The Ground That Fills In Behind You

I.The First Known Thing — the pursuit that is decided by the map

combinatorics · graph searching

The problem entered mathematics from caving. In 1967 Richard Breisch, writing for a speleological bulletin, asked how a party should be organised to find someone lost in a cave when the passages are many and the lost person may be moving. T. D. Parsons formalised it in the mid-seventies as a game on a graph, and it has been studied ever since under the name of graph searching. The passages are edges and the junctions are vertices. Every edge begins contaminated, meaning the fugitive may be in it. Searchers are placed on vertices, moved along edges and removed again. An edge is cleared by a searcher traversing it while its far end is guarded. The fugitive is invisible, knows where every searcher is, and moves at unbounded speed along any path not blocked by a searcher. The search is finished when no edge could hold him.

the resource is simultaneity, not effort

Two consequences follow immediately from the fugitive's speed, and they are what make the problem unlike an ordinary covering problem. The first is that clearing is never accomplished by visiting: a passage swept and then abandoned is contaminated again the moment any searcher-free path connects it to a passage not yet swept, and it does not matter that it was swept thoroughly, or recently, or twice. The second is that what holds cleared ground is not the sweeping but the searchers standing still. At any instant the cleared region is exactly the region separated from the contaminated region by the bodies of the searchers, so the searchers occupied at that instant must form a cut. The quantity that decides whether a cave can be cleared at all with a given party is therefore the number of separators that must be held at once, and total effort — searcher-hours, distance walked, how many times each passage is traversed — does not enter it. The minimum party size that suffices is the graph's search number.

it is a property of the map

The search number is an invariant of the graph. It is untouched by the fugitive's speed, which is already assumed unbounded, and by his cunning, which is already assumed perfect; nothing about him remains to vary. What it responds to is shape, and specifically to width rather than to size. A single corridor, however long, is cleared by one searcher walking it, and lengthening it a thousandfold changes nothing. A cycle needs two. Branching raises the number, because at a junction the searcher who walks on must leave someone behind, and a tree that branches deeply enough will exhaust any fixed party. A square grid needs a number that grows with the side of the grid, not with its area. A graph in which everything meets everything needs all but one of its vertices occupied. The formal statements are exact: for the variant in which searchers clear a vertex by standing on it, the search number is the graph's pathwidth plus one, the quantity Kirousis and Papadimitriou identified in 1986 as the interval thickness or vertex separation number; for the variant in which the fugitive can be seen, Seymour and Thomas showed in 1993 that the answer is the treewidth plus one. These are structural parameters, computed from the map before anything is known about the search, and computing them is hard in general — Megiddo, Hakimi, Garey, Johnson and Papadimitriou proved the search number NP-hard in 1988 — though linear on trees.

recontamination does not help

The result that gives the field its character, and the one that matters here, is less obvious than it looks. A strategy is monotone if it never allows a cleared edge to be recontaminated. It is natural to suspect that a clever strategy might do better by deliberately giving ground — releasing a searcher from a cut, letting a region refill, and using the freed body to open the structure elsewhere, the way a chess player concedes a pawn. Andrea LaPaugh proved that it cannot. Her theorem, circulated in 1982 and published in the Journal of the ACM in 1993 under the title Recontamination does not help to search a graph, says that any strategy clearing a graph with k searchers can be replaced by a monotone strategy clearing it with k searchers. Kirousis and Papadimitriou proved the counterpart for node searching; Bienstock and Seymour later gave a short unified argument covering the mixed case. The practical reading is that the whole benefit of a schedule is in the order it takes ground, never in the ground it hands back, and that the minimum party is reachable by a strategy which, once a passage is clear, keeps it clear until the end.

what the discipline makes of it

Graph searching is treated in the literature as a combinatorial characterisation of pathwidth and treewidth, valuable because it gives those parameters an operational meaning and yields algorithms. Where applications are named they are robotic sweeping of buildings and the decontamination of computer networks. The field has no interest in what clears a body, and no reason to ask what its invariant would mean for a population that hides in tissue rather than in a cave.

