalues of M∈Cn×nM \in\mathbb{C}^{n \times n} be arranged as ∣λ1∣≥⋯≥∣λn∣|\lambda_1| \ge\cdots\ge|\lambda_n| and consider the set of matrices:

PF={M∈Cn×n:λ1(M)∈R+ is simple, R(v)>0}.(2)\mathcal{PF}=\{M\in\mathbb{C}^{n\times n}:\lambda_1(M)\in\mathbb{R}_{+}\text{ is simple},\ \mathfrak{R}(v)>0\}. \tag*{(2)}

In the class of matrices belonging to PF\mathcal{PF}, we know that they have a simple dominant eigenvalue R(v)>0\mathfrak{R}(v)>0. These matrices are relevant in studying the stability of complex-valued linear systems. Specifically, PF matrices are used in power system small-signal stability analysis, consensus, and synchronization in multi-agent networks. However, the PF property of complex-valued matrices alone does not guarantee the steady-state behaviors, because of the complex-valued entries in the dominant eigenvectors. To address this gap, the existing literature has used an additional property of real dominance for dominant eigenvectors [1, 14].

The real dominance property plays a significant role in the analysis of complex-valued matrices. It requires the real part of each dominant eigenvector component to have a greater magnitude than its corresponding imaginary part. The dominant eigenvectors of a complex-valued matrix may or may not satisfy this property. However, owing to the rotations inherent in complex-valued eigenvectors, the components can be suitably rotated to satisfy real dominance. Thus, we consider an angle of rotation ‘ψ\psi’ to account for the rotation by which we rotate the components of eigenvector satisfy real dominance. Nevertheless, such an angle of rotation does not always exist for certain matrices. We provide an example below where the angle of rotation does not exist, and hence we cannot rotate the left dominant eigenvectors so that all their components lie in the permissible region.

Example 1. Consider a matrix MM where,

M=[35−10ι030+10ι100−20ι−35+20ι001−4ι64+4ι].M=\begin{bmatrix} 35-10\iota& 0 & 30+10\iota\\ 100-20\iota& -35+20\iota& 0\\ 0 & 1-4\iota& 64+4\iota \end{bmatrix}.

The eigenvalues of matrix, MM are spec⁡(M)={−33.56+19ι,32.56−5ι,65}\operatorname{spec}(M)=\{-33.56+19\iota,32.56-5\iota,65\}. The dominant right eigenvector vv is real-valued, whereas the normalized left eigenvector is, wH=[0.019−0.22ι 0.04−0.05ι 1.67+0.2ι]w^{H}=[0.019-0.22\iota\ 0.04-0.05\iota\ 1.67+0.2\iota], and the left eigenvector is zH=[−0.009+0.12ι 0.017+0.036ι 0.99]z^{H}=[-0.009+0.12\iota\ 0.017+0.036\iota\ 0.99]. Both the left eigenvectors do not have the property of real dominance when we geometrically imagine each element of wHw^{H} and zHz^{H} as a vector in the complex plane. Then, none of these vectors lie in the region [−45∘,45∘][-45^\circ,45^\circ]. Also, there does not exist a possible angle of rotation ψ\psi since ∠(wH)=[−85∘ −55.2∘ 9.4∘]\angle(w^{H})=[-85^\circ\ -55.2^\circ\ 9.4^\circ] and ∠(zH)=[129.5∘ 64.7∘ 0∘]\angle(z^{H})=[129.5^\circ\ 64.7^\circ\ 0^\circ].

The eigenvectors of complex matrices allow for a degree of rotation, resulting in a broader set of matrices. However, the existence of a rotation angle that ensures real dominance is not always guaranteed. The following lemma addresses this gap and ensures the existence of an angle of rotation.

Lemma 2. Consider a vector ‘mm’ with its components in polar form. Let ϕ1\phi_1 and ϕ2\phi_2 be the smallest and largest phase angles of all the components. Then, there does not exist an angle ‘ψ\psi’ such that meιψm e^{\iota\psi} lies in the region formed by [−π4,π4]\left[-\frac{\pi}{4},\frac{\pi}{4}\right] if ∣(ϕ1−ϕ2)∣≥π2|\left(\phi_1-\phi_2\right)|\ge\frac{\pi}{2} radians.

Using Lemma 2, we conclude that if the components of dominant eigenvector of lie away from each other, then it is not possible to rotate them so that all the components remain in the desired region. Consequently, it cannot be guaranteed that the matrix exhibits rEEP.

