Chapter 2: Social Superlinearity (β = 5/4)

2.1 The Superlinear Puzzle

While infrastructure scales sublinearly, social and economic outputs display the opposite pattern: they grow faster than population. GDP, patents, wages, and even crime all exhibit superlinear scaling:

$$Y_{\text{social}} = Y_0 \, N^{\beta}, \qquad \beta \approx 1.15\text{--}1.25$$

A city twice as large produces more than twice the GDP per capita. This “increasing returns” phenomenon is the engine of urban economic growth. But where does it come from? The answer lies in social interactions: denser populations create more opportunities for people to meet, exchange ideas, and transact.

2.2 Derivation from the Interaction Integral

Setup

Consider $N$ people distributed over a city of area $A$. The density is$\rho = N/A$. Each person interacts with others within a “mixing length” $\ell$, which represents the typical daily travel distance.

Counting Interactions

For a single individual, the number of potential interaction partners within their mixing zone is:

$$n_{\text{contacts}} = \rho \cdot \pi \ell^2 = \frac{N}{A} \cdot \pi \ell^2$$

The total number of unique pairwise interactions across the city is:

$$I = \frac{N \cdot n_{\text{contacts}}}{2} = \frac{\pi \ell^2}{2} \cdot \frac{N^2}{A}$$

Area Scaling

How does city area scale with population? Empirically and theoretically:

$$A \propto N^{2/3}$$

This comes from the idea that cities are approximately isometric in 3D: height, length, and width all scale similarly with $N^{1/3}$, so$A \sim N^{2/3}$. Alternatively, if infrastructure networks constrain the area and$\ell_{\text{infra}} \propto N^{3/4}$, then $A \propto \ell^2 \propto N^{3/2 \cdot 1/d}$ with details depending on the model.

The Superlinear Exponent

Substituting $A \propto N^{2/3}$ into the interaction count:

$$I \propto \frac{N^2}{N^{2/3}} = N^{2 - 2/3} = N^{4/3}$$

However, not all interactions are equally productive. Including network constraints and the mixing length dependence on city structure, Bettencourt (2013) showed that the effective interaction output scales as:

$$\boxed{Y_{\text{social}} \propto N^{1+\delta}, \qquad \delta = \frac{1}{6} \approx 0.167}$$

This gives $\beta = 7/6 \approx 1.167$. The commonly quoted value of $\beta = 5/4 = 1.25$ is an approximate empirical average across many social metrics; the exact theoretical prediction depends on assumptions about mobility patterns and network structure.

2.3 Bettencourt’s 2013 Unified Theory

In his 2013 Science paper, Bettencourt proposed a self-consistent model where a city balances the benefits of social interactions against the costs of infrastructure and transportation. The key equations are:

The Balance Equation

$$G(N) = \underbrace{Y_0 \, \epsilon \, a_0 \, N^{1+\delta}}_{\text{social output}} - \underbrace{W_{\text{transport}} - W_{\text{infra}}}_{\text{costs}}$$

Here $\epsilon$ is the interaction strength,$a_0$ is the per-contact output, and$\delta$ is the excess exponent.

Mixing Length and Mobility

The mixing length $\ell$ is determined endogenously by individual optimisation: people travel farther when the density of opportunities justifies the cost. This gives:

$$\ell \propto A^{1/2} \propto N^{1/3}, \qquad v \propto \ell \propto N^{1/(2+H)}$$

where $H$ is related to the fractal dimension of the transportation network. For $H=1$ (Euclidean travel), $\ell \propto N^{1/3}$.

Resulting Exponents

$$\beta_{\text{social}} = 1 + \delta \approx \frac{7}{6}, \qquad \beta_{\text{infra}} = 1 - \delta \approx \frac{5}{6}$$

Note the elegant symmetry: social and infrastructure exponents are complementary around 1, with $\beta_{\text{social}} + \beta_{\text{infra}} = 2$. This reflects the conservation of the total interaction-infrastructure balance.

2.4 Empirical Superlinear Metrics

From Bettencourt et al. (2007) and subsequent studies, social metrics across US and European cities:

• GDP: $\beta \approx 1.15 \pm 0.03$
• Total wages: $\beta \approx 1.12 \pm 0.04$
• Patents: $\beta \approx 1.27 \pm 0.05$
• R&D employment: $\beta \approx 1.34 \pm 0.07$
• Serious crime: $\beta \approx 1.16 \pm 0.04$
• AIDS cases: $\beta \approx 1.23 \pm 0.05$

The superlinear scaling of both positive outcomes (GDP, innovation) and negative ones (crime, disease) reflects that all forms of social interaction intensify in larger cities—a phenomenon Bettencourt calls the “good, the bad, and the ugly” of urban scaling.

