The Silence That Has A Size

I.The First Known Thing — the parts that give nothing away

cryptography · correlation immunity and resiliency

Adi Shamir published a short paper in 1979 describing how to divide a secret among eleven people so that any six of them together can recover it and any five of them, pooling everything they hold, learn nothing whatever. The construction is a polynomial of degree five: each person is given the value of the polynomial at a distinct point, six points determine it and therefore its constant term, and five points leave every possible constant term exactly as likely as before. George Blakley described a geometric scheme in the same year. The word doing the work in both is nothing. It is not that five shares narrow the field, or make the secret guessable with effort, or leak a bit or two that a larger conspiracy could exploit. The distribution of any five shares is independent of the secret. There is no computation, no quantity of side information and no amount of patience that extracts a signal, because there is no signal in the object being examined.

the property, once it is named, can be measured

Cryptography generalised the property and then learned to grade it. Thomas Siegenthaler, in 1984, was studying attacks on stream ciphers in which an adversary correlates the cipher's output with a single one of its internal shift registers and recovers that register alone, at a cost far below attacking the whole machine. A function resists this exactly when its output is statistically independent of every fixed choice of m of its inputs; such a function is called correlation-immune of order m, and if it is also balanced, m-resilient — a term fixed a year later by Chor, Goldreich, Håstad, Friedman, Rudich and Smolensky, who arrived at the same object from the problem of generating shared randomness among untrusted parties. The order is a single integer and it is the size of the largest group that learns nothing. A correlation attack on a function with resiliency 3 does not run slowly; it returns the same answer it would return on pure noise, however long it is run, and only a search conducted over groups of four or more inputs enters a regime where the function has anything to say.

and the silence turns out to be paid for

The part of this that matters most here is the part Siegenthaler proved in the same paper, and it is a trade-off rather than a definition. Immunity is not free, and what it costs is complexity. For a function of n variables that is correlation-immune of order m, the algebraic degree is at most n − m; if the function is balanced, at most n − m − 1. Every increment of silence at the small scales lowers the ceiling on how intricate the function can be overall. This is the opposite of the intuition that a thing which hides well at every scale one can afford to look must be forbiddingly complicated underneath: within a fixed set of variables, the two properties compete for the same budget, and a designer who buys a wide band of silence has spent the degree that would have bought intricacy. Half of the subsequent literature is the exploration of that frontier — Sarkar and Maitra, Tarannikov, and others constructing functions that meet Siegenthaler's inequality with equality while keeping nonlinearity high, which is to say finding the best that can be had when the budget is fully spent. The discipline's operational habit follows from the bound: a null result from a low-order attack is not filed as evidence of strength in general. It is filed as a measurement of m, and m tells the attacker at what order to resume and tells the designer how much degree remains to be found.

II.The Second Known Thing — the knockout that does nothing

drug discovery · target selection and target validation

Walk to a hall that asks which gene matters. A target is nominated, and the standard instrument for deciding whether it carries the disease phenotype is perturbation: remove the gene, or silence it, or degrade its product, and see whether the phenotype moves. The modern form is a pooled CRISPR screen, in which a library of guides removes each of many thousands of genes one at a time across a population of cells and the phenotype is read as depletion or enrichment. The output is a ranked list of single-gene effects, and the targets that emerge from it are the ones whose removal did something. This instrument is the closest thing the field has to a causal test at scale, and its economics — one gene per cell, the whole genome in one flask — are what made it the default.

what happens to the ones that do nothing

The list has a long, flat middle. Most genes, removed singly, move nothing measurable, and the field's account of why is well developed and largely correct as far as it goes: buffering by paralogues, redundant routes through a pathway, feedback that restores the set point, compensation over the days between the cut and the readout. The practical response has been to raise the order of the probe. Pairwise screens made genetic interaction — the departure of a double knockout from the product of its singles — a measurable quantity at scale, and synthetic lethality became a target class in its own right. CombiGEM and related barcoding schemes industrialised the pairs; three-way combinatorial screens over druggable genes have been run; engineered Cas12a arrays now deliver four, five and six guides to a single cell, and higher-order screens over chromatin regulators are published. The order is climbing because the singles were quiet.

and the thing the quiet is taken to mean

What has not changed is how a quiet result is read. A gene with no single effect is treated as an estimate near zero with an interval around it — a weak effect, a small one, possibly one that better power, more replicates, a cleaner readout or a longer time course would resolve. Interaction scores are reported as continuous magnitudes and analysed as such. The higher-order screens are presented as a way of finding effects that are too small to see at lower order, and the design conversation is about coverage and sensitivity: how many combinations can be afforded, how deeply each must be sequenced. Nowhere in that conversation is there a quantity answering the question a cryptographer would ask first, which is not how large is the effect but at what order does this system begin to speak at all — nor any suggestion that the width of the silence, if it were measured, might itself constrain what is behind it.

