Deutsch
Proof complexity · random 3-SAT · 4 September 2026

Which Lever Actually Moves

Three places where the corpus contradicts itself — and a lever that carries further than any other and dies all the same.

§ 1

The largest lever was on file as having no effect

This project's polynomial layer is width-bounded resolution: derive everything derivable, but discard every clause wider than w. For fixed w there are only O(n^w) of them, so the procedure is polynomial. The question is always the same — how far does it carry?

Two documents in the corpus answer differently, and the contradiction has a strategic consequence.

One measures the ignition density — seed per slot in the level — finds it equal at widths 3 and 4, and files widening the level as closed: “exactly no effect”.

The other measures the same question directly, on 330 instances across eleven values of n, by two methodologically independent routes that agree:

Leveln₅₀transition width
width 352.918.4 in n
width 4108.212.3 in n

Width 4 carries a good factor two further. That is not a small shift but the largest lever measured anywhere in this corpus — larger than extended resolution at factor 1.35.

Why the density argument fails is written in the very document that makes it: the obvious refinement, the branching factor λ as a threshold, is tested there and rejected — an instance with λ > 1 dies, one with λ < 1 ignites. If the round factor is no threshold, the static density certainly is not.

The right order parameter is the reproduction number R as a function of the accumulated closure. And R depends on how easily two clauses fall back into the level: at width 4 one shared position suffices, at width 3 it takes two. The level is not merely larger, it is better connected — and that is exactly what cancels in a density computation.
§ 2

Two points carry two different laws

From the same two numbers the corpus draws a straight line, w*(n) ≈ 0.018·n. But those same two points also carry a geometric law, and the ratio 108.2 / 52.9 = 2.05 is conspicuously close to 2.

Lawpredicts n₅₀(5)and about w*
linear163w* = Θ(n) already in the window
geometric222w* = Θ(log n) in the window

They differ by 36 % and one measurement separates them. The difference is not cosmetic: geometric would mean that across the entire measurable range proofs of size n^O(log n) suffice — quasipolynomial rather than exponential. Asymptotically w* must become linear; where the crossover sits has never been asked.

And the measurement is reachable — the reason given against it was wrong

The corpus holds width 5 to be unreachable because its space already has 10⁸ clauses at n = 55. That is the supply, not the closure. The same text measures elsewhere that the refuting width-4 closure stays cubic while its space grows quartically.

Here is that measurement again, with auxiliary variables and across five support points:

|A| ~ N2.78    r = 0.9998    share of supply constant at 6 to 7 %
Memory follows the closure, not the supply. Anyone who declares a level unreachable because its supply is n^w is using the wrong number.
§ 3

The answer was sitting in a log file

An earlier report concludes that random 3-SAT has neither of the two resources that carry short proofs: no small closure space and no symmetry. The automorphism group is provably trivial up to n = 400.

The obvious objection is that a proof might also lift across an approximate symmetry and pay only for the seam. The corpus takes it seriously, builds the measure, validates it against a positive control — and asks whether the quantity is Θ(1) or Θ(n). Then it says: running.

The run finished. Its output sat unexamined on disk for nine days.

n100150200300400
usable6077101138177
usable / n0.6000.5130.5050.4600.442
usable ≈ 1.53 · n0.790    r = 0.9981

Θ(1) is excluded. It would require usable/n to fall by a factor 4 from n = 100 to 400; the measurement gives 1.36. And the exclusion holds despite thin data, because every value is a lower bound: a lower bound growing faster than any constant rules out Θ(1) no matter how poor the search.

Approximate symmetry does not vanish. The exact symmetry is trivial, the approximate one is not — and the report that declares the door shut says itself what Θ(n) would mean: “a constant fraction of the formula is liftable.”
§ 4

Choosing the definitions is the entire effect

Extended resolution may invent a new variable. A definition

z ⟷ (la ∧ lb)   =   (¬z ∨ la)   (¬z ∨ lb)   (z ∨ ¬la ∨ ¬lb)

constrains the old variables not at all — z is defined, not restricted — and all three clauses fit inside level 3. This is the only direction that leaves the variable set, and therefore the only one the project's own counting argument cannot reach. For it, no lower bound is known for any family of formulas.

The standing verdict is nonetheless that there is no procedure for finding useful extension variables. The established method looks for recurring structure to factor out and finds exactly none on random instances — no surprise, since nothing occurs twice there.

The move that was missing is the reverse one: not search for what is there, but manufacture what is missing.

The rule, derived from the mechanism

The definition yields two binary clauses. One of them merges with every clause C containing ¬la into a width-3 clause over z. Two such descendants merge back into the level when their remainders share a position. Those are the new ignition points, and they are countable before any computation:

points(p, q) = #{(C,D) : p̄ ∈ C, q̄ ∈ D, C and D share one further position}

The result, and four controls

Measured on exactly those instances where level 3 fails without auxiliary variables — and provably fails, since the cap lies above the supply. At identical clause and variable counts, nothing distinguishes the runs but the choice of pairs.

Comparison, k = n definitionstargetedcontrol
vs random variable pairs, n = 8813/200/23
vs random literal pairs, n = 885/90/9
vs spread instead of concentrated choice, n = 806/60/6
vs anti-selection, n = 64 (k = n/2)13/140/14

The third row was a refinement meant to be better, and it falls — which sharpens the mechanism: the definitions must build a hub, not a net. The fourth is the second side of the control: the anti-rule is worse than chance, not merely better than nothing.

Soundness: 438 satisfiable instances extended, 438 remain satisfiable, zero failures. Without this control every rescue on the unsatisfiable side would be worthless.

As a curve over n, with k = n definitions:

n₅₀ rises from 52.9 to 90.9  ·  price: factor 8 in supply

Per unit of cost this is the best lever in the corpus. Width costs a factor n/2 and buys 2.05; the targeted definitions cost a fixed factor 8 and buy 1.72.

§ 5

And then it dies

Whether this is more than one further lever is decided by the number of definitions required at size n. The cost is O((n+k)³) — for any polynomial k* the procedure would stay polynomial. That is measured as a budget response, across two complete curves.

n₅₀(k = n/2) = 80.9     n₅₀(k = n) = 90.9
Doubling the budget buys 10.0 variables. So k* doubles every ten variables: k*(n) ~ 2n/10. The procedure's exponent is therefore 0.10 against 0.049 for an ordinary CDCL solver — extrapolated, twice as expensive.

The finding is now complete, and it is two-sided. Choosing the definitions is a real, large, mechanistically understood effect, controlled in both directions. And it is still no way out: the price grows exponentially, as with every other lever.

The unlikely reading would have been remarkable, since for extended resolution no theorem forbids a polynomial refutation of random 3-CNF. It is now measured and does not hold — at least not for this selection rule in this window.