This paper was converted on www.awesomepapers.org from LaTeX by an anonymous user.
Want to know more? Visit the Converter page.

Syracuse Maps as Non-singular Power-Bounded Transformations and Their Inverse Maps

I. Assani, E. Ebbighausen, A. Hande Idris Assani, University of North Carolina at Chapel Hill assani@email.unc.edu Ethan Ebbighausen, University of North Carolina at Chapel Hill ejebbigh@email.unc.edu Anand Hande, University of North Carolina at Chapel Hill a8675309@live.unc.edu
Abstract.

We prove that the dynamical system (,2,T,μ)(\mathbb{N},2^{\mathbb{N}},T,\mu), where μ\mu is a finite measure equivalent to the counting measure, is power-bounded in L1(μ)L^{1}(\mu) if and only if there exists one cycle of the map TT and for any xx\in\mathbb{N}, there exists kk\in\mathbb{N} such that Tk(x)T^{k}(x) is in some cycle of the map TT. This result has immediate implications for the Collatz Conjecture, and we use it to motivate the study of number theoretic properties of the inverse image T1(x)T^{-1}(x) for xx\in\mathbb{N}, where TT denotes the Collatz map here. We study similar properties for the related Syracuse maps, comparing them to the Collatz map. We also analyze some structural properties of the inverse image in relation to asymptotic density of the set {xk:Tk(x)<x}\{x\in\mathbb{N}\mid\exists k\in\mathbb{N}:T^{k}(x)<x\}.

2020 Mathematics Subject Classification:
Primary 11B75, Secondary 37A40, 37A44

1. Introduction

Many mathematicians have investigated the Collatz Map through a variety of means from number theory to dynamical systems. For a compendium of prior work on the subject, see [7]. However, this paper seeks to investigate a new route of analyzing this notorious map by generalizing the method introduced by the first author in [1] to a larger collection of maps, as well as to elucidate the structure of the pre-image trees of some maps with respect to this process. Mathematicians have also studied several density results related to the Collatz Map. Both Terras [12] and Everett [3] were able to show independently that for almost every xx\in\mathbb{N}, there exists kk\in\mathbb{N} such that Tk(x)<xT^{k}(x)<x, with Korec proving a stronger asymptotic density result, with Tn(x)<xcT^{n}(x)<x^{c} for c>log4(3)c>\log_{4}(3) [4]. Tao obtained results pertaining to logarithmic density [10]. Since this topic deals with the positive integers, we shall consider \mathbb{N} not to contain 0. First, define the truncated Collatz map:

Definition 1.

The Collatz map is a function T:T:\mathbb{N}\rightarrow\mathbb{N} so

(1) T(x)={3x+12xoddx2xevenT(x)=\begin{cases}\frac{3x+1}{2}&x\,\,odd\\ \frac{x}{2}&x\,\,even\end{cases}

The Collatz Conjecture states that {1,2}\{1,2\} is the only cycle of this map, and that the trajectories of the map are bounded. We focus on the latter of these two requirements, which, put rigorously, says that for any nn\in\mathbb{N} there exists mm\in\mathbb{N} such that Tm(n)T^{m}(n) is part of a cycle. Next, consider a broader class of maps as presented in [9]:

Definition 2.

Let a Syracuse-type map V:V:\mathbb{N}\rightarrow\mathbb{N} be a function of the form

(2) V(x)=mix+ridifximod  dV(x)=\frac{m_{i}x+r_{i}}{d}\,\,\,\,\,\,if\,\,\,\,x\equiv i\,\textbf{mod \,d}

where riimimoddr_{i}\equiv-im_{i}\,\,\textbf{mod}\,\,d and i=0,1,2d1i=0,1,2...d-1, and gcd(m0m1md1,d)=1gcd(m_{0}m_{1}...m_{d-1},d)=1

While the Collatz Map is often referred to as the Syracuse map, we distinguish these general cases as Syracuse-type maps or simply Syracuse maps for brevity. A simple sub-class of these maps often called px+1px+1 maps, which are the same as the Collatz map except with some value pp replacing the 33, has often been the focus of study. It is known that several of these maps have cycles, and that some have multiple cycles ([7]).

