[Paper Review] A Note on Power-Laws of Internet Topology

Hongsong Chou|ArXiv.org|Dec 21, 2000

Network Traffic and Congestion Control3 references8 citations

TL;DR

This paper demonstrates the mathematical equivalence between the first two power laws in Faloutsos et al.'s seminal work on Internet topology, showing that the rank-out-degree power law (d ∝ r^R) and the frequency-out-degree power law (f_d ∝ d^O) are interderivable via calculus. It further shows that estimating the exponent R from rank data yields more accurate results with smaller error bars than estimating O from frequency data, especially for nodes with out-degree ≤ 32.

ABSTRACT

The three Power-Laws proposed by Faloutsos et al(1999) are important discoveries among many recent works on finding hidden rules in the seemingly chaotic Internet topology. In this note, we want to point out that the first two laws discovered by Faloutsos et al(1999, hereafter, {\it Faloutsos' Power Laws}) are in fact equivalent. That is, as long as any one of them is true, the other can be derived from it, and {\it vice versa}. Although these two laws are equivalent, they provide different ways to measure the exponents of their corresponding power law relations. We also show that these two measures will give equivalent results, but with different error bars. We argue that for nodes of not very large out-degree($\leq 32$ in our simulation), the first Faloutsos' Power Law is superior to the second one in giving a better estimate of the exponent, while for nodes of very large out-degree($> 32$) the power law relation may not be present, at least for the relation between the frequency of out-degree and node out-degree.

Motivation & Objective

• To clarify the mathematical relationship between the first two power laws proposed by Faloutsos et al. for Internet topology.
• To evaluate the relative accuracy of two data analysis methods for estimating power-law exponents in Internet topology.
• To determine which method—analyzing rank vs. frequency—produces more reliable exponent estimates with smaller error bars.
• To assess the validity of power-law behavior for nodes with very high out-degree (>32).

Proposed method

• Derives the second power law (f_d ∝ d^O) from the first (d ∝ r^R) using first-order approximation and differentiation.
• Reconstructs the first power law from the second via integration, establishing mutual derivability.
• Uses the relation O ≈ 1/R - 1 to link the exponents of the two laws, confirming their equivalence.
• Employs linear regression on log-log plots of rank vs. out-degree and frequency vs. out-degree to estimate exponents.
• Applies error analysis to compare uncertainty in exponent estimates from both methods.
• Analyzes real Internet datasets (Int-11-97, Int-04-98, etc.) to validate theoretical derivations empirically.

Experimental results

Research questions

• RQ1 Are the first two power laws in Faloutsos et al.'s work mathematically equivalent?
• RQ2 Does estimating the exponent R from rank-out-degree data yield more accurate results than estimating O from frequency-out-degree data?
• RQ3 How do error bars compare between the two estimation methods?
• RQ4 Is the power-law relationship between frequency and out-degree valid for nodes with very high out-degree (>32)?
• RQ5 What is the impact of nodes with out-degree 1 and low-frequency outliers on power-law fitting accuracy?

Key findings

• The first and second Faloutsos power laws are mathematically equivalent, with O ≈ 1/R - 1 and C₂ ≈ -1/R × (1/C₁)^(1/R).
• Estimating the exponent R from rank-out-degree data yields more accurate results with smaller error bars than estimating O from frequency-out-degree data.
• For nodes with out-degree ≤ 32, the rank-based method (first power law) provides a superior estimate of the exponent compared to the frequency-based method.
• For nodes with out-degree > 32, the power-law relationship between frequency and out-degree may not hold, suggesting breakdown in the model at high degrees.
• Outliers such as nodes with out-degree 1 and frequency 1 significantly affect linear fitting, particularly in rank-based analysis.
• Empirical data from multiple Internet datasets (e.g., Int-11-97, Rout-95) show relative errors in exponent estimation ranging from 4% to 42%, with higher errors for datasets like Rout-95.