
An Approach to Collatz Conjecture Using Nonlinear Feedback Model
Abstract
Approaching the Collatz conjecture from a hybrid control transfer function model, I realized that it satisfies the conditions for the output trajectory to converge to 0 [dB] on the logarithmic axis. If the probability of even numbers occurring in nonlinear feedback is assumed 62.5% based on the linear congruential method using a modulo function, this is larger than the probability of even numbers occurring under the convergence condition of 61.3% when the initial value is sufficiently large. Simulations confirmed that with three-digit initial values, the probability of even numbers occurring at convergence was approximately 64 to 71%.
1. Nonlinear Feedback Systems
I attempted to prove the Collatz Conjecture, an unsolved problem in number theory proposed by mathematician Lota Collatz, as a convergence problem for hybrid control models in the measurement and control fields. The Collatz Conjecture states that "for any positive integer n, if n/2 is operated when n is even, and 3n+1 is operated when n is odd, it will always reach 1 regardless of the initial value." Numerous attempts have been made to date, but no proof has been achieved. While system modeling techniques such as system identification, which developed from regression models, and Doyle's notation, which combines multistage linear time-invariant system transfer functions, are well established in the measurement and control fields, few techniques address nonstationary and nonlinear system responses. For nonstationary systems, there are Markov process Kalman filters, and for nonlinear systems, there are hybrid control systems with transition responses between multiple states. [1] We approach the Collatz Conjecture from the perspective of hybrid control, where the response changes depending on whether the system input is odd or even.
1.1 Nonlinear Feedback Transfer Function Model
The computational content and conditions of the Collatz Conjecture can be summarized as follows: (1) a range of positive integers, (2) an arbitrary positive integer initial value, (3) nonlinear feedback based on even/odd numbers, and (4) integer addition/multiplication only. Figure 1 shows the Collatz Conjecture expressed as a nonlinear feedback transfer function model with an initial value n. While similar to a simple linear first-order IIR digital filter, this is a nonlinear hybrid control with different feedback processing depending on whether the output is even or odd.

1.2 Trajectory of the Collatz Conjecture
on the Addition/Multiplication Map
Fig. 2(a) shows an example of output n, expressed on a linear axis, converging to "1" after seven operations when the initial input is "3." Fig. 2(b) shows output n expressed on a logarithmic axis, starting from 9.5 dB and reaching 0 dB after seven operations. Since the Collatz Conjecture converges to the lower end of the logarithmic axis, 0 dB, we can simply prove that output n converges toward the lower end of the logarithmic axis by replacing the problem with "a problem of +9.5 dB + addition + 1 when n is odd, and -6 dB when n is even."

Fig. 3 is an addition/multiplication map that shows the trajectory of a number n due to nonlinear feedback in two dimensions. The horizontal direction is linear axis addition, and the vertical direction is logarithmic axis multiplication, showing how the kth value n advances to the next k+1th time on the addition/multiplication map. If n is an even number, it moves approximately -6 [dB] downward on the logarithmic axis, and if n is odd, it moves approximately +9.5 [dB] on the logarithmic axis and diagonally upwards on the addition axis to the right +1.

For the operation "3n+1" when n is an odd number, consider converting the linear axis movement amount "+1" to a movement amount on the logarithmic axis. As shown in equation (1.1), the movement amount on the logarithmic axis changes depending on the value of n. Separating the multiplication and addition components of equation (1.1) gives equation (1.2).

The y(n) in equation (1.3) is the same as equation (1.2) except for the dB conversion unit. If the value n is large enough to be considered a continuous domain operation, differentiating y(n) with respect to n gives equation (1.4).

From equation (1.4), if we convert the linear axis movement amount of the calculation "3n+1" into the logarithmic axis movement amount, we get equation (1.5). However, we assume that the value n is sufficiently large.

When n is a large value, Δy(n) approaches 0 dB, and the increase on the logarithmic axis due to addition can be largely ignored. Specifically, to what extent Δy(n) should be estimated is shown in two cases in equations (1.6) and (1.7).
Like other Collatz conjecture challengers, I have confirmed through simulations that initial values of n converge to 0 dB when they are in the range of 3 to 1000. When Δy(1000) is set to its maximum value, the range of Δy(n) is given by equation (1.6).

Furthermore, as shown in Fig. 2, the value of n oscillates as it converges to 0 [dB]. If we simply consider the maximum value of Δy(n), it becomes n=1, but n=1 is a convergence value, and after convergence, the cycle of "1 → 4 → 2 → 1" repeats. Since n=2 or n=4 are even numbers and correspond to the oscillation cycle after the convergence described above, if we provisionally set n=3 (although in Fi.2 it converges in 7 steps), the range of Δy(n) becomes equation (1.7).

1.3 Odd/Even Probability
in the Collatz Conjecture
Rather than proving the Collatz Conjecture from the regularity of the output trajectory, we explain the convergence to 0 dB on the logarithmic axis by linking the odd/even probability of an odd/even hybrid control model with the logarithmic axis shift.
The hybrid control output is expanded in Fg.4 to show the odd and even probability of occurrence from the kth to the k+1th feedback calculation. If the even/odd probability of the initial value (or the kth value n) is 1/2, respectively, we can estimate that the odd/even probability of occurrence in the k+1th calculation is 5/8 even and 3/8 odd.

Equation (1.8) is a uniform random number generation algorithm based on the linear congruential method using the remainder function. [2] When applied to the even-numbered nonlinear feedback in Fig. 4, m = 4, a = 3, b = 1, and if n is a sufficiently large value, the probability of occurrence of mod(n, 4) is the same for all of them.

