Infinite sequences and convergence

In brief

Ask for the square root of two by an old trick: guess a number, average it with two divided by your guess, and use that average as your next, better guess. Repeat forever and the guesses visibly tighten in on 1.41421 and beyond, never quite arriving but getting closer at every step. Now compare that to a different endless procedure: repeatedly flipping the sign of a number, one, minus one, one, minus one, on and on. Both are infinite lists of numbers produced by an endless rule, but only the first one is heading somewhere. The question this entry answers is how to tell the difference precisely, not by eyeballing a trend, but by a test that either succeeds or fails for a stated destination. That test is convergence, and it is what allows mathematics to say something definite about a process that never actually finishes.

The full treatment

First look: an endless method for a square root

Start with a guess a_1 = 1 for the square root of two. The rule is a_{n+1} equals one half of (a_n plus two divided by a_n). Compute a_2 = 0.5*(1 + 2/1) = 1.5. Then a_3 = 0.5*(1.5 + 2/1.5), which is about 1.41667. Then a_4 comes out to about 1.41422. The terms are visibly crowding toward 1.41421356, the true square root of two, and the amount they still have to travel shrinks fast, roughly by a factor each step. The procedure never stops, there is no final term, yet everyone would agree the sequence is "going to" the square root of two. Making that agreement precise, rather than a matter of eyeballing a table of numbers, is the whole content of this entry.

A definition strict enough to settle any case

A sequence is an ordered infinite list of numbers a_1, a_2, a_3, and so on. The sequence converges to a value L if, however small a tolerance you name, call it epsilon, there comes a point in the list, call its index N, beyond which every single term is within epsilon of L. In symbols: for every epsilon greater than zero, there exists an N such that for all n greater than N, the distance between a_n and L is less than epsilon. This is stronger than saying "the terms get closer to L." It says that for any degree of closeness you demand in advance, no matter how strict, the sequence eventually satisfies that demand and never again violates it. Check the square root example: name any tolerance, say a millionth, and the doubling-precision behavior of the averaging trick guarantees that after a handful of steps, every later term stays within a millionth of the true square root, forever after. That is what makes the claim "the sequence converges to the square root of two" a checkable fact rather than an impression.

Why "keeps getting closer" is not enough

It is tempting to think that a sequence which keeps moving toward some value must converge to it, but this is false, and seeing why sharpens the definition. Imagine a sequence built so that each term closes only half the remaining gap to a supposed limit, but a stray disturbance occasionally pushes it back out by a fixed amount larger than the gap it just closed. The terms can spend most of their time approaching, yet never satisfy the strict demand of staying within every small tolerance forever after, because the disturbances recur without end. Separately, the oscillating sequence one, minus one, one, minus one never approaches any single value at all, despite staying bounded between minus one and one; it fails to converge not because it runs off to infinity but because it refuses to settle on one destination.

Testing convergence without already knowing the destination

The square root example is easy to check because the destination, the square root of two, is known in advance. Most interesting sequences arise from a procedure whose final value is exactly what you are trying to discover, so testing distance to a known L is not available. The way around this, due to Cauchy, is to test whether the terms are getting arbitrarily close to each other, rather than to some named limit: the sequence is called a Cauchy sequence if, for every tolerance epsilon, there is a point beyond which any two terms, however far apart in the list, are within epsilon of one another. A sequence of real numbers converges if and only if it is a Cauchy sequence. The averaging terms above satisfy this: term one hundred and term one hundred and one differ by an amount that keeps shrinking, and so do term one hundred and term one million, even without anyone having first computed the square root of two by other means. This criterion turns convergence into a property you can verify from the sequence's own internal behavior.

A guaranteed case: monotone and bounded

One especially useful sufficient condition needs no tolerance-chasing at all. If a sequence is monotone, meaning it only increases (or only decreases) from term to term, and it is bounded, meaning it never exceeds some fixed ceiling (or falls below some fixed floor), then it is guaranteed to converge to some limit, even if that limit is not obvious in advance. The reasoning is intuitive: a sequence that only ever climbs, but is fenced in below a ceiling, has nowhere left to go except pack itself arbitrarily close to some highest value it never quite exceeds. This theorem is often the fastest way to prove a sequence converges when a direct tolerance argument would be awkward.

Lineage

The trouble with infinite procedures is ancient: Zeno's paradoxes of motion, in fifth-century BCE Greece, turned an endless halving of distance into a puzzle about whether motion could even be completed. Archimedes used a controlled version of an infinite procedure, the method of exhaustion, to compute areas and volumes centuries before calculus existed, effectively working with convergent sequences without a formal theory of them. Isaac Newton's iterative root-finding methods in the seventeenth century relied implicitly on convergence without stating a rigorous criterion. It was Augustin-Louis Cauchy, in his 1821 Cours d'Analyse, who first gave convergence a tolerance-based definition and the criterion that tests a sequence against itself rather than a known limit, and Karl Weierstrass, later in the nineteenth century, who sharpened the epsilon-based language into the form used today.

The strongest case for it

The precise definition of convergence lets mathematics make definite, checkable claims about processes that never terminate, which is otherwise impossible: without it, "the guesses are approaching the square root of two" would be an impression rather than a provable fact. Every numerical method used to approximate roots, integrals, or solutions to differential equations depends on being able to say, rigorously, that repeated refinement eventually stabilizes to within any desired accuracy, and this framework is the reason such a guarantee can be stated and proven rather than merely hoped for.

The strongest case against it

The definition is powerful but purely formal, and formal correctness does not by itself produce intuition; students often absorb the epsilon-N pattern as a ritual to be recited rather than a genuine tolerance-chasing argument, and this defeats its purpose. A related honest limitation: convergence, once defined, still has to be verified case by case, and there is no universal shortcut. Two common misconceptions are worth naming directly. First, a sequence can keep getting closer to a value at every single step and still fail to converge to it, if the amount of remaining approach never shrinks below every tolerance, as certain slowly diverging sequences show. Second, boundedness alone does not guarantee convergence, as the oscillating sequence one, minus one, one, minus one demonstrates; boundedness only guarantees convergence when paired with monotonicity.

Where it stands now

The theory of convergence for sequences of real numbers is completely settled and has been since Cauchy and Weierstrass; nothing about the definition or the monotone bounded sequence theorem is contested. What remains genuinely open, on a case-by-case basis, is whether a specific sequence produced by a specific algorithm converges and how fast, which is an active area of numerical analysis whenever a new iterative method is proposed.

Test yourself

Consider the rule that starts with a guess b_1 = 3 and repeatedly sets b_{n+1} equal to one half of (b_n plus ten divided by b_n), intended to approximate the square root of ten. Compute the first four terms by hand and describe what tolerance-based test you would apply to argue the sequence converges, without simply asserting that it looks like it is settling down. Then invent a second iterative rule, of your own design, that produces a bounded sequence which nonetheless fails to converge, and explain precisely which part of the convergence definition your example violates.