The same terms, in a different order, adding to whatever is asked

The harmonic series $1 + \tfrac12 + \tfrac13 + \tfrac14 + \cdots$ grows without bound, slowly and forever. Now change one thing: alternate the signs.

$$1 - \frac12 + \frac13 - \frac14 + \frac15 - \cdots$$

It converges, and it converges to $\ln 2 = 0.693147\ldots$, which is the first surprise. The second is larger: the same terms, in a different order, converge to something else. Any something else. Pick a number in advance and there is an ordering of these terms that adds up to it.

Why it converges, and why that is easy #

The convergence needs no analysis at all — the picture is the proof.

Every partial sum overshoots and undershoots alternately, because each new term is added or subtracted and is smaller than the last. So the odd partial sums form a decreasing sequence, the even ones an increasing sequence, and every even sum is below every odd one. Two monotone sequences closing on each other from opposite sides have to meet, and whatever they meet at is the sum.

The gap between consecutive partial sums is exactly the next term, $1/(n+1)$, which goes to zero. So the sequences do not merely approach each other; the distance between them is known at every stage, and the error after $n$ terms is smaller than the term not yet added. That is the alternating series test, and it is one of the few convergence criteria that hands over an error bound as a free consequence rather than as extra work.

It is worth pausing on how little was needed. The terms decrease, they alternate, and they go to zero. Nothing about $1/n$ specifically, nothing about logarithms. Any decreasing sequence tending to zero converges when its signs alternate, and the harmonic terms qualify by the narrowest possible margin — they are exactly the sequence that fails to converge without the signs.

Where the logarithm comes from #

That the answer is $\ln 2$ is not obvious from the picture, and the shortest honest route to it goes through an area.

The partial sums of the alternating series can be rewritten by pairing terms and cancelling: $H_{2n} - H_n$, where $H_k$ is the ordinary harmonic sum. And $H_k$ grows like $\ln k$, with an error settling on the Euler–Mascheroni constant, so the difference is $\ln 2n - \ln n = \ln 2$, and the constants cancel because they appear in both.

The alternative route is to integrate the geometric series $1 - x + x^2 - \cdots = 1/(1+x)$ term by term from $0$ to $1$, which turns each power into a reciprocal and produces $\ln 2$ on the other side. That derivation is quicker and it is also where the difficulty is hiding: integrating a series term by term is exactly the operation that stops being safe for a series like this one, and the essay’s second half is about why.

The two halves, each divergent #

Split the series in two: the positive terms and the negative ones.

$$1 + \frac13 + \frac15 + \cdots \qquad\text{and}\qquad \frac12 + \frac14 + \frac16 + \cdots$$

Both diverge. The second is half the harmonic series and so is unbounded; the first is larger than the second term by term, so it is unbounded too.

That is the structural situation, and everything strange follows from it. The series converges not because its terms are small but because two unbounded quantities are cancelling. The finite answer is the residue of an infinite amount of positive against an infinite amount of negative, and a residue of that kind depends on the order in which the cancelling is done.

This is the distinction between absolute and conditional convergence. A series converges absolutely when the sum of the absolute values converges; then the order does not matter and the sum is a number attached to the set of terms. A series converges conditionally when it converges and its absolute version does not; then the order is part of the answer.

Riemann’s theorem, and how to aim it #

Riemann’s rearrangement theorem, published posthumously in 1867: a conditionally convergent series can be rearranged to converge to any prescribed value, or to diverge to either infinity, or to oscillate without settling anywhere.

The construction is greedy and it is entirely explicit. To reach a target, take positive terms in order until the running total exceeds it. Then take one negative term. Then more positives until it exceeds again, then one negative, and so on.

It works for two reasons, and both were established above. The positives are unbounded, so the target can always be exceeded eventually. The terms go to zero, so the overshoot at each crossing shrinks, and the total is squeezed onto the target rather than merely wobbling near it.

Every one of the original terms gets used exactly once, because the positives are consumed in order and so are the negatives, and both supplies are infinite. So the rearranged series has the same terms as the original, each appearing once, and a different sum.

The cheapest rearrangement, and what it costs to notice #

The greedy construction is the general one and it is not the most persuasive, because a reader can always suspect that something underhand is happening in the ordering. A cleaner instance settles that.

Take the terms in the pattern two positives, one negative, forever:

$$1 + \frac13 - \frac12 + \frac15 + \frac17 - \frac14 + \frac19 + \frac1{11} - \frac16 + \cdots$$

Nothing is skipped. The odd reciprocals appear in order, the even ones appear in order, and every term of the original series occurs exactly once. The pattern is regular enough to write down in a line, which is the point — no adversarial choice is being made anywhere.

The sum is $\tfrac32 \ln 2$.

That is a fifty per cent increase produced by a change no reader would think to object to. Nothing was inserted, nothing was dropped, and the ordering is not merely legal but tidy. The general theorem says any target is reachable; this instance says the damage begins immediately, at the first mildly unusual arrangement anybody would write.

The derivation is short enough to state. Write $S$ for the original sum, and consider the series formed by inserting a zero after each pair — $1, 0, -\tfrac12, \tfrac13, 0, -\tfrac14, \ldots$ — which is $\tfrac12 S$ with the terms spread out. Adding that to the original term by term, and grouping in threes, produces exactly the two-positives-one-negative arrangement. So the rearranged sum is $S + \tfrac12 S$.

