Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
41 changes: 41 additions & 0 deletions research/extremal-core-word-obstruction/CREDITS.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,41 @@
# Sources and credit

Ali Romman prepared this proof and verifier with substantial OpenAI
Codex assistance. New work is Apache-2.0. Independent human and
formal-kernel review are unclaimed.

- CubeJerry's [PR #6782](https://github.com/CrocSwap/integer-mult-bounds/pull/6782),
head `685457252b51b9c7932ece363365bdf0b12c9c31`, supplies the
five-dimensional restricted bridge-cancellation theorem and its
associated-graded, star-quotient and chronology arguments. Read the
[exact word proof](https://github.com/CubeJerry/integer-mult-bounds/blob/685457252b51b9c7932ece363365bdf0b12c9c31/research/bit-three-relation-core/FIVE-CORE-WORD.md).
Its notice states that the work was prepared for @CubeJerry with
substantial OpenAI assistance. It credits the retained relation and
augmentation framework; that lineage is not relabelled as new here.
- The full Laurent recurrence is adapted from the same source's
[independent_five_word.py](https://github.com/CubeJerry/integer-mult-bounds/blob/685457252b51b9c7932ece363365bdf0b12c9c31/research/bit-three-relation-core/independent_five_word.py).
The present verifier generalizes its dimension, exponent tuples and
ambient projector factors. This adapted code is not a separately
authored review. The prior k=3 case and wrong-sign bridge control
retain their original credit.
- Ali Romman's [PR #7693](https://github.com/CrocSwap/integer-mult-bounds/pull/7693),
head `dab13e010216570a00e22490b379b31b610f2890`, supplies the all-width
classification and unique ambient link. Read its
[exact proof](https://github.com/aliromman/integer-mult-bounds/blob/dab13e010216570a00e22490b379b31b610f2890/research/edge-extremal-pool-classification/PROOF.md).
That theorem uses the matching bound from
[PR #7692](https://github.com/CrocSwap/integer-mult-bounds/pull/7692),
head `c2870a38841f515db7203a0761cd77f014f7fb7a`,
[exact proof](https://github.com/aliromman/integer-mult-bounds/blob/c2870a38841f515db7203a0761cd77f014f7fb7a/research/edge-matching-image-bound/PROOF.md).
Both are explicit mathematical dependencies; their complete packages
are not rerun here.

The new result is the all-k, all-length word obstruction and its
full-ambient bridge deletion for this exact extremal alphabet.
The common degree-one kernel, five-core argument and known positive
pool family remain prior work. The new one-extra-relation control
shows the limit of the chronology statement without asserting a saving.

No external file is read at runtime. SOURCES.json binds the selected
state and task catalogue on main for provenance. Those records are
not mathematical premises. All required proof dependencies are linked
at their exact public commits.
273 changes: 273 additions & 0 deletions research/extremal-core-word-obstruction/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,273 @@
# Extremal irreducible cores cannot save paid rank at the relation budget

Ali Romman, with substantial OpenAI Codex assistance. Apache-2.0.
This is an all-length scalar-word proof with finite exact checks.
Independent human and formal-kernel review are unclaimed.

## 1. Exact model and result

Let H be the subgroup generated by finitely many paid exponents A_j
in a torsion-free abelian group. Thus H is a finite-rank free abelian
group. Work over `R=F2[H]`, and write [A] for its monomial with exponent A.
The saturated relation lattice is

```
L = kernel(Z^m -> H), e_j -> A_j.
```

On W scalar coordinates, consider a word of calls
`I+([A_j]-1)P_j`. Each P_j is a binary rank-one idempotent with nonzero
off-diagonal pattern. Coordinate projectors are excluded. The span
of the occurring patterns has dimension `k>=2`. Their joint image V
has dimension `2k-1`, and the occurring projectors act irreducibly on V.
Suppose `rank(L)<=k` and the full ambient endpoint is diagonal monomial.