The next result discusses the significance of real dominance for establishing rEEP matrices.

Theorem 1. Let M∈Cn×nM\in\mathbb{C}^{n\times n}. Suppose the normalized dominant right and left eigenvectors vv and ww satisfy R(v)≥∣I(v)∣\mathfrak{R}(v)\ge|\mathfrak{I}(v)| and R(w)≥∣I(w)∣\mathfrak{R}(w)\ge|\mathfrak{I}(w)|. If (M+dI)(M+dI) and (M+dI)H∈PF(M+dI)^{H}\in\mathcal{PF}, then R((M+dI)keιθ)>0\mathfrak{R}((M+dI)^{k}e^{\iota\theta})>0 for all k≥k0k\ge k_0 for some value of θ\theta and d≥0d\ge0.

Proof. We have (M+dI)(M+dI) and (M+dI)H∈PF(M+dI)^{H}\in\mathcal{PF}. Consider the case where the magnitude of real part of all the components of unnormalized left eigenvector zz is greater than or equal to its imaginary part. We consider two cases.

Case (i): Analysis under the condition R(z)≥∣I(z)∣\mathfrak{R}(z)\ge|\mathfrak{I}(z)|

The Jordan decomposition of a matrix is given as M=VJWHM=VJW^{H}, where JJ is the Jordan matrix consisting of all the eigenvalues. Let λ1\lambda_1 be the dominant eigenvalue in the square matrix JJ. Then, MM is partitioned as:

M=[v∣Vn,n−1][λ100Jn−1][wHWn−1,n],M=[v\mid V_{n,n-1}] \begin{bmatrix} \lambda_1 & 0\\ 0 & J_{n-1} \end{bmatrix} \begin{bmatrix} w^{H}\\ W_{n-1,n} \end{bmatrix},

where vv and ww are the right and left dominant eigenvectors of MM, normalized such that wHv=1w^{H}v=1. Since (M+dI)(M+dI) and (M+dI)H∈PF(M+dI)^{H}\in\mathcal{PF}, this implies R(v)>0\mathfrak{R}(v)>0 and R(w)>0\mathfrak{R}(w)>0 from (2). As per hypothesis (M+dI)(M+dI) and (M+dI)H∈PF(M+dI)^{H}\in\mathcal{PF},

(M+dI)k(λ1+d)k=[v∣Vn−1,n][100Jn−1′(λ1+d)k][wHWn−1,n],(3)\frac{(M+dI)^k}{(\lambda_1+d)^k}=[v\mid V_{n-1,n}] \begin{bmatrix} 1 & 0\\ 0 & \frac{J'_{n-1}}{(\lambda_1+d)^k} \end{bmatrix} \begin{bmatrix} w^{H}\\ W_{n-1,n} \end{bmatrix}, \tag*{(3)}

where Jn−1′=diag⁡(λ2+d,…,λn+d)J'_{n-1}=\operatorname{diag}(\lambda_2+d,\ldots,\lambda_n+d) and λi∈spec⁡(M)\lambda_i\in\operatorname{spec}(M).

We have z=cz=αeιθzz=c z=\alpha e^{\iota\theta}z where α∈R+\alpha\in\mathbb{R}_{+}, as both ww and zz are dominant left eigenvectors of MM. Then, zHv=eιθαz^{H}v=\frac{e^{\iota\theta}}{\alpha} has positive real part as R(v)≥∣I(v)∣\mathfrak{R}(v)\ge|\mathfrak{I}(v)| and R(z)≥∣I(z)∣\mathfrak{R}(z)\ge|\mathfrak{I}(z)| holds with the strong complex PF property. For some θ∈R\theta\in\mathbb{R}, the following holds by substituting wH=czH=αe−ιθzHw^{H}=cz^{H}=\alpha e^{-\iota\theta}z^{H}:

lim⁡k→∞(M+dI)k(λ1+d)k=vwH=αve−ιθzH,(4)\lim_{k\to\infty}\frac{(M+dI)^k}{(\lambda_1+d)^k}=vw^{H}=\alpha ve^{-\iota\theta}z^{H}, \tag*{(4)}
lim⁡k→∞(M+dI)keιθ(λ1+d)k=αvzH.(5)\lim_{k\to\infty}\frac{(M+dI)^k e^{\iota\theta}}{(\lambda_1+d)^k}=\alpha vz^{H}. \tag*{(5)}

