The limits of description
The collection's wager is that one construction underlies the limits of formal reasoning. It is time to test the wager at its edge, and the test is unkind: there are two roots of limitation in mathematics, and the diagonal is only one of them. The diagonal bounds what can be proved — a system reasoning about itself runs into Gödel's wall. But there is a second, dual bound, on what can be described: first-order logic cannot pin down a structure, leaving a latitude in which models proliferate, the natural numbers refuse to be unique, and the size of the continuum floats free. This second root — Löwenheim–Skolem opening the slack, forcing exploiting it — is not a diagonal phenomenon. The diagonal appears on its territory only as a borrowed tool, never as the source. Self-reference is one of two pillars of mathematical limitation, and the collection, honestly placed, is the theory of one.
the other limit · the slack · the unfixable ℕ · the second independence · two roots · where the corpus stands
The Staircase climbed the diagonal's syntactic face — incompleteness as it lives in proofs and provable strength, an ordinal tower with no top. This descent turns to the face the collection has least examined, the semantic one: not what a theory can prove, but what it can describe — which structures its axioms succeed in pinning down. The two faces are joined, by Gödel's completeness theorem, at the hinge where provability meets truth-in-all-models. But they come apart in a way that matters to the collection's central claim, because the semantic face has a source of limitation the diagonal does not reach. Where self-reference makes a system blind to its own consistency, the looseness of first-order description makes a theory blind to which world it is talking about — and that blindness, traced to its root, turns out to be a different thing entirely. The honest task of this last descent is to find the edge of the diagonal's dominion, and to name what lies past it.
Beside the limit on proof sits a limit on description — older, and stranger. No first-order theory with an infinite model succeeds in picking out a single structure. There is always more than one world the axioms cannot tell apart.
The diagonal's limit is a limit on proof: a theory rich enough to describe its own provability cannot prove everything true of its intended subject. Set that aside and ask a different question — not what a theory can prove, but what it can specify. You write down axioms hoping to capture a structure: Peano's axioms to capture the natural numbers, the axioms of ZFC to capture the universe of sets. Do the axioms succeed in pinning the structure down — is there exactly one model, up to isomorphism?
This is a limit of a wholly different character from the diagonal's. The diagonal leaves a true sentence unproved; non-categoricity leaves the very subject of the theory underdetermined — the axioms are faithful to many structures at once and cannot prefer any. It is not that something true escapes proof; it is that there is no single "the natural numbers," no single "the universe of sets," that the first-order axioms manage to name. The theory speaks, and a whole class of worlds answers to the name. The question this descent must settle is where that looseness comes from — whether it is, at bottom, the diagonal wearing a semantic disguise, or a second root of limitation standing on its own. To decide it, look at how wide the looseness runs, and what produces it.
The looseness has a precise and astonishing measure. A first-order theory cannot even pin down how many things it is about — and the proof of this owes nothing to self-reference.
The width of the looseness is given by the Löwenheim–Skolem theorems, the founding results of model theory and the purest statement of the semantic slack.
The consequence at the foundations is the famous vertigo of Skolem's paradox. The axioms of ZFC prove the existence of uncountable sets — that is their whole point; P(ℕ) is uncountable, by Cantor's diagonal. Yet by downward Löwenheim–Skolem, if ZFC is consistent it has a countable model M. Inside M, the set playing the role of P(ℕ) is uncountable — there is no bijection with ℕ that belongs to M. Seen from outside, the whole of M is countable, so such a bijection exists in the wider world. The contradiction dissolves into a relativity:
"Uncountable" is not absolute; it is relative to a model. M believes its P(ℕ) is uncountable because the bijection that would refute it lives outside M. The model is genuinely countable and genuinely "thinks" it contains uncountable sets, both at once, because countability is a fact about which functions exist, and that depends on where you stand. The axioms cannot reach out and fix the answer.
Now mark the source, because it is the crux of the whole descent. Nothing here is a diagonal. The slack is a consequence of compactness and the Löwenheim–Skolem construction — features of how first-order satisfaction works, of the logic's finitary, term-built character. No sentence refers to itself; no fixed-point-free map is twisted; there is no naming fed to itself. The diagonal does make a guest appearance — it is Cantor's diagonal that tells M its P(ℕ) is uncountable in the first place — but the diagonal here states the problem (there are more subsets than elements) and contributes nothing to the slack (that the problem's answer is model-relative). The looseness is born elsewhere: in the inability of a finitary logic to constrain the size of its models. A first limit, already, that the diagonal does not author.
