Menu collisions
The one door to robust forgetting is prime recycling.
How wide that door is turns out to be a question about unique
factorization — when two different pairs of menus weigh routes
identically at every temperature — and the answer runs from a collision
so generic it is almost never a failure of factorization at all, down to
a bound on how far a genuine failure can escape a single variable.
The systems are the small grown worlds of
the amnesia certificate. A state is a
positive integer, a move multiplies it by a member of that state's
menu — the moves the state admits — and a route is the
sequence of moves from the start state 1, its product being the state it
ends at. A world is read at a temperature β: a move
m is taken with probability proportional to
m−β, normalized by the state's
normalizer, the sum of m−β over its menu.
A route's probability is the product of its moves' weights divided by the
product of the normalizers it passed through — and since the moves'
product is the endpoint, two routes reaching one endpoint differ only in
those interior normalizers. A world free to reuse a prime it
already holds recycles; one every move of which is coprime to the
state it acts on cannot.
A certificate of
forgetting is robust when a fiber — the routes that reach
one endpoint — is equiprobable at every temperature at once rather than
at a single tuned value, and a coprime
world can never buy that: certified forgetting costs prime recycling.
What is left over is how much room recycling gives, and it is a question
about the normalizers alone. Two routes are flat at every β
exactly when their interior-normalizer PRODUCTS agree as functions of
β; a structural witness for it would be the two MULTISETS
agreeing. So the door's width is the gap between those two conditions —
how far equal products can fall short of equal factors. Call two route
weightings that agree identically in β a collision, and the
question is which collisions are more than bookkeeping.
The normalizer
semiring is not factorial
observation
A menu's normalizer, read as a function of β rather than as a
number, is a Dirichlet polynomial: the sum of
m−β over the menu's members. A route's weight
is then set by a product of such polynomials, and the collisions above
are the pairs of products that agree identically in β. They are
generic rather
than exotic: 4849 colliding pairs, falling into 4487 classes by the
product they share. Measured directly, one scaling family accounts for
4369 of them: writing ZA for the polynomial of a menu
A, and cA for A with every member multiplied by
c, they are the instances of
ZA·ZcB =
ZcA·ZB; another is {2}, {3, 6}
against {3}, {2, 4}, both products
6−β + 12−β. What a collision count
does NOT measure is non-unique factorization, and reading it that way
is a non sequitur: a collision is compatible with unique factorization —
2·6 = 3·4 in the integers — being only a regrouping of one multiset of
factors into two blocks. The scaling family is exactly that, the moved
factor a single monomial, and so is the two-prime example just given.
Where the failure is genuine it is the product's. Write x for
p−β at each prime p and a menu becomes a
polynomial with every coefficient 0 or 1; a factor carrying a negative
coefficient cannot itself stand in a nonnegative semiring, so the
remaining factors pair around it two ways and the product has two
different irreducible factorizations; and since the factors of a
product are those of its two menus together, that negative factor
belongs to one menu or the other, which makes carrying one a property
of a SINGLE menu and never of the pairing. Call a menu that carries one
a seed. At the census's scope that is almost nothing: 4839 of the
4849 pairs sit in a product that factors uniquely and are regroupings
outright, and the other 10 sit in the 6 products that factor
non-uniquely, each in exactly two ways, every one of them generated by
one of the three seeds the scope holds, {2, 16}, {4, 32} and
{2, 8, 32}. The semiring is genuinely not factorial, and the failure is
not confined to a single prime: reading two of them at once as x
and y, {8, 27}·{4, 6, 9} and {2, 3}·{16, 36, 81} span the same six products,
and the negative-coefficient factor the others pair around is
x2 − xy + y2 — the
one-variable case's x2 − x + 1 homogenized.
