Deutsch
Working basis · not an overview, but a tool

Where one could start

The board as it stands: what is proved, what is measured, what remains open — and a checklist that dispatches unworkable ideas in half a minute, so that the time is left for the workable ones.

A What we're actually aiming at

Three questions are easily confused, and the confusion costs weeks.

QuestionWhat would have to be shownOur relation to it
P vs NPa polynomial algorithm, or that none existsnot directly
NP vs coNPa proof system with polynomial-size proofs for all tautologies (Cook/Reckhow)this is our program
Random 3-SATthat this one family is easy or hardour measurement object

Two consequences worth remembering. First: random 3-SAT is neither necessary nor sufficient for P vs NP. P vs NP is a statement about the worst case — were random instances easy, P = NP would not follow; were they hard, P ≠ NP would not follow either. Second: they are nonetheless the standard test family of the Cook/Reckhow program, and the road to NP ≠ coNP runs through lower bounds for ever stronger systems.

Whoever wants the big prize must show that some system has no short proofs — and the family is a free choice. We chose random 3-SAT because that is where most is known.

B The board

Proved — theorems, not measurements

SystemResult for random 3-CNFSource
Resolutionsize 2^Ω(n)Chvátal/Szemerédi 1988
Res(k), small kexponentialAlekhnovich 2005
Polynomial calculus, any fielddegree Ω(n)Ben-Sasson/Impagliazzo 1999
Sums of squaresdegree Ω(n)Grigoriev 2001, Schoenebeck 2008
Rank-1 cuts at x = ½violation 0here, computed exhaustively

Measured

Five closure spaces, all empty: literals · binary clauses · linear forms over GF(2) · LP inequalities · clauses of width ≤ w. Add to that no symmetry (orbit compression exactly 1.00 up to n = 400) and no signal from the recurrence detector (~1% versus 49–73% for Tseitin).

Open

SystemWhat is missing
Cutting planes at k = 3Fleming et al. 2017 covers only k = log n
Bounded-depth Fregeonly Ω(n^(1+ε)) steps known (2024)
Frege, extended Fregeno superpolynomial bound is known for any family

The last point is not a gap in our own work but the open question of proof complexity for forty years.

The cause, in one sentence

Expansion is a GLOBAL property.
Every one of our tools is LOCAL:
     fixing a variable   →  touches 12.8 clauses
     adding a clause     →  touches  1
     a resolution step   →  touches  2

Ω(n) local moves for a global property — hence all costs are linear in the exponent. Proof complexity knows three non-local tools: symmetry (measured absent), linear algebra (proved ineffective), extension (open).

C Checklist

Work through this in order for every new idea. A "no" doesn't mean "wrong" — it means: here is where the burden of proof lies.

  1. Is it a saturation in a polynomial space? Then it's dead, and provably so: a proof whose clauses all lie in a space of size P has at most P distinct nodes; with P polynomial and proof size 2^Ω(n), that's a contradiction. The space must be superpolynomial — or the procedure must not be a saturation.
  2. Is every step local? Does a step touch O(1) clauses? Then you need Ω(n) steps, and choosing each one is itself a search. That is exactly where the detour route and the offset route failed.
  3. Does it require symmetry? Aut(F) is trivial, measured up to n = 400. There is nothing to lift here.
  4. Does it require a small unsatisfiable core? There isn't one: ∂S ≥ ε·|S| for all |S| ≤ δn. Every subset below linear size is satisfiable.
  5. Does it require derivable binary clauses? A starving instance at n = 120 derives 0.1 of them, out of 630 clauses.
  6. Is it width-bounded? w* = 0.018·n. Fixed width w carries up to n ≈ 55w − 110.
  7. Does it compute at x = ½? There, tightness is the same as width 2 — a cutting-planes procedure at this point is resolution.
  8. Would it be faster than 2^0.078·n? That is the measured CDCL cost.

D What extension is

The one non-local avenue still open. Introduce a fresh variable y, with defining clauses:

y ↔ (a ∨ b)     →     (¬y ∨ a ∨ b) ,  (y ∨ ¬a) ,  (y ∨ ¬b)

This preserves satisfiability — y is defined, not constrained. Any number of such definitions may be stacked.

Extension makes width bounds meaningless — by construction.