2. Ergodic Characterization of the General Syracuse Map

2.1. Extending the Hopf Decomposition:


While it is not completely necessary to use the Hopf Decomposition for the primary result of the next section, it is helpful in providing motivation. Colloquially, the Hopf decomposition provides that, in a non-singular system, the ambient space may be separated into sets conservative and dissipative with respect to the map (for a full proof and rigorous statement, see [5]). In the case of the Collatz system, any unbounded trajectory must be in the dissipative part, while any cycle must be in the conservative part. In fact, this may be sharpened further, and an extended version of this decomposition is given in [1] for the Collatz map. It may also be expanded to more general maps on the natural numbers.

Theorem 1.

Consider a non-singular dynamical system (,2,V,ν)(\mathbb{N},2^{\mathbb{N}},V,\nu). There exists a partition of \mathbb{N} into sets C,D1,D2C,D_{1},D_{2} such that:

  1. (1)

    The restriction of VV to CC is conservative. The set CC is VV-absorbing, and is the at-most-countable union of cycles CiC_{i}.

  2. (2)

    The set D1D_{1} is equal to k=1Vk(C)\C\bigcup_{k=1}^{\infty}V^{-k}(C)\backslash C.

  3. (3)

    The set D2D_{2} is the complement of CD1C\cup D_{1} in \mathbb{N}, V1(D2)=D2V^{-1}(D_{2})=D_{2}.

  4. (4)

    Any and all unbounded trajectories of VV lie in D2D_{2}


Proof.

: The Hopf Decomposition provides a partition of \mathbb{N} into sets CC and DD, so that the restriction of VV to CC is conservative and CC is VV-absorbing. Since \mathbb{N} is countable, CC also is, so it can contain at most countable many cycles. This proves (1). Let D1=k=1Vk(C)\CD_{1}=\bigcup_{k=1}^{\infty}V^{-k}(C)\backslash C. Let D2=\[CD1]D_{2}=\mathbb{N}\backslash[C\cup D_{1}]. Then, V1(CD1)=V1(C)k=1Vk1(C)\C=CD1V^{-1}(C\cup D_{1})=V^{-1}(C)\cup\bigcup_{k=1}^{\infty}V^{-k-1}(C)\backslash C=C\cup D_{1}. Therefore, V1(D2)=D2V^{-1}(D_{2})=D_{2}, and (2) and (3) are proven. Finally, given any xx\in\mathbb{N}, a bounded trajectory means that there exists nn so Vn(x)CV^{n}(x)\in C, so either xCx\in C or xD1x\in D_{1}. Given an unbounded trajectory of the point yy, the set {y}\{y\} is wandering, and Vm(y)CV^{m}(y)\notin C for any natural number mm. Thus, yD2y\in D_{2} and (4) is proven. ∎


This characterization sets aside the unbounded trajectories of a map in a set D2D_{2}. Showing that this set is empty would prove half of the Collatz conjecture. This leads to the following characterization that is equivalent to such a case. Furthermore, showing that CC is exactly {1,2}\{1,2\} would prove the entire conjecture.

2.2. Dynamical System Characterization:

Proving that D2D_{2} is empty is essentially the same problem as proving the trajectories bounded, thus it is important to introduce some equivalent criteria for such a case. To do so, consider the following definition.

Definition 3.

: Let (X,𝒜,T,μ)(X,\mathscr{A},T,\mu) be a non-singular dynamical system. It is said to be power bounded in L1(μ)L^{1}(\mu) (or often simply just power bounded) if there exists some M+M\in\mathbb{R}_{+} such that for all sets A𝒜A\in\mathscr{A} and natural numbers nn, μ(Tn(A))Mμ(A)\mu(T^{-n}(A))\leq M\mu(A).


Using this structure from the study of dynamics, the first author presented the following characterization of the Collatz map:

Theorem 2.

Let (,2,T,μ)(\mathbb{N},2^{\mathbb{N}},T,\mu) be the Collatz dynamical system with the counting measure μ\mu. The following are equivalent:

  1. (1)

    There exists a finite measure α\alpha equivalent to μ\mu for which the dynamical system (,2,T,α)(\mathbb{N},2^{\mathbb{N}},T,\alpha) is power bounded in L1(α)L^{1}(\alpha).

  2. (2)

    The set D2D_{2} is empty.

  3. (3)

    The trajectories of each point nn\in\mathbb{N} are bounded.


