Add a long list of numbers on a computer in one order, then add the very same numbers in a different order, and you can get a different total. That unsettling fact is real, and it turns on the distinction this page draws: whether an infinite sum is robust enough to survive being reshuffled, or holds together only by a fragile balance of pluses and minuses.

There are two grades of convergence, and the difference between them is the difference between a bridge and a house of cards. Some series converge the hard way: their terms shrink so fast that even the sizes |a_n| add up to something finite. Strip every minus sign, imagine the worst case where nothing cancels — the sum still lands. That is rock-solid convergence.

Other series converge only by the grace of cancellation. Their positive and negative terms are each, on their own, an infinite reservoir; the series settles on a finite value purely because the two reservoirs keep taking turns wiping each other out. Take the absolute values and the whole thing blows up. That is fragile convergence — alive only thanks to a delicate choreography of signs.

Why care about the distinction? Because it decides what you are allowed to do with a series. Rearrange the terms? Regroup them? Multiply two series together term by term? Swap a sum with an integral? For the rock-solid kind, all of that is safe — an absolutely convergent series behaves like a well-mannered finite sum. For the fragile kind, almost none of it is: as we'll see, merely reordering the terms of a conditionally convergent series can change its sum to any number you fancy. This page is about telling the two grades apart — and about why "check the absolute values first" is the working analyst's opening move on any series with mixed signs.

The definitions

A series \sum a_n converges absolutely when the series of absolute values converges:

It converges conditionally when \sum a_n converges but \sum |a_n| does not. And of course a series may simply diverge. So every series with mixed signs sits on exactly one rung of a three-rung ladder:

(read: each rung is strictly stronger than the one below). The headline theorem — proved in the next card — is that the top rung really does sit above the middle one: stripping the signs and still converging is enough to guarantee the original signed series converges too. The alternating series test you already know supplies plenty of residents for the middle rung.

Absolute convergence implies convergence

Suppose \sum |a_n| converges. We want to conclude that \sum a_n converges. The trick is to bound the messy signed term by something non-negative whose series we already control.

Step 1 — sandwich the shifted term. For every real a_n we have -|a_n| \le a_n \le |a_n|. Add |a_n| throughout to make the left edge zero:

Step 2 — name the non-negative series. Set b_n = a_n + |a_n|. By Step 1 every b_n \ge 0, and each is dominated by 2|a_n|. (Concretely: b_n = 2a_n when a_n \ge 0 and b_n = 0 when a_n < 0 — the shift simply switches the negative terms off.)

Step 3 — compare. The series \sum 2|a_n| = 2\sum |a_n| converges (a convergent series times a constant). Since 0 \le b_n \le 2|a_n|, the comparison test forces \sum b_n to converge as well.

Step 4 — undo the shift. Now solve for a_n: a_n = b_n - |a_n|. Both \sum b_n and \sum |a_n| converge, and the difference of two convergent series converges:

Step 5 — read off the conclusion. We started only with the convergence of \sum |a_n| and derived the convergence of \sum a_n. So absolute convergence is the stronger property — it comes with ordinary convergence for free, plus (as we'll see) a bundle of rearrangement rights that ordinary convergence alone never grants.

The three-tier lineup, worked

Here are three series that look like siblings — same alternating skeleton, different decay — and land on all three rungs of the ladder. Classifying them is a two-question routine: does \sum|a_n| converge? If not, does \sum a_n converge anyway?

Tier 1 — absolutely convergent: \sum_{n=1}^{\infty} \frac{(-1)^n}{n^2}. Strip the signs: \sum |(-1)^n/n^2| = \sum 1/n^2, a convergent p-series (p = 2 > 1; its sum is the famous \pi^2/6). The absolute series converges, so we stamp absolutely convergent and stop — the theorem hands us convergence of the signed series without our ever looking at the signs. The alternation was a bonus, not a lifeline.

Tier 2 — conditionally convergent: \sum_{n=1}^{\infty} \frac{(-1)^{n+1}}{n}. Strip the signs: \sum 1/n, the harmonic series — divergent. So absolute convergence fails, and we must ask the second question. The terms 1/n decrease monotonically to 0, so the alternating series test applies: \sum (-1)^{n+1}/n converges (to \ln 2). Converges, but not absolutely: conditionally convergent. This series is alive only because of cancellation.

Tier 3 — divergent: \sum_{n=1}^{\infty} (-1)^n = -1 + 1 - 1 + 1 - \cdots The terms don't tend to 0 (they bounce between \pm 1 forever), so by the divergence test the series diverges — its partial sums oscillate -1, 0, -1, 0, \ldots and never settle. Alternating signs alone rescue nothing; cancellation only helps when the terms themselves are dying away.

