Combinatorial Directed Laplacian

• Combinatorial Directed Laplacian is a discrete operator on directed graphs defined via in-degree and out-degree formulations, ensuring a constant kernel for diffusion analysis.
• Its rich spectral properties, including biorthogonal decompositions and Jordan forms, enable robust graph Fourier transforms and harmonic analysis on non-symmetric networks.
• The operator underpins practical applications such as signal processing, ranking, and deep learning while supporting combinatorial factorizations like spanning tree enumerators.

A combinatorial directed Laplacian is a discrete differential operator that extends the classical Laplacian from undirected graphs to directed (asymmetric) network structures. Its operator-theoretic and spectral properties underlie much of contemporary spectral and harmonic analysis on directed graphs, supporting applications in signal processing, ranking, flow dynamics, and deep learning. The structure of the combinatorial directed Laplacian admits a range of algebraic forms, variational characterizations, and combinatorial factorizations found across disparate research traditions and problem domains.

1. Definitions and Operator Formulations

On a directed graph $\mathcal{G}=(\mathcal{V},W)$ with $N$ nodes and a (possibly non-symmetric) weight matrix $W\in\mathbb{R}^{N\times N}$ or $\mathbb{C}^{N\times N}$, the two most prevalent conventions for the combinatorial directed Laplacian $L$ are:

• In-degree Laplacian:

$$L = D_{\rm in} - W, \quad (D_{\rm in})_{ii} = \sum_j W_{ij}$$

as in the definition used for graph Fourier transforms on directed graphs (Singh et al., 2016).

• Out-degree Laplacian:

$$L = D_{\rm out} - W, \quad (D_{\rm out})_{ii} = \sum_j W_{ji}$$

which also appears centrally in recent harmonic analysis frameworks (Gokavarapu, 15 Dec 2025).

Both forms yield square matrices whose off-diagonal terms encode negative edge weights and whose diagonals record the respective degree sums, ensuring that each row (in in-degree convention) or column (in out-degree convention) sums to zero. For a graph with adjacency $A$, these Laplacians generally satisfy $L\mathbf{1}=0$, so the constant vector lies in the kernel—a key feature for diffusion and stationary analysis.

For weighted digraphs with possible complex weights or self-loops, the standard Laplacian is $L = D_{\rm out} - A$, with $N$0 collecting out-degree sums defined by modulus or absolute values if complex weights are allowed (Adhikari et al., 2012). Variants such as the signless Laplacian $N$1 and signed Laplacians, as well as deformations (e.g., magnetic or dilation Laplacians), generalize this structure to incorporate orientation, flow, or "magnetic" phases (Dever, 2016, Fanuel et al., 2015, Geisler et al., 2023).

2. Algebraic and Spectral Properties

In stark contrast to the undirected Laplacian, which is always real symmetric and positive semidefinite, the combinatorial directed Laplacian is generally non-Hermitian and may be non-diagonalizable. Its spectrum is fundamentally richer:

• Zero Row Sums and Kernels: For both conventions, the constant vector is an eigenvector with eigenvalue zero: $N$2 (Singh et al., 2016, Gokavarapu, 15 Dec 2025, Adhikari et al., 2012).
• Asymmetry and Diagonalizability: $N$3 is typically asymmetric ($N$4). If $N$5 is diagonalizable, one obtains left and right eigenbases leading to a biorthogonal spectral decomposition (Gokavarapu, 15 Dec 2025).
• Jordan Form: In the non-diagonalizable case, spectral analysis proceeds via the Jordan decomposition $N$6 (Singh et al., 2016).
• Spectrum Location: For nonnegative weights, all eigenvalues $N$7 of $N$8 satisfy $N$9 (Singh et al., 2016).
• Condition Number and Non-normality: The condition number $W\in\mathbb{R}^{N\times N}$0 of the eigenvector matrix governs the energy expansion/contraction in the spectral domain and directly influences the stability of transforms and filters. A normal (but not necessarily symmetric) Laplacian allows unitary bases ($W\in\mathbb{R}^{N\times N}$1), but generic directed graphs lead to $W\in\mathbb{R}^{N\times N}$2, limiting the quality of Fourier-type decompositions (Gokavarapu, 15 Dec 2025).