This characterization may be extended to nearly all maps. Of course, the Collatz map and other similar maps have cycles that define the conjecture surrounding them. It is necessary for these maps to have at least one cycle to use such a characterization, because otherwise the set D2D_{2} cannot be empty in any case. Given this caveat, the next theorem follows.

Theorem 3.

Let (,2,V,μ)(\mathbb{N},2^{\mathbb{N}},V,\mu) be a dynamical system where μ\mu is a finite measure equivalent to the counting measure. Then, it is power-bounded in L1(μ)L^{1}(\mu) if and only if there exists at least one cycle of the map VV and for any xx\in\mathbb{N}, there exists a non-negative integer kk such that Vk(x)V^{k}(x) is in some cycle of the map VV.

Proof.

\Rightarrow) Consider the power-bounded system (,2,V,μ)(\mathbb{N},2^{\mathbb{N}},V,\mu). By contradiction, let there exist some xx so Vk(x)V^{k}(x) is never in a cycle. Then, for all klk\neq l, k,l>0k,l>0, Vk(x)Vl(x)V^{k}(x)\neq V^{l}(x). Let μ(x)=δ>0\mu(x)=\delta>0. Further, since the measure is finite, μ(i=0Vi(x))=i=1μ(Vi(x))=ϵ>0\mu(\bigcup_{i=0}^{\infty}V^{i}(x))=\sum_{i=1}^{\infty}\mu(V^{i}(x))=\epsilon>0, implying that μ(Vk(x))0\mu(V^{k}(x))\rightarrow 0 as kk\rightarrow\infty. Given any M+M\in\mathbb{R}_{+}, we may take some large NN such that μ(VN(x))<δ/M\mu(V^{N}(x))<\delta/M. Hence, VN(VN(x))>δ>Mμ(VN(x))V^{-N}(V^{N}(x))>\delta>M\mu(V^{N}(x)), contradicting that the system is power bounded since MM was arbitrary.

\Leftarrow) Let this property hold. Since the space is countable, there are at most countable many cycles, and every point is in a pre-image of a cycle. Let the cycles be the sets C1,C2,C_{1},C_{2},.... First consider C1C_{1}. The cycle must be of finite length, N, so we construct a measure on \mathbb{N}. Define a function μ1\mu_{1}, which will have a measure value at each specified point, and set the measure of any A2A\in 2^{\mathbb{N}} to be the sum of the measures of the points in AA. Let μ1(c)=12N\mu_{1}(c)=\frac{1}{2N} for any cC1c\in C_{1}, and let so that μ1(C1)=12\mu_{1}(C_{1})=\frac{1}{2}. Next, there are at most countably many points in V1(C1)\C1V^{-1}(C_{1})\backslash C_{1}. The infinite case implies the finite case, so we consider this case. Enumerate these points as {c1,c2,c3}\{c_{1},c_{2},c_{3}...\}. Set μ1(ci)=2i3\mu_{1}(c_{i})=2^{-i-3}, so that μ1(V1(C1)\C1)22\mu_{1}(V^{-1}(C_{1})\backslash C_{1})\leq 2^{-2}. Next, consider V1(cj)V^{-1}(c_{j}), which does not contain cjc_{j} nor any other point with a defined measure since such a case would generate a cycle. It again may be enumerated as {cj,1,cj,2,cj,3}\{c_{j,1},c_{j,2},c_{j,3}...\}, since the set is at most countably infinite. Set μ1(cj,i)=2(j2)+(i2)\mu_{1}(c_{j,i})=2^{(-j-2)+(-i-2)} so that μi(V1(cj))2j3\mu_{i}(V^{-1}(c_{j}))\leq 2^{-j-3} and so μ1(V1(V1(C1)\C1))=23\mu_{1}(V^{-1}(V^{-1}(C_{1})\backslash C_{1}))=2^{-3}. Repeat inductively over the pre-images generated by this set, and set all points not in i=0Vi(C1)\bigcup_{i=0}^{\infty}V^{-i}(C_{1}) to have measure zero.