WHEN factorization fails is graded by the two menus' sizes,
their numbers of members: a product of menus of sizes s1 and
s2 has s1s2
terms counted with multiplicity, and factorization in this semiring is
classified by number of terms (van de Woestijne 2011), so it is
classified by SIZE PAIR — at any bound on the elements, and by theorem
rather than by census. Products at sizes (2, 2), (2, 4) and (3, 3) — 4,
8 and 9 terms — factor uniquely, several variables included. The one
failing pair with both menus of size at most 3 is (2, 3), at 6 terms,
and there non-uniqueness is
forced onto a line: the exponent vectors of the shared product's terms
lie in an arithmetic progression along a single direction, so the
mechanism is an image of the one-variable one — a monomial
substitution where that direction is nonnegative, a homogenization
where it is mixed, and there is no third kind at that size. The two-prime pair above is exactly that
case, a size-2 menu against a size-3 one. Confirmed on this corpus's
own menus, every one of size 2 and 3 with elements in {2..32} and of
size 4 with elements in {2..24}: not one product at 4, 8 or 9 terms
factors non-uniquely, and all 39 found at 6 terms are collinear. The
one-variable reading does not survive size 5 as a statement about
whole products, though. At the
size pair (2, 5), {2, 54}·{2, 6, 10, 30, 90} and
{2, 6}·{2, 10, 54, 90, 810} share the ten-element product
{4, 12, 20, 60, 108, 180, 324, 540, 1620, 4860}, which factors two
ways with exponent vectors that lie on no line at all — the image of
no one-variable polynomial under either operation. What hid it is menu
SIZE: the pair needs a menu of five members and no census above goes
past four. The elements need not be large for it — read at the
cheapest primes the same family is {2, 16}·{2, 4, 6, 12, 24} against
{2, 4}·{2, 6, 16, 24, 96}, largest member 96 against the first
reading's 810. Taken whole, then, that product is past one variable;
taken one edge at a time it is not, and grading that difference is
the block below.
What limits the door is therefore not the
algebra alone but
REALIZATION — whether menus carrying those polynomials assemble into
an actual world — and realization's teeth are exactly one degeneracy:
the unrealized collisions coincide precisely with the all-singleton
ones, 44 of 44, where the four singleton menus of such a pair force
equal move products, hence equal leading moves, hence a conflict at the
first state.
Everything with a non-singleton menu realizes, 4805 of 4849.
Scope. Exhaustive at the collision census's own
scope — menus drawn from {2..16} ∪ {32} of size at most 3, the
realization search's opening moves scanned to scale 8. The singleton
direction is proved, whichever order the menus are taken in along the
routes. The size-pair account carries the tiers of its parts: the
classification by number of terms and the uniqueness at 4, 8 and 9 are
the cited prior work, holding at any element bound and in several
variables; the collinearity at 6 terms is proved here from that
classification together with its own reduction of the several-variable
case to the one-variable one; the sweep is exhaustive over the box
named there. Both multi-prime specimens are hand constructions rather
than census finds: the six-product pair sits outside the collision
census's scope and inside the sweep's box, and the size-5 pair outside
both, each verified exactly on the menus given; nothing is asserted
about the size pair (2, 5) at
large. The same
scan cross-checks, from outside its proof, the coprime half of
the coprimality
dichotomy
the door rests on: zero coprime realizations across all 4849 collisions, against
4805 realizations for worlds free to recycle.
verifiers:
explore_rogue_world.py,
explore_menu_factorization.py,
explore_menu_reach.py
Up to ten terms,
non-uniqueness is one-variable on a face
rule
What the size-5 escape ends is the one-variable reading of a WHOLE
product. A sharper one sits underneath it, and that one survives
everywhere either this corpus or the literature reaches. Collect the
exponent vectors of a shared product — one per term, as above — and
take their convex hull. Looking at that hull from a chosen direction
picks out a face: the terms whose vectors are extreme that way,
a corner or an edge or something of higher dimension. The same look
applied to a factor keeps its own extreme terms, its initial
form, and initial forms multiply exactly as the factors they come
from do. So a collision is never confined to the product it happens
on: it restricts to EVERY face of the hull, each face carrying a
collision of its own in lower dimension. The support-only half of that
is classical and settled — the hull of a product is the Minkowski sum
of the factors' hulls (Ostrowski), and enumerating those
decompositions is a published factorization instrument (Gao and Lauder
2001) — and it is strictly weaker than the collision: {0, 1} + {0, 1,
3} and {0, 1, 2} + {0, 2} share a sumset and a term count with
different convolutions. The productive half is the other one.
