Wiki
Wiki

Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.

Updated

Two odd-greedy termination conventions


Problem 282 uses the least unused allowed denominator whose reciprocal does not exceed the remainder. The usual odd-greedy algorithm allows a denominator to be used again. Pihko's paper defines the latter rule in equations (2.1)--(2.2), and Remark 2.4 identifies the possible repetition. The following elementary reduction compares the universal questions. It does not identify the two trajectories for every input.

Write GG for ordinary odd greedy and UU for the rule requiring unused odd denominators. Both start from a rational xx with 0<x<10<x<1, in lowest terms with odd denominator. Both stop at remainder zero. Subtracting an odd-denominator unit fraction preserves the property that the reduced denominator is odd.

For any remainder 0<r<2/30<r<2/3, let dd be the smallest positive odd integer with 1/d≤r1/d\le r. Then d≥3d\ge3. If d=3d=3, the new remainder is r−1/3<1/3r-1/3<1/3. If d≥5d\ge5, minimality gives r<1/(d−2)r<1/(d-2), and therefore

0≤r−1d<2d(d−2)<1d.0\le r-\frac1d<\frac{2}{d(d-2)}<\frac1d.

Consequently every subsequent denominator in GG strictly increases until termination. In this starting range GG and UU coincide.

If 2/3≤x<12/3\le x<1, GG chooses 33 twice and reaches x−2/3∈[0,1/3)x-2/3\in[0,1/3). Thus universal termination of GG on 0<x<10<x<1 is equivalent to its universal termination on 0<x<2/30<x<2/3.

For UU, consider instead the initial consecutive prefix 3,5,7,…,d3,5,7,\ldots,d. If the remainder after choosing dd is at least 1/d1/d, the next choice is d+2d+2: all smaller odd denominators have been used, and 1/(d+2)<1/d1/(d+2)<1/d is affordable. The prefix must reach remainder less than 1/d1/d, or zero, by d=13d=13 at latest. Indeed, if it has reached 1313, then

∑j∈{3,5,7,9,11,13}1j+113=4648945045>1>x.\sum_{j\in\{3,5,7,9,11,13\}}\frac1j+\frac1{13} =\frac{46489}{45045}>1>x.

At that point every used reciprocal exceeds the remainder. Ordinary greedy from the remainder can use none of those denominators; the strict-increase argument applies thereafter. Hence the rest of UU is an ordinary odd-greedy trajectory from a rational in (0,2/3)(0,2/3) with odd reduced denominator, unless it has already terminated.

It follows that these three assertions are equivalent:

  1. GG terminates for every rational 0<x<2/30<x<2/3 with odd reduced denominator.
  2. GG terminates for every rational 0<x<10<x<1 with odd reduced denominator.
  3. UU terminates for every rational 0<x<10<x<1 with odd reduced denominator.

The reduction is author-recorded. Its only source use is to identify the ordinary algorithm; the inequalities above give the complete argument. No computational experiment or unproved termination premise is used. The precise finite assertion is a prefix of at most six choices before UU reduces to GG, not a bound on total running time. Neither universal assertion has been proved here. For example, at x=2/3x=2/3 ordinary greedy immediately repeats 33, whereas the unused rule cannot do so.

The remaining hard step is termination of ordinary odd greedy in the smaller range. Checking bounded numerators, denominators or trajectory lengths cannot replace it without an additional descent invariant. A useful test of a proposed invariant is whether it decreases across every greedy subtraction in this range; stop the route if it only bounds finitely many initial inputs. The reduction says nothing about other residue classes, square denominators, or an arbitrary allowed set in Problem 282.