Riemann's Rearrangement Theorem

Can rearranging alone manufacture any sum we want?

The greedy rule is simple. Below the target, take a positive term; above it, take a negative one. The terms shrink, so each overshoot shrinks too. That is why any real target can be hit exactly.

For example, think of steering a πŸš— car. Turn more when short of the spot, back when past it. Small corrections land you exactly where you aim.

βˆ‘anΒ convergesΒ conditionallyΒ βŸΉΒ βˆ€β€‰s∈RΒ Β βˆƒΒ rearrangementΒ withΒ sumΒ s\sum a_n\ \text{converges conditionally}\ \Longrightarrow\ \forall\, s\in\mathbb{R}\ \ \exists\ \text{rearrangement with sum } s

Part of these learning paths

Examples & Exercises

Let βˆ‘an\sum a_n be a real series converging conditionally.

With pn=max⁑(an,0)p_n = \max(a_n, 0) and qn=max⁑(βˆ’an,0)q_n = \max(-a_n, 0) (zero terms are harmless, so this is the positive/negative subsequences re-indexed), prove βˆ‘pn=+∞\sum p_n = +\infty and βˆ‘qn=+∞\sum q_n = +\infty.

an=pnβˆ’qn,∣an∣=pn+qna_n = p_n - q_n, \qquad \lvert a_n \rvert = p_n + q_n
βˆ‘pn<βˆžΒ βŸΉΒ βˆ‘qn=βˆ‘pnβˆ’βˆ‘an<βˆžΒ βŸΉΒ βˆ‘βˆ£an∣<∞(contradiction)\sum p_n < \infty \ \Longrightarrow\ \sum q_n = \sum p_n - \sum a_n < \infty \ \Longrightarrow\ \sum \lvert a_n \rvert < \infty \quad \text{(contradiction)}
Target s=1s = 1, N=120N = 120: S120β‰ˆ0.9967S_{120} \approx 0.9967, error β‰ˆ0.00326\approx 0.00326 β€” 78 positive and 42 negative takes, and both blocks recur forever (the proof below is why).
Show proof
  1. Termwise an=pnβˆ’qna_n = p_n - q_n and ∣an∣=pn+qn\lvert a_n \rvert = p_n + q_n (zero terms contribute nothing, so this is the positive/negative subsequences re-indexed). Conditional convergence means βˆ‘(pnβˆ’qn)\sum (p_n - q_n) converges while βˆ‘(pn+qn)\sum (p_n + q_n) diverges.

  2. Suppose βˆ‘pn<∞\sum p_n < \infty. Since βˆ‘an=βˆ‘(pnβˆ’qn)\sum a_n = \sum (p_n - q_n) converges, and qn=pnβˆ’(pnβˆ’qn)q_n = p_n - (p_n - q_n), the series βˆ‘qn=βˆ‘pnβˆ’βˆ‘an\sum q_n = \sum p_n - \sum a_n converges as a difference of convergent series.

  3. Then βˆ‘βˆ£an∣=βˆ‘pn+βˆ‘qn<∞\sum \lvert a_n \rvert = \sum p_n + \sum q_n < \infty β€” so βˆ‘an\sum a_n would converge absolutely, contradicting conditional convergence. Hence βˆ‘pn=+∞\sum p_n = +\infty.

  4. Symmetrically: if βˆ‘qn<∞\sum q_n < \infty, then pn=qn+(pnβˆ’qn)p_n = q_n + (p_n - q_n) gives βˆ‘pn=βˆ‘qn+βˆ‘an\sum p_n = \sum q_n + \sum a_n convergent, again forcing βˆ‘βˆ£an∣<∞\sum \lvert a_n \rvert < \infty β€” contradiction.

    So βˆ‘qn=+∞\sum q_n = +\infty.

    The intuition: a conditionally convergent series harbors both "positive infinity" and "negative infinity", and a rearrangement can shuttle between them to hit any target.

Aim the scene's greedy walk at target s=1s = 1.