II.The Second Known Thing — the response that does not last

drug discovery · resistance and durability of response

A great many treatments work and then stop working. The tumour shrinks and regrows, the bacterial load falls below detection and the patient relapses after therapy is withdrawn, the viral load becomes unmeasurable and rebounds within weeks of interruption. Durability of response — how long the effect lasts, and whether anything was eradicated — is the property that separates a useful drug from a curative regimen, and the discipline has three standing explanations for its absence. Resistant variants may pre-exist and be selected, which is the Luria–Delbrück picture and the basis of the Goldie–Coldman argument for combinations. Cells may survive without genetic change, tolerant or dormant, which is the persister literature. Or the drug may simply fail to arrive somewhere, which is the pharmacological sanctuary.

the field's simultaneity argument is a genetic one

On the first of these the field has a rigorous and well-known result about simultaneity. If resistance to each of n agents arises independently, then giving the agents one after another allows the population to acquire the mechanisms one at a time, each under the selection that the previous agent supplied, whereas giving them together requires the improbable coincidence of all n at once. This is why combination therapy is given as a combination, and the modelling behind it — Bozic and Nowak's among others — makes the case quantitatively. It is worth being clear about what that argument is about. It counts mechanisms. Its state space is genetic, the compartments it concerns are lineages rather than places, and it says nothing about where in the body anything is.

the sanctuary is treated as a gradient, not as a structure

On the third explanation, the spatial one, a great deal is now measurable that once was not. In tuberculosis, drug concentrations are resolved lesion by lesion: rabbit and human lung material shows that the same systemic dose produces very different exposures in a cellular granuloma, in the caseous centre of a necrotic one, and in the wall of an open cavity, and that bacilli in caseum tolerate concentrations that kill them elsewhere. Mechanistic lesion-level models now predict which lesion types a given regimen leaves at partial exposure and where resistance is therefore amplified. In HIV, the lymphoid follicle, the central nervous system and the gut are documented reservoirs with their own penetration characteristics, and scores exist for how well a regimen reaches the brain. In acute lymphoblastic leukaemia the central nervous system was recognised as a sanctuary decades ago, and directed therapy to that compartment became doctrine because relapse kept occurring there. Meanwhile the traffic between compartments — lymphocyte recirculation, dissemination of bacilli from one lesion to another, the seeding of metastases from other metastases rather than from the primary — is itself increasingly measurable, by barcoded strains and by phylogenies reconstructed from sequencing.

what is computed, and what is not

What the discipline does with all of this is arithmetic of totals. A regimen is scored by aggregate exposure against a susceptibility measure, by the number of agents still active, by the number of compartments in which each agent exceeds a threshold. Penetration is assessed drug by drug and site by site, and a site that scores badly is a reason to add an agent that reaches it or to raise a dose. These are quantities of amount and of count. What is nowhere computed is a quantity of arrangement: whether the compartments a regimen actually holds, taken together, separate the compartments it has cleared from the compartments it has not, given the map of exchange between them. The one place where a compartment was singled out and covered deliberately, the nervous system in leukaemia, was reached empirically from the pattern of relapses over years, not derived from any structure. Nor does the field ask whether the failure of an intermittent or sequenced regimen is a failure of adherence and exposure, as it is invariably read, or a failure of a different kind: ground relinquished and refilled.

III.The Bridge

no paper on pathwidth cites a granuloma; no regimen paper cites a search number

The halls do not meet. One is a branch of combinatorics whose literature is graph parameters, monotonicity proofs and the complexity of computing invariants, and whose applied citations run to robot sweeping and computer worms. The other is clinical pharmacology and infectious disease, whose literature is concentrations, thresholds, susceptibility and relapse rates. Neither cites the other; no review spans them; the words they share are ordinary English words used loosely on one side and precisely on the other. And yet both are describing the same object, which is a population that can be eliminated locally, can return to anywhere it is not at that moment excluded from, and is opposed by a force that can be held in only a limited number of places at once.