Thus, due to the assumption on the eigenvector zz, we conclude R(αvzH)>0n×n\mathfrak{R}(\alpha vz^{H})>\mathbf{0}_{n\times n} which in turn ensures R((M+dI)keιθ)>0\mathfrak{R}((M+dI)^k e^{\iota\theta})>0. Case (ii): Analysis under the condition: ℜ(z)≱∣ℑ(z)∣\Re(z) \not\ge|\Im(z)|

Since the real parts of the components of vector zz are less than their imaginary counterparts, we choose a real ψ\psi such that ℜ(zeiψ)≥∣ℑ(zeiψ)∣\Re(ze^{\mathrm{i}\psi}) \ge|\Im(ze^{\mathrm{i}\psi})| holds. Therefore, w=αei(θ−ψ)zw = \alpha e^{\mathrm{i}(\theta-\psi)}z gives wH=αe−i(θ−ψ)zHw^{H} = \alpha e^{-\mathrm{i}(\theta-\psi)}z^{H}, where the angle θ\theta is the phase difference between both the left eigenvectors. From (4),

lim⁡k→∞(M+dI)k(λ1+d)k=vwH,\lim_{k\to\infty}\frac{(M+dI)^k}{(\lambda_1+d)^k}=vw^{H},

Substituting wH=αe−i(θ−ψ)zHw^{H}=\alpha e^{-\mathrm{i}(\theta-\psi)}z^{H}, we get

lim⁡k→∞(M+dI)k(λ1+d)k=vwH=αe−i(θ−ψ)vzH ∀ ψ∈[−ψa,ψb],\lim_{k\to\infty}\frac{(M+dI)^k}{(\lambda_1+d)^k}=vw^{H}=\alpha e^{-\mathrm{i}(\theta-\psi)}vz^{H}\ \forall\ \psi\in[-\psi_a,\psi_b],

where [−ψa,ψb][-\psi_a,\psi_b] contains values of the angles of rotation by which we may rotate the unnormalized left eigenvector such that ℜ(zeiψ)≥∣ℑ(zeiψ)∣\Re(ze^{\mathrm{i}\psi})\ge|\Im(ze^{\mathrm{i}\psi})| follows.

Finally,

lim⁡k→∞(M+dI)kei(θ−ψ)(λ1+d)k=αvzH ∀ ψ∈[−ψa,ψb].(6)\lim_{k\to\infty}\frac{(M+dI)^k e^{\mathrm{i}(\theta-\psi)}}{(\lambda_1+d)^k}=\alpha vz^{H}\ \forall\ \psi\in[-\psi_a,\psi_b]. \tag*{(6)}

From (6), we conclude that the real part of (M+dI)kei(θ−ψ)(M+dI)^k e^{\mathrm{i}(\theta-\psi)} is eventually positive as the magnitude of real part of rotated left eigenvector zHz^{H} is greater than the magnitude of imaginary part in the eigenvectors. □

Theorem 1 shows that real dominance is a property of the chosen eigenvector, not of the eigenspace. A dominant left eigenvector of a complex matrix is determined only up to a factor, so even if wHw^{H} fails real dominance, the rotated eigenvector zH=eiψwHz^{H}=e^{\mathrm{i}\psi}w^{H} may satisfy it [14]. So, a complex matrix that is not eventually positive may become so after a suitable rotation. For real matrices, and for complex matrices with real dominant eigenvectors, the only freedom is a sign, so no such rotation exists.

The next result provides equivalent conditions to characterize rEEP matrices and clarifies their relationship with the class of PF matrices.

Theorem 2. Let M∈Cn×nM\in\mathbb{C}^{n\times n}. Assume ℜ(v)≥∣ℑ(v)∣\Re(v)\ge|\Im(v)| and ℜ(wH)≥∣ℑ(wH)∣\Re(w^{H})\ge|\Im(w^{H})|, where v,wHv,w^{H} are the normalized dominant right and left eigenvectors of MM, respectively. Then, the following statements are equivalent for some scalar d≥0d\ge0:

(i) (M+dI)(M+dI) and (M+dI)H∈PF(M+dI)^{H}\in\mathcal{PF}.

(ii) (M+dI)(M+dI) is real eventually positive for all k≥k0k\ge k_0.

(iii) MM is rEEP.

