祝融说。

The Last Day at Königsberg

September 1930, Königsberg — Kant's native city, a port town woven together by seven bridges over the river Pregel. The Conference on the Epistemology of the Exact Sciences met there for three days: the logical-positivist Vienna Circle, the intuitionist mathematicians, and the representatives of formalism, discussing the foundations of mathematics in one hall. The roundtable discussion was set for the last day, September 7; among the participants were Carnap and Heyting, and von Neumann, who had come from Hamburg. The topic of discussion was the old question: from where does the reliability of mathematics come. Each side stated its position; none would yield; the session was on the point of breaking up as usual.

Just as the discussion was drawing to a close, a young man who had not yet spoken opened his mouth. Kurt Gödel, twenty-four, newly granted his doctorate by the University of Vienna, had come to report the results of his own doctoral dissertation — the completeness of first-order logic. What he said now, however, was not completeness but a few sentences: what you have been discussing is the possibility of mathematical systems, and here a strict proof can be given — even taking every remedy, it cannot be done, within the systems under discussion, to prove their own consistency. The remarks were brief, calm, spoken with an accent; the roundtable scarcely responded. And so the conference closed.

Only one man understood. After the session, von Neumann sought out Gödel, talked with him at length alone, and asked for the paper. A few months later, von Neumann worked out the next result independently on his own and wrote to report it — only to learn that Gödel had already proved it. And no one at the time knew that the drama had a second act: on the day after the roundtable ended, September 8, in that same city, the Society of German Natural Scientists and Physicians was holding its annual meeting, the city of Königsberg conferred honorary citizenship upon Hilbert, and the sixty-eight-year-old leader of the program delivered an address over the radio, closing with the words that would one day be carved on his tombstone: we must know, we shall know. The radio waves and those few sentences at the roundtable, in one and the same city, less than twenty-four hours apart — on the one side, the last mobilization order for capping reason; on the other, the proof that the capping was impossible. Rarely has history arranged its ironies so neatly.

In 1931 the paper appeared in print: "On Formally Undecidable Propositions of Principia Mathematica and Related Systems", in the Monatshefte für Mathematik und Physik. The history of logic has taken it as a boundary stone ever since.

First Understand the Program, Then the Theorem

The incompleteness theorems are often told in isolation, as though they were a wonder fallen from the sky. They are not. They are a result that grew up inside the Hilbert program — and to understand what they refuted, one must first see clearly what that program wanted.

David Hilbert, the mathematical leader of Göttingen, answered the crisis of set theory with more sweep than the theory of types or the axiomatizations. In his lecture of 1925, "On the Infinite", he took his oath: no one shall drive us out of the paradise that Cantor has created for us. The paradise was to be kept, the doubts appeased, and his plan fell into two layers. The object layer: formalize the whole of mathematics — from arithmetic to set theory — thoroughly, into a system of symbols: axioms fixed, rules fixed, inference become the transformation of signs; whoever disputes a result need only re-check it character by character (the apparatus of Chapter Five). The meta layer: then use an absolutely safe species of reasoning — a metamathematics, finite, combinatorial, in principle unfoldable step by step on paper — to prove that this system of symbols would never derive a contradiction. The meaning of the division between the two layers: mathematics may say things as bold as it pleases, provided that a certificate of safety, written in the most conservative means, stands surety for it. This is the ultimate version, within mathematics, of Cartesian doubt, and, on this book's reading, the last campaign to "lock determinacy fast": the object layer locks the convergence paths; the meta layer issues a permanent guarantee for the locking itself.

Chapter Three said that the terminus of logic has a shape that no other field's terminus will show: to question it, one must use it. The Hilbert program pushed this shape into engineering: what he wanted to build was precisely that system in which even the questioning has already been formatted — a building that, seen from inside, has no margin left. What this chapter has to tell is the distance between this project's blueprint and its completion report.

Making the System Talk about Itself

Gödel's proof employed an instrument without precedent: Gödel numbering. The method is almost playfully plain: assign to each sign of the formal system a natural number; a formula is a sequence of signs, and so corresponds to a string of numbers, which can be compressed into a single natural number; a proof is a sequence of formulas, and likewise corresponds to a natural number. In this way, statements about the system — "such-and-such a string of signs is a proof", "such-and-such a formula is derivable from the axioms" — are all translated into arithmetical propositions about the natural numbers: the system acquires the ability to talk about itself. The grammar of Chapter Six is here carried out to the letter: the exchange of the identities of observer and observed, implemented as a pile of numbers.