Compute the partial sums S2,S3,S4,S5,S7,S8S_2, S_3, S_4, S_5, S_7, S_8 at the ends of the first few blocks by hand, and check that each overshoot/undershoot is at most the magnitude of the term that caused the crossing.

S2=43,Β S3=56,Β S4=3130,Β S5=4760,Β S7=13071260,Β S8=10971260S_2 = \tfrac43,\ S_3 = \tfrac56,\ S_4 = \tfrac{31}{30},\ S_5 = \tfrac{47}{60},\ S_7 = \tfrac{1307}{1260},\ S_8 = \tfrac{1097}{1260}
gaps:Β 13,Β 16,Β 130,Β 1360,Β 471260,Β 1631260 ≀ crossingΒ termΒ eachΒ time\text{gaps: }\tfrac13,\ \tfrac16,\ \tfrac1{30},\ \tfrac{13}{60},\ \tfrac{47}{1260},\ \tfrac{163}{1260}\ \le\ \text{crossing term each time}
Target s=1s = 1, N=120N = 120: the bracket tightens 13β†’16β†’130β†’1360β†’471260β†’1631260\frac13 \to \frac16 \to \frac1{30} \to \frac{13}{60} \to \frac{47}{1260} \to \frac{163}{1260}, reaching S120β‰ˆ0.9967S_{120} \approx 0.9967 with error β‰ˆ0.00326\approx 0.00326.
Show proof
  1. The rule: take the next positive term while t≀st \le s, the next negative term while t>st > s. For s=1s = 1 with t0=0t_0 = 0, the first take is positive: S1=0+p1=1S_1 = 0 + p_1 = 1 β€” the walk lands exactly on the target (gap 0).

  2. Since S1=1≀1S_1 = 1 \le 1, one more positive: S2=1+p2=1+13=43S_2 = 1 + p_2 = 1 + \frac13 = \frac43 β€” the first positive block ends with overshoot 13=p2\frac13 = p_2 (equality holds when crossing upward from exactly ss; see the lemma below). Then S3=43βˆ’q1=43βˆ’12=56S_3 = \frac43 - q_1 = \frac43 - \frac12 = \frac56 β€” the first negative block, undershoot 16<q1=12\frac16 < q_1 = \frac12.

  3. Continuing: S4=56+p3=56+15=3130S_4 = \frac56 + p_3 = \frac56 + \frac15 = \frac{31}{30} (overshoot 130<p3=15\frac1{30} < p_3 = \frac15), S5=3130βˆ’q2=3130βˆ’14=4760S_5 = \frac{31}{30} - q_2 = \frac{31}{30} - \frac14 = \frac{47}{60} (undershoot 1360<q2=14\frac{13}{60} < q_2 = \frac14).

    The next positive block takes two terms: S6=4760+p4=389420<1S_6 = \frac{47}{60} + p_4 = \frac{389}{420} < 1 (not yet across), S7=389420+p5=13071260S_7 = \frac{389}{420} + p_5 = \frac{1307}{1260} (overshoot 471260<p5=19\frac{47}{1260} < p_5 = \frac19).

    And S8=13071260βˆ’q3=10971260S_8 = \frac{1307}{1260} - q_3 = \frac{1097}{1260} (undershoot 1631260<q3=16\frac{163}{1260} < q_3 = \frac16).

  4. Every value matches the model computation (verified as exact rationals in the tests).

    The gap sequence 13,16,130,1360,471260,1631260\frac13, \frac16, \frac1{30}, \frac{13}{60}, \frac{47}{1260}, \frac{163}{1260} is not monotone, but each gap is at most its crossing term, and the crossing terms tend to 0.

    After 120 terms the walk sits at S120β‰ˆ0.9967S_{120} \approx 0.9967 with error β‰ˆ0.00326\approx 0.00326 β€” the bracket is closing in on the target.

