Taku Toguchi
tnotes.com / notes
Research  /  Collatz

A number called 4/3
Just looking at the characteristic function
of the Syracuse variable

A random variable sits at the heart of the Collatz conjecture. Where does the peak of its characteristic function stand? A record of pushing n up to 21 by computer and simply watching. There is no proof here. Only an observation.

This is a note, not a paper.

Between work, I play with number theory as a hobby. I like problems that wear a simple face and yet no one can crack — primes, Collatz — and on my days off I let the computer run, a little at a time. This was that kind of thing: weekend carpentry, you could say. I just wanted to see, with my own eyes, the characteristic function of the random variable at the heart of Tao's proof, so I started raising n from 1 — and before I knew it I was at n=21, a computation over seven billion points.

What I saw was cleaner than I expected. The position of the peak walks straight ahead, at slope 4/3. Since I'd come this far, I tidied up the observation and put it, code and numerical data and all, on Zenodo with a DOI (link below) — so that anyone who wants to carry the computation further can reproduce it as is.

I have proven nothing. This contributes nothing to the Collatz conjecture. I just found what I glimpsed inside interesting enough to write down.

§ 01

First, the 3N+1 story

Pick a positive integer. If it's even, halve it. If it's odd, triple it and add one. Repeat — and no matter where you start, you eventually reach 1. Or do you?

This is the Collatz conjecture. Since 1937, no one has proven it. The question could hardly be simpler, and yet it has defeated everyone.

In 2019, Terence Tao blew a large hole in the problem. He showed that "almost all" orbits reach almost bounded values. At the technical heart of that proof sits an object called the Syracuse random variable.

What the Syracuse random variable isThe object you get by extracting only the "odd steps" of the Collatz map and smoothing them probabilistically. At each step, how many times 2 divides out is given by a geometric distribution Geom(2), and their composition is viewed inside the finite world ℤ/3n. Tao's proof relied on this variable being "well spread out" (equidistribution).

What does "well spread out" mean? In Fourier's language: the peak of the characteristic function (the Fourier transform) is low. If the peak is high, probability mass is concentrated at that frequency. If it's low, the mass is spread flat. Tao proved that this peak sinks with n "faster than any polynomial."

Tao's upper bound says how fast the peak sinks.
But it doesn't say where the peak stands.
This "where" is what I went looking for, by computer. That's all this note is.
§ 02

2 is the vowel of the world of 3

To talk about the "position" of the peak of the characteristic function, you need a coordinate. Here a classical fact of elementary number theory comes into play.

For every n, 2 is a primitive root modulo 3n. That is, every frequency ξ not divisible by 3 can be written in exactly one way as a power of 2, ξ = 2k mod 3n. Keep cycling the powers of 2 and you step through every frequency coprime to 3, each exactly once, leaving none out.

So I stop viewing frequencies by ξ and instead view them by the exponent k. This is the whole viewpoint of the note. I reindex the characteristic function as a function of k, and call it ψn(k). Then I write the k at which its absolute value is largest as k*(n).

The powers of 2 make one loop through the world of 3n (n=2, mod 9)
124875k : 0 1 2 3 4 52ⁿ : 1 2 4 8 7 5Six ×2 steps cycle oncethrough the 6 residues coprime to 3.→ Each frequency ξ can be  numbered by an exponent k.
Fig.1 2 is a primitive root modulo 3n. So every frequency coprime to 3 can be given a one-dimensional coordinate, the exponent k. That hands us a language for talking about the "position" of the peak.
§ 03

The peak walks at slope 4/3

Raising n from 1 to 21, I computed ψn over its whole domain. At n=21 the domain reaches 2·320 ≈ 7 billion points. Record the peak position k*(n) at each n and line them up, and — neatly, almost in a straight line — they climb. The slope is right around 4/3.

The drift of k*(n) — all 21 points computed
0510152025k*(n)15101520nslope 4/31820,21k*(21)/21 = 4/3
Fig.2 The black dots are the observed k*(n). The blue dashed line is the reference at slope 4/3. The vermilion points are the "outliers" that come up later (n=2, 3, 18, 20, 21). At the far right, n=21 lands exactly on k*(21)/21 = 28/21 = 4/3.
From n = 4 to 21, over 17 steps, k* grew by exactly 24.
24 ÷ 18 = 4/3. With zero error.
This is the kind of agreement chance doesn't produce.
§ 04

A formula right 14 times in a row breaks at 18

Drop the slope 4/3 into an integer prediction and you get a simple floor function, ⌊4(n−1)/3⌋. Line it up and at first it's almost eerily accurate. From n=4 to 17, 14 exact matches in a row. One look and you want to believe it's a structural identity.

But at n=18, for the first time, it missed. Against a prediction of 22, the measured value was 23. A gap of just 1. Then +1 at n=20, and +2 at n=21. The set of n where it misses is {2, 3, 18, 20, 21} — clustered at the ends of the range.

