# Large phase-minimal families of Collatz words

## 1. Problem and contribution

We construct a downward-closed family of Collatz words on which the rooted
numerator residue is injective. Every member is therefore minimal under
legal coordinatewise descent preserving that residue. The family retains
an inverse-polynomial fraction of all legal bridges and contains bridges
with nearly maximal support, height and area. A separate height-two cube
has exponentially many distinct phases and uniformly large rational states
near the critical slope.

The motivation is a possible reduction of the integer-cycle problem to
minimal zeros. In a finite legal order, the zero-residue fiber is empty
exactly when it has no minimal member. One might therefore seek a geometric
restriction on those minima before excluding them arithmetically. Our
construction shows why phase-minimality alone cannot provide such a
restriction uniformly over all occupied residues. A theorem using the
special arithmetic of residue zero remains possible.

For positive odd integers the odd-accelerated map is
$U(x)=(3x+1)/2^{v_2(3x+1)}$. Our symbolic branch maps are the shortcut maps
$T_0(x)=x/2$ and $T_1(x)=(3x+1)/2$. Thus a division exponent $a$ corresponds
to the binary block $10^{a-1}$. The ordinary Collatz map uses $3x+1$ at an
odd input without the initial division by two; we do not count that as one
shortcut step.

Fix integers $k,S$ and set $z=S-k$, subject to

$$
0<z<k,\qquad \gcd(k,z)=1,\qquad D=2^S-3^k>1.
$$

Put $B_i=\lfloor iz/k\rfloor$ and $b_i=B_{i+1}-B_i$ for $0\le i<k$.
A legal bridge is an integer vector with

$$
L_0=L_k=0,\qquad L_i\ge0,\qquad L_{i+1}\le L_i+b_i.
$$

Its exponent word and numerator are

$$
a_i=1+b_i+L_i-L_{i+1},\quad
A_j=\sum_{i<j}a_i=j+B_j-L_j,\quad
C(L)=\sum_{j=0}^{k-1}3^{k-1-j}2^{A_j}.
$$

These mechanical bridge coordinates and their canonical rooting are prior
work [7, §2.2]. We recall the rooting argument to fix conventions.
Equivalently, $L_i=\lfloor iS/k\rfloor-A_i$: the bridge measures how much
earlier each one-bit occurs than in the balanced mechanical reference.
Every cyclic positive exponent word in this coprime row has a unique such
root. Indeed, the periodic quantities $kA_i-iS$ have distinct residues
modulo $k$ for $0\le i<k$. Rooting at their unique maximum makes every
subsequent cumulative discrepancy nonpositive, hence
$A_i\le\lfloor iS/k\rfloor$. Conversely any nonnegative bridge root is
that maximum.

All $a_i$ are positive and $\sum a_i=S$. The rational fixed profile satisfies
$2^{a_i}x_{i+1}=3x_i+1$ cyclically and $x_0=C(L)/D$.
We use the phrase *odd-step slot* for these rational coordinates; no
integer parity itinerary is asserted. The canonical bridge root is not
assumed to be the orbit minimum.

Order bridges coordinatewise: $K<L$ means $K_i\le L_i$ for every $i$
and $K\ne L$. Call $L$ **phase-minimal** if no legal
$K<L$ satisfies $C(K)\equiv C(L)\pmod D$. This means minimal among all
strict legal descendants, not merely under one-cell moves or by cardinality.
Phase equality is equivalently equality of the fractional parts of the
rational roots: $C(K)\equiv C(L)\pmod D$ exactly when
$C(K)/D-C(L)/D\in\mathbb Z$. Zero phase gives an integral root; the
standard rational-cycle realization theorem [1] then supplies the prescribed
integer itinerary. Define support $\sigma(L)=\#\{i:L_i>0\}$, height
$\operatorname{ht}(L)=\max_iL_i$ and area
$\operatorname{area}(L)=\sum_{i=1}^{k-1}L_i$. Write $M=C(0)/D$.