Prove the greedy rearrangement converges to the target ss: both blocks recur forever, each crossing overshoot/undershoot is at most the crossing term's magnitude, and since the term magnitudes tend to 0, ∣SNβˆ’sβˆ£β†’0\lvert S_N - s \rvert \to 0.

t≀sβ‡’Sβˆ’s≀p(crossingΒ up),t>sβ‡’sβˆ’S<q(crossingΒ down)t \le s \Rightarrow S - s \le p \quad (\text{crossing up}), \qquad t > s \Rightarrow s - S < q \quad (\text{crossing down})
∣SNβˆ’sβˆ£β‰€max⁑(qlast,pcross)β€…β€Šβ†’Nβ†’βˆžβ€…β€Š0\lvert S_N - s \rvert \le \max(q_{\text{last}}, p_{\text{cross}}) \;\xrightarrow[N\to\infty]{}\; 0
Target s=1s = 1, N=120N = 120: S120β‰ˆ0.9967S_{120} \approx 0.9967, error β‰ˆ0.00326\approx 0.00326 β€” as the term magnitudes die out, so does the walk's amplitude around the target.
Show proof
  1. The alternating harmonic series βˆ‘(βˆ’1)n+1n\sum \frac{(-1)^{n+1}}{n} converges conditionally (Leibniz gives convergence; βˆ‘1n\sum \frac1n diverges, so no absolute convergence). For the positive terms pk=12kβˆ’1p_k = \frac{1}{2k-1} and negative magnitudes qk=12kq_k = \frac{1}{2k}, we have βˆ‘pk=βˆ‘qk=+∞\sum p_k = \sum q_k = +\infty (a corollary of the previous problem, or directly via 12kβˆ’1β‰₯12k\frac{1}{2k-1} \ge \frac{1}{2k} and comparison with the harmonic series).

  2. Both blocks recur forever.

    If positives were taken only finitely often, then eventually only negatives are taken: the running sum drops by qjq_j each step and βˆ‘qj=∞\sum q_j = \infty forces tβ†’βˆ’βˆžt \to -\infty β€” eventually t≀st \le s, and the rule would take a positive: contradiction.

    Symmetrically for negatives (otherwise tβ†’+∞t \to +\infty would force a negative take).

    Since each block consumes the next unused terms in order, every pip_i and every qjq_j is eventually used exactly once β€” a genuine rearrangement.

  3. Crossing lemma: crossing upward from t≀st \le s by adding pp, the new gap is (t+p)βˆ’s=pβˆ’(sβˆ’t)≀p(t + p) - s = p - (s - t) \le p (equality only when t=st = s).

    Crossing downward from t>st > s by subtracting qq, the new gap is sβˆ’(tβˆ’q)=qβˆ’(tβˆ’s)<qs - (t - q) = q - (t - s) < q.

    So every crossing overshoot/undershoot is at most its crossing term's magnitude.

  4. Now look at one positive block (after the first crossing; finitely many initial steps don't affect convergence).

    Starting from an undershoot gap Ξ΅\varepsilon (with Ρ≀qlast\varepsilon \le q_{\text{last}}, the last negative taken, by the lemma), the block increases monotonically, so before the crossing step sβˆ’S≀Ρs - S \le \varepsilon, and at the crossing step Sβˆ’s≀pcrossS - s \le p_{\text{cross}}.

    Hence throughout the block ∣SNβˆ’sβˆ£β‰€max⁑(qlast,pcross)\lvert S_N - s \rvert \le \max(q_{\text{last}}, p_{\text{cross}}).

  5. Since the blocks recur forever (step 2), both term indices tend to ∞\infty: qlastβ†’0q_{\text{last}} \to 0 and pcrossβ†’0p_{\text{cross}} \to 0 (the crossing positives are a subsequence of the pip_i).

    Hence ∣SNβˆ’sβˆ£β†’0\lvert S_N - s \rvert \to 0, i.e. SNβ†’sS_N \to s.

    Negative blocks are symmetric.

    This is the greedy proof of Riemann's rearrangement theorem for finite ss.

    (The ±∞\pm\infty cases lie outside this statement's scope and need a separate construction: for target +∞+\infty, for instance, replace the fixed target ss by stage targets mβ†’+∞m \to +\infty β€” at stage mm, take positive terms until the partial sum exceeds mm.)

8. Power Series