The Axiom the Theorem Gives Back

I.The First Known Thing — the proof that must be run backwards

mathematical logic · reverse mathematics and the calibration of theorems

Almost all of ordinary mathematics can be written down in the language of second-order arithmetic, which speaks of natural numbers and of sets of them and of nothing else. Inside that language sit subsystems of different strengths, distinguished by how much they are prepared to assume about which sets exist. In 1975 Harvey Friedman proposed a use for this hierarchy. Instead of asking whether a theorem can be proved, ask how much is needed to prove it, and answer with an exact measurement rather than an upper bound. The programme that grew from the proposal is called reverse mathematics, and Stephen Simpson’s Subsystems of Second Order Arithmetic remains its reference.

why one proof is never the measurement

The procedure has two halves and the second is the one that does the work. To place a theorem T, a logician first proves T from some set-existence axiom S. That is the ordinary direction and it establishes only that S is enough. The measurement comes from the reversal: working inside a fixed weak base system, prove S from T. If that succeeds, T and S are equivalent over the base, and the conclusion is no longer that S suffices but that nothing weaker than S could have done, because T itself delivers S back. A theorem’s strength is therefore not a number attached to the theorem. It is a statement about three things at once — the theorem, the axiom, and the base over which the two are found to be interchangeable.

the base must be weak, and is chosen to be

That third thing is the part of the method most easily overlooked and the part most deliberately engineered. The standard base is RCA₀, built from basic arithmetic, induction for the simplest formulas, and comprehension only for sets that are computable from what is already to hand. It is chosen under a double constraint: strong enough to state the theorems and carry out routine definitions, and weak enough to prove almost none of them. Both halves matter. A base too weak cannot even express the question. A base too strong proves the theorems outright, at which point every candidate axiom suffices, no reversal is available, and the distinctions the programme exists to draw simply vanish. The weakness is not an austerity; it is the instrument.

the finding nobody designed, and the exception that was a result

What makes the programme more than a filing system is an empirical result that none of its definitions predicted. Theorems drawn from analysis, algebra, topology and combinatorics — fields with no contact between them — turn out again and again to be equivalent to one of five systems, now called the big five: the base itself, then weak König’s lemma, arithmetical comprehension, arithmetical transfinite recursion, and Π¹₁ comprehension. The Heine–Borel covering theorem on the unit interval, the Brouwer fixed-point theorem and the Jordan curve theorem land on the second. Bolzano–Weierstrass and Ascoli’s theorem land on the third. Hundreds of statements collapse onto five strengths, and the collapse is the finding. Equally instructive are the statements that refuse: Ramsey’s theorem for pairs sits outside the five, and establishing where it does sit was a major result rather than an embarrassment. One further caution the field has learnt is that a theorem must be stated before it can be measured, and two faithful renderings of the same informal statement can have different strengths, so the robustness of a placement is itself something to be checked and not assumed.

II.The Second Known Thing — the target that is called required

drug discovery · target selection and target validation

A programme begins by naming a molecule it intends to act on. Validation is the work of establishing that acting on it would change the disease, and it is done, almost without exception, in one direction. The target is deleted, silenced, degraded, or inhibited with a tool compound, and a disease-relevant readout is observed to move. Human genetic evidence is sought in the same direction: a variant that lowers the protein’s activity is found alongside less disease. A programme that can show several of these is held to be well validated, and the field’s own audits report that targets with human genetic support carry roughly double the chance of surviving the clinic.

what the forward direction establishes, exactly

Every one of those experiments establishes a sufficient condition. Removing this protein is enough to move this readout in this system. None of them establishes that the disease could not have been reached, or relieved, by a route that leaves the protein alone; and none of them is a statement about the disease rather than about the preparation. The field knows this, and the knowledge is held as a list of cautions rather than as a measurement: the knockout may be compensated, the phenotype may be driven by something upstream, the tool compound may be acting elsewhere, the line may be an outlier. Each caution is handled case by case, and none of them is a quantity.

the two instruments that come closest

Two existing practices come close enough that they must be named and not waved past. The first is rescue. Re-express the protein and see the phenotype return: this is the discipline’s converse experiment and its standard of rigour. But what rescue restores is the very thing that was taken away, so it demonstrates again, with the sign reversed, that the protein’s activity is sufficient to set the readout. It does not ask whether the disease state, arrived at by any route whatever, carries the protein’s activity with it. The second is ordering. Classical epistasis places genes in sequence along a pathway by which perturbation’s phenotype prevails when both are made, and genome-scale Perturb-seq has since grouped thousands of perturbations into modules by the similarity of their transcriptional signatures. Both are real orderings, and both are built from similarity or precedence within a system; neither is a derivation, and neither declares the background it holds fixed.

the background, which is chosen for the opposite reason

That background deserves its own paragraph, because it is selected by a criterion nobody defends. The cell line, the strain, the medium, the inducible construct: these are chosen because they grow, because they give a large and reproducible window, because the phenotype is robust in them. Robust means the system has what it needs to produce the phenotype. The practice is therefore to measure necessity inside the most permissive preparation available, and to record the preparation in the methods rather than in the claim.

