Rotation swaps, numerator gaps, and central structure in Collatz parity vectors

I read MathKook’s “Collatz High Cycles Do Not Exist” and wondered under what conditions a similar argument could produce pure power-of-3 numerator gaps, rather than the power-of-2 gap used there.

The underlying mechanism turns out to be a local 01↔10 swap between rotations of the same parity word. Such swaps occur precisely when the word contains a central (palindromic bispecial) factor in the Sturmian sense.

Primitive upper Christoffel words form a particularly rigid subclass of this situation, but the swap identity itself holds at the more general level of central words.


1. Parity vectors and affine form

Let s(v)=\#\{\text{ones in }v\}.

The composition associated with v is an affine map

T_v(x)=\dfrac{3^{s(v)}x + B(v)}{2^{l(v)}},

where B(v)\in\mathbb Z is the numerator constant.

The corresponding fixed point is

x_v=\dfrac{B(v)}{2^{l(v)}-3^{s(v)}}.

If an integer cycle existed with parity vector v, then every rotation w of v would correspond to another starting point of the same cycle. In particular,

(2^{l(v)}-3^{s(v)}) \mid B(w)\qquad\text{for every rotation }w\text{ of }v.


2. The local swap identity

Let A be any binary word, and define

r=s(A)

Consider the two parity vectors

w=01A,\qquad u=10A

They have the same length and the same weight.

Lemma (prefix swap)

B(w)-B(u)=3^{r}.

Proof (direct composition)

T_{01}(x)=U(D(x))=\dfrac{3x+2}{4},\; T_{10}(x)=D(U(x))=\dfrac{3x+1}{4}.

T_A(x)=\dfrac{3^{r}x+B(A)}{2^{l(A)}}.

T_{01A}(x)=T_A\!\left(\dfrac{3x+2}{4}\right) =\dfrac{3^{r}(3x+2)+4B(A)}{2^{l(A)+2}},

\Rightarrow\ B(01A)=2\cdot3^{r}+4B(A).

T_{10A}(x)=T_A\!\left(\dfrac{3x+1}{4}\right) =\dfrac{3^{r}(3x+1)+4B(A)}{2^{l(A)+2}},

\Rightarrow\ B(10A)=1\cdot3^{r}+4B(A)

\Rightarrow\ B(01A)-B(10A)=3^{r}


3. Terminal swaps and powers of two

If two parity vectors differ only by a terminal swap

w=A01,\qquad u=A10,

then the same local mechanism yields a pure power of two gap:

B(w)-B(u)=2^{t},

for an exponent t determined by the position of the swap.

Thus:

  • power-of-two and power-of-three gaps arise from the same local swap 01\leftrightarrow10,
  • the position of the swap determines whether a factor 2^t, 3^r, or a mixed 2^a3^b appears.

In particular, MathKook’s argument corresponds to the terminal-swap case of this general identity.


4. Central words and when prefix swaps occur for rotations

To exploit the prefix-swap lemma inside a cycle, the two vectors
01A\quad\text{and}\quad10A
must occur as rotations of the same parity vector v.
Equivalently, v must admit two rotations
A01\quad\text{and}\quad A10.
In combinatorics on words, this means that A is bispecial.
When A is also a palindrome, it is called a central word.

For an aperiodic upper Christoffel word, one has the classical decomposition
v=1A0,\qquad v^{R}=0A1,
with A central.

Left-rotating once gives
\rho(v)=A01,\qquad \rho(v^{R})=A10,
a terminal-swap pair.

Right-rotating once gives
\rho^{-1}(v)=01A,\qquad \rho^{-1}(v^{R})=10A,
a prefix-swap pair, hence
B(01A)-B(10A)=3^{s(A)}.
Thus, Christoffel structure does not create the swap identity, but it guarantees the existence of the required conjugate rotations within the same parity vector.

More generally, the relevant condition is central (palindromic bispecial) structure; Christoffel words form a particularly rigid, primitive subclass.


5. Consequence for integer cycles

Let v be a parity vector and w,u two of its rotations. If an integer cycle existed, then

(2^{l(v)}-3^{s(v)}) \mid (B(w)-B(u)).

  • Terminal-swap rotations force a power-of-two divisor, contradicting the oddness of 2^{l(v)}-3^{s(v)} except in trivial cases.
  • Prefix-swap rotations force a power-of-three divisor, but occur only when v contains a central factor.

Outside the central/Christoffel setting, swap-induced gaps are generally mixed
2^𝑎3^𝑏 , and do not force an immediate contradiction with the denominator (2^{l(v)}-3^{s(v)}) .


6. Example: central Sturmian word not yielding a Christoffel vector

Consider the palindromic Sturmian word A=0100010,
with s(A)=2.
Form the vectors
w=01A=010100010,\qquad u=10A=100100010.
Then
B(w)-B(u)=9=3^2.