But the numbering is only the prop; the true lever is the diagonal method. With it Gödel constructed a sentence G whose arithmetical content says exactly this: G is not provable in this system. The construction of G is entirely legal — every step is a transformation within the rules of the system — and yet it reopens, inside the system, that self-referential crack of Chapter Six. The liar says "I am false", and gets a dead knot of truth values; Gödel's sentence says "I am unprovable", and gets not a dead knot but a live slipknot: if the system is consistent, G is indeed unprovable — for suppose G were provable; then its content (G is unprovable) holds, and the system has proved an unprovable proposition to be true, which makes the system inconsistent; and suppose the negation of G were provable; then the system endorses both "G is provable" and "G is unprovable", likewise inconsistent. Hence, in a consistent system, neither G nor its negation can be proved: G is undecidable. And we, outside the system, can see that G is in fact true — it speaks its own unprovability, and it is indeed unprovable. True, but unprovable: "truth" and "provability", two coins that in weaker systems look struck from the same die, are assayed for the first time as two.

This is the first incompleteness theorem: any consistent formal system containing enough arithmetic has propositions that it can neither prove nor refute. (Gödel originally required the slightly stronger condition of ω-consistency; Rosser in 1936 weakened it to simple consistency.) It was soon found that undecidability is not the monopoly of artificial constructions: in 1977, Paris and Harrington proved that a quite natural combinatorial proposition (a strengthened form of Ramsey's theorem) is independent of Peano arithmetic. The boundary does not lie in the far suburbs; it lies in the city blocks of mathematics itself.

The Second Theorem, and the Program's End

The first theorem already hurt; the second is what passed the sentence on the program. Arithmetize the very argument of the first theorem — the reasoning above, "if the system is consistent, then G is unprovable", can itself be translated into a formula inside the system — and so the system can formally write out, within itself: if the system is consistent, then G is unprovable. But "G is unprovable" is precisely G itself. That is to say, if the system could prove its own consistency, it could prove G, which by the first theorem it cannot. Conclusion: any consistent formal system containing enough arithmetic cannot prove its own consistency from within itself.

This is the second incompleteness theorem. What it strikes is not the reliability of mathematics but the original form of Hilbert's plan: the guarantee cannot be issued by the very party that is guaranteed. The system's freedom from contradiction — the most important certificate of safety — is precisely the one certificate that cannot be signed from inside the system. The finite metamathematics can still testify from outside; but the means employed by the outside proof no longer carry a guarantee of the same strength; guarantees either regress infinitely, or come to rest upon some stratum that certifies itself no further. The program did not come crashing down; it was rewritten with precision: Hilbert's grand goal — to prove the consistency of all classical mathematics by finite means — could not be achieved in its original form, and in the years that followed, stronger meta-means bought local consistency proofs — Gentzen in 1936 used induction up to transfinite ordinals to write arithmetic an insurance policy — but that was already a contract of another kind: the guarantor stronger than the thing guaranteed, not cleaner than it. The absolute, internal, self-completing proof of safety was from then on struck off the list of possible projects.

Tarski: Truth Stands outside the System

The same boundary has a twin scale. In 1933 the Polish logician Tarski published "The Concept of Truth in Formalized Languages": within a consistent formal language containing arithmetic, the truth predicate of that very language cannot be defined — there exists no formula T(x) inside the system that holds of a sentence x exactly when that sentence is true. (The original appeared in Polish; the German version followed in 1935.) To say "which sentences of this system are true", you must stand outside the system and say it in a stronger, or at least a different, metalanguage; the metalanguage can define the truth of the object language, but the truth of the metalanguage itself must in turn be spoken from one level higher still.

The lesion of the liar paradox received from Tarski its final localization: truth is a predicate that crosses levels. On this book's reading, this theorem is the semantic version of Chapter Six's "the theory of types is an artificial stratification of observation" — except that the stratification is no longer an optional engineering decision but a proven necessity: in any sufficiently strong system, the position from which its own truth could be uttered does not exist inside the system. For an observer to describe its own observation, it must rise one level; and this rising has no terminus, because every level of observer has its "truth" on the level still above. Part Two here touches its deepest structure: the observer cannot be completely locked by its own observation — and this is no poetic lament; since 1933 it has been a strictly proved theorem.