The process above gives that μ1(i=0Vi(C1))=μ1(C1)+μ1(V1(C1)\C1)+μ1(V1(V1(C1)\C1))+12+14+18+=1\mu_{1}(\bigcup_{i=0}^{\infty}V^{-i}(C_{1}))=\mu_{1}(C_{1})+\mu_{1}(V^{-1}(C_{1})\backslash C_{1})+\mu_{1}(V^{-1}(V^{-1}(C_{1})\backslash C_{1}))+...\leq\frac{1}{2}+\frac{1}{4}+\frac{1}{8}+...=1. Repeat this inductively again over the CiC_{i}, and we may generate a new measure by the countable sum of measures. Let μ=i=12i1μi\mu=\sum_{i=1}^{\infty}2^{-i-1}\mu_{i}, so μ()1\mu(\mathbb{N})\leq 1 and μ\mu is a finite measure. Further, every point in \mathbb{N} is in one of the i=0Vi(Cj)\bigcup_{i=0}^{\infty}V^{-i}(C_{j}), so every point has nonzero measure under μ\mu. Next, we need to show that the measure allows our point transformation to be power bounded. By construction, μi(V1(A))μi(A)\mu_{i}(V^{-1}(A))\leq\mu_{i}(A) for any set AA such that ACi=A\cap C_{i}=\emptyset. For a set BB intersecting the cycle CiC_{i}, separating B=(BCiC)(BCi)B=(B\cap C_{i}^{C})\cup(B\cap C_{i}) gives μi(Vn(B))2nμi(BCiC)+2μi(BCi)2μi(B)\mu_{i}(V^{-n}(B))\leq 2^{-n}\mu_{i}(B\cap C_{i}^{C})+2\mu_{i}(B\cap C_{i})\leq 2\mu_{i}(B) for all nn\in\mathbb{N}. Therefore, μ(Vn(A))=i=12i1μi(Vn(A))i=12iμi(A)=2μ(A)\mu(V^{-n}(A))=\sum_{i=1}^{\infty}2^{-i-1}\mu_{i}(V^{-n}(A))\leq\sum_{i=1}^{\infty}2^{-i}\mu_{i}(A)=2\mu(A). The measure μ\mu is finite and power-bounded in 1\mathscr{L}^{1}, as well as equivalent to the counting measure, as desired.


Remark 1: Given the previous result, it is helpful to note which Syracuse maps have cycles and thus can use this characterization. For example, consider a map

(3) T(x)={px+12xoddx2xevenT(x)=\begin{cases}\frac{px+1}{2}&x\,\,odd\\ \frac{x}{2}&x\,\,even\end{cases}

where pp is an odd number. It is easy to see that T(1)=p+12T(1)=\frac{p+1}{2}, and if p=2k1p=2^{k}-1, there exists a cycle {1,2k1,2k2,2}\{1,2^{k-1},2^{k-2},...2\}. The Collatz map falls into this category. It is also known that cycles exist in the case of 5n+15n+1, and 181n+1181n+1 [2].

3. Chain Structure in The Collatz Map

Looking at the Collatz and generalized Collatz maps in terms of a power-bounded system as above leads naturally to examining the inverse map. The action of the Collatz map is rather erratic when looking in the forward direction, but in the backward direction, some patterns begin to arise, specifically when classifying values as nodes modulo 3. However, these patterns don’t necessarily extend to general maps, allowing for classification of these maps.

3.1. The Collatz Map:

While the map may seem unpredictable at first, looking at values modulo 3 presents some regularity. A simple check shows that T1(3p)={6p}T^{-1}(3p)=\{6p\}, T1(3p+1)={6p+2}T^{-1}(3p+1)=\{6p+2\}, and T1(3p+2)={2p+1,6p+4}T^{-1}(3p+2)=\{2p+1,6p+4\}. For shorthand, we designate three sets.

Definition 4.

A value nin\in\mathbb{N}_{i} is said to be in i\mathbb{N}_{i} if ni(mod  3)n\equiv i\,\,(mod\,\,3).

The pre-image of a node in 0{\mathbb{N}}_{0} is again in 0{\mathbb{N}}_{0}, the pre-image of a node in 1{\mathbb{N}}_{1} is in 2{\mathbb{N}}_{2}, and the pre-image of a node in 2{\mathbb{N}}_{2} includes one node either in 1{\mathbb{N}}_{1} or 0{\mathbb{N}}_{0} and another in 1{\mathbb{N}}_{1}. The latter behavior is the most interesting, and may be characterized in a more lucid way, using a simple lemma.