III.The Bridge

no paper on subsystems of second-order arithmetic cites a paper on target validation; both halls issue verdicts of the form this is required, and only one of them has a procedure for earning one

The halls do not meet. The literature of reverse mathematics is mathematical logic and proof theory; the literature of validation is molecular pharmacology and genetics. Neither cites the other, no review spans them, and the one phrase that appears to join them — reverse pharmacology — is a collision of names only, denoting the practice of working back to a mechanism from a clinical or traditional observation. Four further neighbours must be set aside. Causal inference has a probability of necessity, defined on a counterfactual for a single variable; it quantifies the same intuition but has no base system and no equivalence structure, so it cannot say that two causes are interchangeable or that a set of them collapses. Rescue, as above, is sufficiency with the sign flipped. Epistasis and signature clustering order and group, but within a fixed preparation and by similarity rather than derivation. Closest of all, and the most interesting to set aside, is the microbiologist’s minimal medium: a background deliberately stripped so that a requirement can be seen at all. That tradition understands perfectly that a weak background is an instrument, but it uses one to assay a single gene’s own requirement, never to establish that two genes are equivalent over the background, and so it has the base without the reversal.

X: a verdict of necessity is a relation among three things, and the third is a background chosen in each hall for opposite reasons

Call it calibration by reversal over a declared base. Where a claim is established by exhibiting a condition sufficient for it, the necessity of that condition is not thereby shown and can be shown only by a derivation in the opposite direction, carried out inside a declared background too weak to supply the claim by itself; so required is a three-place relation among the claim, the condition and the base, and no measurement of the condition’s effect, however large or however often repeated, can stand in for the missing derivation. Four things follow, and they are not one thing said four times. The first is that the base is a free parameter whose setting decides the answer, and it must be set against the practice rather than with it. In the strong background every candidate suffices, no reversal is available, and all candidates look alike; in the weak one the derivations become possible and the distinctions appear. One hall chooses its base for weakness and names it in the theorem; the other chooses its base for robustness and names it in the methods. The second is that the reversal is a different experiment from the converse. The converse restores the condition and watches the claim follow; the reversal starts from the claim as presented, by whatever route it arose, and recovers the condition from it. For a target the counterpart is not re-expression but a construction that takes an instance of the disease state in the weak background and returns an alteration of that target’s activity — the assertion that the state cannot be occupied without the target having gone through its step. The third, and the reason the method is worth borrowing rather than admiring, is that calibration of this kind has an empirical payoff that is not in its definition: in mathematics, hundreds of unrelated theorems proved equivalent to five axioms, so that the strengths available turned out to be far fewer than the statements. An ordering of targets built by reversal can be asked the same question, and the answer is not settled by constructing it. The fourth is that the statements which refuse placement are the valuable ones. Ramsey’s theorem for pairs sits outside the five and locating it advanced the subject; a target for which no reversal can be constructed and which is equivalent to nothing is the case worth the attention, not the case the method failed on.

IV.The Question That Becomes Possible

neither discipline would ask this alone

The logician has no reason to ask it. A cell is not a formal system, nothing in a screen is a proof, and the programme’s interest ends with the hierarchy. The pharmacologist has no reason either, because the question requires treating the preparation as a declared parameter of the claim rather than as a condition of the experiment, and requires wanting a partition of the target list where the field wants a ranking of it. Held together, one question becomes available and it is a counting question.

the calibration hypothesis

For a disease with a long list of nominated targets and genome-scale perturbation data, define a relation on the list by reversal in a background deliberately chosen to be minimally permissive: one target is at least as strong as another if the second’s phenotype can be recovered from the first’s within that fixed background. The hypothesis is that the resulting equivalence classes are far fewer than the targets and stop multiplying as the list lengthens; that class membership is preserved when the phenotype is restated or the background changed within the permitted range, while the ordering of individual targets is not; and that the class is what a programme’s fate follows, so that two programmes against targets in one class share outcomes more closely than two against targets merely in one pathway.

why it would matter

Because it would make a quantity of something the field now treats as a narrative. If a disease admits ten strengths and four hundred nominated targets, then the length of the target list is not a measure of opportunity, the choice within a class is a question of chemistry and exposure rather than of biology, and a programme whose target is equivalent to one already failed in the clinic has inherited that failure before it starts. It would also convert the standard caution that validation is background-dependent from a disclaimer into a parameter that is set, reported, and varied on purpose.

V.How It Could Be Proven — or Broken

a candidate must be breakable, or it is merely poetry

An intercognix candidate earns nothing until it names its own test. This one needs no new instrument and no new chemistry. Genome-scale perturbation screens with rich readouts already exist in multiple backgrounds for several diseases, together with decades of recorded programme outcomes against named targets. What the test requires is that the background be treated as the declared base and deliberately weakened, and that the relation be built by recovery rather than by similarity.

the honest status

Note the narrowness. The claim is about a partition of one nominated target list in one disease under one declared background, and it says nothing about whether any member of any class can be drugged, which is the question that usually decides. Nothing here asserts that biology has five strengths, or any particular number; the number is the thing to be counted, and a finding that it is large would be a finding and not a failure of nerve. The mathematical result being borrowed is a fact about what is provable from what, not a theorem about cells, and no formal correspondence is claimed between a proof and a perturbation. What transfers is narrower and older than either hall: that a verdict of necessity is owed a derivation in the opposite direction, that such a derivation is only available over a background weak enough not to supply the answer, and that a discipline which selects its backgrounds for robustness has been arranging, without intending to, never to be able to earn the verdict it issues.