The Law of Large Numbers: Why Casinos Win

"The Law of Large Numbers is not magic. It's the rigorous version of the sentence: 'do this enough times and the noise averages out.'"

The Statement, Plainly

The Law of Large Numbers (LLN) says: in its familiar iid form, if you take independent samples from the same distribution with finite expected value μ, the sample mean converges to μ as the number of samples grows large. In symbols:

(X₁ + X₂ + ... + Xₙ) / n → μ as n → ∞

For a fair-coin model, the proportion of heads gets close to 0.5 with high probability as the number of flips grows. The model supplies the probability; the observed long-run frequency checks its calibration rather than defining fairness by itself.

What It Does Not Say

The LLN is precise about what it guarantees, and citizens routinely overclaim it.

It does not say that individual outcomes become less random with more trials.
It does not say rare events stop happening.
It does not say the variance shrinks to zero (it does shrink, but for the sample MEAN, not for individual draws).
It does not say the distribution shape changes.

The most common misreading is the gambler's fallacy: 'red has come up five times in a row, so black is due.' The LLN does not back this up. Each spin is independent; past outcomes do not affect future probabilities. The long-run mean converges, but every single trial is still its own independent event.

The Casino Is the LLN Made Visible

Standard casino rules usually give the house positive expected value. European single-zero roulette has a house edge of about 2.7%; blackjack's edge varies materially with rules and strategy, and slot returns vary by machine and jurisdiction. Individual outcomes are completely random and unpredictable. But the casino doesn't bet on individual outcomes — it lets thousands of customers play millions of hands per month, and the LLN guarantees that the realized average per hand converges to that positive house edge.

The casino is not gambling. The customers are. The casino is letting the law of large numbers do its work, while collecting a small positive expected value on each bet. With many comparable bets, average profit per bet can stabilize around its expectation. Cumulative profit still fluctuates in absolute terms, and capital, limits, dependence, and changing play conditions matter.

The Citizen Lesson

Long-run averages require a stable process and appropriate dependence conditions. Positive expected value alone does not guarantee realized profit or survival: tail risk, capital constraints, costs, model error, and strategy decay can dominate. The LLN is the citizen's friend on the right side of the expected value and the enemy on the wrong side.

Code

Cumulative outcome of a negative-EV bet·python

import numpy as np
rng = np.random.default_rng(33)

# Simulate a single 'gambler' playing roulette with a 2.7% house edge.
# Net result per bet: +1 with probability 18/37, -1 with probability 19/37.
# Expected value per bet ≈ -0.027.
def play(n_bets):
    spins = rng.random(n_bets) < 18/37   # win or lose
    return np.where(spins, 1, -1).cumsum()

for n in (100, 10_000, 1_000_000):
    final_position = play(n)[-1]
    expected = -0.027 * n
    print(f"after {n:>9,} bets: realized = {final_position:>+9}   "
          f"expected ≈ {expected:>+9.0f}")

# Small N: the gambler may be winning or losing — pure noise.
# Large N: realized result converges to the expected value: a near-certain loss.
# The casino is on the opposite side of every bet, accumulating the inverse.
# This is the LLN: long-run mean dominates short-run noise.

Exercise

Pick a recurring decision in your life with a clear expected value (taking the subway vs walking when in a rush, ordering the same dish vs trying new ones, skipping breakfast vs eating one). Estimate the EV per occurrence, then estimate how many occurrences you've already accumulated. Does the LLN suggest the realized result has converged to the EV by now? If yes, you should be acting on EV. If no, you're still in noise territory.

Hint

Roughly speaking, the LLN's convergence is meaningful once N is large enough that random variation is smaller than the effect size you care about. A coin flip needs ~100 trials to be convinced of the EV; a small EV difference needs many more.

The Statement, Plainly

What It Does Not Say

The Casino Is the LLN Made Visible

The Citizen Lesson

Code

External links

Exercise

Progress

Comments 0