Our whole hardness picture hangs on width: proofs need wide clauses, wide clauses are costly. Extension lets one name a wide fact and then keep computing at width O(1). That is why it isn't merely the last door — it is the door our own argument cannot, in principle, close.

Cook's proof of the pigeonhole principle in extended resolution is polynomial, where ordinary resolution needs 2^Ω(n). The pattern: extension encodes a self-reduction — one names the instance of the smaller problem and runs an induction.

Why it is hard to use

E The target, and the trap in front of it

Find polynomially many definitions y₁…y_p such that the extended formula has small closed clause sets — i.e., low expansion in the extended variable space.

Because closed sets (∂S = 0) are exactly what random instances lack, and their count demonstrably predicts the ignition point (p ~ 2·10⁻¹³).

And now the trap. Compute the boundary of the extension gadget itself:

S = { (¬y ∨ a ∨ b) , (y ∨ ¬a) , (y ∨ ¬b) }
Variables:  y (3×),  a (2×),  b (2×)      →      ∂S = 0

The gadget is itself a closed set. Low expansion can therefore be forged at will, without gaining anything — the defining clauses are satisfiable, no contradiction ever follows from them. Two binary clauses come free with it, but they are definitional, not constraining.

Closedness is necessary, not sufficient. Everything except the contradiction can be forged.

Anyone proposing an extension therefore has to say why the resulting closedness is load-bearing and not merely definitional. That is the sharpest test that can be formulated — it has already taken apart one idea here.

F Six seed ideas

IdeaAssessment
Name the interface of a subformula. If everything in S is resolved away, a clause over ∂S remains. Name it. Cost 2^|∂S| = 2^εk. For k ~ log n the gadgets are polynomial. Close to what CDCL learning already does anyway. Testable.
Name a coarsening. Group variables into blocks, one variable per block for the block state. Probably no — randomness is not self-similar. But testable in an hour, and the "no" would itself be informative.
Name learned clauses. Replace every learned clause of width w with a variable. This is at its core "extended learning"; empirically it gives nothing on random instances. But why was never measured — with the expansion profile that's now possible. New measurement.
Count beyond x = ½. If a procedure cuts elsewhere, the optimum shifts. The only one of the three open doors reachable empirically. What's missing is a measure of tightness at an arbitrary polytope corner. A tooling problem, solvable.
Exploit the SAT/UNSAT asymmetry. WalkSAT solves every satisfiable threshold instance. A strong hint that the issue is the certificate, not the search. Fits squarely into the Cook/Reckhow frame.
A lower bound instead of an algorithm. Bounded-depth Frege on random 3-CNF. Open, well-defined, publishable — and the expansion data are exactly the tool with which such bounds are usually established. The most realistic contribution.

G What can be re-measured in minutes

The instruments for every idea are ready to hand.

QuestionToolDuration
Does an extension actually lower expansion?expansionsprofilminutes
Does a seeded or extended formula ignite?wechselkurs, umwegminutes
How many detours does it really need?existenz (proof DAG)hours
Does a procedure reach further than n₅₀ = 52.9?pvier, schnellp4minutes
Does a family have a small closure space?recurrence detectorminutes
Does it have symmetry?bahnen (1-WL)seconds
How many binary clauses arise?zweierseconds
And the rule that has saved us three times over: a negative finding without a working positive control is not a finding.

For every idea, first find a family where it must work (pigeonhole, Tseitin, 2-SAT), and check that the instrument fires there. A cutting-planes test once reported "nothing found" here — and the control revealed that the pigeonhole family had already stopped firing.

H Three questions to start with

  1. What exactly do we want to name? Extension is a namer, and every success came from there being something recurring to name. Random instances have no recurrence — unless we simply haven't looked for it. Our detector measures exactly that and stays silent at ~1%. But it only sees one kind of recurrence, namely variable supports. Is there another?
  2. What is the analogue here of induction on n? For PHP it is the reduction to PHPn−1. A random instance contains no smaller random instance — but after fixing a few variables it contains an instance of higher density, and higher density is easier (n₅₀ = 16.5α − 18). That is a genuine reduction direction, and the price is the case split. Anything that pushes it below 2^0.059n is new.
  3. Why is the UNSAT side hard and the SAT side not? Both have the same expansion. So the difference must sit somewhere else. This is the question where it's least clear what the answer could even be — and precisely for that reason the place I would start.