If you play a perfectly fair game—say, coin flips where heads wins 1 unit and tails loses 1 unit—ten thousand times, what would your profit/loss curve look like?
Intuition says: since the game is fair, the curve should oscillate around zero, sometimes winning, sometimes losing, roughly balanced overall.
Wrong.
William Feller, in his two-volume work hailed as the "bible of probability theory," *An Introduction to Probability Theory and Its Applications*, proved a chilling result: in a fair random walk, the most common situation is not frequent switching between winning and losing, but long stays on one side.
You will most likely stay on the winning side or on the losing side for a long time, rarely crossing zero. This is not about luck—it is mathematics.
1. Who Was Feller?
William Feller (1906–1970) was a Croatian-American mathematician and professor at Princeton University. His *An Introduction to Probability Theory and Its Applications* comes in two volumes: Volume 1 (1950) covers discrete probability; Volume 2 (1966) covers continuous probability.
These books occupy a position in probability theory comparable to *On the Origin of Species* in biology—not because they are the easiest to read (they are famously difficult), but because they integrated probability theory, for the first time, from a collection of scattered tricks into a rigorous mathematical discipline.
Feller's uniqueness: he was never satisfied with proving theorems—he insisted on finding the intuition behind them. His books are full of heuristic reasoning about "why this result is true," not just cold proofs. This style influenced an entire generation of probabilists.
> When asked why he wrote the book, Feller said: probability intuition is too easily misled, and without mathematical rigor the field would be exploited by charlatans—including those teaching "guaranteed winning strategies."
2. The Random Walk: Simplest Model, Most Counterintuitive Results
One of the classic topics in Feller's book is the random walk: a particle starts at zero and moves left or right one step with equal probability each turn—the mathematical model of a fair game (50% win/lose each round).
Intuition says that after many steps the particle should hover near zero. But Feller proved several counterintuitive results:
Result 1: The particle returns to zero, but far less often than you'd think.
In a 1D random walk, the particle returns to zero almost surely (probability 1). But the expected waiting time for a return is infinite. You will eventually break even—but "eventually" may outlast your lifetime.
Result 2: You spend most of the time either winning or most of the time losing.
The most counterintuitive result is the arcsine law: the fraction T/n of time the particle spends on the positive (winning) side has probability density
f(t) = 1 / (π · √(t · (1-t)))
This function is minimized at t = 0.5 and diverges as t → 0 or t → 1. In other words: the probability of spending ~50% of the time on the winning side is the smallest; the probability of spending nearly 0% or nearly 100% is the largest.
Concrete numbers: the probability of spending more than 97.5% of the time on the positive side is about 10%; more than 85% of the time, about 25%. Spending 40%–60% of the time has probability only about 30%.
> In a fair game of 10,000 rounds, you have roughly a 25% chance of essentially always winning (positive side >85% of the time) and a 25% chance of essentially always losing. Only a 30% chance that your profit curve "normally" oscillates around zero.
3. Why Gamblers Can't Stop
The arcsine law explains a deep trap in gambler psychology.
A gambler wins 70 of the first 100 hands, feels "hot," and keeps playing. But he has merely landed on the positive branch of a random walk—not luck, but an inevitable product of the probability distribution.
Conversely, a gambler who loses 70 of the first 100 hands thinks he is "due for a comeback"—surely mean reversion must kick in? But Feller's theorems say: not necessarily. In a fair random walk you may keep losing for a long time. The expected return time to zero is infinite, so a "comeback" may never happen within your lifetime.
> This is why gambler's ruin is so lethal: not because the house is unfair (we assumed perfect fairness), but because the geometry of the random walk makes you likely stuck on the losing side, while your intuition whispers that a comeback is near.
Feller also proved something harsher: with finite capital against an infinitely capitalized house, even with perfectly fair odds, your probability of eventual ruin is 1. On an infinite time scale, a finite-bankroll gambler must go broke—the gambler's ruin theorem.
4. Stock Markets and the Illusion of "Trends"
The arcsine law applies not just to casinos but to any sequence that can be modeled as a random walk—including short-term stock price movements.
If daily price changes are random (weak-form efficient markets), then the fraction of time a stock spends above some baseline also follows the arcsine law.
This means: a stock can spend years "mostly rising" with no fundamental support whatsoever—it is just the normal behavior of a random walk.
> This explains why trend-following strategies look effective for a while and then suddenly stop working: you thought you captured a trend, but you were just standing on the positive side of a random walk. When it finally crosses zero (it will—timing just uncertain), your "trend" vanishes.
Under the random-walk framework, much of technical analysis—"support levels," "resistance," "trend lines"—is geometric illusion of noise. Feller gave the mathematical proof in the 1950s, yet seventy-plus years later, countless people still draw trend lines.
5. Pólya's Theorem: Dimension Changes Everything
Another classic in Feller's book is Pólya's recurrence theorem (1921):
- 1D random walk: probability of returning to the origin = 1 (almost surely)
- 2D random walk: probability of returning = 1 (almost surely)
- 3D random walk: probability of returning ≈ 0.34 (may never return)
The deeper meaning: dimension changes everything. One and two dimensions are recurrent; three and above are transient. Physically, this explains why molecular diffusion in 3D space is irreversible—molecules that wander off do not come back.
The theorem also connects to machine learning: data sparsity in high-dimensional spaces (the curse of dimensionality) is essentially the other side of Pólya's theorem. In high dimensions, "neighbors" are no longer near and "return" no longer happens. The math Feller taught in the 1950s remains a key to understanding high-dimensional machine learning today.
6. Why Feller's Book Is Hard but Worth Reading
Feller's book is famously difficult. It is not textbook-style "definition–theorem–proof"; it is dialogic—Feller gives you an intuition, shows why it is wrong, gives the correct intuition, then the rigorous proof. This narrative demands active thinking throughout; you cannot passively absorb it.
But that is exactly what makes it a classic. Feller doesn't just teach theorems—he teaches how to think probabilistically: how to build correct intuition in an uncertain world.
> One review says: "After reading it, the way you see the world changes. You start noticing how many 'patterns' are actually noise, how many 'trends' are actually randomness, how many 'certainties' are merely probability."
Another feature: extremely rich examples, drawn from physics, genetics, epidemiology, actuarial science, and gambling—each demonstrating how probability theory solves real problems. This bridge from mathematics to reality is missing from many modern textbooks.
7. The Humility of Probability
The deepest takeaway, similar to reading ESL-type books: mathematics teaches humility.
Probability theory tells you your intuition is systematically biased under uncertainty. You think "I'm due for a comeback"—but random walks have no memory. You think "I'm on a hot streak"—you've just landed on the positive side. You think diversification lowers risk—but under heavy-tailed distributions, it can make everything fail together.
Feller's book doesn't teach you to compute probabilities—that's the engineer's job. It teaches reverence for uncertainty. In a noisy world, the greatest danger is not lacking the answer, but believing you have it.
The first edition came out in 1950—76 years ago—but its core insight has never aged: randomness is more random than you think, and genuine patterns are scarcer than you think. In the AI era, when large language models generate seemingly reliable content, this probabilistic intuition matters more than ever—you must distinguish what the model actually learned from the illusion of a random walk.
---
Book: *An Introduction to Probability Theory and Its Applications* (Vol. 1, 3rd Ed., 1968; Vol. 2, 2nd Ed., 1971) Author: William Feller (Princeton University) Publisher: Wiley
Link mentioned in the post for obtaining the books: https://b23.tv/4vCEQYn