Same skeleton, three fates. The decay rate of |a_n| — faster than 1/n, exactly harmonic-slow, or not decaying at all — is what moves a series up and down the ladder.

When the signs are a mess: \sum \sin(n)/n^2

The alternating series test needs a tidy +,-,+,- pattern. But consider

The numerator \sin(n) wobbles through positive and negative values in a pattern with no period at all (because \pi is irrational, the signs never repeat): \sin 1 > 0, \sin 2 > 0, \sin 3 > 0, \sin 4 < 0, … The alternating series test is useless here. This is precisely where absolute convergence earns its keep: take absolute values and the sign chaos evaporates.

The series \sum 1/n^2 converges, so by the comparison test \sum |\sin(n)/n^2| converges — the series converges absolutely, and therefore converges, and we never had to understand the signs at all. That is the real power of the theorem: it converts a hard question about a wildly signed series into an easy question about a positive one.

This suggests the working strategy for any series with mixed or messy signs:

1. Check absolute convergence first. Run your positive-series toolkit (comparison, ratio, root, p-series) on \sum|a_n|. If it converges — done: the series converges absolutely, and every rearrangement right comes bundled in.
2. If \sum|a_n| diverges, don't despair — go conditional. Try the alternating series test (or subtler cancellation tests) on the signed series itself. Success means conditionally convergent: handle with care, no shuffling.
3. If the terms don't even tend to 0, stop. The series diverges, signs or no signs.

See conditional convergence at work

Take the alternating harmonic series \sum (-1)^{n+1}/n. The bold curve is its partial sum S_N = \sum_{n\le N}(-1)^{n+1}/n, which settles down toward \ln 2 \approx 0.693. The faint curve is the partial sum of the absolute values \sum_{n\le N} 1/n — the harmonic series, which climbs without bound. Drag N: the signed sums converge while the absolute sums run away.

That widening gap between the two curves is conditional convergence, drawn. The upper curve says the raw material — the total quantity of stuff being added — is infinite. The lower curve says the signed bookkeeping nevertheless nets out to a finite number. The series works like a bank account with infinite total deposits and infinite total withdrawals, carefully interleaved so the balance stays tame. And that image is exactly the loose thread Riemann pulled.

Riemann's rearrangement theorem: any sum you like

Conditional convergence is not merely a weaker stamp of approval; it is structural instability. Bernhard Riemann proved one of the most startling facts in analysis: if \sum a_n converges conditionally, then for any target L \in [-\infty, +\infty] — any real number, or even \pm\infty — there is a reordering of the very same terms whose series converges to L.

The mechanism is exactly the failure of absolute convergence. Split the terms into positives and negatives. Because \sum|a_n| diverges while \sum a_n converges, the positive part and the negative part must each diverge — an infinite reservoir of positive sum, an infinite reservoir of negative sum, and individual terms shrinking to zero. With those three ingredients you can steer the partial sums anywhere, by a greedy algorithm: below target? spend positives. Above target? spend negatives. Repeat forever.

Watch it run on the alternating harmonic series, aiming at L = 1.5 (remember, in its usual order this series sums to \ln 2 \approx 0.693):

Add odd reciprocals until you first poke above 1.5; then one even reciprocal drags you below; then back to the odds. Every term gets used exactly once — this really is a rearrangement, not a selection. Two facts make it converge to the target: the reservoirs never run dry (each part diverges, so you can always climb or descend as far as needed), and the overshoot at each turn is at most the size of the last term used — which tends to 0. So the partial sums squeeze onto 1.5 exactly.

Want +\infty instead? Climb to 1, spend one negative term, climb to 2, spend one negative, climb to 3… The negative terms trickle in but can never keep up. Every target is reachable. The symbol "\sum a_n" for a conditionally convergent series secretly carries its ordering as part of its meaning.

Steer the series yourself

The chart below runs the greedy algorithm live on the terms \pm 1/n of the alternating harmonic series. Pick a target L with the slider; the curve shows the partial sums of the rearranged series — climbing on positive terms, dipping on negative ones — as they zigzag onto your chosen value. Slide L from -1 to 2.5 and back: the terms never change, only their order, yet the series obediently converges wherever you point it.

Notice how the zigzags get finer as you move right: the overshoot at each reversal is one term, and the terms shrink like 1/n. Notice too how a high target makes the curve spend long stretches climbing — the positive reservoir pays out ever more slowly, but it never runs out. An absolutely convergent series could never be steered like this: with \sum|a_n| finite, every rearrangement is forced to the same sum.

See it explained