Concept study / Tests & research methodsConceptual illustration. Shapes, colors, and motion are not measured data.
A log-log graph transforms a relationship y = A xᵇ into a line when the ideal model applies. That makes it useful for exploring scale dependence, but it does not justify linear regression automatically. Additive measurement error in y becomes a different, often uneven error after taking logarithms. Zeros and negative readings cannot be logged, and silently deleting them changes the sample. For the Hill comparison tools, a plotted slope should therefore be labeled descriptive unless the user has specified a valid sampling and error model. The graph is an interface to a hypothesis, not a self-contained test of fractality.
Normalize before taking logarithms
Use ratios such as x/x₀ and y/y₀ so the logarithm’s argument is dimensionless. State the reference values and units in the export. Their choice changes the intercept but should not change a correctly specified exponent. Do not add an arbitrary small constant to make zeros fit without explaining what that does to the model. If a detection threshold creates censored values, represent the censoring explicitly. The proposed tool should distinguish invalid input, missing data and below-threshold readings instead of converting all three into a conveniently positive number.
Match the fit to the noise
If variability is approximately multiplicative, a log-domain model can be appropriate. If uncertainty is primarily additive in the original measurement, fitting in the original domain may better match the experiment. Errors in both axes require additional care, especially when x spans only a short range. Inspect residuals against fitted value, scale and acquisition order. A high coefficient of determination does not remove curvature, drift or dependence. Report uncertainty in the exponent and investigate whether a few extreme observations dominate it before presenting it as a cross-scale invariant.
Show the missing information
Display the original-scale plot alongside the log plot, with rejected points and the fitting interval documented. For image-derived measurements, retain pixel resolution and preprocessing settings. For correlation curves, note that adjacent distances often share the same underlying pairs and are therefore correlated. Treating every plotted point as an independent sample can make error bars misleadingly small. The project’s export should include the actual data and analysis settings so another researcher can try a different interval and determine whether the claimed scaling survives a defensible alternative treatment.
The mathematical lens
A log-domain statistical model. Its suitability depends on the distribution and dependence of ε, not only on the visual straightness of points.
Turn it into a question.
- Question
- Is the fitted slope robust to a defensible error model and range?
- Method
- Compare original-domain and log-domain fits, inspect residuals and vary prespecified range boundaries.
- What would count against it?
- Curvature, influential points or incompatible noise assumptions undermine a single-exponent interpretation.
A straight line can be informative without being a universal law.
Follow the original work.
These sources support the mechanisms and limits discussed here. They do not endorse Hill’s framework. Notes distinguish research proposals, reviews, and observations.
Continue the investigation
Edition 02 · 18 October 2026 · AI-assisted synthesis. See editorial method.