Grade a colliding pair by that restriction. The descent
dimension δ is the smallest dimension of a face on which the two
factorizations already differ — each initial form stripped first of
any single-term factor, since a one-term factor is an atom that may
sit beside either factorization and moving it is not a difference of
mechanism. Small δ says the mechanism is INHERITED from a face rather
than new at full dimension. δ ≥ 1 always: at a corner every initial
form is a single term — a point decomposes only into points — and its
coefficient divides the 1 the product carries there, so stripping
leaves nothing on either side, a factor that has become 1 being no
factor at all. No corner ever grades anything. The size-(2, 5) escape
reads δ = 1, and what sits on the edge it descends to is (1 +
v3)(1 + v + v2) = (1 +
v)(1 + v2 + v4) — the
six-term identity itself, the case that escape was supposed to have
left behind. Up to ten terms that is the rule: δ ≤ 1 throughout, so
every mechanism there is one-variable on some edge. The term count
does the work. Products of 4, 8 and 9 terms factor uniquely and every
other count up to 10 is trivial or prime, save 6 and 10; what those
two leave is the six-term form, three sporadic ten-term identities and
two ten-term two-parameter families. The six-term form and the three
sporadics spend one of their two free exponents on a translation, so
their exponent vectors are collinear however they are lifted, and the
families descend because the independence of their two parameters
supplies a direction cutting an edge that carries the six-term
identity. They are not all the SAME mechanism, though, and the rule
does not say they are: those three sporadic identities regroup the
three cyclotomic factors of 1 + x + ⋯ + x9 —
one of them bare, two with a further cyclotomic factor riding along —
and are not images of the six-term one. Whether any
collision anywhere reaches δ ≥ 2 — full dimension, inherited from no
face — is the question those counts locate, and it is
answered below. No count below twelve can
produce one: 11 is prime
and so factors uniquely, which makes 12 the first COMPOSITE count past
the classification. At twelve terms the menus can pair only as
(2, 6) or (3, 4), and the qualifier is load-bearing: a menu's
polynomial has every coefficient 0 or 1, so a factor carrying a
coefficient above 1 is outside the menu frame altogether — the
literature's own twelve-term example is such a factor — and those two
size pairs exhaust the menu question rather than the polynomial one.
The twelve-term instance this corollary names reads δ = 1 like
everything below it.
The (3, 4) half has since been swept, and a sweep is a search over a
box where everything above is a theorem. The box is this corpus's own
menus — sizes 2 and 3 with elements in {2..32}, size 4 in {2..24} —
walked wherever either side is a seed, which is the only way a pair can
collide at all: 85,253 pairs. It returns 336 non-unique products, and
the frame cuts them before anything is graded. A three-member menu
against a four-member one gives twelve terms exactly when no two
products coincide, and 265 of the 336 come out at ten terms or eight —
a coefficient above 1, outside the frame just named. The in-frame
population is 71, every one reads δ = 1, and the dimension column is
what makes that a measurement rather than an artifact: all 71 sit at
product dimension 2, so δ was free to read 2 at each of them and read 1
instead. The two products where δ ≤ dim forces the answer are both
outside the frame.