Proof. (i) ⟹\Longrightarrow (ii): The Jordan decomposition of a matrix is given as M=VJWHM=VJW^{H}, where JJ is the Jordan matrix consisting of all the eigenvalues. Let λ1\lambda_1 be the dominant eigenvalue in the square matrix JJ. Then, MM is partitioned as:

M=[v∣Vn,n−1][λ100Jn−1][wHWn−1,n].M=[v\mid V_{n,n-1}] \begin{bmatrix} \lambda_1 & 0\\ 0 & J_{n-1} \end{bmatrix} \begin{bmatrix} w^{H}\\ W_{n-1,n} \end{bmatrix}.

where vv and ww are the right and left dominant eigenvectors of MM, normalised such that wHv=1w^{H}v=1. Since (M+dI)(M+dI) and (M+dI)H∈PF(M+dI)^{H}\in\mathcal{PF}, this implies ℜ(v)>0\Re(v)>0 and ℜ(w)>0\Re(w)>0 from (2). Because (M+dI)(M+dI) and (M+dI)H∈PF(M+dI)^{H}\in\mathcal{PF},

(M+dI)k(λ1+d)k=[v∣Vn−1,n][100(Jn−1′)k(λ1+d)k][wHWn−1,n],(7)\frac{(M+dI)^k}{(\lambda_1+d)^k}=[v\mid V_{n-1,n}] \begin{bmatrix} 1 & 0\\ 0 & \dfrac{(J'_{n-1})^k}{(\lambda_1+d)^k} \end{bmatrix} \begin{bmatrix} w^{H}\\ W_{n-1,n} \end{bmatrix}, \tag*{(7)}

where Jn−1′=diag⁡(λ2+d,…,λn+d)J'_{n-1}=\operatorname{diag}(\lambda_2+d,\ldots,\lambda_n+d) and λi∈spec⁡(M)\lambda_i\in\operatorname{spec}(M). Thus, by taking the limit k→∞k\to\infty on the either sides of (7) and solving yields,

lim⁡k→∞(M+dI)k(λ1+d)k=vwH.\lim_{k\to\infty}\frac{(M+dI)^k}{(\lambda_1+d)^k}=vw^{H}.

Considering the real part, we have

lim⁡k→∞ℜ((M+dI)k(λ1+d)k)=ℜ(αvwH)>0n×n.(8)\lim_{k\to\infty}\Re\left(\frac{(M+dI)^k}{(\lambda_1+d)^k}\right)=\Re(\alpha vw^{H})>0_{n\times n}. \tag*{(8)}

From (8) and the definition of rEP matrices, we conclude that (M+dI)(M+dI) is rEP.

(ii) ⟹\Longrightarrow (i): Assume that ℜ((M+dI)k)>0\Re((M+dI)^k)>0 and define B=(M+dI)B=(M+dI). The scalar d≥0d\ge0 can be chosen such that BB is a non-nilpotent matrix [14]. Thus, it follows that B,BH∈PFB,B^{H}\in\mathcal{PF} which can be proved on the same lines as in Theorem 2.3, [14].

(ii) ⟹\Longrightarrow (iii): We have ℜ((M+dI)k)>0n×n\Re((M+dI)^k)>0_{n\times n}. Since,

e(M+dI)τ=I+(M+dI)τ+(M+dI)2τ22!+…+(M+dI)k0τk0k0!+h.o.t.(9)e^{(M+dI)\tau}=I+(M+dI)\tau+\frac{(M+dI)^2\tau^2}{2!}+\ldots+\frac{(M+dI)^{k_0}\tau^{k_0}}{k_0!}+\text{h.o.t.} \tag*{(9)}

At a large time instant (τ≥τ0)(\tau\ge\tau_0), the higher order terms (h.o.t.)(h.o.t.) dominate and hence the whole summation becomes real positive as (M+dI)(M+dI) is rEP. On the other hand, eMτ=e−dτe(M+dI)τe^{M\tau}=e^{-d\tau}e^{(M+dI)\tau}, where e−dτ≥0e^{-d\tau}\ge0. For all τ≥τ0\tau\ge\tau_0, the h.o.t.h.o.t. in (9) dominate the sum of the terms until and including k0k_0. Note that k0k_0 is depends on τ0\tau_0. Hence, (M+dI)(M+dI) is rEEP, as the whole summation e(M+dI)τe^{(M+dI)\tau} is positive. Also, MM inherits the property of rEEP from (M+dI)(M+dI). This follows directly from the fact that eMτe^{M\tau} can be expressed as e−dτe(M+dI)τe^{-d\tau}e^{(M+dI)\tau} where e−dτe^{-d\tau} is positive. Hence, MM is rEEP.