The following table highlights the algebraic distinctions:

Property	Undirected Laplacian $W\in\mathbb{R}^{N\times N}$3	Directed Laplacian $W\in\mathbb{R}^{N\times N}$4
Symmetry	$W\in\mathbb{R}^{N\times N}$5	$W\in\mathbb{R}^{N\times N}$6
Diagonalizability	Always (real spectrum)	Possibly defective
Spectrum	$W\in\mathbb{R}^{N\times N}$7	$W\in\mathbb{R}^{N\times N}$8 in $W\in\mathbb{R}^{N\times N}$9
Orthonormal Eigenbasis	Yes	Only if $\mathbb{C}^{N\times N}$0 is normal

3. Harmonic Analysis and Graph Fourier Transforms

On directed graphs, harmonic analysis builds on Laplacian eigenstructure. The core notions are:

• Jordan Eigenvectors as Harmonics: The columns of $\mathbb{C}^{N\times N}$1 in the Jordan decomposition are the graph harmonics, with the corresponding eigenvalues $\mathbb{C}^{N\times N}$2 interpreted as graph frequencies (Singh et al., 2016).
• Graph Fourier Transform (GFT): For $\mathbb{C}^{N\times N}$3, the GFT is $\mathbb{C}^{N\times N}$4. The inverse is $\mathbb{C}^{N\times N}$5. If $\mathbb{C}^{N\times N}$6 is non-normal, biorthogonal GFTs use dual bases:

$\mathbb{C}^{N\times N}$7

with energy in the GFT domain controlled by $\mathbb{C}^{N\times N}$8 and subjected to precise Parseval-type inequalities (Gokavarapu, 15 Dec 2025).

• Frequency Ordering: The total variation of the $\mathbb{C}^{N\times N}$9th harmonic $L$0 is $L$1, so the natural frequency ordering is in terms of $L$2 (Singh et al., 2016).
• Non-orthogonality: Except in the normal case, the Fourier harmonics are not orthogonal and the transform does not preserve $L$3-norm, making spectral filtering sensitive to $L$4 (Gokavarapu, 15 Dec 2025).

A robust alternative for Hermitian Laplacians (e.g., magnetic Laplacians) uses complex-valued but symmetric operators, guaranteeing real spectra and an orthonormal eigenbasis even for directed edge structures (Geisler et al., 2023). This formulation supports deep learning applications requiring stable spectral node features.

4. Variational Principles, Total Variation, and Filtering

The combinatorial directed Laplacian generalizes discrete differentiation and energy minimization:

• Gradient and Total Variation: For a graph signal $L$5, the natural difference is $L$6. The $L$7 total variation is $L$8 (Singh et al., 2016). The $L$9 version, $$L = D_{\rm in} - W, \quad (D_{\rm in})_{ii} = \sum_j W_{ij}$$0, serves as a directed smoothness semi-norm, directly measuring signal variation along edges (Gokavarapu, 15 Dec 2025).
• Quadratic Forms: For $$L = D_{\rm in} - W, \quad (D_{\rm in})_{ii} = \sum_j W_{ij}$$1,

$$L = D_{\rm in} - W, \quad (D_{\rm in})_{ii} = \sum_j W_{ij}$$2

with generalizations handling negative or complex weights (Adhikari et al., 2012).

• Spectral Filtering: A graph filter $$L = D_{\rm in} - W, \quad (D_{\rm in})_{ii} = \sum_j W_{ij}$$3 is LSI if it commutes with the shift $$L = D_{\rm in} - W, \quad (D_{\rm in})_{ii} = \sum_j W_{ij}$$4, in which case it is a polynomial in $$L = D_{\rm in} - W, \quad (D_{\rm in})_{ii} = \sum_j W_{ij}$$5, $$L = D_{\rm in} - W, \quad (D_{\rm in})_{ii} = \sum_j W_{ij}$$6 (Singh et al., 2016). In biorthogonal frameworks, filtering is spectral-domain multiplication by $$L = D_{\rm in} - W, \quad (D_{\rm in})_{ii} = \sum_j W_{ij}$$7, with worst-case energy amplification given by the eigenbasis condition number (Gokavarapu, 15 Dec 2025).

