[Paper Review] Reduced Google matrix

Klaus M. Frahm, Dima L. Shepelyansky|arXiv (Cornell University)|Feb 7, 2016

Quantum Mechanics and Applications3 references18 citations

TL;DR

This paper introduces the reduced Google matrix $G_R$, a novel method to model effective interactions among a selected subset of $N_r$ nodes within a large directed network of size $N$, by accounting for all indirect paths via the remaining $N_s = N - N_r$ nodes. Using a quantum scattering analogy, $G_R = G_{rr} + G_{rs}(\mathbf{1} - G_{ss})^{-1}G_{sr}$ efficiently captures long-range dependencies, enabling accurate analysis of node group dynamics without full network computation.

ABSTRACT

Using parallels with the quantum scattering theory, developed for processes in nuclear and mesoscopic physics and quantum chaos, we construct a reduced Google matrix $G_R$ which describes the properties and interactions of a certain subset of selected nodes belonging to a much larger directed network. The matrix $G_R$ takes into account effective interactions between subset nodes by all their indirect links via the whole network. We argue that this approach gives new possibilities to analyze effective interactions in a group of nodes embedded in a large directed networks. Possible efficient numerical methods for the practical computation of $G_R$ are also described.

Motivation & Objective

• To model effective interactions between a small subset of nodes in a large directed network, accounting for indirect links via the rest of the network.
• To overcome the computational infeasibility of analyzing full network dynamics when only a subset of nodes is of interest.
• To develop a mathematically rigorous and numerically efficient framework for studying group-level dynamics in complex networks.
• To draw an analogy with quantum scattering theory to interpret the reduced matrix as a propagator for effective interactions.
• To enable practical computation of $G_R$ using matrix inversion and modified damping to ensure convergence.

Proposed method

• Formulate the full Google matrix $G$ in block form: $G = \begin{pmatrix} G_{rr} & G_{rs} \\ G_{sr} & G_{ss} \end{pmatrix}$, where $r$ denotes the reduced subset and $s$ the complementary nodes.
• Derive the reduced Google matrix via $G_R = G_{rr} + G_{rs}(\mathbf{1} - G_{ss})^{-1}G_{sr}$, ensuring column stochasticity and unit eigenvalue for the PageRank vector.
• Use the geometric series expansion $ (\mathbf{1} - G_{ss})^{-1} = \sum_{k=0}^\infty G_{ss}^k $ to enable numerical computation when $ \|G_{ss}\| < 1 $.
• Introduce a modified Google matrix $G_{\text{mod}}$ with an additional damping factor $\eta$ to accelerate convergence of the series expansion.
• Propose $G_{\text{Rmod}} = G_{rr} + (1-\eta)U_{rs}G_{sr} + \eta[G_{rs} + (1-\eta)U_{rs}G_{ss}](\mathbf{1} - \eta G_{ss})^{-1}G_{sr}$ for improved numerical stability.
• Leverage existing numerical methods for eigenvalue computation of $G_{ss}$ and matrix inversion to enable efficient $G_R$ evaluation.

Experimental results

Research questions

• RQ1 How can effective interactions between a subset of nodes in a large directed network be modeled while preserving indirect path contributions?
• RQ2 What mathematical framework allows for the construction of a reduced Google matrix that retains the stochastic and spectral properties of the original?
• RQ3 Can the analogy with quantum scattering theory be rigorously applied to derive a physically meaningful reduced matrix for network dynamics?
• RQ4 What numerical strategies ensure efficient and stable computation of $G_R$ when $G_{ss}$ has eigenvalues close to unity?
• RQ5 How does the modified Google matrix $G_{\text{mod}}$ improve convergence and numerical feasibility in practical implementations?

Key findings

• The reduced Google matrix $G_R = G_{rr} + G_{rs}(\mathbf{1} - G_{ss})^{-1}G_{sr}$ correctly models effective interactions between a subset of $N_r$ nodes, preserving the Google matrix properties including unit eigenvalue and column stochasticity.
• The matrix $G_R$ allows for the accurate computation of the PageRank vector of the subset, which reflects the stationary distribution of a random surfer confined to the subset with indirect transitions via the rest of the network.
• The geometric series expansion $ (\mathbf{1} - G_{ss})^{-1} = \sum_{k=0}^\infty G_{ss}^k $ enables efficient numerical computation when $ \|G_{ss}\| < 1 $, with convergence speed improved by introducing a damping factor $\eta$.
• The modified matrix $G_{\text{Rmod}}$ with $0.5 \leq \eta < 1$ ensures rapid convergence even when $G_{ss}$ has eigenvalues close to unity, making the method robust for real-world networks.
• For $N < 20,000$, exact diagonalization of $G_{ss}$ allows efficient inversion via $({\bf 1} - G_{ss})^{-1} = Q[{{\bf 1}}/({\bf 1} - D_\lambda)]Q^{-1}$, enabling precise computation of $G_R$.
• The method is applicable to both original and inverted networks, allowing construction of $G_R^*$ for CheiRank analysis, extending its utility to multiple network analysis tasks.