(iii) ⟹\Longrightarrow (ii): Refer to Theorem 3, [1]. □

Theorem 2 states that a matrix MM is rEEP if the translated matrix (M+dI)(M+dI) follows the strong complex PF property. This analysis directly extends to Laplacian matrices. Consequently, for any Laplacian matrix satisfying the rEEP property, the matrix, B=dI−LB=dI-L satisfies the strong complex PF property. The example below considers a translated Laplacian matrix and illustrates the possible rotation in dominant eigenvectors, discussed in previous results:

Example 2. Consider

M=[0.75−1.125i0.588−0.147i1.073+1.419i−0.55−i2.176+0.705i2.147−0.41i2.75−0.125i0.58−0.147i−0.926+0.419i].M= \begin{bmatrix} 0.75-1.125\mathrm{i} & 0.588-0.147\mathrm{i} & 1.073+1.419\mathrm{i}\\ -0.55-\mathrm{i} & 2.176+0.705\mathrm{i} & 2.147-0.41\mathrm{i}\\ 2.75-0.125\mathrm{i} & 0.58-0.147\mathrm{i} & -0.926+0.419\mathrm{i} \end{bmatrix}.

For this matrix, λ1=3\lambda_1=3 is the dominant eigenvalue with a complex-valued dominant right eigenvector. But the dominant left eigenvector wH=[0.5−0.43i  0.65+0.14i  0.65+0.14i]w^{H}=[0.5-0.43\mathrm{i}\ \ 0.65+0.14\mathrm{i}\ \ 0.65+0.14\mathrm{i}], ∠(wH)=[−41.28∘ 12.7∘ 12.4∘]\angle(w^{H})=[-41.28^\circ\ 12.7^\circ\ 12.4^\circ] and $\angle(z^{H})=[0^\circ 53.1° − 53.1°. Geometrically, none of these vectors lie in the region [−45°, 45°]. Thus, the assumption that R(zH)≥∣I(zH)∣\mathfrak{R}(z^{H}) \geq|\mathfrak{I}(z^{H})| is violated. However, assume ψ=20∘\psi= 20^\circ. Then, ∠(zHeiψ)=[20∘−33.1∘−33.1∘]\angle(z^{H}e^{\mathrm{i}\psi}) = [20^\circ- 33.1^\circ- 33.1^\circ], and again resorting to the geometric intuition, we note that zz follows the property of real dominance.

Analysis of Complex-valued Pseudoinverse Laplacians

The rEEP property ensures consensus in complex-valued Laplacian flows [1]. In contrast, the rEEP property in the case of pseudoinverse Laplacian flows implies consensus solely for undirected and weight-balanced digraphs [12]. Hence, the spectral properties and the steady-state behavior of pseudoinverse Laplacian flows in non-weight-balanced digraphs remain unexplored. We now address this gap. We examine the eigenspace corresponding to the zero eigenvalue of the Laplacian under the pseudoinverse operation. This analysis subsequently helps us to comment on the convergence properties of pseudoinverse Laplacian flows.

Lemma 3. The set of eigenvectors corresponding to the zero eigenvalue of the Laplacian matrix remains same as of the pseudoinverse Laplacian matrix.

Proof. For non weight-balanced digraphs, we have L1n=0nL\mathbf{1}_{n} = \mathbf{0}_{n} and wHL=0nT\mathbf{w}^{H}L = \mathbf{0}_{n}^{T}. Further, from Def.3,

(L†)H=L†L,(10)(L^{\dagger})^{H} = L^{\dagger}L, \tag*{(10)}
(LL†)H=LL†.(11)(LL^{\dagger})^{H} = LL^{\dagger}. \tag*{(11)}

Using (10), 1nTL†=1nTL†LL†=wH(L†L)HL†\mathbf{1}_{n}^{T}L^{\dagger} = \mathbf{1}_{n}^{T}L^{\dagger}LL^{\dagger} = \mathbf{w}^{H}(L^{\dagger}L)^{H}L^{\dagger},

0nT(L†)HL†=0nT,\mathbf{0}_{n}^{T}(L^{\dagger})^{H}L^{\dagger} = \mathbf{0}_{n}^{T},

Similarly using (11), L†w=0nL^{\dagger}\mathbf{w} = \mathbf{0}_{n}.