The gap between prediction ⌊4(n−1)/3⌋ and measured k*(n)
n=121+1+214 exact matches in a row (n=4…17)outliers cluster at the ends
Fig.3 The value of k*(n) − ⌊4(n−1)/3⌋ at each n. The thin blue line is zero (exact match). The vermilion bars are the gaps. After 14 matches in a row over n=4–17, it breaks at the top of the range — 18, 20, 21 — by +1, +1, +2. Meanwhile n=19 returns to a match.

What's interesting is that even when the formula misses, the slope 4/3 doesn't break. The gap at n=21 is +2, yet k*(21)/21 = 4/3 still holds exactly. A +1 "wobble" shows up only at the boundary of the floor function; the underlying linearity is preserved. The formula is only an approximation, but the slope looks real.

On honestyIn the first draft I started to write that this floor function "always holds for n ≥ 4." Until I computed n=18, that's how it looked. But the computer returned 23. The prediction was 22. Keeping the fact that it missed, rather than erasing it, is what makes this a note of observation rather than a paper. Not hiding the one loss after fourteen wins seemed worth more than boasting about the fourteen wins.
§ 05

How the height of the peak sinks

Not just the position — I looked at the height too. Plot the peak value Mn = max|ψn| against n and it decays smoothly. Fit it to a line on a log-log plot, and over the finite range (n=9–21) a power law Mn ∼ C·n−2.1 fits well.

But this is a finite-range observation, not an asymptotic claim. Tao's upper bound guarantees superpolynomial decay — faster than any fixed power n−A. The "effective exponent 2.1" seen through a finite window is consistent with that exponent slowly growing as n increases. At the present range, you can't see that far.

The decay of the peak value Mn (log-log)
log Mₙlog n∝ n⁻²·¹
Fig.4 A log-log plot of the peak value Mn. From n=9 on, it fits a power law of slope about −2.1 well (vermilion dashed line). This is an effective, finite-range behavior, consistent with the superpolynomial decay Tao guarantees. The asymptotics lie beyond this range.
§ 06

The numbers themselves

I turned it into figures, but here's the raw table too. The whole of the observation is in it. gn/Mn2 is a measure of how far the peak pulls ahead of the runner-up.

Table 1 | Characteristic function of the Syracuse variable, n = 1…21
nMnk*(n)⌊4(n−1)/3⌋gapgn/Mn2
15.7735e−100·
23.7792e−121+10.654
32.5224e−132+10.406
41.7700e−144·0.421
51.2927e−155·0.394
69.6106e−266·0.106
77.5870e−288·0.105
86.0891e−299·0.220
94.8026e−21010·0.119
103.8278e−21212·0.027
113.1944e−21313·0.133
122.6458e−21414·0.111
132.2052e−21616·0.025
141.9128e−21717·0.139
151.6284e−21818·0.014
161.4409e−22020·0.109
171.2511e−22121·0.023
181.1187e−22322+10.103
199.8157e−32424·0.009
208.8846e−32625+10.110
217.8721e−32826+20.012

All in IEEE 754 double-precision complex arithmetic. Cross-checked against an independent implementation (recursion in the original space → discrete Fourier transform) for n≤8, agreeing to a relative error below 10−10. n=20, 21 were computed on a cloud machine with 961 GB of memory.

§ 07

What I don't know

This note has no answers. It sets down an observation and leaves the questions open. That, exactly, is a numerical open question.

Does k*(n)/n converge as n→∞? Is the limit really 4/3? At n=21 it landed exactly on 28/21 = 4/3, but 18 points of observation can't tell 4/3 apart from a nearby value such as 24/17 ≈ 1.412.
Is there a structural reason behind the outliers {2, 3, 18, 20, 21}? (18, 20) are even with +1; 21 is the only odd one, with +2. Coincidence, transient, or governed by a common arithmetic mechanism? I have no idea yet.
For how many n is the floor function ⌊4(n−1)/3⌋ correct — density 1, density 0, or somewhere between? After 14 matches in a row, it missed 3 times in the last 4 of the range. With this material, I can't draw the picture.
What is the exact asymptotic exponent of the decay? Does the transition to the exp(−cn) that Tao suggests happen within reach of direct computation?
Extend the computation to n ≥ 22. I tried n=22 once and it ran out of memory while building the phase-factor array and crashed. It needs a rewritten algorithm, or a bigger machine.

I leave this observation as a record, not a conclusion. To the Collatz conjecture itself, it contributes nothing. I just wanted, once, to look properly at how the object at the heart of Tao's proof behaves inside a computer. Push n up to 21 and the peak walks at slope 4/3; the floor function staggers by 1 somewhere along the way, then comes back to 4/3. That's all of it.

Do this kind of thing all the time and people occasionally ask why. The answer is simple: I want to see, with my own eyes, the regularity hidden behind a simple question. It isn't useful for anything. That seven billion points of computation turned up the plain fraction "4/3" — that single point — is, I think, worth a few weekends. Maybe something will come of it someday; maybe combining things that seem unrelated will lead to a big discovery.

If anyone can run the computation past n=22, I'd love to see how it continues. The code and the data are there for exactly that.

Taku Toguchi
Paper: Numerical observations on extremal Fourier coefficients of the Syracuse random variable, 2026.
All data and verification code released under MIT; the text under CC BY 4.0. Reproducible.
All figures drawn directly from the observed data.