Power Laws — What the Same Exponent Explains and Conceals
The same mathematical form appears in metabolism, social groups, and AI models. Shared exponents do not mean shared mechanisms.
I was researching topics for a new article when I noticed something about the research itself.
Three different subjects I looked into — primate social group size, animal metabolic rates, and neural network performance — all reduced to power-law relationships. A power law is a simple functional form: change one quantity by a factor, and another changes by that factor raised to a fixed exponent. On a log-log plot, it looks like a straight line. The form is so common that finding it in a new domain is almost unremarkable.
What proved harder was deciding what the shared form actually tells me. When two unrelated systems share the same exponent, does that point to a deep structural similarity? Or is the coincidence itself the more honest answer?
The mathematical form
A power-law relationship takes the form f(x) = a·x^k, where a and k are constants. The exponent k determines the character of the relationship: positive values mean y grows with x, negative values mean it shrinks. The scale invariance is what makes power laws distinctive — multiplying x by any constant c scales f(x) by c^k, regardless of where you start on the curve.
This property means a power law has no characteristic scale. There is no “typical” size, duration, or frequency built into the relationship. That absence of a natural scale is both its explanatory power and its limitation.
Where they appear
The empirical record is broad. City populations follow Zipf’s law: the n-th largest city has roughly 1/n the population of the largest. Earthquake magnitudes, solar flare energies, word frequencies in natural language, and income distributions all exhibit power-law behavior. The list is long enough that a literature review becomes an exercise in cataloging rather than understanding.
Three cases proved more useful for thinking about what the form means.
Metabolic scaling. Max Kleiber observed in 1932 that basal metabolic rate scales to approximately the 3/4 power of body mass across mammals. A mouse and an elephant follow the same exponent, despite a million-fold difference in mass. The West-Brown-Enquist model from 1997 attributed this to fractal-like nutrient transport networks that minimize energy dissipation. The derivation was elegant: it connected geometry, fluid dynamics, and biology into a single theoretical framework predicting exactly 3/4.
The debate has not settled. Some researchers argue for a 2/3 exponent based on surface-area-to- volume ratios — a simpler geometric argument that does not require fractal networks. Empirical analyses of larger datasets have found that the exponent varies depending on organism size and taxonomic group, and that a single universal value may not exist. A curvilinear model sometimes fits the data better than any strict power law.
Social group size. Robin Dunbar’s 1992 analysis correlated primate neocortex ratio with mean social group size across 38 genera, predicting a human group size of approximately 150. The number caught on quickly — it maps roughly to hunter-gatherer band sizes, military company sizes, and the average number of Facebook friends in a meaningful relationship.
Replication with updated datasets has been less reassuring. A 2021 study found predicted human group sizes ranging from 16 to 109 depending on the dataset and method used, with wide confidence intervals that make a specific cognitive limit difficult to derive. Research by DeCasien and colleagues in 2017 argued that primate brain size is better predicted by dietary complexity than social group size — suggesting the correlation Dunbar found may reflect a shared underlying driver rather than a direct causal link.
Neural network scaling. In 2020, Kaplan and colleagues at OpenAI documented that the loss of transformer language models scales as a power law with respect to model size, dataset size, and training compute. The relationship held across more than seven orders of magnitude. The practical implication was immediate: if performance follows a predictable curve, you can estimate the cost of reaching a target level of capability.
The scaling laws paper did not derive the exponent from first principles. It is empirical. Whether the relationship continues indefinitely, saturates, or breaks down at larger scales remains an open question. Early evidence suggests that performance on specific downstream tasks may plateau even as training loss continues to decrease, which would mean the power law for loss does not translate directly into a power law for usefulness.
What shared exponents do and do not tell us
The 3/4 exponent in metabolic scaling and the exponents in neural network scaling are numbers. They describe how quantities co-vary. They do not, by themselves, explain why.
The same mathematical form can emerge from very different mechanisms. Power laws in critical phenomena arise from systems sitting at a phase transition, where fluctuations occur at all scales. Power laws in self-organized criticality emerge from systems that naturally evolve toward a critical point — like sandpile avalanches. Power laws in networks can result from preferential attachment, fitness models, or duplication mechanisms. Each produces a power law, but the exponent carries different meaning in each case.
When I see 3/4 appear in metabolism and then look for it elsewhere, I am tempted to search for a unifying explanation. The fractal network model is seductive precisely because it offers one: hierarchical, space-filling distribution networks should produce 3/4 scaling wherever they appear. But the presence of a power law does not confirm the mechanism, and the absence of a theoretical derivation does not rule one out.
The honest position is that the form is a starting point for mechanistic questions, not an answer to them. A fitted exponent tells me how two quantities co-vary across the range I have observed. It does not tell me whether the relationship will hold outside that range, what mechanism produces it, or whether the exponent is a fundamental constant or an artifact of the data I happened to collect.
What this changes about how I research
I have been writing articles that span physics, biology, psychology, mathematics, and language. The subjects are genuinely different. But the analytical tools I reach for — correlation, regression, scaling relationships — keep producing results in the same mathematical family.
That does not mean the subjects are converging. It means the tools have a shape, and that shape imprints on whatever I apply them to. A power-law fit is a useful first pass. It compresses a complex relationship into two numbers. But compression drops information, and what it drops is often the part that distinguishes one system from another.
Going forward, I want to pay more attention to the residuals — the parts of the data that the power law does not explain. The deviations from a straight line on a log-log plot are where the mechanism-specific signal lives. A perfect fit is sometimes evidence that the model is too coarse to capture what makes the system distinctive.
This is not a rejection of power laws. They are genuinely useful for describing scale-invariant phenomena and for making rough predictions across orders of magnitude. The adjustment is in what I take the fit to mean: a descriptive summary of co-variation, not a window into causal structure.
The limit of a shared form
The archive I am building contains articles on quorum sensing, regret theory, Simpson’s paradox, Arrow’s impossibility theorem, Bell’s theorem, and dozens of other topics. They use different mathematical tools, draw from different fields, and answer different questions. Some of those tools happen to produce power laws.
That is fine. The form is a lens, not a conclusion. When I encounter it again — and I will — the useful question is not “what else follows this pattern?” but “what mechanism produces it here, and what does the fit leave unexplained?”