Lemma 4.

Any node in 2{\mathbb{N}}_{2} may be written as 3a2bh13^{a}2^{b}h-1 where a>0a>0, b0b\geq 0, and h1h\geq 1 is not a multiple of 22 or 33.

Proof.

For arbitrary n2n\in{\mathbb{N}}_{2}, it may be written as 3p+23p+2 be definition, and thus as 3(p+1)13(p+1)-1. Using the Fundamental Theorem of Arithmetic, p+1p+1 has the desired representation 3a12bh3^{a-1}2^{b}h. ∎


Under this representation, T1(3a2bh1)={3a12b+1h1,2(3a2bh1)}T^{-1}(3^{a}2^{b}h-1)=\{3^{a-1}2^{b+1}h-1,2(3^{a}2^{b}h-1)\}. In particular, we may characterize a family of nodes by either the image or part of the pre-image, where T(2ah1)=3(2a1)h1T(2^{a}h-1)=3(2^{a-1})h-1, T(3(2a1)h1)=322a2h1T(3(2^{a-1})h-1)=3^{2}2^{a-2}h-1, and this repeats until Ta(2ah1)=3ah1T^{a}(2^{a}h-1)=3^{a}h-1. Define this as a family structure, when such a repeating form occurs.

Refer to caption
Figure 1. Inverse Image Tree generated by 3ah13^{a}h-1 up to the atha^{th} level

In this final case, the node is even and its image is a 1{\mathbb{N}}_{1} node. The pre-image of the 1{\mathbb{N}}_{1} node is again a 2{\mathbb{N}}_{2} node, starting the process again. Connecting these families forms a chain. This chain may extend in the other direction as well. In the case that 2ah12^{a}h-1 is an 1\mathbb{N}_{1} node, its pre-image 2a+1h22^{a+1}h-2 is in 2\mathbb{N}_{2}, and so it takes the form 3rk13^{r}k-1 following the conditions of the lemma as well. It has been useful to distinguish the nodes of the form 3rk13^{r}k-1 as “chain head nodes,” although they are on equal footing of importance with the 1\mathbb{N}_{1} nodes of the form 2ah12^{a}h-1.

Refer to caption
Figure 2. A Family and a Chain Illustrated with Nodes

3.2. More General Maps:

Many of the more general maps with similar forms to the Collatz Map have extremely different structures and properties, for example, that many currently are not known to have a cycle. Consider a sub-class of these maps, where

(4) V(x)={px+r2xoddx2xevenV(x)=\begin{cases}\frac{px+r}{2}&x\,\,odd\\ \frac{x}{2}&x\,\,even\end{cases}

where pp is an odd integer and rr is taken to be an integer so |r|<p|r|<p and rr is relatively prime to pp. In this case, the idea of a family structure includes some repeating form over a number of nodes. For example, if T(pα2βk1)=pα+12β1k1T(p^{\alpha}2^{\beta}k-1)=p^{\alpha+1}2^{\beta-1}k-1. We then have the following.

Theorem 5.

A map V:V:\mathbb{N}\rightarrow\mathbb{N} has a chain structure precisely when r=p2r=p-2 or r=2pr=2-p.

Proof.

This relies on a couple building blocks of concepts that led to the chain structure in the Collatz map. First, the family structure gives a node of the form pα2βklp^{\alpha}2^{\beta}k-l (where kk is neither a multiple of 22 or pp) so that V(pα2βkl)=pα+12β1klV(p^{\alpha}2^{\beta}k-l)=p^{\alpha+1}2^{\beta-1}k-l. This requires that ll be odd, and so

(5) p(pα2βkl)+r2=pα+12β1kl\frac{p(p^{\alpha}2^{\beta}k-l)+r}{2}=p^{\alpha+1}2^{\beta-1}k-l

Hence, rpl2=l\frac{r-pl}{2}=-l and r=(p2)lr=(p-2)l. Since ll is an integer, rp2\frac{r}{p-2} is, and since |r|<p|r|<p, r=±(p2)r=\pm(p-2). Note that these calculations also show that elements in the class of r2p(modp)\frac{r}{2-p}\,\,(mod\,\,p) have two pre-images. These are, in fact, the only nodes with 22 pre-images.

