Notes on the proof of the φ^k lower bound

One organizing idea

The theorem proves that a naive heuristic is correct to within constants: consecutive sums in a cardinality layer behave like independent random pairs, so a C(k−j,j)/C(k,j) fraction of the layer’s adjacencies should be disjoint. The entire difficulty of the paper is that blockers are the obstruction to this heuristic, and the proof’s two mechanisms are precisely the two ways a blocker can try to defeat it. Every note below hangs on that frame.

The window scale makes the heuristic legible. The layer holds C(k,j) sums whose fluctuation scale is √j, so the typical spacing between consecutive sums is √j / C(k,j) — and the proof’s window θ is chosen at exactly that spacing. The small-gap probability P(|Δ| ≤ θ) ≈ 1/C(k,j) is then not an estimate but the statement that the pair lands within one typical gap of each other.

Seven ideas that must not lose the exponential scale

The proof’s difficulty comes from making several delicate ideas work together without losing exponential precision at any stage:

1. translating adjacent divisors into disjoint adjacent subset sums;
2. realizing the random-weight model with actual primes, without invoking any shrinking-interval prime result;
3. separating the cardinality layers exactly, via Q > (1+η)^k;
4. proving that a fixed disjoint pair is consecutive with probability of order 1/C(k,j);
5. splitting blockers into external and internal classes;
6. recognizing that internal blockers require a codimension-two estimate;
7. choosing the saddle layer so the combinatorial term has exponential base exactly φ.

Why internal blockers are the hard part

The external-blocker estimate is fairly standard once formulated correctly: an external R keeps at least one coordinate outside S∪T, that coordinate survives the conditioning, and convolution smoothing bounds the conditional density. For internal blockers this route is not merely weaker — it is structurally unavailable. An internal R has no coordinate outside the pair, so after conditioning on the endpoint sums there is no independent randomness left to convolve with.

The proof succeeds by changing mechanism, not by sharpening the old one. The swap decomposition writes the gap as Δ = U + V with U and V independent signed sums, and blocking pins both into (0,θ]. One factor of θ becomes two, and the internal total drops to 2^{−(H₂(β)−2β)k} of the gap probability. That single extra factor of θ is the entire difference between a proof and a false start, and it is the genuinely inventive step.

Quantifiers and the order of limits

The proof is unusually sensitive to quantifiers. One must keep k fixed while taking Q → ∞, then let η → 0, and only afterward study large k. The inequality of Theorem 3.1 is exact for every k once the limits are done, so asymptotics in k enter only in Theorem 5.1, where they belong. Reversing or blurring those limits could silently introduce an unproved prime-distribution assumption. As it stands, the number-theoretic input is the prime number theorem in fixed proportional intervals and nothing more.

The saddle constant is algebraic, not numerical

It is natural to assume F(β*) = log φ is a coincidence verified to some precision. It is not. The critical equation is 5β² − 5β + 1 = 0, so at β* one has β(1−β) = (1−2β)² = 1/5. Substituting, F(β*) collapses in three lines to log(1−β*) + ½ log 5 = log((√5+1)/2). The golden ratio enters through this identity, not through the Fibonacci sum Σ_j C(k−j,j) = F_k+1, which the single-layer proof never uses.

Where to read with hostility

The proof is plausible-looking even at points where a small mistake would destroy the conclusion, so referee difficulty is substantial. Three joints carry the load and reward hostile reading:

1. the order of limits and the total-variation coupling in Section 3 — this is where a hidden prime-distribution assumption would live, and the fixed-k discipline is what excludes it;
2. the independence of U and V, which rests entirely on the swap algebra producing disjoint coordinate supports, and should be verified by hand;
3. the fact that only the upper bound on the hypergeometric blocker sum is ever used — a matching lower bound appears nowhere and is not needed.

Everything else is standard machinery correctly assembled.

The theorem is checkable by experiment

Theorem 5.1 admits direct Monte Carlo falsification: draw the weights, enumerate the layer, sort, and count disjoint consecutive pairs. If the ratio E[N] / C(k−j,j) decayed with k, the theorem’s constant could not exist. In practice the ratio holds near 0.72 for k from 14 to 29 (layers up to 4.3 million subsets), with no drift. An interactive browser companion runs the experiment live and draws the layer itself. The stability of the constant also suggests it converges to an explicit limit — an open question the paper does not address.

Level and character

Conceptual difficulty

High.

Technical difficulty

Moderate to high.

Length

Comparatively short for what it proves.

Prerequisites

Analytic and combinatorial number theory, probability, Fourier density estimates, asymptotic combinatorics.

Referee difficulty

Substantial, for the reasons above.

This is not a proof built on a very deep external theorem. It is a difficult assembly proof: the challenge is discovering the right representation, the right probabilistic scale, and the right decomposition so that tools no stronger than the prime number theorem and a local limit estimate become sufficient.