5. Combinatorial Factorizations and the Matrix-Tree Theorem

The directed Laplacian encodes deep combinatorial information:

• Spanning Trees and Principal Minors: Determinants of principal submatrices of $$L = D_{\rm in} - W, \quad (D_{\rm in})_{ii} = \sum_j W_{ij}$$8 correspond to enumerators of rooted spanning trees, generalizing Kirchhoff's matrix-tree theorem to directed graphs (Biane et al., 2015).
• Lifting to Tree Graphs: Given $$L = D_{\rm in} - W, \quad (D_{\rm in})_{ii} = \sum_j W_{ij}$$9, one defines a tree-graph $$L = D_{\rm out} - W, \quad (D_{\rm out})_{ii} = \sum_j W_{ji}$$0 whose vertices are rooted spanning trees. The lifted Laplacian $$L = D_{\rm out} - W, \quad (D_{\rm out})_{ii} = \sum_j W_{ji}$$1 admits a canonical factorization:

$$L = D_{\rm out} - W, \quad (D_{\rm out})_{ii} = \sum_j W_{ji}$$2

where $$L = D_{\rm out} - W, \quad (D_{\rm out})_{ii} = \sum_j W_{ji}$$3 counts certain combinatorial objects defined via exploration algorithms (Biane et al., 2015).

• Random Walks and PageRank: As the deformation parameter (in dilation/magnetic Laplacians) is taken to infinity, the system interpolates to random walk or PageRank dynamics on directed graphs, connecting spectral analysis to Markov process theory (Fanuel et al., 2015).

6. Algorithmic Aspects and Applications

Linear algebraic solvers for directed Laplacian systems are critical for computing stationary distributions, personalized PageRank, and escape probabilities in non-reversible Markov chains:

• Sparsified Block Elimination: Efficient solution of $$L = D_{\rm out} - W, \quad (D_{\rm out})_{ii} = \sum_j W_{ji}$$4 for directed Laplacians (and Eulerian specializations) can be achieved via a Schur-complement-based block elimination and sparsification scheme, attaining nearly-linear time complexity (Peng et al., 2021).
• Regularity and Symmetrization: For Eulerian Laplacians (where row and column sums match), the symmetrization $$L = D_{\rm out} - W, \quad (D_{\rm out})_{ii} = \sum_j W_{ji}$$5 is positive semi-definite, enabling the use of combinatorial preconditioners and facilitating efficient parallelization (Peng et al., 2021).
• Deep Learning: The magnetic Laplacian enables principled, direction-aware positional encoding for transformer architectures on directed graphs, guaranteeing a real spectrum and global structure encoding (Geisler et al., 2023).

7. Variants, Generalizations, and Limitations

Numerous deformations and extensions have emerged:

• Magnetic Laplacians: Decorate undirected edges with complex phases encoding directionality, producing Hermitian operators with real spectra—effective for both theoretical and applied contexts (Dever, 2016, Geisler et al., 2023).
• Dilation Laplacians: Employ a continuous parameter to interpolate between undirected and strongly directed regimes, providing a powerful tool for spectral ranking (Fanuel et al., 2015).
• Hypergraph Laplacians: For directed hypergraphs, the Laplacian $$L = D_{\rm out} - W, \quad (D_{\rm out})_{ii} = \sum_j W_{ji}$$6 is symmetric but may lose positivity, stochasticity, or diagonal dominance, requiring combinatorial control for contractive heat flows and stochasticity (e.g., in equipotent/equitable cases) (Mugnolo, 20 Oct 2025).

A central limitation remains non-normality: when $$L = D_{\rm out} - W, \quad (D_{\rm out})_{ii} = \sum_j W_{ji}$$7 is far from normal, the eigenbasis becomes ill-conditioned, and filter stability and spectral approximations degrade rapidly. Design choices that promote normality or regularization are necessary for robust signal analysis and spectral learning on complex directed networks (Gokavarapu, 15 Dec 2025).

---

Key citations:

• (Singh et al., 2016): Graph Fourier Transform based on Directed Laplacian
• (Gokavarapu, 15 Dec 2025): Harmonic Analysis on Directed Networks: A Biorthogonal Laplacian Framework for Non-Normal Graphs
• (Biane et al., 2015): Laplacian matrices and spanning trees of tree graphs
• (Dever, 2016): Eigenvalue Sums of Combinatorial Magnetic Laplacians on Finite Graphs
• (Fanuel et al., 2015): Deformed Laplacians and spectral ranking in directed networks
• (Adhikari et al., 2012): Laplacian matrices of weighted digraphs represented as quantum states
• (Mugnolo, 20 Oct 2025): The heat flow driven by the Laplacian of a directed hypergraph
• (Peng et al., 2021): Sparsified Block Elimination for Directed Laplacians
• (Geisler et al., 2023): Transformers Meet Directed Graphs