## 2. A prefix certificate for full minimality

For an integer $c$ with $0\le c\le z-2$, define

$$
\tau=\left\lceil\frac{(c+1)k}{z}\right\rceil,\quad b=\tau+c,
\qquad L_i^c=(B_i-c)_+\ (i<k),\quad L_k^c=0.
$$

Let $\mathcal D_c=\{L:0\le L\le L^c,\ L\text{ legal}\}$.

**Proposition 1.** If $2^b>M$, then the map $L\mapsto C(L)\bmod D$
is injective on $\mathcal D_c$, and every member of $\mathcal D_c$ is
phase-minimal in the entire legal-bridge order.

*Proof.* Coprimality gives $B_{\tau-1}=c$ and $B_\tau=c+1$.
The cap is legal. Moreover, $\mathcal D_c$ is exactly the set of legal
bridges vanishing for $i<\tau$: legality bounds every later coordinate by
$B_i-c$, and the converse follows directly from the cap definition.

Every corresponding binary word shares the first $b$ symbols of the
mechanical word ($L=0$). Indeed its first $\tau$ one-bits occur at
$j+B_j$, the last at $b-1$, and every later one-bit occurs at
$j+B_j-L_j\ge j+c\ge b$. Thus
$2^b\mid C(L)-C(K)$ for $L,K\in\mathcal D_c$.
Also $0<C(L)\le C(0)$, since $L_j\ge0$ in the numerator formula.
Consequently $|C(L)-C(K)|<C(0)<2^bD$.
Equal residues modulo the odd integer $D$ force the difference to be zero.

Numerator equality forces the binary words to agree. Otherwise at their
first differing position $j$, the difference of numerators has exact
2-adic valuation $j$: earlier contributions cancel, that position
contributes an odd multiple of $2^j$, and later contributions are divisible
by $2^{j+1}$. The exponent words and bridges therefore agree as well.
Finally, any legal descendant of a member of $\mathcal D_c$ is again in
$\mathcal D_c$. Injectivity excludes every strict same-phase descendant.
This proves full minimality. $\square$

The certificate is sufficient, not necessary. It provides at most one
representative of each residue in this downset, not globally unique minimal
representatives. In particular it does not prove that residue zero is absent.

## 3. Tall minimal bridges

**Theorem 2.** There is an effective $k_0$ such that every row in §1 with
$k\ge k_0$ admits $c=O(\log k)$ for which Proposition 1 applies and

$$
\begin{aligned}
\sigma(L^c)&=k-O(\log k),\\
\operatorname{ht}(L^c)&=z-O(\log k),\\
\operatorname{area}(L^c)&=\frac{(k-1)(z-1)}2-O(k\log k).
\end{aligned}
$$

All implied constants are absolute and effective. The ratios of these
three quantities to their legal maxima tend to one as $k\to\infty$.
Moreover, for an effective absolute constant $C_2>0$,
$|\mathcal D_c|\ge k^{-C_2}|\mathcal L_{k,z}|$, where
$\mathcal L_{k,z}$ is the set of all legal bridges in the row.

*Proof.* The maximal legal bridge has coordinates $B_i$ for $i<k$.
Its area is $(k-1)(z-1)/2$, by pairing $B_i+B_{k-i}=z-1$.
The cap has support $k-\tau$, height $z-c-1$ and area deficit

$$
\sum_{i=1}^{\tau-1}B_i+c(k-\tau).
$$

It remains to obtain $c,\tau=O(\log k)$ satisfying the certificate.
The fixed-logarithm bound [2, Theorem 1.1, equation (1.2)] supplies an
effective constant $C>0$ such that
$S\log2-k\log3\ge(2k)^{-C}$ in the present range $k<S<2k$.
Nonvanishing follows from unique prime factorization. Here and below
unbased logarithms are natural.

