Localization Phenomena in Large-Scale Networked Systems: Robustness and Fragility of Dynamics
Poorva Shukla, Bassam Bamieh · 2024
Read Localization Phenomena in Large-Scale Networked Systems: Robustness and Fragility of Dynamics by Poorva Shukla, Bassam Bamieh (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-00252.txt.
Introduction
Poorva Shukla and Bassam Bamieh study in 2024 a fragility that hides inside large networked systems. In some networks, key vibration patterns of the graph stay trapped in small corners instead of spreading. Physics knows a cousin effect as Anderson localization. Using spectrum perturbation tools, the authors show how trapped patterns shrink the safety margin against unmodeled dynamics at single nodes or links.
Simple-looking networks can carry the features that trigger the trap. The result means a system can look robust in bulk yet fail from a small local surprise. The authors briefly weigh what this means for infrastructure and control design. Their examples show low structural complexity is no guarantee of safety.
Catalog record
| CREATOR | Poorva Shukla, Bassam Bamieh |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 49 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-00252
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Localization Phenomena in Large-Scale Networked Systems: Robustness and Fragility of Dynamics Poorva Shukla, Bassam Bamieh Abstract We study phenomena where some eigenvectors of a graph Laplacian are largely confined in small subsets of the graph. These localization phenomena are similar to those generally termed Anderson Localization in the Physics literature, and are related to the complexity of the structure of large graphs in still unexplored ways. Using spectral perturbation theory and pseudo-spectrum analysis, we explain how the presence of localized eigenvectors gives rise to fragilities (low robustness margins) to unmodeled node or link dynamics. Our analysis is demonstrated by examples of networks with relatively low complexity, but with features that appear to induce eigenvector localization. The implications of this newly-discovered fragility phenomenon are briefly discussed. I Introduction and Problem Setting The phenomenon of Anderson localization is well known in condensed-matter physics. The original work [ 2 ] from the late 50’s received the Physics Nobel prize in 1977, and it remains an active and important area of research to this day. In physics, it refers to the abrupt change of the nature of eigenfunctions in the presence of disorder (randomness) in the governing potentials in quantum-mechanical or classical models. For certain models, arbitrarily small amounts of disorder can cause eigenfunctions to change from being “delocalized” (having their mass distributed globally in space) to having exponential decay, i.e. becoming “localized” in space. This can translate to abrupt changes in material properties. In this sense, localization phenomena can be broadly understood as those of “fragilities” of certain mathematical models of reality. It has recently been recognized that localization phenomena are much more ubiquitous than previously thought, and can occur not only due to randomness in media, but also due to complex geometry of boundaries in non-random problems [ 3 ] . It has also been observed in acoustic systems where localization occurs due to fractal boundaries [ 4 ] , or inhomogeneity in the medium [ 5 ] ; and bio-molecular vibrations [ 6 ] where localization is caused by to the spatial complexity of 3D protein structures. Therefore it may be suspected that localization phenomena can occur in other mathematical models of the dynamics of large-scale systems and networks such as in neuroscience, power grids, multi-agent systems, vehicular and robotic networks among others. In this paper we introduce the concept of eigenvector localization in the context of graph Laplacians. Characterizing which graphs poses this property requires a much lengthier analysis and will be reported elsewhere [ 7 ] . In this paper we instead address the following system-theoretic question: Given a graph where some of its Laplacian’s eigenvectors have the localization property, what effects does this property have on robustness/fragility of networked dynamical systems defined over this graph . In particular, we examine 2nd-order “consensus-type” systems such as the swing equations of power networks, vehicular formations, or networked 2nd-order linear oscillators in general. We consider various types of dynamical node and edge perturbations and examine their effects on overall system stability using large-scale examples. A spectral perturbation/sensitivity analysis is included to give some theoretical guidance. This analysis indicates that the fragility phenomena observed in these examples are likely ubiquitous in networked dynamical systems whose graph Laplacians exhibit eigenvector localization. We now give an informal description of the localization phenomenon by way of the example given in Figure 1 . Formal definitions will be given in later sections. Figure 1(a) shows a network with N=60 nodes, and Figure 1(c) shows an overview of the Laplacian’s eigenvectors for a network with the same structure but with N=200 . Roughly about 80 of those eigenvectors (depicted in red) appear to be “localized”, i.e. most of their mass is concentrated over a very small number of nodes, while the remaining eigenvectors (depicted in blue) are spread out over a large subdomain. This contrast in structure of localized versus delocalized eigenvectors is shown more clearly for two specific eigenvectors in Figure 1(b) . The main feature is that localized eigenvectors decay rapidly (with respect to node index) away from their peak. Note that in Figure 1(c) the eigenvectors indices are ordered in increasing magnitude of eigenvalues, and the localized ones appear to be towards the bottom of the spectrum. Other examples, not reported here due to lack of space, exhibit localization in other portions of the spectrum depending on the graph structure. Furthermore, the localized eigenvectors appear to have most of their masses in particular regions of the graph (depicted with the red bars in Figure 1(d) ), with the complementary subregion having delocalized eigenvectors whose mass is widely spread in that region (depicted with the blue bar). We call such regions of the graph the “localized” and “delocalized” regions respectively. (a) A banded Laplacian with equal edge weights ( left ), and the corresponding graph ( right ). A small network size of N=50 and band size 20 is shown for ease of visualization. (b) Same example with {\scriptstyle N}=1000 and band size 400 . ( Left ) A localized eigenvector plotted with absolute and semi-log scales. ( Right ) A typical delocalized eigenvector that has substantial magnitude over a large subdomain. In both plots the x -axis is the node index. (c) A plot of all the eigenvectors of the same network structure as (a) with N=200 and band size \approx 80 . The red lines show localized eigenvectors, while the blue lines are delocalized eigenvectors. The eigenvectors are plotted in increasing magnitude of eigenvalues, and the localized ones are towards the bottom end of the spectrum. (d) All the eigenvectors on a single plot for a larger network N=6000 . The localized eigenvectors (red) have O(1) peaks and remain confined to the red region of the graph (roughly 1\leq k\leq 800 and 5200\leq k\leq 6000 ), while the delocalized eigenvectors (blue) are spread out over the blue region (roughly 800\leq k\leq 5200 ) with negligible magnitude over any particular node. (e) The eigenvalues for a network with N=100 and band size \approx 40 . Localized and delocalized eigenvalues are in red and blue respectively. There are approximately 40 localized eigenvalues, about the same number as the band size of the Laplacian. With the exception near the spectrum boundaries, the localized eigenvalues tend to have larger inter-value gaps compared with the delocalized ones. This is consistent with the analysis in [ 7 ] . Fig. 1 : An example illustrating Laplacian eigenvector localization. A subset of the eigenvectors are localized, while others are not. The basic “banded” Laplacian structure is shown in (a). Different network sizes N are used in the various subplots for ease of visualization. The question we investigate in this paper concerns robustness of dynamical systems defined on graphs whose Laplacians posses localized eigenvectors like those of the example of Figure 1 . Is the system more fragile when there is uncertainty in node/link dynamics within a localized region compared to those in a delocalized region ? We show with examples that the answer is in the affirmative, and provide a perturbation analysis to support this conclusion for any graph with localized Laplacian eigenvectors. The perturbation analysis also implies that this fragility contrast (between localized and delocalized perturbations) can become arbitrarily