The following result states the relation between the Laplacian and pseudoinverse Laplacian matrices using Lemma 3.

Theorem 3. Consider a strongly connected digraph G(A)\mathcal{G}(\mathcal{A}) and for 1≤i,j≤n1 \leq i,j \leq n. Let v=α1n,α∈C\mathbf{v} = \alpha\mathbf{1}_{n},\alpha\in\mathbb{C} and w∈Cn\mathbf{w} \in\mathbb{C}^{n} be the right and left eigenvectors associated with the zero eigenvalue of LL, respectively. Suppose the phase angles, ϕvi=∠(v)i\phi_{v_i} = \angle(v)_i and ϕwi=∠(w)i\phi_{w_i} = \angle(w)_i lie in [−45∘,45∘][-45^\circ,45^\circ]. Then, the following statements are equivalent:

  1. −L-L is rEEP.

  2. −L†-L^{\dagger} is marginally stable.

  3. −L†-L^{\dagger} is rEEP.

Proof. (i) ⟹\Longrightarrow (ii): From Lemma 1, if −L-L is rEEP, then it is equivalently marginal stable. Hence, spec⁡(−L)={0,−λ2,−λ3,…,−λn}\operatorname{spec}(-L) = \{0,-\lambda_{2},-\lambda_{3},\ldots,-\lambda_{n}\}. Then, we have the spectrum of L†L^{\dagger} given in (12),

spec⁡(−L†)={0,−1/λ2,−1/λ3,…,−1/λn},(12)\operatorname{spec}(-L^{\dagger}) = \{0,-1/\lambda_{2},-1/\lambda_{3},\ldots,-1/\lambda_{n}\}, \tag*{(12)}

where λ1,λ2,…,λn\lambda_{1},\lambda_{2},\ldots,\lambda_{n} are the eigenvalues of LL [19]. Hence, L†L^{\dagger} is marginally stable.

(ii) ⟹\Longrightarrow (iii): From the previous proof, −L†-L^{\dagger} is marginally stable. Using Jordan canonical form and substituting λ1=0\lambda_{1} = 0,

−L†=[v(1)Vn,n−1][000−D−1][w(1)HWn−1,n],-L^{\dagger} = \begin{bmatrix} \mathbf{v}^{(1)} & V_{n,n-1} \end{bmatrix} \begin{bmatrix} 0 & 0 \\ 0 & -D^{-1} \end{bmatrix} \begin{bmatrix} \mathbf{w}^{(1)H} \\ W_{n-1,n} \end{bmatrix},

where −D-D is the diagonal matrix containing eigenvalues (except 0) of −L-L. Then,

e−L†t=[v(1)Vn,n−1][100e−tλj][w(1)HWn−1,n](13)e^{-L^{\dagger}t} = \begin{bmatrix} \mathbf{v}^{(1)} & V_{n,n-1} \end{bmatrix} \begin{bmatrix} 1 & 0 \\ 0 & e^{-\frac{t}{\lambda_{j}}} \end{bmatrix} \begin{bmatrix} \mathbf{w}^{(1)H} \\ W_{n-1,n} \end{bmatrix} \tag*{(13)}

(13) holds for all the cases of degenerate and non-degenerate eigenvalues as the exponential decay is faster than the growth of the corresponding polynomial. In case of −L†-L^{\dagger}, the eigenvectors corresponding to zero eigenvalue remain same. Hence, −L†-L^{\dagger} is rEEP.

(iii) ⟹\Longrightarrow (ii): This can be proved on the same lines as in the proof of Theorem 3 of [1]. □

An example is provided to validate the previous result.

Example 3. Consider LL and L†L^{\dagger}, whose real parts are shown in Figure 1 via heatmaps.