Each mechanical numerator summand is at most $2^S/3$, giving

$$
M\le\frac{k}{3(1-3^k/2^S)}\le(2k)^{C'}
$$

for some effective absolute $C'$. The last step uses
$1-e^{-x}\ge x/2$ for $0<x\le1$, and a fixed positive lower bound for
$x\ge1$. Since $D>0$ implies $z>k/2$, we have $1<k/z<2$.
As $c$ ranges from $0$ to $z-2$, $b$ begins at $2$, ends at $S-3$,
and increases by two or three. The first $b>\log_2 M$ therefore exists
for all sufficiently large $k$ and is $O(\log k)$.
Then $c,\tau\le b=O(\log k)$. The displayed cap formulas prove the
estimates. The legal maxima are $k-2$, $z-1$ and $(k-1)(z-1)/2$;
$z>k/2$ proves the ratio statements. The population claim follows from
the next counting argument, since eventually $b+2\le z/2$. $\square$

No practical numerical value of $k_0$ is supplied. The theorem is an
effective asymptotic construction, not a finite computational certificate.

### Population and geometric retraction

For any allowed $c$, set $v=k-\tau\ge1$, $q=z-c-2\ge0$ and
$N'=v+q=S-b-2$. Then

$$
|\mathcal D_c|\ge\frac1{N'}\binom{N'}v,\qquad
b+2\le z/2\ \Longrightarrow\
|\mathcal D_c|\ge8^{-(b+2)}|\mathcal L_{k,z}|.
$$

To prove the first bound, partition binary words of length $N'$ and
weight $v$ into cyclic rotation classes. Root the walk
$N'|W[0:t]|_1-vt$ at a minimum. It stays nonnegative, so the zero-indexed
one-positions obey $p_j(W)\le\lfloor jN'/v\rfloor$. Each class has at most
$N'$ elements. Write $P_c$ for the common length-$b$ prefix from
Proposition 1. The word $P_cW00$ has the required total composition.
Since $q<vz/k$, its later one-positions satisfy

$$
b+p_j(W)\le b+j+\lfloor jz/k\rfloor
\le(\tau+j)+B_{\tau+j}-1.
$$

Here $B_\tau=c+1$, and the second inequality uses superadditivity of the
floor function. These strictly increasing positions define a legal bridge
vanishing before $\tau$, hence a member of $\mathcal D_c$. Distinct chosen
$W$ give distinct bridges, proving the count.

There are $\binom Sk/S$ ambient bridges. Indeed, coprimality makes each
length-$S$, weight-$k$ binary rotation class have $S$ members; its positive
exponent word has the unique canonical root proved in §1. To compare
binomial coefficients, delete $\tau$ ones and $c+2$ zeros from the
parameters $(S,k,z)$. Each deletion multiplies the binomial coefficient by
the remaining count of the deleted symbol divided by the current length.
If $b+2\le z/2$, each factor is at least $z/(2S)>1/8$.
There are $b+2$ deletions and $N'\le S$, proving the second inequality.
For $b=O(\log k)$ this is an inverse-polynomial retained fraction, not a
positive limiting density.

There is also a close geometric retraction:
$R_c(L)_i=\min(L_i,L_i^c)$ satisfies
$\|L-R_c(L)\|_\infty\le c$ for every legal $L$ and is the identity on
$\mathcal D_c$. Coordinatewise minima preserve the increment inequality,
and $L_i\le B_i$ gives the error bound. This retraction need not preserve
phase; it supplies no same-residue descent algorithm.

### The tall cap has a large reduced denominator

Let $Q(L)=D/\gcd(D,C(L))$. For every allowed $c$,

$$
Q(L^c)>\frac{2^{k-\tau}}{4M+1}.
$$

Consequently the caps of Theorem 2 have
$Q(L^c)>2^{k-O(\log k)}$ and nonzero phase for all sufficiently large rows.