This illustrates the local power-of-three identity at the level of central words.
However, the vector 1A0=101000100
is not a (primitive or non-primitive) Christoffel word, showing that central structure is strictly more general than Christoffel structure.


Closing remark

The local swap identity explains all known strong numerator gaps in Collatz parity vectors.
Christoffel words provide a canonical setting where the required rotations are guaranteed, but the underlying mechanism operates at the more general level of central (palindromic bispecial) Sturmian words.

1 Like

Right! In the figure below, 485-421=2^6 but also 638-557=3^4. And you’re right that the difference between the nth cycle member and the (n+5)th cycle member is always of the form 2^i 3^j. Nice!

For the more general case of vector v with some LRS of length A, the “difference method” cancels 2’s with members w=AC and u=AB, while it cancels 3’s with members w=CA and u=BA. You’d think canceling 3’s would be better (for the goal of defeating \delta(v) = 2^{l(v)}-3^{s(v)}); however, you cancel l(A) 2’s, but only s(A) 3’s.

For example,

\beta(101.011) - \beta(101.110) = 2^3 \cdot 5
\beta(110.101) - \beta(011.101) = 3^2 \cdot 5

where the uncancellable part (5) is the same either way.

1 Like

Maybe this would be a good way to analyze the known integer cycles, particularly in 5n + 1 and 3n - 1. Since I already wrote them out with their rotations in my notebook I though I would put them here:

5n+1

17 cycle - minimum rotation 1100100

B(w) = 1001 100 (space for LRS emphasis)

B(v) = 1001 001

13 cycle - minimum rotation 1110000

B(w) = 11 10000 or 000 0111

B(v) = 11 00001 or 000 1110

1 cycle - minimum rotation 11000

B(w) = 1 1000 or 00 011

B(v) = 1 0001 or 00 110

3n-1

17 cycle - minimum rotation 11110111000

B(w) = 1110 1110001

B(v) = 1110 0011110

5 cycle - minimum rotation 110

B(w) = 1 10

B(v) = 1 01

1 cycle - too small to analyze

My observations were firstly that for each integer cycle the non-prefix portions of each rotation pair are themselves rotations of the same vector. I believe this occurs when the repeated substrings are adjacent or overlapping. In fact, in these cases you can keep recursively trimming these substrings until you are left with ‘10’. It could be that these are just common properties of small vectors as I couldn’t think of a reason that this would help with divisibility.

Another observation (or possible red herring) is that the 3x-1 cycles follow the following rule where B(v_{min}) corresponds to the minimum rotation, d_1 corresponds to the denominator of the full cycle, and d_2 corresponds to the denominator of the remaining trimmed vector:

d_2 - d_1 = B(v) + B(v_{min})

or since B(w) - B(v) just happens to be exactly equal to d_1 in both cases, maybe it’s

-d_1 \pmod{B(v)} = B(v_{min}) \pmod{d_2}

I almost didn’t include this since I wasn’t able to replicate it with the 5x+1 cycles, but I’ll leave it up to you whether to discard it. I don’t fully understand the combinatorics terminology yet so I’ll come back to this later after I read up sometime. I just thought it would be a good excuse to show off the integer cycle vectors since it seems relevant.

I’m really fascinated by the use of rotations to look at divisibility. I’ve assumed for a long time, and I think maybe others have too, that since it only takes one rotation to determine divisibility, it’s best just to keep it as simple as possible. Personally I’m happy right now to leave it to others to investigate what this means for eliminating cycles, but it’s a tool that I will be trying to apply to cycle hunting. My first thought was that B(v) \equiv B(v) * 2^n \pmod{d_1} or just 2^n \equiv 1 \pmod{d_1} for when the remaining trimmed vectors are rotated by only a series of n even steps creates an easy opportunity for the 5n+1 cycles to have freedom to “choose” an n to satisfy the congruence (if 2 is a primitive root of d_1), but this is obviously only possible for small cycles and comes with a few caveats. I would be interested if the rotational symmetry and recursive “trimmability” of these vectors have anything to do with their divisibility.

Keep it up. I just joined the forum but it’s nice to see new threads and new ideas pushing in directions where we have more questions than answers. I don’t have a huge level of familiarity with the literature but if an idea doesn’t seem to be tracing the footsteps of a PhD or professor that already took it to its conclusion, I’m interested. Of course it’s important to be aware of these trodden paths but at least speaking for myself there’s no use in trying to push at the same spot in the same direction with less expertise. That being said if this is one of those paths, it would be nice to know. It’s hard to have a bird’s eye view of these things. I don’t know why I said all this, it’s only tangential to your post but I needed an excuse to offload some thoughts.

1 Like