The slack reaches the natural numbers themselves: Peano's axioms describe not one ℕ but a whole zoo of others. And here, in the sharpest theorem about that zoo, the diagonal at last appears — entering, tellingly, as hired help.
Apply the slack to arithmetic. Take PA and add a new constant c with the axioms c > 0, c > 1, c > 2, … By compactness — every finite subset is satisfied by the standard ℕ, choosing c large enough — the whole infinite theory is consistent, so it has a model. In that model there is an element c greater than every numeral: a nonstandard number, with a whole tail of others above the standard ones. First-order PA cannot pin down ℕ; in fact it has 2^{ℵ₀} countable models pairwise non-isomorphic, only one of which is the genuine ℕ.
How wild are the impostors? It is tempting to imagine a nonstandard model as ℕ with a few extra numbers tacked on — something you could in principle write down. Tennenbaum's theorem destroys that hope, and draws the sharpest possible line between the true ℕ and every counterfeit.
Now watch how the theorem is proved, because the lesson of the whole descent is written in its machinery. The proof fuses two ingredients of opposite character. From the diagonal's side it takes a pair of recursively inseparable sets — two disjoint computably-enumerable sets with no computable set separating them, an object that exists precisely because of the Gödel–Rosser construction, the diagonal in its sharpest recursion-theoretic form. From model theory's side it takes overspill — the fact that a property holding of every standard number must, in a nonstandard model, hold of some nonstandard element too. Combine them: overspill produces a nonstandard element coding a set that separates the inseparable pair, hence a non-computable set living in the model; and if + or × were computable, that set would be computable after all — contradiction.
Here, for the first time on the semantic side, the diagonal is genuinely present — recursive inseparability is pure Gödel–Rosser, pure diagonal. But notice its role. It does not produce the nonstandard models; compactness does that, with no self-reference anywhere. The diagonal is brought in only to measure the models once they exist — to prove they are wild. It is the instrument that detects the slack, not the cause of it. The semantic limitation was already there; the diagonal merely takes its sharpest reading.
There are two great independence phenomena in mathematics, and they have two different sources. Gödel's is the diagonal's. Cohen's is the slack's — and the size of the continuum is its monument.
Cantor's diagonal opens a gap: |ℕ| < |P(ℕ)|, the reals outnumber the integers. It then asks the obvious question — is there a set of intermediate size? — the continuum hypothesis, CH: that 2^{ℵ₀} = ℵ₁, no cardinality strictly between. The diagonal poses the question and falls silent. Its answer came in two halves, twenty-five years apart, and the answer is that there is no answer.
Gödel showed in 1938 that if ZFC is consistent, so is ZFC + CH — he built the constructible universe L, a model where CH holds. Cohen showed in 1963 that if ZFC is consistent, so is ZFC + ¬CH — and for that he invented forcing. Start with a countable model M of set theory and adjoin a generic object G, built to meet every dense condition in M; the extension M[G] is again a model of ZFC, but now with so many new subsets of ℕ that 2^{ℵ₀} overflows ℵ₁. Between them, CH is independent: true in some models, false in others, settled by none.
And the kind of independence is not Gödel's. This is the distinction the whole descent has been driving at, and it is sharp. Gödel's unprovable sentence Con(T) is true — true in the standard model — and merely escapes proof; the limit falls on proving a determinate fact. CH has no such determinate fact behind it: it is a switch, forced true or false at will, with no preferred value that the slack permits us to call correct. One is a truth that outruns proof; the other is a question with no model-independent answer at all. The literature names the difference without ceremony: the independence of the continuum hypothesis is not self-referential the way Gödel's is.
Forcing does use a diagonal-flavored move — the generic G exists only because M is countable, so its dense sets can be enumerated and dodged one by one, a genericity construction in the diagonal's family. But again the diagonal is the instrument, not the cause. What makes CH independent is not that construction; it is the freedom to build the model either way — the same semantic latitude that let ℕ have impostors now lets the continuum take any size. Genericity dodges the dense sets; the slack is what leaves room to dodge.