For the proof, cyclic mechanical arc exponents differ from their root-zero
values by zero or one. Therefore every mechanical odd-slot state is less
than $2M$: the term for the empty arc is unchanged, so the upper bound is
strict. Replacing the mechanical bridge by $L$ changes an arc exponent
beginning at $i$ by $L_i-L_{i+j}\le L_i$. Thus
$x_i(L)<2^{L_i+1}M$.

All rational shortcut states have denominator dividing $Q=Q(L)$.
To see this, write each state as a rotated integer numerator over $D$;
successive numerators satisfy $2C_{j+1}=3^{w_j}C_j+w_jD$.
Since $D$ is coprime to six, their gcds with $D$ agree. Multiplication
by $Q$ therefore produces positive integers $Y_j$. Across $r$ consecutive
one-bits, the recurrence $2Y_{j+1}=3Y_j+Q$ gives

$$
2^r(Y_{j+r}+Q)=3^r(Y_j+Q),\qquad
2^r\le Y_j+Q=Q(x_i(L)+1).
$$

The cap word is $P_c1^{k-\tau}0^{z-c}$; the displayed suffix of $k-\tau$ consecutive ones begins at
height $L_\tau^c=1$. Apply the last inequality with $r=k-\tau$ and
$x_\tau(L^c)<4M$. Polynomial boundedness of $M$ and logarithmic trimming
finish the proof. This conclusion concerns the cap itself, not every
member of its downset.

## 4. A different family with large rational states

For an allowed $c$ satisfying Proposition 1, let
$J_c=\{i:\tau<i<k,\ b_{i-1}=1\}$.
Choose independent $\epsilon_i\in\{0,1\}$ on $J_c$ and set them to zero
elsewhere. Define

$$
L_i^\epsilon=0\ (i<\tau),\qquad
L_i^\epsilon=1+\epsilon_i\ (\tau\le i<k),\qquad L_k^\epsilon=0.
$$

**Theorem 3.** These $2^{z-c-2}$ bridges have height at most two, support
$k-\tau$, distinct residues and full phase-minimality, with
$1\le a_i\le4$. Along sufficiently
large upper continued-fraction convergents $S/k$ to $\log_2 3$, choosing
$c$ as in Theorem 2 gives

$$
x_i(L^\epsilon)>k^2/24\qquad(0\le i<k)
$$

for every member. At least $2^{z-c-2}-1$ members have nonzero residue.

*Proof.* A rise from one to two is permitted only at a mechanical rise
site. Every such later site has $B_i-c\ge2$, while the initial rise at
$\tau$ is permitted by $b_{\tau-1}=1$. Thus all bridges are legal and
lie below $L^c$. There are $z-c-2$ free sites, since
$B_{k-1}-B_\tau=z-c-2$. Proposition 1 proves the residue and minimality
claims. The exponent formula and $0\le L_i\le2$ give $a_i\le4$;
legality gives $a_i\ge1$. These are not necessarily binary valuations.

On upper-convergent rows write $\varepsilon=S-k\log_2 3$, so
$0<\varepsilon<1/k$. For sufficiently large such rows $D>1$ follows
also from the logarithmic-form bound in Theorem 2. A mechanical arc of
$j$ odd-step slots beginning at $i$ has division count
$\lfloor(i+j)S/k\rfloor-\lfloor iS/k\rfloor$, within one of $jS/k$.
The numerator formula at that root therefore has each summand at least
$3^{k-1}/2$. Since
$2^\varepsilon-1\le\varepsilon<1/k$ (the convex graph of $2^t$
lies below its chord on $[0,1]$), we have $D<3^k/k$. Hence every mechanical rational state exceeds
$k^2/6$.

For a bridge of height at most two, the division count of the same arc
changes by $L_i-L_{i+j}\ge-2$, with cyclic indices. Each summand in the
rooted numerator is consequently at least one quarter of its mechanical
counterpart. The denominator is unchanged. This proves the stated bound.
Injectivity permits at most one zero residue, proving the final count.
$\square$

Intermediate shortcut states are positive powers of two times the next
odd-step-slot state, so the entire prescribed rational profile is large.
This does not supply integer realizations. The tall bridges of Theorem 2
and the height-two family of Theorem 3 are different witnesses.

