WHY THE GAUSSIAN WINS

Start with the photograph. Blur it a little, then blur the blurred copy, and keep going. Two things happen. The picture becomes a featureless smudge — expected — and, more interestingly, you can no longer tell what it was a picture of. Every photograph ends at the same place. The detail is not merely hidden; past a certain point it is genuinely gone, which is why the film-detective instruction to “enhance” a hopeless image is fiction.

Now a fact from a different subject. Add up a lot of small random quantities — the errors in many measurements, the noise from many independent sources — and the total almost always has the same characteristic shape, the bell curve. It matters remarkably little what the individual quantities looked like. Nearly everything, averaged enough times, becomes a bell.

And a third, from physics. Take a complicated system and squint at it: describe it in coarser and coarser terms, deliberately throwing away fine detail, and ask which features survive the squinting and which wash out. That procedure — coarse-graining, and asking what is left — is one of the central techniques of twentieth-century physics.

In the setting studied here these are not three analogies. They are the same statement. The thing being blurred is a correlation structure; the blurring is the operation the framework uses to turn it into something observable; and when you work out what repeated blurring converges to, what averaging converges to, and what squinting converges to, you get the same object — the Gaussian — by what is recognisably the same proof written in three vocabularies. That coincidence is the subject of this page.

Two consequences follow that are worth having even if you never look at the mathematics. First, you cannot un-blur past a point. Running the process backwards fails — not because the computation is difficult, but because the information is no longer there. In a framework where deeper structures give rise to shallower ones, that means the chain of “and what did that emerge from?” terminates for ordinary structures, at a depth you can compute. It is not a universal law: one special class — structures with no fine detail to begin with — can be run backwards forever. Section 05 gives both cases.

Second, and this is the one that cost something: repeated blurring destroys sharp features so completely that a sharp feature can never appear in anything you can measure. The framework had spent years predicting exactly such a sharp feature. Once this was noticed, the prediction was not merely unsupported — it had been impossible from the start, for reasons internal to the framework itself.

Start with a positive semi-definite kernel — a correlation structure — and push it through a smoothing operation, then read the result as a new correlation structure. Since the output is again positive semi-definite, nothing stops you doing it again. The construction defines a map from kernels to kernels, and three structural questions follow immediately.

First, fixed points: is any kernel distinguished by the map itself, or is the choice of a Gaussian just modelling convenience? Second, regress: if one structure emerges from a deeper one, can that one emerge from a deeper one still, indefinitely? Third, poles: what sits at the degenerate ends of iterated smoothing and of iterated descent?

In the translation-invariant setting all three have clean answers, and the answers are governed by three pieces of classical analysis — Bochner's theorem, the central limit theorem, and the ill-posedness of the backward heat equation. That is itself the point. The structure of the tower is not available for choosing; it is dictated by harmonic analysis.

In Fourier variables the smoothing acts as a multiplier: the kernel's spectral density is multiplied by the square of the smoothing kernel's transform. For Gaussian smoothing that multiplier is exp(−σ²k²), which is exactly the heat semigroup evaluated at time σ². One application of the emergence map is one step of heat flow on the kernel profile.

That single identification does most of the work below. Iterating the map is running heat flow forward. Asking what a kernel emerged from is running heat flow backward, which is the classical example of an ill-posed problem. Asking what survives infinite iteration under rescaling is asking for the heat flow's self-similar profile, which is a central-limit question. Three different-sounding questions collapse onto one operator.

The Gaussian shows up as the answer to three questions asked for entirely different reasons. It is worth separating them, because their independence is what makes the coincidence meaningful.

Three arguments, three different fields, one object. Four distinct things are nevertheless being called Gaussian here, and they are not interchangeable, so it is worth separating them: the projection is Gaussian by choice — an input to the construction, derived from nothing; the kernel is Gaussian as a maximum-entropy default given a fixed second moment; the Gaussian is an attractor for iterated smoothing, but only inside the stated basin; and it is the fixed point of the rescaled flow, again only on the finite-variance cone. Selections 2 and 3 coincide exactly. Selection 1 is weaker, and the projection choice is not a selection at all. The honest summary is that the kernel previously treated as an assumption sits where three of these overlap — a reason the modelling choice is not arbitrary, not evidence that the physics is right.

