Stalls
Where an exact descent stops short of the bottom,
including where no measurement tie sits anywhere near it: which stall
species exist, what cures each, and how far an escape is when one
exists. What decides whether a lowered patience improves anything at
all is the geometry underneath them —
nesting.
The thing being descended is a commitment policy over a cover of
intervals (descent in reader space), and what
follows states what it needs of one. A stream is the
continued-fraction expansion of a real number, arriving digit by digit,
and a map is a function applied to it. After n digits the
value is pinned to a cylinder, and what the reader is certain of
about the output is the image interval, that cylinder's picture
under the map; these nest and shrink as digits arrive. A row is
one (map, stream) pair.
What a reader commits to are cells drawn from one fixed
family, the cover: the tree cells, the Stern–Brocot cells,
each the set of positive reals whose expansion opens with a given run of
digits, and the straddle chains standing at their
vertices, a cell's vertex being the mediant of its two
endpoints — the fraction whose numerator and denominator are their sums.
Both kinds are graded by rank, which steps by one along every
child relation (the redundant cover).
A commitment is a ratchet: never undone, and permanently sound,
since a cell containing the current image contains every later one. What
a policy chooses is the route past a vertex — the tree child or
the chain — and how long to wait before committing: four coordinates, a
route preference at each of the two cell kinds and a patience for
each. A kind at patience p commits a candidate only once it has
contained the image for p+1 consecutive steps; patience 0 is
greedy, committing on sight, and patience ∞ refuses that kind
outright. A kind's reference is the image interval its patience
makes a candidate contain, the one from p steps back at patience
p, so a policy with equal patience on both kinds reads one
reference at both.
Two exact rulers price a policy over a counted window,
a fixed span of input steps, the same for every policy compared; its
length is the horizon. The deficit at a step is the scale
the reader is short — writing a cell's scale for ln(1/length), the image
interval's scale less the committed cell's — and the clock ruler
is that same lag counted in whole steps. Both compare by
cross-multiplying big integers, so no float enters any decision.
Descent is the matching finite-graph notion: a move is taken only
when it strictly lowers the loss — the descent fixes one of the two
rulers and runs against it alone — and a stall is a policy with
no strictly better neighbour. Descent runs on the behavioural
quotient — policies identified when their committed-cell traces over
the window agree, since two policies with identical traces cannot be
ordered by any loss reading that window; the resulting classes
are what descent moves between. Resources, where they are in force,
enter as settings rather than as coordinates: a rank budget
B of cover-rank units per input step, and a bank of
capacity W holding what a step does not spend, against which the
policy gains a fifth coordinate for how much of the bank a single step
may draw. What they are spent against is the row's demand, the
rank per input step that row asks for. A setting with neither in force
is unresourced, and there the four coordinates are all of it.
A stall is a fact about a move set before it is a fact about a
landscape, and the move set here is the unit step. Neighbouring
policies differ by one notch in one coordinate, or by both route
preferences flipped together, and nothing below is a claim about a
reader who may do more. Let a reader move a patience two notches at a
time and the corpus's stall count over its ten landscapes — a
landscape being one loss surface, a row and a setting fixed — falls
from 10 to 1; at three notches it is zero — not at the ten specimens only but
over every one of the 601 off-bottom classes those landscapes hold — and
a reader free to set patience outright adds nothing beyond that
(explore_move_set.py).
So every stall named on this page is a reader who may change patience
ONLY one notch at a time, and every distance below is counted in those
notches.
The stall
census pattern
Every stall found at the scanned scopes, with the cure the
species takes. The
cures name three instruments: a refinement of the comparator, the
quotient the descent already runs on, and lookahead — accepting
a short sequence of moves whose net effect improves where no single
move does.