**Theorem.** The occurring alphabet is the complete edge-pair-and-link
family in Section 2. All link calls lie in one consecutive block.
Their exponent sum is zero. Deleting that block preserves the full
ambient linear map, including every spectator-input column. Each
original target exponent is a sum of a disjoint subset of paid
exponents. If the exponents are rational address matrices, the sum
of target ranks is at most the sum of paid ranks.

Word length, ambient width and address ranks are unrestricted within
these hypotheses. This is not a theorem for coordinate calls, reducible
or smaller joint images, or a larger relation budget. No native
supplier, complete multiplier invoice or new kappa is supplied.

CubeJerry's [PR #6782](https://github.com/CrocSwap/integer-mult-bounds/pull/6782),
[exact word proof](https://github.com/CubeJerry/integer-mult-bounds/blob/685457252b51b9c7932ece363365bdf0b12c9c31/research/bit-three-relation-core/FIVE-CORE-WORD.md),
already proves the five-dimensional restricted bridge result. The
argument below extends its degree-two and chronology proof to every k.
The full ambient deletion uses the later ambient-link uniqueness in
Ali Romman's [PR #7693](https://github.com/CrocSwap/integer-mult-bounds/pull/7693),
[exact classification](https://github.com/aliromman/integer-mult-bounds/blob/dab13e010216570a00e22490b379b31b610f2890/research/edge-extremal-pool-classification/PROOF.md).
It is stronger than deleting calls only in a restricted core, and
is asserted only for the full alphabet specified here.

## 2. Forced alphabet and the endpoint after shear

Use the cited all-width classification, which depends on the matching
theorem in [PR #7692](https://github.com/CrocSwap/integer-mult-bounds/pull/7692),
[exact proof](https://github.com/aliromman/integer-mult-bounds/blob/c2870a38841f515db7203a0761cd77f014f7fb7a/research/edge-matching-image-bound/PROOF.md).
Put `t=k-1`. After a coordinate permutation the occurring projectors are

```
P_i = e_a_i(e_a_i+e_b_i)^T,
Q_i = (e_a_i+e_b_i)e_b_i^T, i=1,...,t,
B = (h_x+a)(h_y+b)^T,
h_x = sum_i e_b_i, h_y = sum_i e_a_i,
```

with disjoint edge endpoints, fresh a,b and `a dot b=1`. Every one of
these `2t+1` types occurs, because this is the set of actual word types.
The classification forces the entire set to be complete. In particular,
all B calls use the same ambient matrix, not merely the same restriction.

Let E be the coordinate span of the endpoints. Then
`V=E direct-sum F2*a`. The shear
`S=product_i(I+E_a_i,b_i)` acts only on E. In the basis of E and a,
it sends the restricted alphabet to the first 2t coordinate projectors
and `J_n`, the all-one matrix of odd size `n=2t+1`.

The full diagonal endpoint preserves V after extension of scalars to R.
Its action on a has fresh support only, so lies in R*a. All coordinates
in the support of a therefore have the same endpoint monomial. Before
the shear, the endpoint on E and a is diagonal monomial. After the
shear it can have off-diagonal entries at the t directed edge positions.
These positions are allowed throughout the proof. They must not be
silently assumed zero.

## 3. The entire linear relation kernel

Introduce one independent Laurent variable T_j for each call, with
augmentation variable `u_j=T_j-1`. The quotient onto the paid-generated
group H induces the linear map

```
F2^m -> H/2H, u_j -> A_j mod 2H.
```

Its kernel is L/2L. Indeed, the exact sequence with quotient H splits
over the integers because H is free. Tensoring with F2 keeps it exact.
Thus the kernel dimension is rank(L). An arbitrary larger ambient
exponent group cannot replace H in this step: its mod-two map can
lose the rank of paid generators that are even in that larger group.

The associated graded algebra of R at its augmentation ideal is
`Sym(H/2H)`. Choose a basis of H: R is a Laurent polynomial algebra,
and each basis monomial minus one gives an independent degree-one
generator. This proves the stated associated-graded description.