The manoeuvre that makes this legal is adding two series term by term, which requires them to be aligned — and the alignment is where the whole effect lives. The same terms, offset, added to themselves, produce a different total. This is the same category of trap as a Taylor series that converges everywhere and to the wrong thing: an operation that is unimpeachable for finite sums has been extended by analogy, and the analogy does not hold.

What has actually broken #

It is tempting to say addition is not commutative for infinite series, and that phrasing is slightly wrong in a way worth fixing.

Addition of any two numbers is still commutative, and any finite rearrangement leaves the sum alone. What fails is the exchange of two limits. An infinite sum is not an addition at all — it is the limit of a sequence of finite additions — and rearranging the terms rearranges the sequence of partial sums, which is a different sequence with a different limit. Nothing about addition broke; a limit was taken of a different thing.

That is the recurring shape of every paradox in this area. A function that is $1$ on the rationals and $0$ elsewhere has no Riemann integral not because area is ill-defined but because two ways of taking a limit disagree. A Taylor series can converge and converge to the wrong function for the same reason. In each case an operation that is obviously safe on finitely many things has been applied to infinitely many, and the safety did not come along.

The general principle behind the safe cases is worth stating because it is the one that gets used: absolute convergence is what licenses rearrangement, term-by-term integration, and multiplication of series. All three of those are exchanges of limits, and absolute convergence is the hypothesis that makes them legal. A series that converges only conditionally has none of those privileges, and the theorem above is what happens when one of them is taken anyway.

What it costs #

There is a practical residue, and it is not academic.

Summing a conditionally convergent series numerically is an unstable operation: the answer depends on the order, and floating-point addition reorders nothing but rounds every partial sum, which for a series whose partial sums oscillate widely means the rounding errors do not cancel. The alternating harmonic series is a standard example of a series nobody should sum directly — reaching six correct digits needs about a million terms, and the accumulated rounding is comparable to the answer.

The fix is to transform rather than to sum harder. Euler’s transformation, or any of the acceleration schemes descended from it, converts an alternating series into one that converges geometrically, and gets those six digits in a dozen terms. This is a general lesson about slow convergence: the remedy is almost never more terms, because the arithmetic degrades at roughly the rate the series improves.

There is a cheaper trick available here and it is instructive because it is a rearrangement used deliberately. Averaging consecutive partial sums places the result between the two envelopes rather than on one of them, which removes most of the oscillation at once; averaging the averages removes most of what is left. Three rounds of that on twenty terms beats a million terms of direct summation. What makes it legal is that averaging partial sums is not reordering the series — it is a statement about the sequence of partial sums, which is the object that actually converges.

That distinction is the practical form of the whole essay. Operations on the sequence of partial sums are safe; operations on the terms are not, unless the series converges absolutely. Every acceleration method is on the first side of that line and every paradox in this essay is on the second, and knowing which side an operation is on is most of numerical analysis. It is the same discipline as knowing that a Riemann sum’s limit depends on more than the rectangles shrinking.

Where it does not apply #

The theorem needs conditional convergence, and it is worth being precise about what is being ruled out.

An absolutely convergent series is immune: every rearrangement gives the same sum, and the proof is that the tail can be made small independently of the order. That is what makes absolute convergence the right notion, and it is why analysis defines it at all rather than treating it as a curiosity.

The theorem also does not extend to series of vectors in the way one might expect. For a conditionally convergent series in the plane, the set of achievable sums is not everything — the Lévy–Steinitz theorem says it is an affine subspace, a point or a line or the whole plane, and never anything more complicated. So the real line is the case where rearrangement is maximally destructive, and adding dimensions makes the damage more structured rather than less.

What the picture cannot show #

The rearrangement figure walks a hundred and twenty terms toward its target and stops. The theorem is about a limit, and the picture shows a beginning: nothing in it rules out the total wandering off after term ten thousand. What the generator checks is the part that can be checked at finite length — that every term used is one of the originals, that none is used twice, and that the walk is near the target when it stops.

Nor can any of these figures show the essential fact, which is that the two halves are separately infinite. That is a statement about what happens beyond every frame, and it is the reason for everything the frames do show. A reader who has not accepted the divergence of the harmonic series has no reason to believe any of this, and the divergence is precisely what a picture cannot supply — it is invisible at every finite size.

The ladder from here #

Rungs on this anchor: Oresme’s grouping proof, already below this one. The integral test, which is the general machine that decides these questions and which turns the sum into an area under a curve. The Euler–Mascheroni constant, whose irrationality is open. The block-stacking problem, where the harmonic sum is an overhang. The $p$-series and the boundary at $p = 1$. The prime harmonic series, whose divergence is a stronger statement about the primes than their infinitude. The Basel problem, where the squares converge and $\pi$ appears from nowhere. And the zeta function, which is what this series becomes when the exponent is allowed to be complex.

The moral is about hypotheses, not about infinity #

The usual reading of this theorem is that infinity is strange. That reading is available and it is not the useful one.

The useful reading is that the theorem is a demonstration of what a hypothesis is for. “Absolutely convergent” appears in the statement of a dozen results that would otherwise be about rearrangement, integration, differentiation and multiplication of series, and to a reader meeting those results it looks like a technicality — a phrase to be checked and moved past.

This theorem is what the phrase is holding back. Drop it and the conclusion does not weaken slightly; it fails completely, and it fails in the strongest imaginable way, with the sum becoming an arbitrary parameter. A hypothesis whose removal costs everything is not a technicality, and the way to find out which hypotheses those are is to try removing them and see what walks through the gap.