| species | what stalls | cure |
| the plateau (space-born) |
a blind region of unbounded width — total refusal, where
every neighbour prices infinite and the comparator has no
signal |
a SIGNAL: lexicographic refinement by the loss's own finite
part; radius-free, so it scales with the plateau |
| the degeneracy (space-born) |
a flat in the quotient — behaviourally identical policies
tied together, manufactured by resource abundance (a full
bank smooths two patiences into one trace) |
the QUOTIENT: descend counted-window behaviours rather than
coordinates, on which the species cannot form |
| the ruler disagreement (loss-born) |
two exact losses over one window ordering an edge oppositely
— the coarse ruler dams the fine one's gradient, so the
deficit's own exit is clock-blocked and the class stalls under
the clock order alone |
a two-move lookahead route, MEASURED; or descending the finer
ruler across the coarse ruler's plateaus |
| the burst trap (resource-born) |
a strict local minimum with no tie in it — under a binding
budget a demand burst makes one patience pay a little early to
win a lot late, a paired trade every single move breaks |
a two-move lookahead route, MEASURED, which at one of the six
is the quotient's walk and costs a reader three moves; neither
refining the ruler nor quotienting the space touches any of
them |
| the horizon-cut stall (window-born) |
a strict local minimum with no budget anywhere in it — the
counted window ends inside a trade's transient stretch (two
traces part, the early loser collects late, they re-merge),
freezing the late win in place at the cut |
a two-move lookahead route, MEASURED — or the count itself:
the census specimens dissolve once the window outgrows the
trade |
The three lookahead rows say MEASURED because two-move lookahead
also has a proved cure attached to it, and that cure's
hypothesis holds at none of the ten finite-loss stalls these three
rows hold between them. What reaches them is
an existence claim carrying no such hypothesis, plus a route someone
counted: the escape radius below.
One more species once held a seat and lost it: coordination
blindness,
the trap a binding budget seemed to open under the unthrottled
optimum itself, cured by a diagonal move — on the quotient the
trapped class merges with its behavioural twin one route flip from
the bottom, so the diagonal cure was two ordinary moves through a
twin-flat and the space, rightly quotiented, never had the
disease.
What the survivors share was once an open conjecture — every stall
is a tie artifact, a measurement tie being
neighbouring policies whose exact losses read equal, so that the
coincidence dissolves under a refined ruler or the quotient — and it
settles as a dichotomy. The spine
is the policies whose two kinds share one reference: equal patience
on both, unresourced, every patience shorter than the counted window's
start. On the spine the conjecture is a theorem: the cover cells
containing an interval have an inclusion minimum, that minimum is
monotone in its interval, so committed cells nest pointwise along the
patience diagonal and lexicographic deficit descent on the behavioural
quotient stalls nowhere off the bottom, which is greedy — at any nested
strictly shrinking stream and any horizon. Off the spine the counterexamples are
genuine hills with no measurement tie anywhere in their neighbourhoods,
so the instrument pair does not suffice in general: the burst trap
under a binding budget, and the horizon-cut stall with no budget at
all — the latter three specimens: two found by census in plain
period-2 digit streams under the
squaring map at horizons 11 and 12, one built from three digit bursts
at horizon 16, where the same census over 177 generic streams finds
none. The census-found pair dissolves once the window outgrows its
trade, so
a stall-free landscape is a property of the battery it was drawn from, generic
streams and long counts and slow maps, never of the unresourced
space; what says
where nesting still holds there is
the decision lemma.
Scope. A census over the settings scanned plus
a designed battery, not a classification of exact reader spaces, and
every row of it read at the unit patience step. The
two space-born species and their cures are rules at the scopes
the funnel and
the resource laws state;
the ruler disagreement is a single specimen, at
(B, W) = (2, 0), and whether it is knife-edge or a family
is open; the burst trap is six designed specimens (spike digit streams
under the doubling map at budget 2); the horizon-cut stall is three
specimens in exact arithmetic, one hand-verified at each horizon; the
spine theorem is proved for this cover and move set.
