The Subexponential Class

The mathematical property that defines fat tails — when the maximum dominates the sum.

The subexponential class provides the mathematical foundation for understanding fat tails. Its defining property — that the sum is dominated by the maximum — captures exactly why fat-tailed phenomena behave so differently from what our intuition (trained on thin tails) expects.

The Defining Property

Definition

Subexponential Distribution

A distribution F with support on [0, ∞) is subexponentialif for independent X₁, X₂ with distribution F:

Read: The limit as x approaches infinity of P(X₁ + X₂ exceeds x) over P(X₁ exceeds x) equals 2

The probability that the sum exceeds x is asymptotically twice the probability that one term exceeds x

Why “2”? Because the sum exceeds x primarily through one of two mutually exclusive events: either X₁ > x OR X₂ > x. The probability that both are large is negligible compared to one being large.

Contrast with Light Tails

For light-tailed distributions (like Gaussian), the sum exceeds x through a collective contribution of many moderate values:

Light Tails (Gaussian)

If X₁ + X₂ = 100, typical scenario:

  • X₁ ≈ 50, X₂ ≈ 50
  • Both contribute moderately
  • No single term dominates

Fat Tails (Pareto)

If X₁ + X₂ = 100, typical scenario:

  • X₁ ≈ 98, X₂ ≈ 2
  • Or: X₁ ≈ 3, X₂ ≈ 97
  • One term dominates

The Catastrophe Principle

The subexponential property extends to sums of any size:

More strikingly, this is equivalent to:

Definition

Catastrophe Principle

For subexponential distributions:

Read: The probability that the sum exceeds x is asymptotically equal to the probability that the maximum exceeds x

The sum is big when and only when the maximum is big — one term causes the 'catastrophe'

Key Insight

The Single Event Dominance

In fat-tailed systems, large outcomes come from single extreme events, not accumulation of many moderate events. This is why:

  • One earthquake causes more damage than 100 smaller ones
  • One book sale (Harry Potter) dwarfs thousands of average books
  • One trading loss can exceed all previous profits

Explore: Sum vs Maximum

Sample from different distributions and observe how the sum relates to the maximum. In fat-tailed distributions, they track closely.

Click “Run Simulation” to generate samples and visualize the relationship between sum and maximum.

Properties of Subexponential Distributions

Closure Properties

  • Products: If X is subexponential, so is cX for any c > 0
  • Mixtures: Mixtures of subexponential distributions are subexponential
  • Convolutions: The sum of independent subexponentials with the same tail index is subexponential

Important Examples

DistributionTail BehaviorSubexponential?
ParetoYes
Log-normalYes (borderline)
Weibull (β < 1)Yes
ExponentialNo (boundary)
GaussianNo

Implications for Practice

Example

Insurance and Risk

Consider an insurance company with 1000 policies, each with independent claim sizes.

If claims are Gaussian:

  • Total claims ≈ sum of many similar-sized claims
  • Easy to predict with Law of Large Numbers
  • Diversification works perfectly

If claims are Pareto:

  • Total claims ≈ one massive claim
  • Historical averages are misleading
  • Diversification provides limited protection
Example

Investment Returns

Annual returns on an aggressive strategy over 20 years:

If returns are Gaussian:

  • Cumulative return = steady accumulation
  • Each year contributes roughly equally

If returns are fat-tailed:

  • A few exceptional years dominate your wealth
  • Missing the best 5 days can destroy decades of returns
  • Similarly, one crash can wipe out decades of gains
Key Insight

Why Diversification Has Limits

In Gaussian world, adding more independent assets reduces risk as 1/√n. In fat-tailed world, diversification helps less because single events dominate. You can't diversify away the maximum.

Key Takeaways

  • Subexponential =
  • The sum exceeds x when ONE term exceeds x, not through accumulation
  • Catastrophe principle:
  • Pareto, log-normal, and Weibull (β < 1) are subexponential
  • In subexponential domains, single events dominate — diversification has limits