The Common Misreading: Sliding from Boundary to Relativism

The theorems are so famous that misreadings breed in their wake. The most common is a triple step into relativism: every system is incomplete — therefore no system can monopolize truth — therefore all systems are alike, anything goes, and to speak of "truth" at all is naïve. This inference served as ornament in the humanist writing of the mid-century, and is still commonly met today in citations of the "Gödel proved that all knowledge is limited" kind.

It is wrong on every step of the stairs. First step: incompleteness delimits a specific capacity — completeness of internal proof, and self-certification of consistency — of a specific class of objects: consistent formal systems containing enough arithmetic; it says nothing about "any system whatsoever", still less about knowledge being limited in general. Second step: the theorems say precisely not that systems cannot be compared with one another. Quite the opposite: the proofs of Gödel and of Tarski are themselves triumphs of metamathematics — a stronger framework gave a strict, testable characterization of a system's boundary, one acknowledged by mathematicians as a body. If "anything goes", this characterization could not exist; the boundary theorems are themselves specimens of cross-system rigor. Third step: from "there is a boundary" it does not follow that "all are equal". A map has a border; that does not make all maps equally accurate. Part Three will return to this point on the side of mathematics; for now let the verdict stand: the incompleteness theorems are precise knowledge about the boundary, and relativism is the abandonment of knowledge about the boundary — to confuse the two is to read a medical report as a proof that medicine does not exist.

There is a family of abuses to be dismissed in one sentence as well: incompleteness did not prove the existence of God, did not prove the immortality of the soul, and did not prove artificial intelligence impossible — what it proved is a structural fact concerning formal systems and arithmetic. The articles that use it as a cure-all have for the most part never read even the conditions of the first theorem.

On RC's Reading: The Margin of What Is Locked

Now comes the most crucial transcription of Part Two — a restatement of Gödel on RC's reading.

Step one: a formal system is a locked structure of convergence. RC discusses observational locking in its axiomatic system (Axiom A3): determinacy is locked only when observation occurs. Chapter Five delivered the transcription: a formal system is the symbolization of an already-locked structure of convergence — axioms and rules are the explicitly locked achievements of convergence. Gödel numbering then gave this locked structure the capacity to point at itself: the system became its own object of observation. This is precisely an implementation, in the world of symbols, of what RC discusses in its axiomatic system as the symmetry of observer and observed (Axiom A4) — and what the second theorem shows is that what this implementation struck was not an accident but structure.

Step two: the second incompleteness theorem read as the inner limit of A4. The system cannot, from inside a single point, quiet the observational divergence concerning itself — the self-assertion "I am consistent" is exactly the observer's attempt to take a total snapshot of itself; Chapter Six already proved that the total snapshot cannot catch the hand that presses the shutter, and this chapter makes that result exact: consistency and truth are predicates that can be spoken only from one level higher. No self-guarantee of determinacy exists; guarantee is always a conferment across levels. RC discusses consensus reinforcement in its axiomatic system (Axiom A6): observations across subjects and levels verify one another, forming a positive reinforcement loop. On this book's reading, Gödel's result is not a counterexample to A6 but A6's inner limit: convergence cannot complete itself from inside a single point; verification must cross levels — a system's determinacy was never a certificate it issued to itself, but a letter issued by the structure of convergence above it. This squares, word for word, with the verdict of Chapter Three: the chess player's rules are the hardest, but the chess player cannot use the rules of chess to prove the rules of chess unflawed.

