The Word That Forgets the Start

I.The First Known Thing — the sequence that works from anywhere

automata theory · a word that ends in one place whatever it started from

Take a machine with a finite set of internal states and a small set of inputs, each input carrying every state to some state. Suppose the machine is running and nobody knows which state it is in. There is a question that sounds as though it should need a sensor and does not: is there a sequence of inputs which, applied blind, leaves the machine in the same state no matter where it began? Such a sequence is called a synchronising word, or a reset sequence, and Černý put the question in this form in 1964. His four-state example takes two inputs and needs nine moves; applied from any of the four starting states it ends in one. Nothing is observed while the word runs. The uncertainty about the state is not resolved but discarded, because the word maps the whole set of possibilities onto a single point.

the parts orienter: sensors replaced by a sequence of obstacles

The idea was given a physical form by Natarajan in 1986 and developed by Eppstein in a 1990 paper on monotonic automata. Parts arrive along a conveyor in arbitrary orientations. Instead of a camera and an actuator, the line carries a series of fixed passive obstacles, each of which turns a part by a quarter if it meets the obstacle one way and leaves it alone if it meets it another. A sequence of such obstacles, correctly chosen, delivers every part at the end in the same orientation regardless of how it arrived. The device has no idea what it is doing to any individual part. The order of the obstacles is the word, and the guarantee is structural: the machine ends in a known state because the map the word induces has a one-element image.

what is easy to decide, and what is hard

Two results govern the practice. Whether a reset word exists at all is easy: an automaton can be synchronised if and only if every pair of states can be brought together by some word, and the pairwise condition can be checked in polynomial time, which is Eppstein's greedy construction — merge two states, then merge the result with a third, and so on, giving a word of cubic length. Finding the shortest word is a different matter and is NP-hard. The length of the shortest one is the automaton's reset threshold, and how large it can be remains, sixty years on, an open question: Černý conjectured a bound of the square of one less than the number of states, and exhibited the family that meets it, while the best proved general bounds are cubic and have been lowered by small increments, most recently in work of Szykuła. The relevant point for anyone outside the field is what the threshold is a property of. It is fixed by the merging structure of the transitions and by how many states must be collapsed. It is not fixed by any rate, because the machine has none.

and the two facts that are easy to miss

Two further features matter, both of which read as pedantry inside automata theory and as something else outside it. First, the state the word arrives at is not privileged. It is not the machine's initial state, or its rest state, or an empty state; it is simply wherever that word happens to funnel everything, and a different word may reset the same machine to a different place. Synchronising is not the same as returning to the beginning. Second, doing nothing is itself a word — the idle input, repeated — and there is no reason in the structure why the idle input should be the synchronising one. A machine can be resettable by a short sequence of active inputs and not resettable at all by waiting. And some machines are not resettable by anything: if there is any partition of the states that every input preserves, no word can cross it, and the distinction between two states survives every sequence that can be applied. For those, no amount of input removes the uncertainty, and the right conclusion is not that the sequence was too short.

II.The Second Known Thing — the pause before the second period

drug discovery · absorption, distribution, metabolism, excretion

A crossover trial gives each subject both treatments in turn and compares them within the person, which is why bioequivalence studies and a great many early-phase and drug-interaction studies are built that way. The design rests on one assumption: that by the time the second period begins, the subject is in the same condition as at the start of the first, so that nothing of the first treatment is carried into the estimate of the second. The instrument for securing that assumption is the washout, and the rule for setting it is a rule about the drug's terminal half-life — a gap of at least five of them, sometimes seven, occasionally more where a metabolite is slower. Statisticians have been sharp for decades about what happens when the assumption fails and about the practice of testing for carryover and then acting on the test, which Senn showed inflates the error it claims to control. The remedy they urge is a washout long enough that the question does not arise.

what a half-life actually synchronises

Five half-lives is a statement about one quantity: the concentration of the parent compound in plasma, which is the quantity the washout rule was derived from and the quantity a sample can report. It is not a statement about the state of the subject. A body that has received a drug for two weeks differs from one that has not in ways that have nothing to do with how much compound remains in it — the amount of a metabolising enzyme, which induction raises and which falls back only as the enzyme protein turns over; the amount of a transporter; the fraction of an enzyme irreversibly inactivated, which recovers only by resynthesis; what is held in a deep tissue compartment and released slowly; the composition of the gut flora; the density of a receptor; in the case of a biologic, whether antibodies to it have formed. Each of these is a component of the state that determines what the next dose will do, and none of them is visible in a plasma concentration that has fallen to zero.

the hidden states run on clocks that are not the drug's

The decisive point is that these components do not relax at the drug's rate. Rifampicin leaves the plasma in a few hours and the induction of CYP3A4 it causes takes one to two weeks to disappear, because what is decaying is a population of enzyme molecules with its own turnover, not the inducer. A mechanism-based inactivator may be cleared overnight while the enzyme it destroyed returns only as new protein is made. Carbamazepine induces its own metabolism over weeks. So the washout rule is calibrated to a clock that in these cases governs none of the state that matters, and a washout that is entirely adequate for the concentration may leave the subject's metabolic capacity where the first period put it. The field is not unaware of this — interaction guidance gives separate advice for inducers — but the advice is a set of exceptions appended to a rule, not a different way of choosing the gap.

and the pre-period inputs that already exist, used for something else

Trials already put non-empty inputs in front of a measurement period. There are placebo run-ins, lead-in periods on background therapy, standardised diets, stabilisation periods, and fixed loading schedules. But these are used to screen for adherence, to establish a baseline, or to bring a clinical variable into range; they are not designed as sequences chosen to make the subject's unobserved condition independent of what came before, and their length is not set by asking what would be sufficient for that. The one input the field does treat as a reset is the empty one, and it is treated as a reset because the observable it governs converges, not because anything was shown about the rest.

