[Paper Review] The accuracy of roughness exponent measurement methods

Jan Øystein Haavig Bakke, Alex Hansen|ArXiv.org|Jul 12, 2007

Surface Roughness and Optical Measurements2 references3 citations

TL;DR

This paper evaluates the accuracy of common methods for measuring the self-affine roughness exponent in 1D profiles, finding that power spectrum density analysis and averaged wavelet coefficient methods yield the most reliable estimates with errors <0.03 for profiles longer than 256 points. It further identifies that power-law noise and long-range correlations in sign changes can mimic self-affine scaling, leading to systematic biases in other methods.

ABSTRACT

We test methods for measuring and characterizing rough profiles with emphasis on measurements of the self-affine roughness exponent, and describes a simple test to separate between roughness exponents originating from long range correlations in the sign signs of the profile, and roughness exponents originating from L{é}vy distributions of jumps. Based on tests on profiles with known roughness exponents we find that the power spectrum density analysis and the averaged wavelet coefficients method give the best estimates for roughness exponents in the range 0.1 to 0.9. The error-bars are found to be less than 0.03 for profile lengths larger than 256, and there are no systematic bias in the estimates. We present quantitative estimates of the error-bars and the systematic error and their dependence on the value of the roughness exponent and the profile length. We also quantify how power-law noise can modify the measured roughness exponent for measurement methods different from the power spectrum density analysis and the second order correlation function method.

Motivation & Objective

• To assess the accuracy, error bars, and systematic biases of various roughness exponent measurement techniques for self-affine profiles.
• To distinguish between roughness exponents caused by long-range correlations in profile sign changes versus those from Lévy-distributed jumps.
• To quantify how power-law noise distorts roughness exponent measurements across different methods.
• To provide guidelines for selecting robust methods in experimental and simulation studies of fracture and surface roughness.

Proposed method

• Profiles with known roughness exponents were generated using the Voss method and wavelet-based techniques to create controlled test data.
• Roughness exponent was measured using multiple methods: power spectrum density analysis, averaged wavelet coefficients, detrended fluctuation analysis, local window methods, and second-order correlation functions.
• Systematic errors and statistical uncertainties were quantified by comparing measured exponents to true values across a range of exponents (0.1–0.9) and profile lengths.
• Power-law noise was introduced into random walks with varying tail exponents (α ∈ {0.5, 3.0}) to simulate real-world distortions and assess method sensitivity.
• Multi-affine corrections and scaling deviations were analyzed using the Rk/Rk^G ratio to detect non-self-affine behavior at small scales.
• Theoretical scaling relations such as w(l) ∝ l^{1/α} for Lévy flights were used to validate expected roughness exponents under power-law jump distributions.

Experimental results

Research questions

• RQ1 Which roughness exponent measurement methods yield the most accurate and unbiased estimates across the range 0.1 ≤ ζ ≤ 0.9?
• RQ2 How do power-law distributed jumps in profile increments affect the measured roughness exponent, and which methods are most sensitive to this noise?
• RQ3 To what extent do long-range correlations in the sign of profile increments mimic self-affine scaling and bias measurement outcomes?
• RQ4 How do error bars and systematic biases in roughness exponent estimation vary with profile length and true exponent value?
• RQ5 Can multi-affine corrections due to power-law noise be detected, and how do they affect the validity of self-affine scaling assumptions?

Key findings

• Power spectrum density analysis and the averaged wavelet coefficient method produced the most accurate roughness exponent estimates, with error bars consistently below 0.03 for profile lengths ≥256.
• These two methods showed no significant systematic bias across the tested range of roughness exponents (0.1–0.9), making them the most reliable for quantitative studies.
• Local window and averaged wavelet coefficient methods exhibited systematic errors when applied to profiles with power-law distributed jumps, especially for α < 1.2, where the true Lévy-flight exponent ζ_LF = 1/α exceeds 1.
• The power spectrum and second-order correlation function methods measured ζ = 0.5 regardless of the Lévy exponent α, as they detect uncorrelated increments rather than long-range correlations.
• Profiles with α = 1.5 showed strong multi-affine corrections below scale l ≈ 20, rendering them non-self-affine at small scales, while α = 3.0 showed weaker deviations.
• The study confirms that power-law noise and long-range sign correlations can produce similar roughness exponents via different physical mechanisms, necessitating method selection based on noise characterization.