Two readings sharpen that null result, and only the second leaves
the box. The first says the box was not what constrained it, and it
reads where the collisions are NOT. A pair's reachable dimension is the
rank of the directions its two menus span and needs no factorization to
compute; 96.5% of the walked box reaches dimension 3 or 4; and not one
of those 82,249 pairs collides. Where dimension is plentiful there are
no collisions at all — and it is not the seed holding the products
down, since
split by which half carries the seed both populations reach dimension 3
or 4 at better than 95%, the partner being unbounded. What collapses
the dimension is the collision. Negativity cannot cancel across
disjoint variables, so a second factorization must set the seed's
negative factor beside a factor sharing its variables; where the seed
is collinear — 64 of the 71 — the partner's core, its polynomial
with its monomial content divided out, supplies one on the seed's own
line at every one of them; and where that partner is not itself a seed
— 48 of the 64 — the term count prices its core as two 0/1 binomials,
one of them spent on that absorption and one direction left over.
Dimension 2 by the law rather than by the box. Of the other 23, the 16
whose partner is itself a seed are one family, two-dimensional by its
shape — one line and one direction off it — and mixing rather than
sharing a factor
(the mixing theorem);
7 are measured and unexplained. The graded instance itself, {2, 8, 32}
against {2, 4, 16, 32}, sits outside the box and reads dimension 1 for
a reason the seeds give: both halves are seeds, each the unique
collinear one of its size in {2..32}, and their product is the only
place either box holds where a collinear seed meets a collinear seed.
All of that is still inside the boxes. The second reading is what
carries out of them, being proved rather than searched, and it is a
fact about seeds alone: a seed of n members has core dimension
at most n − 2, at most n − 3 from five members on, and at
most ⌊n/2⌋ at every size
(the half-size
ceiling). At four members that is 2, so the (3, 4) seed side is
capped at every bound and not merely inside this box, while the same
law puts a dimension-3 seed at six members or more, and most six-member
seeds reach it. Whatever room the one surviving corridor has is
therefore in the (2, 6) half, past every menu size swept here.
Scope. The face restriction and the polytope
identity are the cited prior work; the descent dimension is a
definition and δ ≥ 1 is proved from it. δ = 1 for the size-(2, 5)
escape is exact on the one pair given. The bound up to ten terms and
the twelve-term corollary are proved from the term-count
classification cited in the block above, read in full, together with
the grading of each family it leaves standing. Those are theorems and
need no box; the twelve-term instance named with the corollary is a
single graded object rather than a search. Everything from the (3, 4)
sweep down is an observation over the box named there and carries its
tier: exhaustive inside that box, silent outside it, and silent about
the pairs the box excludes — among them the graded instance itself,
whose four-member half reaches 32. The seed-dimension bound quoted at
the end is proved and is
the seeds page's; the
collinear-seed reading of the graded instance is its census's. Each δ
is read off a
sampled set of directions, which can only overstate it by missing a
face that differs, so the readings are upper bounds — the safe side for
a search whose target is δ ≥ 2.
verifiers:
explore_menu_faces.py,
explore_descent_hunt.py,
explore_seed_confine.py
The door's width, then. Collisions are everywhere and almost all of
them are bookkeeping — one multiset of factors banked two ways. The
genuine failures are graded by menu size and settled outright at the
small pairs. Every mechanism proved up to ten terms, and every one the
(3, 4) sweep at twelve found, is a one-variable identity riding on some
face of the product. The other pairing at twelve is not. The seeds
said where to look — a seed of dimension 3 needs six members at least,
by the half-size
ceiling, and most six-member seeds have it — and walking the (2, 6)
pairing over {2..24} turns a full-dimension collision up.
Full dimension occurs,
and in the widest box walked it occurs once
observation
Twelve terms, every coefficient 0 or 1, four 0/1 factors, and every
proper face of its Newton polygon — corner and edge alike — induces the
same factors on both sides. Only the whole polygon tells the two
factorizations apart, so δ = 2 and the mechanism is inherited from no
face. One witness settles an existence question, and this one is checked
without a sampled search: for a two-dimensional polygon the faces can be
enumerated outright, and they were. So the reading that survives the
whole classification is false at the first term count where it could
fail, and the classification of these collisions is structurally
incomplete rather than merely unfinished.
