Concept study / Tests & research methodsConceptual illustration. Shapes, colors, and motion are not measured data.
The null model is a concrete account of what the data could look like without the proposed new mechanism. It should not be an obviously unrealistic world of featureless noise. A galaxy catalogue has a survey boundary, selection function and ordinary clustering. A material time series has autocorrelation, drift and known response times. A rendered spiral has choices imposed by its artist. The project’s task is to distinguish its claimed pattern from these alternatives. Surrogate-data methods offer a useful methodological precedent, but the exact null must be designed for each observable and cannot be transferred unchanged between systems.
Preserve what is not being tested
For a temporal-memory test, shuffling every sample may destroy autocorrelation and make the original sequence appear exceptional for an ordinary reason. A suitable surrogate might preserve the spectrum or blocks of temporal dependence while removing a specified higher-order relationship. For a spatial test, controls should reproduce the observation window, point count and detection limits. List every preserved and randomized property. That list defines the null hypothesis. A result can reject this particular null while remaining compatible with several other conventional mechanisms that were not included in it.
Use a ladder of alternatives
Begin with a simple baseline to verify implementation, then add plausible sources of structure. The proposed ladder for a scaling analysis is uniform sampling, correlated sampling, a finite-range nonfractal process, and a physically motivated baseline. For memory, compare instantaneous response, delayed relaxation, a state-space model and a trainable model. Record whether the candidate improves prediction at each step. If its apparent advantage disappears after including a known lag or calibration drift, the result identifies a missing baseline component rather than a discovery of universal adaptation.
Run the analysis on its own controls
Generate synthetic datasets with known answers and process them through the exact pipeline used for observations. Include negative controls, weak injected signals and difficult near-matches. This reveals whether a selected statistic spuriously labels ordinary data as fractal or adaptive. A catalogue of failed controls is valuable: it identifies the boundaries of the method. Freeze the analysis after these development checks, then evaluate independent observations. Report the number of candidate statistics and transformations tried, because extensive searching can manufacture an apparently rare pattern even when each individual calculation is correct.
The mathematical lens
An upper-tail Monte Carlo comparison for B exchangeable null simulations, when the null and statistic were specified appropriately.
Turn it into a question.
- Question
- Can an ordinary process reproduce the claimed signature?
- Method
- Simulate several plausible baselines through the same observation and analysis pipeline.
- What would count against it?
- If realistic baselines routinely reproduce the signature, that signature does not distinguish the proposed mechanism.
The baseline is part of the scientific claim.
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.