2.5 Python: Multi-Metric Superlinear Fits

We fit power laws to three social metrics (GDP, patents, crime) simultaneously, demonstrating superlinear scaling for each.

Superlinear Scaling: GDP, Patents, Crime

Python

script.py51 lines

import numpy as np

np.random.seed(123)
n_cities = 40
log_pop = np.random.uniform(np.log10(1e5), np.log10(2e7), n_cities)
population = 10**log_pop

# True exponents for three social metrics
metrics = {
    'GDP (billion $)':   {'beta': 1.15, 'Y0': 1.2e-4, 'noise': 0.10},
    'Patents/year':      {'beta': 1.27, 'Y0': 3.0e-6, 'noise': 0.15},
    'Serious crimes':    {'beta': 1.16, 'Y0': 5.0e-3, 'noise': 0.12},
}

def ols_fit(x, y):
    n = len(x)
    Sx, Sy = np.sum(x), np.sum(y)
    Sxx, Sxy = np.sum(x**2), np.sum(x*y)
    b = (n*Sxy - Sx*Sy) / (n*Sxx - Sx**2)
    a = (Sy - b*Sx) / n
    yp = a + b*x
    R2 = 1 - np.sum((y-yp)**2) / np.sum((y-np.mean(y))**2)
    return b, 10**a, R2

print("=" * 62)
print("  Social Superlinearity: Multi-Metric Power-Law Fits")
print("=" * 62)
print(f"  {'Metric':<22} {'beta_true':>10} {'beta_fit':>10} {'R^2':>8}")
print("-" * 62)

for name, params in metrics.items():
    noise = np.random.normal(0, params['noise'], n_cities)
    log_y = np.log10(params['Y0']) + params['beta'] * log_pop + noise
    beta_fit, Y0_fit, R2 = ols_fit(log_pop, log_y)
    print(f"  {name:<22} {params['beta']:>10.3f} {beta_fit:>10.3f} {R2:>8.4f}")

print("-" * 62)
print()
print("  All exponents are > 1  =>  SUPERLINEAR scaling")
print("  Larger cities produce disproportionately more output")
print()

# Demonstrate the per-capita gain
print("  Per-Capita Advantage (city of 10M vs 1M):")
for name, params in metrics.items():
    ratio = (10e6)**params['beta'] / (1e6)**params['beta']
    per_cap = ratio / 10  # ratio of per-capita values
    print(f"    {name:<22}: {per_cap:.2f}x per capita")

print()
print("  The 'urban premium': bigger cities are more productive per person")

Click Run to execute the Python code

Code will be executed with Python 3 on the server

2.6 The Complete Interaction Integral

For completeness, let us write the full interaction integral. Place person$i$ at position$\mathbf{r}_i$ in the city. The total social output is the sum over all pairs, weighted by a kernel$K(|\mathbf{r}_i - \mathbf{r}_j|)$ that decays with distance:

$$Y = \frac{a_0}{2} \sum_{i \neq j} K(|\mathbf{r}_i - \mathbf{r}_j|) \approx \frac{a_0}{2} \int \int \rho(\mathbf{r}) \rho(\mathbf{r}') K(|\mathbf{r} - \mathbf{r}'|) \, d\mathbf{r} \, d\mathbf{r}'$$

For a uniform density $\rho = N/A$ and a step-function kernel $K(r) = \Theta(\ell - r)$:

$$Y = \frac{a_0}{2} \left(\frac{N}{A}\right)^2 \int_A \int_{B(\mathbf{r},\ell) \cap A} d\mathbf{r}' \, d\mathbf{r} = \frac{a_0}{2} \cdot \frac{N^2}{A^2} \cdot A \cdot \pi\ell^2 = \frac{\pi a_0 \ell^2}{2} \cdot \frac{N^2}{A}$$

For a more realistic kernel with algebraic decay$K(r) = r^{-\alpha}$, the integral introduces logarithmic corrections but preserves the essential superlinear scaling:

$$Y \propto \frac{N^2}{A} \cdot \ell^{2-\alpha} \propto N^{2 - 2/3} \cdot N^{(2-\alpha)/3} = N^{(8-\alpha)/3 - 2/3}$$

For $\alpha = 2$ (gravity-model decay), this recovers $Y \propto N^{4/3}$; for weaker decay, the exponent increases.

2.7 Summary & Key Takeaways

• Social metrics scale superlinearly: $Y \propto N^{\beta}$ with $\beta > 1$
• Superlinearity emerges from the interaction integral: more people in denser space = disproportionately more contacts
• Bettencourt (2013) derives $\delta = 1/6$ from self-consistent city geometry
• Both positive (GDP, innovation) and negative (crime) social outputs are superlinear
• Infrastructure ($\beta < 1$) and social ($\beta > 1$) exponents are complementary

← Infrastructure Scaling Next: Entropy Production →