If the kth time is odd, then the k+1th time will 100% be an even number, but if the kth time is even, then the k+1th time can be either an even or odd number. If the remainder function output of 4 from equation (1.8) has the same probability of occurring, when the remainder is "0", n is a multiple of 4 and the k+1th time will be an even number. Therefore, the probability of an odd/even number occurring is 3/8 (37.5%) for odd numbers and 5/8 (62.5%) for even numbers.
1.4 Probability of Even Numbers Occurring
in Nonlinear Feedback
and Convergence Conditions
Next, we use the odd/even probability of occurrence to prove that the value n converges to 0 [dB] on the logarithmic scale.
If the probability of an even number occurring in nonlinear feedback calculation is p and the probability of an odd number occurring is 1-p, then the logarithmic scale shift [dB] for even numbers is given by Equation (1.9), and the logarithmic scale shift [dB] for odd numbers is given by Equation (1.10).

From equations (1.9) and (1.10), if the average logarithmic axis shift is "negative," it will converge to 0 [dB], and in order for this to happen, equation (1.11) must be satisfied.

From equation (1.11), we can see that convergence occurs at approximately p=61.3% or more when n=∞, and p=62.3% or more when n=7.
This supports the fact that the 62.5% even number occurrence probability model in Fig. 4 satisfies the convergence conditions.
(Additional information) In equation (1.11), p=63.7% or more when n=3, and p=62.8% or more when n=5. These do not appear to satisfy the 62.5% even number occurrence probability model in Fig. 4, but in fact they converge. The reason for this is thought to be that n is not large enough, so the random number generation using the modulo-4 function is not accurate enough, resulting in bias.
2. Simulation Verification
To verify the validity of the odd/even probability of the feedback transfer function model and the convergence condition of Equation (1.11), we performed the Collatz conjecture calculation using four randomly selected 3-digit initial values. Figure 5(a) shows the convergence for initial values of 156, 444, 613, and 929. The number of steps required for convergence is 36, 71, 38, and 129, respectively. After convergence, the "1 → 4 → 2 → 1" loop is repeated. Figure 5(b) shows the cumulative even probability p for the four calculation processes. At the 999th step of each calculation, p is nearly the same, ranging from 66.4% to 66.8%, and is thought to have increased slightly from 62.5% in the "1 → 4 → 2 → 1" loop after convergence. Furthermore, the p values at the time of convergence for each calculation were 69.4%, 66.2%, 71.1%, and 64.3%, respectively, which exceeded the strictest convergence condition of 63.7% when n=3.

Conclusion
As described above, we have considered the calculation of the Collatz conjecture to be a hybrid control transfer function model, and have discovered that it satisfies the conditions for the numerical trajectory to converge to 0 [dB] on a logarithmic axis. If we assume that the probability of an even number occurring in nonlinear feedback is 62.5% from the linear congruential method using a modulo 4 function, this is larger than the 61.3% or higher probability of even numbers occurring as a convergence condition when the initial value is sufficiently large. Simulations have confirmed that with a three-digit initial value, the probability of even numbers occurring at convergence is approximately 64-71%.
3. Additional Considerations
Although it appears that the Collatz Conjecture has been proven using the hybrid control transfer function model, the author believes that this is still insufficient. We will make another prediction using the convergence principle of the Collatz Conjecture in Equation (1.11).
Fig. 6 shows a partial Collatz Conjecture calculation changed from "3n+1" to "3n-1." Because the degree of increase on the logarithmic axis is smaller, we can estimate that the convergence will be faster than the Collatz Conjecture. We simulated the convergence of the same four types of three-digit initial values as in Fig. 5 and the probability of even numbers appearing. The results are shown in Fig. 7.

The number of convergence steps for initial values of 156, 444, 613, and 929 are respectively 36 → 15 (p = 80.0%), 71 → 15 (p = 73.3%), 38 → 48 (p = 68.75%), and 129 → 46 (p = 65.2%), and a tendency for the convergence speed to increase in accordance with the convergence principle of logarithmic axis movement can be confirmed. However, for initial values of 444 and 613, although the value approaches 0 [dB] when the value of n is large, as n becomes smaller, convergence stops at 17 and 37, after which periodic oscillations (18 cycles, 14 cycles) are repeated.

In Fig. 7(a), when n is sufficiently large, the system converges toward 0 dB. However, when n is not sufficiently large, oscillations similar to the limit cycle of a digital filter (linear system) occur. The cause of limit cycles is due to insufficient accuracy of the feedback coefficients and the quantization accuracy of the internal signal processing. [3, 4] In fact, in Fig. 5(a), three limit cycles occur even after the Collatz prediction converges to 0 dB. In the "3n-1" model in Fig. 6, the convergence speed itself is fast, but it gets caught in a limit cycle loop just before the convergence value of 0 dB. Looking at it from another perspective, when n is sufficiently large, the probability of an even number occurring in nonlinear feedback is approximately p = 62.5%, but when n is not large enough, the parameter itself becomes small, making p more susceptible to fluctuations and resulting in a quasi-stable state.
References
[1] Tatsuro Baba, Nobuo Ohmae, Koichi Osuka, The Optimization of Ultrasound System Doppler Velocity Range using Hybrid Control, The Society of Instrument and Control Engineers, Vol.44, No.9(2008), 760--765.
[2] Donald E. Knuth, THE ART OF COMPUTER PROGRAMMING Volume 2 / Seminumerical Algorithms, Addison-Wesley Publishing Company, Massachusetts,1981.
[3] V. Cappellini, A. G. Constantinides and P. Emiliani, Digital filters and their applications, Academic Press, London, 1978.
[4] Tatsuro Baba, Investigation of gain optimization technique in Doppler ultrasound system, Acoustic Science and Technology, Vol.30, No.2(2009), 67—71.