This case where r=±(p2)r=\pm(p-2) gives a further sub-class of maps with the family structure. Next, we most consider the connection of families to create chains. Using the previous notation, let mm be the largest positive integer so 2m2^{m} divides pα+βk+r2pp^{\alpha+\beta}k+\frac{r}{2-p}. We would require that

(6) Vm+1(pα+βk+r2p)r2p(modp)V^{m+1}(p^{\alpha+\beta}k+\frac{r}{2-p})\equiv\frac{r}{2-p}\,(mod\,\,p)

Such a condition would connect two families, since Vm+1(pα+βk+r2p)V^{m+1}(p^{\alpha+\beta}k+\frac{r}{2-p}) would then have a representation pα22β2k2rp2p^{\alpha_{2}}2^{\beta_{2}}k_{2}-\frac{r}{p-2}.

The above equation says that

(7) p(pα+βkrp22m)+r2r2p(modp)\frac{p(\frac{p^{\alpha+\beta}k-\frac{r}{p-2}}{2^{m}})+r}{2}\equiv\frac{r}{2-p}\,(mod\,\,p)

Since pp is odd, this occurs exactly when

(8) (2p)(pα+β+1kprp2+2mr)2m+1r(modp)(2-p)(p^{\alpha+\beta+1}k-\frac{pr}{p-2}+2^{m}r)\equiv 2^{m+1}r\,(mod\,\,p)

which reduces to the trivial

(9) 2m+1r2m+1r(modp)2^{m+1}r\equiv 2^{m+1}r\,(mod\,\,p)

We thus have that the families connect for this form of map whenever they occur. ∎


Remark 2: We may extend further to the case of

(10) V(x)=mix+ridifximod  dV(x)=\frac{m_{i}x+r_{i}}{d}\,\,\,\,\,\,if\,\,\,\,x\equiv i\,\textbf{mod \,d}

where riimimoddr_{i}\equiv im_{i}\,\,\textbf{mod}\,\,d and i=0,1,2d1i=0,1,2...d-1, and gcd(m0m1md1,d)=1gcd(m_{0}m_{1}...m_{d-1},d)=1, as proposed in [9] and [8]. In this case, since ri=imir_{i}=im_{i}, when rimid\frac{r_{i}}{m_{i}-d} is an integer, T(miαdβkrimid)=miα+1dβ1krimidT(m_{i}^{\alpha}d^{\beta}k-\frac{r_{i}}{m_{i}-d})=m_{i}^{\alpha+1}d^{\beta-1}k-\frac{r_{i}}{m_{i}-d}. These families may occur in one such class modulo dd, or in several depending on this condition. Future results using the chain structure on the Collatz map may be generalized to such a class of maps.

4. The Set D2D_{2} and Previous Density Results

The properties of the set L={xk:Tk(x)<x}L=\{x\in\mathbb{N}\mid\exists k\in\mathbb{N}:T^{k}(x)<x\}\subset\mathbb{N} are of interest, as the Collatz Conjecture is equivalent to the statement L={1}L=\mathbb{N}-\{1\}. The set LL was introduced by R. Terras in [11], where it was also shown to have density 1. The works of Korec generalized this set to Mc={xk:Tk(x)<xc}M_{c}=\{x\in\mathbb{N}\mid\exists k\in\mathbb{N}:T^{k}(x)<x^{c}\} for c>log3(4)c>log_{3}(4) in [4], which also has density 1, and further generalizations follow in Lagarias [6], Everett [3], and Tao [10].

However, these density results do not have immediate implications to the structure of D2D_{2}, the elements of \mathbb{N} with unbounded trajectories, or D2\mathbb{N}-D_{2}, since McM_{c} and LL could contain elements which have unbounded trajectories as long as they have some lesser iterated image.

Proposition 6.

Assume that D2D_{2}\neq\emptyset. Then,

  • 1) There exists yD2Mcy\in D_{2}\cap M_{c} for any 0<c10<c\leq 1.

  • 2) Assume D2D_{2} has density 0, and that there are finitely many cycles C1,,CnC_{1},...,C_{n}. Let f:f:\mathbb{N}\to\mathbb{R} satisfy f(x)>max1in(min(Ci))f(x)>\max_{1\leq i\leq n}(\min(C_{i})) for all xx\in\mathbb{N}. Then Lf:={xk:Tk(x)<f(x)}L_{f}:=\{x\in\mathbb{N}\mid\exists k\in\mathbb{N}:T^{k}(x)<f(x)\} has density 11.