verifiers:
explore_bootstrap_cures.py,
explore_scale_clock.py,
explore_stall_tie.py,
explore_stall_assembly.py,
explore_move_set.py
The slow-map refusal
and the one-cell wall observation
Of the battery's three legs — generic streams, long counts, slow
maps — the slow map is the one an adversary
without a budget cannot buy back — the burst trap already buys the
doubling map back with one. Widen the digit range to 256 and take
horizons 9–20: the identity and doubling maps refuse across some 2,800
distinct censused landscapes and 70,000 adversarial search
evaluations, under the same detector that reproduces every
squaring-map specimen. A large digit buys one step of shrink under any
map, but only the squaring map compounds it along the window, so the
best slow-map stall margin — the signed gap, in the scale's
log units, by which a class's best neighbour fails to improve on it,
positive exactly at a stall — approaches zero inversely in the digit
cap
and never crosses. Behind the approach stands a wall. The two frontier
records — doubling at digit cap 64, identity at 256 — each equal
−ln(1 + 1/R) exactly: the price of one further
step down the cover, from a cell to a child of it, at the binding
cell's own lopsidedness R, the ratio of its endpoint
denominators, each record's R just under the bound its digit
cap
sets. The doubling
record traces to its cells — a single mediant step, the other counted
step cancelling — and the identity record is a single-counted-step
ratio by construction. That one-cell
wall is
proved for any reader class holding a one-move neighbour whose
committed cells sit nested-or-equal at every counted step. That
domination condition holds for all 9,401 classes measured under
the identity map, without exception; the doubling map
dents it in 632 classes, of which 540 keep a neighbour shorter
step for step without nesting and 92 still lose only on aggregate; the
squaring map's stalls break it outright — every neighbour of a stall
concedes somewhere inside the window.
Scope. The refusal is an adversarial search at
the stated digit caps and horizons, so it is an observation and not a
law about slow maps. The one-cell wall is proved for this cover,
conditional on the domination condition; that condition is exhaustive
at scope under the identity map (9,401 off-bottom classes at horizons
9, 10, 12 and 16, zero exceptions) and measured, not derived, under the
others. Toy scale.
verifiers:
explore_stall_maprate.py,
explore_stall_domination.py
The escape
radius rule
A stall's escape radius is the length of the shortest
improving move sequence out of it, and the answer depends on who
counts. Through the cure graph — whose edge exists as soon as
some policy in a class carries the move — all ten finite-loss
stalls above have radius exactly 2. Counted as one reader's own
moves, nine of the ten agree and the tenth needs three. At the burst
trap on the spiked stream under doubling, the one class supporting a
two-edge walk out of the stall is entered at three policies and left
at two others, and those sets are disjoint, so no reader takes the
walk in two — nor does changing identity inside that class help: one
change plus two moves is three, and three moves reach a better class
outright.
Both figures are measured at these ten rather than inherited from
the proved two-move cure, because that cure does not reach them.
Lowering both patiences by one is an exact delay and improves
strictly wherever it changes the class, on any map at once
(the nesting
neighbour), but its hypothesis is a policy with both patiences
finite and positive whose counted window still refines, and no stall
in the table above holds one. Every
one of the ten has a patience pinned to the current image — a
patience of 0, whose reference is the image itself — and no
species straddles the split: the three horizon-cut stalls sit pinned
at the chain, the ruler disagreement and all six burst traps at the
tree. So the proved cure is a theorem about the classes that do NOT
stall. What reaches the stalls carries no patience hypothesis and no
map fence, and it is an existence claim rather than a route — that a
strictly better class exists from every class off the bottom at
finite loss, the floor
theorem, proved at the three stalls carrying no budget and
measured at the seven that do. How far away it is, is what the radius
adds.
Scope. Measurements over the stalls the census
holds, exhaustive at those ten under each reading of the count and at
the unit patience step, the pinning anatomy likewise; a rule at that
scope and not a theorem. The route behind each radius-2 figure, and
the nesting that picks it out, are
the reference geometry's. Toy
scale.
verifiers:
explore_charged_radius.py,
explore_scale_clock.py
Nesting
What
decides whether a lowered patience improves anything at all is one
question: do the two readers' committed cells nest? Where no resource
cap is in force, a greedy reader commits the floor of the
current image — the inclusion minimum of the cover cells containing it,
a closed form that reads the image and runs no policy — and that floor
sits inside every cell any policy commits, which is why a strictly
better class exists from every class off the bottom at finite loss,
with no map fence and no patience hypothesis on it. Which single moves
nest is measured, and the answer is narrow: patience-down moves and
nothing else. One corner of it is proved — a run preferring the straddle
chain at both kinds, under the identity map — and that proof's
hypothesis turns out to be two inequalities on the indices of the
references a run reads at a single step, the step where it can leave a
chain for the cell the chain sits in. Off the spine, nesting dies only
where the commit loop had a route preference to consult, though not at
every such place:
nesting.