X: what holds cleared ground is what is held, not what was done

Call it sweep width. Where a population can be eliminated within a region but re-enters any region it is not at that moment excluded from, and the eliminating pressure can be maintained in only a limited number of regions at once, clearing the whole is not a matter of how much pressure is applied or for how long, but of whether the regions held at each instant separate what has been cleared from what has not. Three things follow, and they are the content of the bridge. First, the number of regions that must be held simultaneously is a property of the map of exchange between regions and not of the population in them: it is untouched by growth rate, by the number of escape mechanisms, and by the potency of the pressure, all of which set how long a region takes to clear and none of which set how many must be held. Second, it responds to the width of that map rather than to its size, so that a body with many refuges arranged in a chain is easier to clear than one with fewer refuges all exchanging with each other, and adding refuges along a line costs nothing while adding cross-connections costs a great deal. Third — and this is the part that no dose-response intuition supplies — relinquishing cleared ground cannot lower the number, because any schedule that permits re-entry can be replaced by one that never does, at no extra cost in simultaneous holding; so a sequenced or rotating regimen has no combinatorial advantage whatever over one that never gives ground, and may buy tolerability and time but can never reach a clearance that simultaneity could not. The consequence for the clinic is that coverage is the wrong thing to count: a regimen holding four refuges that do not separate is strictly worse than one holding two that do, at equal exposure, and the durability of a response becomes a prediction from the topology of exchange rather than a summary of the amount of drug delivered.

IV.The Question That Becomes Possible

neither discipline would ask this alone

The combinatorialist has no reason to ask it because there is no cave in the question, and the pharmacologist has no reason to ask it because nothing in the training suggests that a property of a regimen might be computable from a map with no concentrations on it. Held together, they produce a question with a definite shape.

the sweep width hypothesis

For a population persisting in anatomically distinct refuges between which it demonstrably moves — bacilli across cellular granulomas, caseous lesions and cavity walls; virus across lymphoid follicles, gut and the nervous system; leukaemic cells across marrow and the meninges; tumour cells across metastatic sites that seed one another — can a sweep number be computed from two things that are separately measurable, namely the graph of exchange between refuges and the set of refuges in which a given regimen holds the population above its killing threshold; and does that number predict relapse better than total exposure, than the count of active agents, or than the count of refuges reached? The sharp form of the question is the arrangement claim: among regimens matched on total exposure and on the number of refuges covered, those whose covered set separates the cleared refuges from the uncovered ones should be the durable ones, and the difference should be large. The second form is the monotonicity claim, which is the more surprising and the more easily wrong: that an intermittent or rotating schedule cannot achieve a clearance that the same agents held continuously could not, so that the failure of such schedules is not only a matter of exposure but of ground conceded.

why it would matter

Because it would make durability partly computable before a regimen is assembled, from a map that can be drawn from barcoding and phylogeny and from penetration data already collected drug by drug; because it would identify which refuges are worth the effort of reaching — those that sit on a cut — rather than treating every poorly penetrated site as an equal liability, which is how the problem is currently costed; and because it would give a structural reason, distinct from adherence, for why intermittent regimens disappoint. And if it fails — if relapse tracks the totals and is indifferent to the arrangement — then the sanctuary is a matter of concentration after all, exactly as the field treats it, and the cave has nothing to say to the clinic.

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 has the unusual property that its most likely break condition can be reached on paper, from data already published, before any animal is treated.

the honest status

Note the narrowness. The claim is about persistence by spatial refuge with measurable exchange between refuges, and about nothing else. It says nothing about resistance arising genetically within a single site, which is the dominant mechanism in many settings and is the subject of the field's existing and much better established simultaneity argument; it does not compete with that argument and does not replace it. It says nothing about latency, nothing about immune-mediated durability, and nothing about the great many diseases in which no compartment structure is demonstrable. It does not claim that combinations are unnecessary, nor that exposure does not matter; it claims that among regimens of comparable exposure an arrangement quantity may be doing work that nobody measures. And the analogy is imperfect in a way that is worth stating plainly rather than burying: the searchers in the cave are indivisible bodies and the party size is an integer, whereas drug coverage is graded, partial and decaying, so the crisp invariant of the graph becomes at best a weighted quantity in the body, and it is entirely possible that the weighting destroys the sharpness that made the idea interesting. The whole thing rests first on a map that can be drawn from published data, and that map may end the claim in an afternoon. Until it has been drawn, 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.