  • 2) D2D_{2} can be written as a countable disjoint union of sets of the form Bx:=nk0Tn(Tk(x))B_{x}:=\bigcup_{n\in\mathbb{N}}\bigcup_{k\geq 0}T^{-n}(T^{k}(x)) where each xx is in D2D_{2}. Furthermore, each BxB_{x} is invariant under TT.

Proof.

(1) Let D2D_{2} be nonempty. By the well-ordering of the natural numbers, there exists some minimum element dd. Consider that 2adD22^{a}d\in D_{2} for any aa\in\mathbb{N} and in particular Ta(2ad)=dT^{a}(2^{a}d)=d. Fix some c(0,1]c\in(0,1]. Then, Take aa large enough such that (2ad)c>d(2^{a}d)^{c}>d, and then Ta(2ad)<(2ad)cT^{a}(2^{a}d)<(2^{a}d)^{c}.

(2) Since D2D_{2} has density 0, D2\mathbb{N}-D_{2} has density 11. So it suffices to show that D2Lf\mathbb{N}-D_{2}\subset L_{f}. If xx has bounded trajectories, then by definition for some jj\in\mathbb{N} we have Tj(x)CiT^{j}(x)\in C_{i} for some i=1,,ni=1,...,n. By taking more iterates, we have Tk(x)=min(Ci)<f(x)T^{k}(x)=\min(C_{i})<f(x).


(3) We clearly have D2=xD2BxD_{2}=\bigcup_{x\in D_{2}}B_{x} which is a countable union since D2D_{2}\subset\mathbb{N} is countable. Since D2D_{2} is invariant under TT and yBxy\in B_{x} implies ByBxB_{y}\subset B_{x}, we can discard some elements in the above union to yield a disjoint union. By definition, BxB_{x} is the smallest set that contains all preimages of all images of xx under TT so BxB_{x} is invariant.

Remark 3: With these assumptions, we would obtain stronger density results for the likes of functions ff previously analyzed. In addition, while this set D2D_{2} is not as well-understood as LL, Theorem 1 shows that it has invariance under the forward and reverse Collatz map, T(D2)D2T(D_{2})\subset D_{2} and T1(D2)D2T^{-1}(D_{2})\subset D_{2}. Similarly, if any number xx has bounded trajectories, then so does T(x)T(x) and so do the values in the set T1(x)T^{-1}(x). Such invariance properties are not immediate for LL: If Tk(x)<xT^{k}(x)<x for some x,kx,k\in\mathbb{N} and xx is even, then it is not immediate that Tk(x)<x2T^{k}(x)<\frac{x}{2}, or if any iterate is less than x2\frac{x}{2}. If x=3p+22x=3p+2\in\mathbb{N}_{2}, then Tk(x)<xT^{k}(x)<x does not imply Tk(x)<2p+1T1(x)T^{k}(x)<2p+1\in T^{-1}(x). This motivates the study of the structure of D2D_{2}, assuming it is non-empty.


Part 3 of the above proposition gives a way to organize D2D_{2} in relatively nice, invariant chunks BxiB_{x_{i}} which generate D2D_{2}. Following the family and chain structures introduced previously, the values in 2\mathbb{N}_{2} and how they behave under T1T^{-1} determine the structure of the sets BxiB_{x_{i}} and hence D2D_{2}. Given that the preimages of 3kh13^{k}h-1 expand under the action 2x13\frac{2x-1}{3} exactly kk times before reaching a non-2\mathbb{N}_{2} node, we then focus on this preimage level.

Definition 5.

We refer to the nodes in i=0kTi(3kh1)\bigcup_{i=0}^{k}T^{-i}(3^{k}h-1) as the triangle generated by the node 3kh13^{k}h-1 and the nodes nTa(3kh1)n\in T^{-a}(3^{k}h-1) as the aa-th level preimages.