Heatmaps showing the matrix structure of L and L dagger for real and imaginary parts

Figure 1. Heatmaps showing the matrix structure of LL and L†L^{\dagger} for real and imaginary parts respectively

The following remarks are in order for the Laplacians, pseudoinverse Laplacians, and real parts of exponential maps shown in Figure 2a and Figure 2b, respectively. First, notice that the real parts vary from positive to negative values for the Laplacian as well as the pseudoinverse Laplacian. Since the spectrum of the Laplacian matrix has a simple zero eigenvalue, the negated Laplacian matrix as well as the negated pseudoinverse Laplacian is rEEP. Thus, it is evident from Figure 2 that the rEEP property is preserved under the operation of pseudoinverse for Laplacian matrices, as proved in Theorem 3. In contrast to the results in [12], the pseudoinverse Laplacian flows do not converge in the case of non-weight-balanced digraphs. It happens so because from Theorem 3, we conclude that the property of rEEP is conserved under the pseudoinverse operation. But referring x(t)=e−L†tx0x(t)=e^{-L^{\dagger}t}x_0 as the solution of pseudoinverse Laplacian flow given by x˙(t)=−L†x(t)\dot{x}(t)=-L^{\dagger}x(t):

Figure 2. The matrix structure of R(e−Lt)\mathfrak{R}(e^{-Lt}) and R(e−L†t)\mathfrak{R}(e^{-L^{\dagger}t}) at t=3t = 3 and t=4t = 4 respectively. Values on the colorbar represent the entire range that encompasses all the values of a matrix, and all of these values are positive implying rEEP property.

lim⁡t→∞e−L†t=wvH⋅x0\lim_{t\to\infty}e^{-L^{\dagger}t}=w v^{H}\cdot x_0

where vH⋅x0v^{H}\cdot x_0 is constant =k=k.

lim⁡t→∞x(t)=kw⟹x(t)∉span⁡{1n} because w∉span⁡{1n}.\lim_{t\to\infty}x(t)=kw\Longrightarrow x(t)\notin\operatorname{span}\{1_n\}\text{ because }w\notin\operatorname{span}\{1_n\}.

Next, we demonstrate that the eigenspace, along with the zero eigenvalue remains same when we rotate all the elements of the Laplacian matrix with some angle of rotation, ψ\psi. In particular, a global phase rotation of the Laplacian matrix rotates its non-zero spectrum but leaves the zero eigenvalue and the eigenspace completely invariant.

Corollary 1. The set of eigenvectors corresponding to the zero eigenvalue of remains invariant under rotation of the Laplacian matrix; i.e., if all the elements of the matrix are rotated by a rotation angle of ψ\psi.

Proof. Let L∈Cn×nL\in\mathbb{C}^{n\times n} be a complex-valued Laplacian matrix. And Lψ=eιψLL_{\psi}=e^{\iota\psi}L be the rotated Laplacian with an angle ψ∈C\psi\in\mathbb{C} and t∈Rt\in\mathbb{R}. Let v∈Cnv\in\mathbb{C}^{n} such that Lv=0Lv=0. It then follows that

Lψv=eιψLv=0(14)L_{\psi}v=e^{\iota\psi}Lv=0 \tag*{(14)}

Hence vv is the eigenvector corresponding to the zero eigenvalue of LψL_{\psi}. Conversely, because eιψ≠0e^{\iota\psi}\ne0, Lψv=0L_{\psi}v=0 implies Lv=0Lv=0. Thus ker⁡(Lψ)=ker⁡(L)\ker(L_{\psi})=\ker(L). In other words, eigenspaces corresponding to the rotated and actual Laplacians coincide. Similar argument holds for left eigenspaces corresponding to the zero eigenvalue as well.

Corollary 1 implies that the rEEP property is also preserved under a uniform rotation. This is because the rEEP property depends only on the eigenvectors corresponding to the zero eigenvalue of the complex-valued Laplacian.

Remark 1. The property of rEEP is not always invariant under the global phase rotation. This is because of the change in the non-zero eigenvalues in the spectrum. While the zero eigenvalue and its eigenspace remain unchanged, the non-zero eigenvalues are rotated by the angle, ψ\psi, which can shift them in the left half of the s-plane. Thus, the rEEP property is preserved only if the non-zero eigenvalues of the negated Laplacian remain in the right half of the complex plane after rotation. Hence, the property of rEEP depends on the full spectrum and does not guarantee consensus in all cases.

CONCLUSIONS AND FUTURE WORK

We extend real eventual exponential positivity (rEEP) to pseudoinverses of complex-valued Laplacians on digraphs. We show that rEEP is preserved under the pseudoinverse, and that it is determined entirely by the relative phases of the edge-weights. We also show that the null space of the Laplacian is invariant under a uniform rotation of the edge-weights. Altogether, these results provide a foundation for further analysis of consensus and stability in complex-valued networked systems. Our future studies will include deriving the conditions to achieve rEEP Laplacians, without imposing the real dominance on the eigenvectors corresponding to the zero eigenvalue of complex-valued Laplacians.