The descent's findings assemble into a single picture. Mathematical limitation has two roots, dual to each other — and the diagonal is the whole of one and a mere visitor to the other.
Set the two limitative phenomena side by side and their independence becomes unmistakable. They differ in what they bound, in the character of what they leave open, and — decisively — in where they come from.
A system reasoning about itself cannot prove all it should. The unprovable statement is true and merely escapes derivation. Source: the diagonal, which needs nothing of the answer-space — pure syntax. Cantor, Gödel, Tarski, Turing, the whole catalogue of no.
A first-order theory cannot pin down its world. The open statement has no model-independent answer — a switch, true here, false there. Source: compactness and Löwenheim–Skolem — pure semantics. Non-categoricity, nonstandard ℕ, the floating continuum.
The duality runs deep enough to be worth stating as the descent's thesis. The diagonal's limit is internal — a system turned on itself, the loop of self-reference closing into a wall. The slack's limit is relational — the gap between a theory and the structures it hopes to name, the looseness of a finitary language stretched over infinite worlds. One is a fact about a system's relation to itself; the other a fact about a theory's relation to its models. No amount of self-reference produces the second, and no amount of semantic looseness produces the first. They meet, by the completeness theorem, on a single bridge — but they are raised from opposite banks, and the collection has built its whole edifice on one of them.
The descent ends at the edge of the collection's dominion, and marks it honestly. The diagonal is the root of one of the two great limitations — and the deeper, more portable of them — but a complete account of what mathematics cannot do needs the other root too, and the collection does not contain it.
Here is the boundary, stated plainly, in the collection's own discipline of separating what is proved from what is hoped. The wager has been that one construction underlies the limits of formal reasoning. The descent shows the wager is true of half the limits and false of the other half. The diagonal is the root of the limits of proof — incompleteness, undecidability, paradox, the entire syntactic catalogue of no — and that root the collection has traced to its bottom, six times over, in every disguise. But the limits of description — that no theory names its world, that ℕ has uncomputable impostors, that the continuum floats free — spring from a different root, the semantic slack of first-order logic, and on that ground the diagonal works only as hired help, sharpening readings of a looseness it did not cause.
This is the deflation of The Provenance, carried one storey higher. There, the diagonal was right-sized from "the master fixed-point theorem" to "the fixed-point theorem of logic" — one source of fixed points among three. Here it is right-sized again, from "the root of mathematical limitation" to "the root of the syntactic limitation" — one pillar of two. Each deflation costs the collection a claim of universal reach and pays it back in honesty and precision, which is the trade the whole project is built to make. The diagonal does less than the grandest reading wants. What it does, it does exactly, and the boundary of what it does is now drawn.
Two cautions keep the boundary from being read as a demotion. First, the second root is no removable defect. By Lindström's theorem, first-order logic is the maximal logic possessing both compactness and the Löwenheim–Skolem property — so the slack is not a flaw a cleverer logic would repair but the standing price of any logic with a workable proof theory; flee it into second-order logic and you regain categoricity only by surrendering completeness and effective proof altogether. The slack is as deep as the diagonal, and as permanent. Second, the two roots are not equals in portability. The slack is the signature of a particular kind of formalism — logics with good model theory. The diagonal is not tied to any formalism: it surfaces in set theory and arithmetic and computation and the λ-calculus and the category of any cartesian closed world, wherever a thing can be made to refer to itself. So of the two roots, the collection has built upon the one that travels — the limitation that is not about logic but about self-reference, and so appears everywhere self-reference does. That is the more fundamental root in the only sense that matters: it is the one that is everywhere.
So the collection stands, at the last, exactly placed. It is the complete and rigorous theory of self-reference — the internal limit, the loop closing into a wall, the construction that draws every boundary of proof and, climbed, builds the whole staircase of strength. It is not the theory of all mathematical limitation, because mathematical limitation has a second root, relational rather than internal, the gap between a theory and its worlds, and that root is a different subject with a different source. The diagonal bounds what a system can prove of itself; the slack bounds what a theory can say of its world. Two limits, two banks, one bridge of completeness between them — and Quodlibet, honestly, is the map of one shore, drawn to its every edge, with the far bank in view across the water.
The diagonal bounds what a system can prove of itself; the slack bounds what a theory can say of its world. Two limits, raised from opposite banks, joined by a single bridge. The collection is the map of one shore — and now it knows which.