Whether that failure is a curiosity or the rule is a second question,
and the same pairing hunted through a wider box is what bears on it.
Six-member menus with elements up to 32 number 736,281, against the
100,947 up to 24 that turned the witness up — 7.3 times the population,
on a seed population nearly four times larger. That box carries 41
in-frame collisions, and they are 22 distinct pairs of cores:
multiplying every member of a menu by one fixed integer changes its
polynomial by a monomial and its core not at all, so one pair of cores
wears several
menu suits and counting at the menu counts one mechanism repeatedly. Of
the 22, 21 read δ = 1 and exactly one reads δ = 2 — the identity above,
wearing ten of those 41 suits. So the wider box turns up no
full-dimension collision but the one already found: in the population
walked it is a RARITY rather than the generic behaviour past the
classification, and the incompleteness the witness opens is real and
narrow.
The same box reads where the collisions are NOT, and it is the
reading the (3, 4) half could not make: there the half-size ceiling caps
a seed's core at dimension 2, so no seed could carry a third dimension
into a pair by itself. At six members the cap is 3, and the wide box's
seeds mostly reach it — 177 of its 203 six-member seeds have core
dimension 3, against 26 at 2. A pair's
reachable dimension contains its seed's core dimension and needs no
factorization, so every one of the 177 × 465 = 82,305 pairs joining
those seeds to a two-member menu reaches dimension 3 or more, and not
one of them carries an in-frame collision: all 41 in-frame collisions of
the wide box sit at product dimension 2, as the narrow box's 16 did. The
room was there, seven times over, and went unused, which makes the
confinement the surviving structure rather than the suspected one. And
the witness's own seed has core dimension 2, so full dimension is not
bought by the dimension a six-member seed can add — whatever selects it,
it is not that.
Scope. The witness is exact and settles
existence at any bound: the identity is checked as written, and the
proper faces of its polygon are enumerated outright rather than sampled.
That is what a δ of 2 needs and a δ of 1 does not — a sampled set of
directions can only overstate δ, by missing a face that differs, so it is
the safe side for reporting 1 and the unsafe side for reporting 2. The
rarity and the confinement are observations over one box and are
nothing more. A
box bounds the SIZE of a menu's members and nothing the mathematics
names, so alone in {2..32} can never mean alone — a wider box is a
larger sample of an unbounded population, and every count here is a
statement about the sample. The confinement is read over the in-frame
population only: the wide box's products outside the frame — 4 of its
26 objects, where the narrow box had none — have no dimension reading
here, so whether it extends to them is untested rather than settled.
The menu frame is not the whole of the
question either: a 0/1 polynomial can factor through a factor that is
not 0/1, as the literature's own twelve-term example does, and nothing
here has graded one.
verifiers:
explore_descent26.py,
explore_descent26_wide.py
The mechanisms
What the
second factorization IS is settled three ways. Where the two
factorizations share a factor, δ is derived rather than read: a face
separates them exactly when the negative factor's initial form has more
than one term and the two factors beside it do not read the same stripped
form there — at product dimension 2, an edge of the negative factor's
polygon parallel to an edge of one of theirs — a criterion proved both
ways, and the full-dimension witness is where that parallelism fails at
every edge. Where they share no factor, a pair with a three-term factor
on either side is one explicit family or one-variable, a theorem at
twelve terms, and every such pair either box holds is that family. And at a seed
whose core is two irreducible factors, whether a pair collides at all is
one sign scan on the product of two of them:
Collision mechanisms.
The seeds
Which menus are seeds has a closed
form at two and three members — a pair whose members, divided by their
gcd, are d-th powers for a d with an odd prime factor, and a
collinear trinomial whose exponents {0, a, b} meet all
three residues modulo 3 with {a, b} not {1, 2} — and a
seed's
core has dimension at most HALF the menu's size, a theorem at every
size that caps the seed side of every pairing above and puts the first
dimension-3 seed at six members: Menu
seeds.