III.The Bridge

no automata paper mentions a washout

The halls do not meet. Synchronising words belong to automata theory, to the testing of finite-state machines and to a small literature on parts orienters; washout belongs to clinical pharmacology and to the statistics of crossover designs. No review spans them, and the observation that both concern systems with memory predicts nothing and is not the bridge. Three neighbours must be set aside. It is not the discharge order of an earlier specimen, which concerns the age of what leaves a store and the geometry of withdrawal; there the question is what comes out, here it is whether what remains can be made the same for everyone. It is not the hiatus of another, about rates averaged over dormant intervals. And it is not identifiability, the established pharmacometric question of whether a state can be inferred from a record, because the point here is the opposite one: the state is not inferred at all. What the halls share is a single structural move. Where later behaviour depends on an internal condition that is not observed, one may either estimate the condition or apply an input that carries every condition the system might be in onto one — and the second route needs no estimate, no sensor, and no knowledge of which condition obtained.

X: a state can be removed by an input instead of by a measurement or a wait

Call it open-loop reset. Where a system's response depends on an internal state that is not observed, the state can be discarded rather than determined, by applying an input under which every state the system might be in arrives at the same state. Four things follow. Whether such an input exists is decided by the transition structure and pairwise — every two admissible states must be capable of being brought together — and not by any of the system's rates. The state reached is not privileged and need not be the empty or initial one; it is wherever the chosen input funnels everything. Doing nothing is only one candidate input, synchronising by accident if at all, so an active input may reset a system that waiting does not, and may do it sooner. And where some distinction between states is preserved by every available input, no input resets the system, and the failure is a property of the system rather than a shortfall of duration. Read into the clinical hall, each has a consequence the hall does not state. The washout is a candidate reset input whose synchronising property has been argued for one observable and assumed for the rest, and its adequacy is a question about the transitions of the hidden components, not about the drug's clearance. Because the reset state need not be the untreated one, a conditioning input that drives every subject to a shared ceiling — induction taken to its plateau, a depot filled, a binding site saturated — may merge subjects that decay to an individual floor would not, and may merge them in less time, since a ceiling is common where a baseline is personal. The length of an adequate input is therefore a parameter of the design in its own right, not recoverable from any half-life, and the conventional gap is an unexamined guess at it. And where the first treatment leaves something no input can undo, the two periods cannot be made comparable at all, which is a finding about the admissibility of the design rather than a reason to lengthen the gap.

IV.The Question That Becomes Possible

neither discipline would ask this alone

The clinical hall asks how long a washout must be and answers with a multiple of a half-life, because the half-life is what it can measure and what its rule was built on. The automata hall has no subjects, but it settled long ago that resetting an unobserved system is a question about which inputs merge which states, that the empty input has no special claim, and that some systems admit no reset at all. Neither hall asks whether the gap before a crossover period is best chosen as a designed input rather than a duration, nor whether the shortest input that makes a subject's response independent of prior assignment bears any relation to the quantity by which the gap is currently set.

the open-loop reset hypothesis

For a given prior treatment and a given probe, is there a pre-period input — a specified schedule, not necessarily the empty one — after which the probe's exposure is independent of which prior treatment the subject received; is the shortest such input's length predicted by the turnover of the hidden components rather than by the drug's terminal half-life, so that a saturating conditioning schedule can achieve independence in less time than the conventional abstinence; and are there prior treatments for which no such input exists, so that the crossover cannot be reset in principle?

why it would matter

Because the washout is a substantial part of the cost and the duration of crossover and interaction studies, and the rule that sets it is derived from a quantity that in the hysteretic cases governs nothing that matters. If a designed input resets sooner, studies shorten and subjects are retained. If the required abstinence turns out to be governed by enzyme and transporter turnover rather than by clearance, the rule should be restated accordingly, and drugs with short half-lives and long hidden clocks — the inducers above all — should stop being allocated short gaps by a rule that was never about them. And if for some compounds no input resets, a crossover should not be run at all and a parallel design chosen at the outset, which is cheaper to decide in advance than to discover in the analysis.

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 a study that can be built from instruments already in routine use: a strong inducer with a short plasma half-life, a validated probe for the affected pathway, and an endogenous marker of the pathway's activity so that the hidden component is read directly rather than inferred.

the honest status

Note the narrowness. The claim concerns crossover and drug-interaction designs in which the first treatment is known to perturb a durable component — induction, inactivation, transporter expression, a deep compartment — and not crossover designs in general, where the washout rule works because there is nothing slower than the drug. It does not claim that washouts are generally too short or too long; it claims that their length is set by the wrong quantity in the hysteretic cases, and that the empty input has no structural claim to be the reset. It does not claim that an active conditioning schedule is safe or acceptable, only that it is a candidate that the current framing does not generate. It is worth saying what does not transfer. An automaton has finitely many states and exact merging; a person does not, and a reset in the clinical hall can only ever be a merging within a tolerance that someone must choose, which means the sharpest results of the formal hall — the pairwise criterion, the reset threshold, the hardness of finding the shortest word — arrive as intuition and not as theorems. The whole thing rests on a comparison that a single study can make: whether a designed input reaches independence from prior assignment sooner than waiting does. If it does not, this is an elaborate way of saying that a washout should be long enough, and should be said so. Until that comparison is made, 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.