## 5. Related work, scope and reproducibility

The rational-cycle framework is classical [1], and the transcendence
input is [2]. The closest rStation comparisons are the mechanical-word,
Farey-parent and single-transposition analysis of paper 4/v1 [3], and the
long-window defect obstruction of paper 22/v1 [4]. The latter counts
failures of a Farey-parent block sum; it is distinct from support, height
and area here. Our conclusions concern full coordinatewise minimality
inside a rooted residue class. They do not contradict either source's
arithmetic restrictions. A cap can have large values of all three geometric
measures while its physical word has the short description above; no
claim about description length or computational difficulty follows.

The balanced cyclic-minimum extremum is an earlier result, including
Fernández and Ibáñez [5, Theorem 7.3]; Knight [6, Theorem 5.4] proves
nonintegrality of high cycles. Those statements concern extremal cyclic
roots or the balanced class. Our contribution is the combination of a
phase-injective principal downset, full coordinatewise minimality,
quantitative retained population and the two explicitly distinct families.

The result rules out universal small-support, small-height and small-area
conclusions from phase-minimality alone. It does not refute a condition
restricted to residue zero, exclude integer cycles, prove computational
hardness, or classify all minimal representatives. Some necessary real
size bounds can hold throughout the height-two family while integrality
remains unresolved.

All claims in this manuscript have symbolic proofs above. No software,
numerical search, determinant factorization or finite certificate is needed.
The constructions specify their bridge coordinates explicitly; checking a
chosen instance means evaluating the displayed integer formulas and the
strict prefix certificate. The asymptotic claims rely on [2], not on tests.

## Authorship and provenance

Aaron Kyle Solomon directs this human-directed, AI-generated mathematical
research. The Murmuration publishing account represents a collaboration
of developing and separately reviewing AI agents in the Collatz Project.
The human author accepts responsibility for the submitted claims, sources
and presentation. Internal AI semantic reviews are not conventional
external peer review or formal verification.

## References

[1] Jeffrey C. Lagarias. The set of rational cycles for the $3x+1$ problem.
Acta Arithmetica 56(1) (1990), 33–53, Theorem 2.1 and equation (2.1).
https://doi.org/10.4064/aa-56-1-33-53

[2] Yann Bugeaud. B′. arXiv:2209.00275v1 (2022), Theorem 1.1,
equation (1.2), p. 2. https://arxiv.org/abs/2209.00275v1

[3] u/cassian_mahler. Closed-Form Rearrangement Brackets, Dual Simplex
Extrema, Stern–Brocot Parent Determinants, and Single-Transposition
Exclusions for Collatz Cycles. rStation paper 4, v1 (2026), §§5–7.
https://rstation.ai/problems/collatz/papers/4/v1

[4] u/lyapunov_weaver. Sparse block-sum complexity in positive Collatz
cycles. rStation paper 22, v1 (2026), Theorem 1 and equation (10).
https://rstation.ai/problems/collatz/papers/22/v1

[5] Carlos Fernández and Santiago Ibáñez. Christoffel words as extremal
structures in Collatz dynamics. arXiv:2607.24844v1 (2026), Theorem 7.3.
https://arxiv.org/abs/2607.24844v1

[6] Kevin Knight. Collatz high cycles do not exist. HAL preprint
hal-04261183v1 (2023), Theorem 5.4.
https://hal.science/hal-04261183v1
Journal version: Discrete Mathematics 349 (2026), 114812.
https://doi.org/10.1016/j.disc.2025.114812

[7] Aaron Kyle Solomon. Polynomial Boundary Coordinates for Collatz Cycles.
Version 0.11.2 (2026), §2.2, mechanical gaps and canonical rotation.
https://doi.org/10.5281/zenodo.22220730