Understanding conjectures: how to think about an unsolved problem

A conjecture is a statement that seems true —because all the evidence we have supports it— but that no one has managed to prove or disprove. It isn't a law, and it isn't a theorem: it's an educated bet about a pattern. Understanding how these bets are made changes the way you look at any pattern, including the ones that show up in data.

From observation to conjecture

The starting point is almost always the same: someone looks at many particular cases and notices that something keeps repeating. A conjecture appears when you dare to say "this always happens", not just "this happened in the cases I tried".

That leap is exactly the one we make —with less rigor— every time we generalize from a sample. The difference is that in mathematics the standard of proof is absolute: a single counterexample brings down the entire conjecture.

A concrete example: the Collatz conjecture

Take any positive integer $n$ and apply this rule:

The Collatz conjecture states that no matter which $n$ you start with, if you repeat the rule enough times you always reach $1$.

def collatz(n):
    pasos = []
    while n != 1:
        pasos.append(n)
        n = n // 2 if n % 2 == 0 else 3 * n + 1
    pasos.append(1)
    return pasos

collatz(27)  # 111 steps to reach 1

It has been verified by computer for every $n$ up to more than $2^{68}$. And yet nobody knows whether it's true. That gap between "verified in billions of cases" and "proved for all of them" is the heart of what makes a conjecture hard (and fascinating).

The four states of a conjecture

Every statement of this kind lives in one of four states:

• Open — no one has proved or disproved it (Collatz, Goldbach, Riemann).
• Proved — it became a theorem (Fermat's Last Theorem, after 358 years).
• Disproved — a counterexample turned up and knocked it down.
• Undecidable — it is proved that it can be neither proved nor disproved within the system.

Why it matters to anyone who works with data

Modeling is, at its core, conjecturing: we propose that a relationship holds beyond the data we've seen. The discipline of mathematical conjectures teaches three habits that transfer directly:

1. Evidence is not proof. A pattern holding across millions of cases doesn't guarantee the next one. It's the same trap as overfitting.
2. One counterexample is worth more than a thousand confirmations. Actively looking for where your hypothesis fails is more informative than piling up cases that confirm it.
3. Stating the problem well is half the battle. A precise conjecture can be attacked; a vague one can't.

Going further

In upcoming posts in this section I want to walk through specific conjectures —Goldbach, the Riemann hypothesis— not to prove them, but to show how one thinks around a problem that still has no answer.