What is a mathematical proof? How mathematicians know for sure

A mathematical proof is an argument that shows a statement must be true, by a chain of steps in which each one follows by strict logic from agreed starting points, called axioms, or from results already proved. Once proved, a statement is true in every case it covers, not just the cases anyone checked. That is what sets mathematics apart: a proof settles a question for good, where experiments and examples can only make an answer likely.

In short

• A proof starts from axioms and definitions and moves by valid logical steps to a conclusion that cannot be false if the starting points are true.
• Common kinds are direct proof, proof by contradiction and proof by induction, each a different way of building the chain.
• Checking examples, even billions of them, is evidence and not proof: some claims hold for hundreds of millions of cases and then fail.

What makes a proof different from evidence?

In science, a claim earns trust by surviving tests, and a better test can always overturn it. In mathematics a claim about all whole numbers can't be tested on all of them, because there are infinitely many. Instead it is proved: shown to follow, step by step, from things already accepted. Each step must be a valid inference, one where the conclusion is forced by what came before. That is deductive reasoning, and its rules are the subject of logic.

The result is a different kind of certainty. A proof shows that a statement is true for every case, including the infinitely many no one will ever check. The Pythagorean theorem has a proof in Euclid's Elements, written around 300 BCE, and no right triangle drawn on a flat surface since has broken it, or ever will.

What are axioms, and why does a proof need them?

Every argument has to start somewhere, and a proof can't prove its own starting points without going in circles. So mathematics states them openly. Axioms are basic statements accepted without proof, chosen because they seem self-evident or because they define the system being studied. Definitions fix exactly what words mean, such as "even" or "prime".

Euclid's Elements is the model. It begins with 23 definitions, 5 postulates and 5 common notions, such as "things equal to the same thing are equal to each other", and builds hundreds of propositions from them (Euclid's Elements, Clark University edition). Today most of mathematics can be built from one set of axioms about collections of objects, Zermelo–Fraenkel set theory, usually with one more axiom, the axiom of choice, added (ZFC).

Proof is therefore relative to its axioms. Change Euclid's fifth postulate, about parallel lines, and you get geometries where the angles of a triangle don't add to 180 degrees, like the surface of a sphere. Both kinds of geometry are consistent; which one describes physical space is a question for physics, not proof.

What does a proof look like?

Three short proofs show the three most common shapes.

Direct proof: the sum of two even numbers is even

Start from the definition. An even number is one that equals 2 times a whole number. So take any two even numbers and write them as 2a and 2b, where a and b are whole numbers. Their sum is 2a + 2b, which is 2 × (a + b). Since a + b is a whole number, the sum is 2 times a whole number, which is the definition of even. Done. Nothing was checked by example; the algebra covered every pair of even numbers at once.

Proof by contradiction: the square root of 2 is not a fraction

Assume the opposite and show it leads to something impossible. Suppose √2 could be written as a fraction p/q in lowest terms, so p and q share no common factor. Squaring gives 2 = p²/q², so p² = 2q². That makes p² even, and an odd number squared is odd, so p itself is even: p = 2r for some whole number r. Then 4r² = 2q², so q² = 2r², and by the same reasoning q is even too. But then p and q share the factor 2, contradicting "lowest terms". The assumption must be false, so √2 is not a fraction. That fact, known to the ancient Greeks, is why the real numbers include irrationals. A proof of the same shape shows there are infinitely many prime numbers.

Proof by induction: 1 + 2 + … + n = n(n + 1) ÷ 2

Induction proves a statement for every whole number in two moves, like knocking over a line of dominoes. First, show it is true for 1. Second, show that whenever it is true for some number n, it must also be true for n + 1. Then it is true for 1, so for 2, so for 3, and so on without end.

The base case: for n = 1 the left side is 1, and the right side is 1 × 2 ÷ 2 = 1. The step: suppose 1 + 2 + … + n = n(n + 1) ÷ 2. Adding the next number, n + 1, to both sides gives n(n + 1) ÷ 2 + (n + 1), which simplifies to (n + 1)(n + 2) ÷ 2. That is exactly the formula with n + 1 in place of n. So the formula holds for every whole number; for 1 to 100 it gives 100 × 101 ÷ 2 = 5,050.

Despite its name, mathematical induction is deduction, not the inductive reasoning of science that generalizes from examples. The second step is proved for every n at once.

Why isn't checking lots of examples enough?

Because patterns can hold for a very long time and then stop. These are real cases:

Claim	How far it was checked	What happened
Collatz conjecture: repeatedly halve an even number, or triple an odd one and add 1, and you always reach 1	Every start up to 2⁷¹, about 2.36 × 10²¹, by 2025	Still unproved
The Pólya conjecture of 1919: up to any number, at least half the numbers have an odd count of prime factors, counted with repetition	Held for every number below 906,150,257	Fails at 906,150,257 (Minoru Tanaka, 1980)
Euler's sum of powers conjecture: at least five fifth powers are needed to add up to a fifth power	Believed for almost two centuries	False: 27⁵ + 84⁵ + 110⁵ + 133⁵ = 144⁵ (Lander and Parkin, 1966)
Fermat's Last Theorem: no whole numbers solve aⁿ + bⁿ = cⁿ for n above 2	Checked for huge exponents by computer	Proved by Andrew Wiles in 1994, after more than 350 years

The Collatz conjecture has been checked by computer for every starting number up to 2⁷¹, according to a 2025 paper by David Barina, and every one reaches 1. That is overwhelming evidence, and it is still not a proof: the Pólya example shows a pattern can survive the first 906 million cases and then break. Mathematicians believe the Collatz conjecture, but no one has proved it.

Who checks a proof, and can computers do it?

Traditionally, other mathematicians. A new proof is read line by line by referees before a journal publishes it, and important ones are picked over for years. Wiles's first announced proof of Fermat's Last Theorem, in 1993, had a gap that took him and Richard Taylor about a year to repair.

Computers now help in two ways. Some proofs use them to check too many cases for people: the four color theorem, that any map can be colored with four colors so no neighbors match, was proved in 1976 by Kenneth Appel and Wolfgang Haken with computer checks of more than 1,400 configurations. More recently, proof assistants check every logical step themselves. In 2005 Benjamin Werner and Georges Gonthier rebuilt the four color proof inside the Coq proof assistant, and a community library for the Lean proof assistant, mathlib, is building a single library of mathematics in which every proof has been checked by machine.

There are limits to what any proof can do. In 1931 Kurt Gödel showed that any consistent formal system able to do basic arithmetic contains statements it can neither prove nor disprove (Stanford Encyclopedia of Philosophy). Proof gives certainty, but only within a system, and no single system captures every truth about numbers.

Check yourself

• Without looking back, explain why checking a claim for a billion numbers doesn't prove it, using one of the examples in the table.
• In the proof that √2 isn't a fraction, which assumption leads to the contradiction, and what is the contradiction?
• What are the two steps of a proof by induction, and why do they together cover every whole number?

Sources

• Euclid's Elements — David E. Joyce, Clark University
• Gödel's incompleteness theorems — Stanford Encyclopedia of Philosophy, revised 2025
• Mathlib overview — Lean community
• Collatz conjecture — Wikipedia
• Pólya conjecture — Wikipedia
• Mathematical proof — Wikipedia