For example, the triangle generated by 9h19h-1 is just the first 33 levels of the tree generated by 9h19h-1, i.e {9h1},{6h1,18h2},{4h1,12h2,36h4}\{9h-1\},\{6h-1,18h-2\},\{4h-1,12h-2,36h-4\}. By a quick computation, one can see that the only possible elements of the triangle which are not in LL are iterates of the 2x13\frac{2x-1}{3} action (specifically, the nodes 9h1,6h1,4h19h-1,6h-1,4h-1). For brevity, let us consider the two possible actions of the inverse map T1T^{-1}. Call the 2x13\frac{2x-1}{3} the left action, and the action 2x2x the right action. Thus, the iterates of left actions are on the leftmost branch of the triangle (see Figure 1). This leads to the following claim:

Claim: For all xx in a triangle generated by 3kh13^{k}h-1, we can find mm\in\mathbb{N} such that Tm(x)<xT^{m}(x)<x, if xx does not lie on the left branch of the triangle.

This statement has been verified for several different values of a,ha,h\in\mathbb{N} and h0h\notin\mathbb{N}_{0}. In particular, we have verified this statement for all possible combinations of a{1,2,,20}a\in\{1,2,...,20\} and h{1,2,,300}0h\in\{1,2,...,300\}-\mathbb{N}_{0}, through the code in the appendix.

The numerical data suggests that this claim is true, which would give a more precise characterization of the elements of D2D_{2}.

5. Code Appendix

def t(n):
if n % 2 == 0:
return n/2
else:
return (3*n + 1)/2
def t_iter(n,k):
ls = [n]
next = n
for i in range(1,k+1):
var1 = t(next)
next = var1
ls.append(var1)
return ls
def inv_t(n):
if n % 3 != 2:
return [2*n]
else:
p = (n-2)/3
return [2*p + 1, 2*n]
def flatten(l):
return [item for sublist in l for item in sublist]
def generateTree(a,h):
num = 4*(3**a * h - 1)
flag = True
tree = [[num]]
current_lvl = [num]
for i in range(1, a+1):
next = flatten([inv_t(x) for x in current_lvl])
tree.append(next)
current_lvl = next
return tree
def leftBranch(a,h):
num = 3**a * h - 1
return [3**(a-k) * (2**k * h) - 1 for k in range(0,a+1)]
###Run this method to check statement for different a,h values.
def testTree(a,h):
tree = generateTree(a,h)
branch = leftBranch(a,h)
for k in range(2, len(tree)):
lvl = tree[k]
for node in lvl:
t_image = t_iter(node, k)[1:]
if (min(t_image) > node and node != lvl[0]):
#Checks if the node doesnt shrink and if its not on the left hand branch
return (False, node, k)
return (True, None, None)

References

  • [1] Idris Assani. Collatz map as a power bounded nonsingular transformation. https://arxiv.org/abs/2208.11675, 2022.
  • [2] R. E. Crandall. On the “3x+13x+1” problem. Mathematics of Computation, 32(144):1281–1292, 1978.
  • [3] CJ Everett. Iteration of the number-theoretic function. The Ultimate Challenge: The 3x+ 1 Problem, page 225, 2010.
  • [4] Ivan Korec. A density estimate for the 3x+13x+1 problem. Mathematica Slovaca, 44(1):85–89, 1994.
  • [5] Ulrich Krengel. Ergodic Theorems. De Gruyter, 2011.
  • [6] Jeffrey C. Lagarias. The 3x + 1 problem and its generalizations. The American Mathematical Monthly, 92(1):3–23, 1985.
  • [7] Jeffrey C. Lagarias. The 3x+1 Problem: An Overview. https://arxiv.org/abs/2111.02635, 2021.
  • [8] Keith R. Matthews and Robert N. Buttsworth. On some Markov matrices arising from the generalized Collatz mapping. Acta. Arithmetica, 55:43–57, 1990.
  • [9] Keith R. Matthews and Anthony M Watts. A generalization of Hasse’s generalization of the Syracuse algorithm. Acta. Arithmetica, 43:167–175, 1984.
  • [10] Terence Tao. Almost all orbits of the collatz map attain almost bounded values. https://arxiv.org/abs/1909.03562, 2019.
  • [11] Riho Terras. A stopping time problem on the positive integers. Acta Arithmetica, 30(3):241–252, 1976.
  • [12] Riho Terras. On the existence of a density. Acta Arithmetica, XXXV:85–89, 1979.