Skip to content
Why 60% Earn Below Average Income: The Multiplicative Central Limit Theorem
Fundamentals 10 min read

Why 60% Earn Below Average Income: The Multiplicative Central Limit Theorem

The article explains why most people earn below the average income using a multiplicative random walk simulation and log-normal distribution theory, showing that even with equal starting points and luck-based returns, the average exceeds the median due to compounding volatility, and validates this with China's 2025 income data.

Model Perspective
Model Perspective
Model Perspective

01 A Thought Experiment

In 2025, China's per capita disposable income was 43,377 yuan, while the median was 36,231 yuan — the average exceeds the median by about 20%. Half the population earns less than 36,000 yuan, far below the "per capita" figure. The common explanation is that the rich pull up the average, but the article asks a deeper question: even if everyone starts with the same wealth and follows the same luck-based rules, would the average still outrun the median?

The answer is yes. The reason lies in the multiplicative version of the Central Limit Theorem.

A simulation: 100,000 people start with wealth = 1. Each year for 40 years, wealth is multiplied by a random factor uniformly drawn between 0.8 and 1.25. No skill differences, pure luck. After 40 years, 65% of people fall below the average wealth . The average is 2.68, the median only 1.94. The distribution has a long right tail; the top 10% hold 32% of total wealth. No cheating — just repeated multiplication.

Simulation result: 65% below average
Simulation result: 65% below average

02 Additive vs. Multiplicative Central Limit Theorems

The classic Lindeberg–Lévy CLT states that the sum of i.i.d. finite-variance random variables converges to a normal distribution. If we switch to additive increments (each year add a random number between -0.1 and +0.1), after 40 years the average and median nearly coincide, and exactly half are below average — the familiar bell curve.

Multiplication can use the same theorem by taking logarithms:

ln(wealth after 40 years) = ln(factor₁) + ln(factor₂) + … + ln(factor₄₀)

The right side is a sum of 40 i.i.d. variables, so the CLT applies: the logarithm of wealth is approximately normal , meaning wealth itself follows a log-normal distribution .

Log-normal distribution fit after log transform
Log-normal distribution fit after log transform

The previously skewed distribution, after log transformation, matches the normal curve almost perfectly.

03 Why the Average Pulls Ahead of the Median

Let ln X ~ N(μ, σ²). Then:

Median = e^μ (log transform preserves order; normal median is μ).

Mean = e^(μ + σ²/2) (using the normal moment-generating function E[e^Z] = e^(μ + σ²/2)).

The ratio Mean/Median = e^(σ²/2), depending only on volatility σ. Larger volatility pushes the average further above the median.

The proportion below the mean is P(X < E[X]) = Φ(σ/2), which is always > 0.5 and increases with σ.

Plugging the simulation parameters: annual log-factor variance ≈ 0.0164, so after 40 years σ ≈ 0.81. The theoretical predictions match the simulation:

10 years: Mean/Median = 1.09, Below Average = 58%

20 years: Mean/Median = 1.18, Below Average = 61%

40 years: Mean/Median = 1.39, Below Average = 66%

The simulation gave 65% — consistent. The gap widens over time because σ² accumulates linearly with years.

An overlooked detail: the simulation's average annual growth is 2.5%, yet the median person experiences only 1.7% growth. In a multiplicative world, "average return" and "typical person's return" are two different numbers.

04 Fitting Real Data

Using the official figures: 43,377 / 36,231 ≈ 1.197 = e^(σ²/2) → σ ≈ 0.60.

Then Φ(0.60/2) = Φ(0.30) ≈ 62%.

If household per capita disposable income is approximately log-normal, about 60% of people earn below the "per capita" average . Separate urban/rural calculations give 59% and 61%, both around 60%.

05 Where the Model Falls Short

The very top does not follow log-normal. Extreme incomes/wealth follow a Pareto distribution with a fatter tail; log-normal severely underestimates billionaires.

Annual multipliers are not independent. Success compounds — more capital and opportunities — widening gaps faster. Conversely, taxes, social security, and transfers pull gaps back.

The data is household per capita disposable income, not individual income. Applying the model to it is only a rough approximation. The 62% figure is derived from two summary statistics, not an official census count.

The 0.8–1.25 range is illustrative.

06 The Takeaway

When you see "average", ask whether the quantity is additive or multiplicative.

Height, test scores (additive) → mean ≈ median. Income, housing prices, city populations, follower counts, stock returns (multiplicative) → mean > median; to understand a typical person, look at the median.

A puzzle: If ln X ~ N(0, 1), what is P(X > E[X])? Answer in the comments tomorrow.

Model Card

Model: Log-normal distribution (multiplicative Central Limit Theorem)

Assumption: Annual wealth multiplied by i.i.d. random factor

Conclusion: Mean/Median = e^(σ²/2); Proportion below mean = Φ(σ/2) > 0.5

Failure modes: Top tail is Pareto; multipliers correlated; redistribution reduces gaps

References

[1] National Bureau of Statistics. 2025 Resident Income and Consumption Expenditure. 2026-01-19.

[2] Gibrat R. Les Inégalités Économiques. Paris: Sirey, 1931. (Proportional effect and log-normal)

[3] Limpert E, Stahel W A, Abbt M. Log-normal Distributions across the Sciences: Keys and Clues. BioScience , 2001, 51(5): 341–352.

Code example

[1] 国家统计局. 2025年居民收入和消费支出情况. 2026-01-19.
[2] Gibrat R. Les Inégalités Économiques. Paris: Sirey, 1931.(比例效应与对数正态)
[3] Limpert E, Stahel W A, Abbt M. Log-normal Distributions across the Sciences: Keys and Clues.
Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

statistical modelingcentral limit theoremincome distributionmultiplicative processesChina income datalog-normal distributionmean vs median
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.