Log-normal distribution
A log-normal distribution is a continuous probability distribution of a random variable whose logarithm is normally distributed. If \( X \) is log-normally distributed, then \( Y = \ln X \) follows a normal distribution. The distribution is defined by two parameters — the mean \( \mu \) and standard deviation \( \sigma \) of the underlying normal — and it produces right-skewed, heavy-tailed distributions in which most values cluster near zero but a long tail permits rare extreme events.
The log-normal is one of the most common distributions in nature and society, yet it is routinely confused with the power law. Both are heavy-tailed. Both produce rare extreme events. But they arise from fundamentally different generative processes, and mistaking one for the other has led to serious errors in fields from finance to network science.
Why the Log-Normal Appears
The log-normal emerges from multiplicative processes. When a quantity is the product of many independent positive random variables — each multiplying the previous result — the central limit theorem applies to the logarithms, and the resulting distribution is log-normal. This is in contrast to the normal distribution, which arises from additive processes (sums of random variables).
Examples abound:
- Particle sizes in crushing and grinding processes, where repeated fracture multiplicatively reduces size
- Stock prices in geometric Brownian motion, where daily returns compound multiplicatively
- Income and wealth distributions, where individual wealth is the product of many multiplicative shocks (investment returns, job changes, inheritance)
- Neuron firing rates and gene expression levels, where biological amplification cascades are multiplicative
- Network degrees in certain growth models where attachment probabilities depend on multiplicative fitness factors
The key insight: addition produces normality; multiplication produces log-normality. Any system in which effects compound rather than accumulate will tend toward log-normal statistics.
Log-Normal vs. Power Law
This distinction is critical for network science. The Barabási–Albert model and preferential attachment produce pure power-law degree distributions with exponent \( \gamma = 3 \). But real networks rarely exhibit clean power laws. Clauset, Shalizi, and Newman's 2009 statistical analysis showed that many networks previously claimed to be scale-free are better fit by log-normal distributions — particularly when the network growth process involves multiplicative fitness factors, geographic constraints, or saturation effects.
The two distributions are superficially similar on log-log plots over limited ranges, but they diverge in their tails. A power law has no characteristic scale: events of any magnitude can occur, with probability decaying as a polynomial. A log-normal has an implicit scale set by \( \sigma \): beyond a certain point, the tail decays faster than any power law (specifically, like \( \exp(-(\ln k)^2) \)). This means that the largest events in a log-normal distribution are much less extreme than in a power-law distribution with comparable bulk behavior.
The policy implications are severe. If a financial market follows a power law, extreme crashes are not just possible but expected at any scale. If it follows a log-normal, the probability of catastrophic crashes is exponentially smaller. The difference between these models is not academic — it determines how much capital banks must hold, how insurance is priced, and how regulators think about systemic risk.
Similarly in network science: if the Internet's degree distribution is a power law, targeted attacks on hubs are devastating and random failures are harmless. If it is log-normal, the network is less extreme in both its robustness and its fragility. The scale-free network paradigm has been applied with the confidence of a power-law universality that the data do not support.
Generative Models
Several network growth models produce log-normal degree distributions:
Preferential attachment with multiplicative fitness: When each node has an intrinsic fitness \( \eta_i \) and the attachment probability is proportional to \( \eta_i k_i \), the degree distribution becomes log-normal if fitnesses themselves are log-normally distributed. The hubs are not just lucky in timing; they are intrinsically better at attracting connections.
Growth with saturation: When nodes have a finite capacity for connections, the degree distribution develops a cutoff that transforms a power law into something closer to log-normal. Social networks, where individuals cannot maintain arbitrarily large numbers of relationships, often show this pattern.
Random multiplicative processes on networks: When edge weights or node strengths evolve through multiplicative noise, the resulting distribution is log-normal even when the topological structure is fixed.
Statistical Identification
Distinguishing log-normal from power-law data is statistically challenging. On a log-log plot, a log-normal can look linear over a wide range before rolling over in the tail. Maximum likelihood methods and model comparison (e.g., Vuong's test) are necessary. Visual inspection is not enough. The literature is littered with claims of power-law behavior that collapse under rigorous statistical scrutiny.
The log-normal distribution is the power law's quieter cousin. It produces similar-looking heavy tails without the infinite variance and scale-invariance that make power laws mathematically dramatic. In many real systems, the log-normal is not merely a better fit — it is the right mechanistic model, because the generative process is multiplicative, not accumulative. The persistence of power-law claims in the face of log-normal alternatives is a case study in how mathematical elegance can outrun empirical warrant.