If structures emerge from deeper structures, an obvious worry is infinite regress. The heat-flow identification answers it, and the answer is more interesting than either “yes” or “no”.

The rank-one endpoint deserves a plain-language warning, because it is regularly misread. A structure in which everything is perfectly correlated is a state of maximal constraint, not of no constraint: every distance has collapsed to zero. The state of no constraint is the opposite pole — the diagonal kernel, where every point is infinitely far from every other. Getting those two backwards inverts every interpretive claim built on top of them, which is why the classification is stated as a lemma rather than left to intuition.

The minimal way to give the construction a time evolution is to postulate that the structure evolves by its own emergence semigroup and by nothing else — continuous-time heat flow on the kernel profile, with the smoothing width setting the diffusivity and no other constant entering. It is the least creative possible postulate, which is exactly why it is worth auditing rather than decorating.

The audit is written to give the failures the same space as the successes, and there are two that matter.

If the heat flow is the minimal candidate, the obvious question is what happens with less minimal ones. A pre-registered scan took two one-parameter families of alternatives — porous-medium flows, where the diffusion depends on the density itself, and stable-law flows, which generalise diffusion by changing the exponent in the Fourier multiplier — and ran each through four tiered filters: positivity of the profile, positive-definiteness of the evolved kernel, mass conservation, and entropy monotonicity. Rules, thresholds and the spec were hashed before the scan existed.

The porous-medium failures are not numerical noise: the negative spectral minima run from four percent to thirteen percent, orders of magnitude beyond the tolerance, and they grow monotonically with the exponent. Every tested porous-medium cell violated positive-definiteness under the registered discretisation and tolerances — the evolved object stops being a correlation structure, which is exactly the failure the bound in the companion note measures from the other side. That is a statement about the cells that were run. There is no general analytical proof here that all porous-medium flows must fail, and the note claims none; what the scan establishes is that the property cannot be assumed, and that it failed in every case checked so far.

The structural finding is the one worth carrying elsewhere. A positivity requirement that had been treated as kinematic bookkeeping — a property a correlation structure must have — turns out to be a filter on admissible dynamics, and on this evidence a severe one. It eliminated every tested member of a nonlinear family without any appeal to observation. Whatever nonlinear evolution one proposes next has to preserve positive-definiteness by construction, which sharply narrows the search: flows defined on the spectrum rather than on the profile, or flows coupled to a source term, rather than flows written directly on the density.

The mass flags on the stable family are the opposite case, and were classified as such under the frozen rules: stable semigroups conserve mass exactly, so the measured loss is heavy-tailed leakage past the finite computational domain — its magnitude decreases with the index exactly as truncation predicts, while positivity and definiteness hold at machine precision. Those cells are recorded as non-adjudicable on mass rather than as failures, with the caveat attached to every number quoted from them.

The scan was also asked a selection question: among admissible flows, which is most consistent with the structure the framework has already locked, and which is most consistent with an inflationary cosmological clock?

The answer is a clean, unwelcome result. Consistency with the locked register selects the heat flow uniquely — it is the only scanned law whose attractor shape is the required one, with the deviation measure driven to zero. But the heat flow forces a radiation-like clock. An inflation-compatible clock would require the stable index below about 0.1, and by that point the attractor is maximally far from the required shape; the deviation is already 0.70 at index 0.5. The Pareto front contains no point with both properties.

That upgrades an earlier statement in the parent programme from “a contradiction under an assumed identification” to “the selected consequence of consistency, within the scanned families”. If the locked structure really is the dynamical fixed point, the emergence epoch cannot be inflationary. That in turn independently reinforces a decision taken elsewhere on completely different grounds — the transition precedes inflation and imprints through initial conditions — which is the sort of convergence that is worth more than either argument alone, because neither was aimed at the other.

A scan like this has an obvious failure mode: if you set the fitness targets from your own current claims, you will rediscover your own current claims. The protocol handles it with an asymmetric rule that is worth copying.

Strip the physics away and what remains is a set of statements about smoothing operators on positive semi-definite kernels — objects that turn up in kernel methods, in Gaussian-process regression, in spatial statistics, in signal processing and anywhere covariance structures are coarse-grained or evolved. Three of the results transfer directly.