Step three, the point on which this chapter lands: Gödel is the strictest proof, inside the world of symbols, of the thesis that determinacy equals what is locked. Look closely at what the theorem guarantees: so long as the system is consistent, G is true — the territory of truth is strictly greater than the territory of proof. Why? Because "being proved" means being taken in by this particular procedure of locking: from these axioms, by these rules, along this path. The taking-in has an aperture, and the aperture was chosen (Chapter Five: axioms and rules are explicitly locked achievements of convergence); whatever aperture is chosen, there are things it cannot take in. What is locked necessarily carries the margin that "it could have been locked otherwise" — RC's axiomatic system lays this down as Axiom A7 (conservation of margin): in locking determinacy, observation does not exhaust the Ground; there always remains available margin not yet locked. The undecidable proposition G and the unreachable truth predicate are the registered residence of this axiom in the world of symbols: they are not defects of the system but the mode of existence of the act of locking — a system that truly locked away every margin would precisely be unable to lock out even its own freedom from contradiction; it could not so much as exist as a consistent system. Gödel is thus a supporting case in this series' casebook: not an obituary of reason running aground, but the first time that the thesis this book's Part One has been laying down since Chapter One — determinacy is the locked achievement, and locking carries margin — has received, inside mathematics, a proof without breach. The boundary is not the failure of reason; the boundary is a definitional component of the very concept of determinacy — what has no boundary is not omniscience but nonexistence.

Two Boundaries, Clarified as Usual

The customary clarifications, two of them.

First, this chapter is not saying that mathematics is unreliable. Quite the opposite: the theorems inside the system hold with iron validity — every proved proposition has a proof checkable character by character within the system, its force not diminished by a hair; incompleteness has not shaken a single existing mathematical theorem. What proof has drawn is the difference between the true and the provable: there are true propositions that proof does not take in. Engineers go on using the calculus; number theorists go on proving theorems; the boundary theorems set no checkpoint on any road of proof. In the ledger of Chapter One: not one deposit in the account is diminished; the theorem says only that there exists in the world money that is true nonetheless though the ledger cannot record it.

Second, this chapter is not saying that one may leap out of the system at will. That G is true is seen from outside the system — but the meta-system that "sees" it is itself a structure of convergence: it has its own axioms, its own rules, its own undecidable propositions. To leap out is only to go upstairs, not to ascend to heaven; every flight of Tarski's ladder leads one level higher, and there is no top. Whoever reads "the system is incomplete" as "therefore I may freely choose what to believe" mistakes the level above for the absence of levels — the observer on the level above is bounded by its own observation in just the same way. There is no view from nowhere; what exists is only a testable ladder between one standpoint and another. And this happens to be the conclusion that RC's paper, discussing theoretical dimensional reduction in Section 3.3, reaches: every theory is a dimension-reducing projection under a finite horizon, its validity constrained by the triple boundary constraint of level, subject, and time. The meta-system is no exception — it, too, is an interpretive system awaiting re-observation by a level higher still.

The Close of Part Two

Now the four chapters of Part Two are closed.

Chapter Four, the walking path of Athens: inference became for the first time an object of observation, and form was stripped out of content — the explicitation of pure structures of convergence had its first specimen. Chapter Five, Jena and Cambridge: inference was compressed into mechanically executable transformation rules, consensus reinforcement was fitted with industrial gears, and logicism attempted to lock the whole of mathematics into one explicit path. Chapter Six, a letter of a few lines: a sufficient capacity for self-observation let the observer fall into his own range of observation, total observation was proved an illegal act, and the community relaid the foundation with the two patches of stratification and authorization. Chapter Seven, Königsberg: the limit of locking was proved — consistency cannot certify itself, truth cannot be defined from within, and locked determinacy necessarily carries margin.

Strung together, the four chapters are the complete redemption of what Chapter Three announced: "what a convergence system in dialogue with itself meets when it stabilizes to its limit" — the answer: it meets itself, and measures, with the rigor of mathematics, the distance between itself and itself. The chess player's rules are the hardest; and precisely because they are the hardest, their boundary has been proved the most strictly. This is not the disgrace of logic but exactly the coming of age of logic: of the three forms of reason, it is the first to know, in the form of a theorem, where its own boundary lies.

The next part turns to mathematics — the architect's enterprise. The incompleteness theorems will reappear from the other side: there they are not logic's boundary but a crack in mathematical construction; together with non-Euclidean geometry and the paradoxes of set theory, they will constitute the answer to the question "at the cost of which cracks does a self-enclosed freedom of expansion come". Gödel's two pages are on file at both construction sites — and this is in itself one more working of Chapter Three's thesis: one and the same mechanism of convergence, leaving one and the same signature in two directions of dialogue.