References

  1. [1]Aditi Saxena, Twinkle Tripathy, and Rajasekhar Anguluri. Laplacian flows in complex-valued directed networks: Analysis, design, and consensus. arXiv preprint arXiv:2509.04196, 2025.arxiv.org/abs/2509.04196
  2. [2]Lucas Böttcher and Mason A Porter. Complex networks with complex weights. Physical Review E, 109(2):024314, 2024.DOI
  3. [3]Jiu-Gang Dong and Qin Lin. Laplacian matrices of general complex weighted directed graphs. Linear Algebra and its Applications, 510:1–9, 2016.DOI
  4. [4]Jiu-Gang Dong and Li Qiu. Complex Laplacians and applications in multi-agent systems. arXiv preprint arXiv:1406.1862, 2014.DOI
  5. [5]Florian Dorfler and Francesco Bullo. Synchronization and transient stability in power networks and nonuniform Kuramoto oscillators. SIAM Journal on Control and Optimization, 50(3):1616–1642, 2012.
  6. [6]Reza Olfati-Saber, J Alex Fax, and Richard M Murray. Consensus and cooperation in networked multi-agent systems. Proceedings of the IEEE, 95(1):215–233, 2007.
  7. [7]Aditi Saxena and Twinkle Tripathy. How do emotions spread and affect decision-making in social networks? an opinion dynamics perspective. arXiv preprint arXiv:2610.22334, 2026.arxiv.org/abs/2610.22334
  8. [8]Angela Fontan, Lingfei Wang, Yiguang Hong, Guodong Shi, and Claudio Altafini. Multiagent consensus over time-invariant and time-varying signed digraphs via eventual positivity. IEEE Transactions on Automatic Control, 68(9):5429–5444, 2022.DOI
  9. [9]Claudio Altafini. Investigating stability of Laplacians on signed digraphs via eventual positivity. In 2019 IEEE 58th Conference on Decision and Control (CDC), pages 5044–5049. IEEE, 2019.DOI
  10. [10]Angela Fontan and Claudio Altafini. On the properties of Laplacian pseudoinverses. In 2021 60th IEEE Conference on Decision and Control (CDC), pages 5538–5543. IEEE, 2021.arxiv.org/abs/2109.14587
  11. [11]Zhiyun Lin, Wei Ding, Gangfeng Yan, Changbin Yu, and Alessandro Giua. Leader–follower formation via complex Laplacian. Automatica, 49(6):1900–1906, 2013.DOI
  12. [12]Aditi Saxena, Twinkle Tripathy, and Rajasekhar Anguluri. Are the flows of complex-valued laplacians and their pseudoinverses related? In 2025 European Control Conference (ECC), pages 653–658, 2025.arxiv.org/abs/2411.09254
  13. [13]Yu Tian, Mason A Porter, and Lucas Böttcher. Generalizing perron–frobenius theory and eigenvector-based centralities to networks with complex edge weights. arXiv preprint arXiv:2606.12026, 2026.arxiv.org/abs/2606.12026
  14. [14]Dimitrios Noutsos and Richard S Varga. On the Perron–Frobenius theory for complex matrices. Linear algebra and its applications, 437(4):1071–1088, 2012.DOI
  15. [15]Aditi Saxena, Twinkle Tripathy, and Rajasekhar Anguluri. Real eventual exponential positivity of complex-valued laplacians: Applications to consensus in multi-agent systems. In 2024 Tenth Indian Control Conference (ICC), pages 226–231, 2024.arxiv.org/abs/2410.13700
  16. [16]Yu Tian and Renaud Lambiotte. Structural balance and random walks on complex networks with complex weights. SIAM Journal on Mathematics of Data Science, 6(2):372–399, 2024.arxiv.org/abs/2307.01813
  17. [17]João Barata, Carlos Alves, and Mahir Saleh Hussein. The Moore–Penrose pseudoinverse: A tutorial review of the theory. Brazilian Journal of Physics, 42:146–165, 2012.DOI
  18. [18]Francesco Bullo. Lectures on network systems, volume 1. CreateSpace, 2018.
  19. [19]Angela Fontan and Claudio Altafini. Pseudoinverses of signed Laplacian matrices. SIAM Journal on Matrix Analysis and Applications, 44(2):622–647, 2023.DOI

Paper details

vlnc:W_vs5gd5x7
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 6

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents