Ugly Windows

The Collatz Conjecture is almost insultingly simple. Take any positive integer. If it is even, cut it in half. If it is odd, multiply by three and add one. Repeat. The claim is that you always reach 1. Computers have checked ranges so large they sound like jokes. Nobody has proved it for every positive integer. Erdős said mathematics may not be ready. Terence Tao has warned that the problem eats people.

I am one of those people, carefully. Not as a prize hunter. As someone who needed a clean mirror for a question I already lived with: does enough awareness eventually bring you home, or can a streak of ugly steps break every fixed way of measuring progress?

I wanted a score. Call it energy. Bigger when the number is big. Smaller when a lot of halving is about to happen. If every step made the score fall, you would eventually have to hit the floor. That kind of argument is called a Lyapunov argument. It is a standard dream in dynamical systems: invent a height function that always goes down, and you have trapped the motion.

I also wanted look-ahead. Not only the next step. A short window of future steps, so a present that looks like growth could still be redeemed by collapse queued just ahead. I wrote that intuition down in February 2025 before I could defend it. The hunch had a name forming around it. The theorem did not.

Here is the version you can hold without notation. Imagine you are allowed to peek exactly k steps into the future of the sequence, where k is a fixed number you chose in advance. Two steps. Five. Eight. Whatever. You design a reward system that pays you for how much halving shows up in that window, and you hope that for every starting number the score goes down as you walk.

Fixed depth means you do not get to say, "for this ugly number I will look farther." The depth is locked. The weights on the rewards are locked. You are betting that one finite window size is enough for the whole infinite zoo of starting points.

That bet is emotionally familiar. A lot of life advice is fixed depth. Give it ninety days. Give the company four quarters. Give the grief six months. Give the identity one more version of the plan. Sometimes that works. Sometimes you meet a streak that is longer than the window you budgeted, and every scorecard you built for that window lies.

I did not prove Collatz. Say that out loud with me. I am not on the doorstep of a proof.

What I did prove, in clean form, is an obstruction. For every fixed look-ahead depth k at least 2, and for every choice of reward weights, there are infinitely many odd starting numbers where that candidate energy does not fall. The score can stay flat in the reward terms and still rise because the number itself grew. The inequality reverses by a uniform margin. The method fails as a method, not because I picked bad coefficients.

The explicit witnesses are a family of Mersenne odds: 7, 15, 31, 63, and on. In the dynamics I was studying, these numbers produce long streaks of "ugly" steps, low collapse by twos, one after another, longer than your fixed window. While that streak is running, every fixed-depth score of that shape gets broken. You can check them. That matters. A result you cannot exhibit is a story. A result with an infinite family of examples is a finding.

After the odd map, people often write the Syracuse step: from an odd m, form (3m+1) divided by the highest power of two that divides it. The 2-adic valuation ν counts how many factors of two come off. My look-ahead summed those valuations over a window of length k. The energy looked like log of the number, minus a weighted immediate collapse, minus a weighted look-ahead sum.

On an L-step bad streak with L bigger than k, the reward terms cancel between consecutive states. What remains is roughly log((3n+1)/(2n)), which approaches log(3/2), which is positive. Energy goes up. Forever many such n exist. Increasing k from 2 to 8 in computation barely saved the idea. The pathological family does not care that you bought a slightly longer flashlight.

If you want the formal writeup, it lives in the Going Home project notes under the bounded-k obstruction finding. This essay is the human surface of that file.

The first writeup killed a two-parameter energy: pick alpha and beta, pick a fixed look-ahead depth k, hope the score always falls. Digging past that binary search, the obstruction is wider. Replace the linear reward with any function of a fixed window of upcoming 2-adic valuations. On a long enough ugly streak, the window looks identical before and after one Syracuse step, so the reward cancels, and you are left with a pure log increase larger than log(3/2). The coefficients were never the point. Fixed finite look-ahead was.

There is parallel work in 2026 on residue-only Lyapunov certificates that also die near integers congruent to -1 modulo high powers of two. Same villains, different ansatz. Honesty means citing that lane and claiming only what we actually closed: look-ahead window rewards of fixed depth.

What stays open is adaptive depth with a real contraction test. Not another grid search. Not a slogan that pretends nu minus log2(3/2) already encodes descent. One-step growth needs roughly nu at least 2. The flashlight has to earn its length.

For a stretch of years I built companies, carried debt, kept teams shipping, and still had nights where the interior score would not fall. The calendar looked full. The body knew better. Pain is often just a long ugly window: more expansion than collapse, longer than the story you told yourself about how soon relief should arrive.

The math did not heal that. It named a pattern I already knew. If your measure of "things are getting better" is locked to a fixed depth, there exist trajectories that will violate it no matter how clever the weights are. In life that is not a license for despair. It is a warning against fake certainty. Some seasons are Mersenne seasons. They are longer than the dashboard.

The Santana Principle I keep on the Going Home page is still a philosophical bet, not a theorem: sufficient awareness guarantees return. The obstruction taught me the cheap version of awareness is not enough. Fixed k is cheap awareness. Adaptive depth, honesty about streaks, willingness to keep looking when the first window fails: that is closer to the principle I actually mean.

The original bounded-k energy plan is dead. Good. Dead plans make room for adult work.

What remains is an open, harder framing: for each number, is there some finite depth, depending on that number, where surplus finally beats expansion? That question is essentially Collatz again, restated to emphasize adaptivity over uniformity. Restating is not solving. It is aiming the flashlight at the right mountain.

Also remaining: an open notebook with dated field notes, a small result I can defend, and a refusal to sell poetry as a lemma. The philosophy can stay. The overclaim cannot.

If your life feels like 31 right now, a streak of steps that refuse to look like home, I will not tell you the score must fall by Friday. I will tell you that fixed-depth stories are brittle, that ruling out a bad method is real progress, and that going home, if it happens, may require a longer look than the one you budgeted.

I am still a student of the problem. The companies still fund the hours. The notebook is public on purpose. The point was never to be the one who proves Collatz. The point was to stay awake long enough that the climb, and the failures inside it, change what I am allowed to say out loud.

Collatz reminds me of that every day. Every step is worth the look. I am going to die. Enjoy the view.