III.The Bridge

no screening paper cites Siegenthaler

The halls share nothing. Boolean function theory for stream ciphers and functional genomics have no common journals, no shared citations and no review that spans them; the cryptographers' degree is the size of the largest monomial in an algebraic normal form and the biologists' is a count of edges, and the traffic that does exist between cryptography and biology runs to privacy of genomic data, which is a different matter entirely. What the two halls have in common is a situation. Both interrogate a system they cannot open, by fixing a small number of its components and watching an output. Both find, very often, that the output does not move. One of the two halls has a name for the width of that stillness, a way to measure it, and a theorem saying what it costs the system to maintain it.

X: silence is an integer, and the integer is a ceiling

Call it immunity order. Where an outcome is determined by a set of components arranged so that every function of fewer than k of them is statistically independent of the outcome, a probe of order below k does not return a weak signal but no signal, so the null it returns is a fact about the arrangement and not about the size of anything; the failure is of order rather than of power, and no depth, replication or sensitivity recovers it, because the quantity being estimated is genuinely zero. From which follows the part that is easy to miss, and is the whole of the interest here: the silence is paid for out of the same budget as the intricacy, so a determinant immune to probing up to order m cannot also be of arbitrarily high degree within its own support, and the width of the silence is a ceiling on the complexity of what is behind it rather than a warning of depth. Two consequences are worth stating exactly, because they are not the ordinary caution that biology is combinatorial. First, the null has a use it is not currently put to: an estimate of m is a direct instruction about the order at which the search should resume, and a system that has been probed to order two and is silent is telling the experimenter something numerical, which the field currently files as an absence and discards. Second, the direction of the inference is inverted. A determinant that hides from singles and pairs is, on this account, not the most intricate kind but a constrained kind, and constrained in a way that makes the order at which it becomes visible predictable in advance — which is exactly the design parameter that higher-order screening now spends its whole budget guessing at.

IV.The Question That Becomes Possible

neither discipline would ask this alone

The screening hall asks how large an interaction is and climbs in order until it finds one. The cryptographic hall has never looked at a cell. Neither asks whether the silence of the low orders is itself a measurement with a distribution — whether quiet genes come in two kinds rather than one, and whether the kind can be told apart before the expensive high-order experiment is designed.

the immunity-order hypothesis

Does a phenotype's determinant have an immunity order — the largest combination size at which every perturbation is indistinguishable from unperturbed — such that among gene sets with a strong high-order effect the low-order effects are distributed in two modes, a graded shadow and an exact silence, rather than on one continuum; and does the silent class behave as the bound predicts, becoming visible sharply at a single order and requiring no further orders beyond it, where the graded class continues to gain from each increment?

why it would matter

Because a screen's null results are currently the cheapest and most abundant data the field produces, and they are used for nothing. If the two-mode structure is real, then a gene set that is exactly silent at orders one and two is a different object from one whose effects are merely small, and belongs in a different queue: the first is a candidate for a combination whose order can be estimated rather than searched, and the second is a power problem to be solved with depth. It would also change what a validation failure means. A target that fails validation in a single-gene experiment is at present retired as unsupported, and on this account some proportion of those retirements are readings of an order — a report that the experiment was run at the wrong size, from which the right size can be computed. And it bears directly on how the higher-order screens are designed, because the budget of a combinatorial screen is set almost entirely by the order it is run at: a prior estimate of the order at which a module becomes visible converts an exhaustive search into a targeted one, and the estimate would come from the silent data already in hand.

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 experiment for its first and decisive pass: the published combinatorial screens contain, for thousands of gene sets, the lower-order measurements the hypothesis makes a claim about.

the honest status

Note the narrowness. The claim is about the shape of one distribution — the low-order effects of gene sets that have a strong high-order effect — and about whether that shape has two components. It does not say that most quiet genes are hiding something; the expectation is that the silent class, if it exists at all, is a small minority, and the claim is worth making only because that minority would be systematically discarded by every screen now run. It proposes no new assay, no change to how interaction scores are computed, and no revision to any published screen: only that the ladder of subsets be assembled for the sets that already have one, and the distribution of its lowest rungs be looked at as a mixture rather than summarised by a mean. It is also worth saying what the import does not carry. Cryptography's functions are small, exactly specified, and adversarially designed, and a cell's are none of those things; what transfers is a definition, a way of measuring it, and one inequality, not the reason the property exists in a cipher. The whole thing rests on an assumption that is checkable and may fail at once — that a biological null can be distinguished from a small effect at all, given the floors and the mosaicism, in which case there is no silence to measure and the question does not arise. Until the subset ladders are assembled and the mixture fitted, this remains exactly what its stamp says: a candidate. A structural echo between two halls of the library, waiting for someone with the tools of both to walk the corridor.