The t edge patterns and the bridge pattern are independent. At first
order the original diagonal endpoint therefore puts the following
k forms in the linear kernel:

```
L_B = sum of u_j over bridge calls,
M_i = sum of u_j over calls of P_i or Q_i, i=1,...,t.
```

Their supports are nonempty and disjoint. They are independent. Hence
`rank(L)=k`, and these forms give the entire linear kernel. The ideal
of the induced map between symmetric algebras is exactly
`(L_B,M_1,...,M_t)`. This is an associated-graded assertion; the actual
Laurent relation ideal need not be linear.

## 4. The general star quotient

In the sheared n-dimensional core, quotient the off-diagonal matrix
space by

```
Q = span(off(J_n), E_a_1,b_1, ..., E_a_t,b_t).
```

The universal constant and linear word terms vanish in this quotient
before exponent substitution. Consequently nonlinear substitution
terms cannot alter its degree-two image. Only the linear map from
Section 3 acts on that degree-two term.

For each old coordinate r, define
`K_r=off(J_n E_rr+E_rr J_n)`. The 2t matrices K_r are independent
modulo Q. Suppose their linear combination equals lambda times
`off(J_n)` plus allowed edges. Entries in the fresh row and old
column r force each coefficient to be lambda. At the reverse position
`(b_i,a_i)`, which is not an allowed edge, the left side is
`lambda+lambda=0`, so lambda is zero. All coefficients are zero.
This argument works even for t=1. It does not treat the fresh-coordinate
star as an additional independent matrix.

Bridge/bridge quadratic products are multiples of J_n, since n is odd.
Products of coordinate projectors are diagonal. Only mixed products
remain. For an old coordinate type r put

```
q_r = sum over calls j of type r
u_j times (sum of bridge variables before call j).
```

Modulo L_B, bridge prefix and suffix sums are equal. The degree-two
term in the quotient is therefore the sum of `K_r q_r`. Independence
of the K_r implies that every q_r lies in
`(L_B,M_1,...,M_t)`.

## 5. Chronology forces a single bridge block

Eliminate the last bridge variable by L_B=0. Let qbar_r be the resulting
polynomial. It uses only the first h-1 bridge variables and call
variables of coordinate type r, where h is the number of bridge calls.
It belongs to `(M_1,...,M_t)`.

This polynomial subring meets that ideal only in zero. In the pair
that contains r, choose a call variable of the partner type, which
exists because every type occurs. In each other pair choose any call
variable. Solve M_i=0 for these t chosen variables. This substitution
kills every M_i and fixes every variable in qbar_r. Hence qbar_r=0.

Every coordinate call has its own universal variable. Its coefficient
in qbar_r is its bridge prefix after elimination. A proper nonempty
prefix is a nonzero linear form in the first h-1 bridge variables.
It is zero only before the first bridge or after the last bridge.
Thus no coordinate call occurs strictly between the first and last
bridge calls. All bridge calls are consecutive. At h=1 this conclusion
has an empty interior and is immediate.

Combine the block on the restricted core. With Z the sum of its paid
exponents, the sheared word is

```
D_left [I+([Z]-1)J_n] D_right,
```

where both surrounding matrices are invertible monomial diagonals.
A fresh-row/old-column entry is an invertible monomial times `[Z]-1`.
This is outside the allowed endpoint edge positions, so it is zero.
Distinct exponents index distinct group-algebra monomials; hence Z=0.

## 6. Full ambient deletion and the rank invoice

The original ambient B is a single idempotent. For any A,C in H,

```
[I+([A]-1)B][I+([C]-1)B] = I+([A+C]-1)B.
```

The consecutive bridge block therefore equals the ambient identity
when Z=0. Its deletion preserves every input column, including dirty
spectator columns outside V. This conclusion uses ambient uniqueness;
it would not follow merely from identical restrictions on V.

All remaining projectors act only on E. In the sheared basis, old
coordinate r has diagonal label equal to the sum of exponents of
calls of that type. All fresh and other spectator coordinates have
label zero. Undoing S creates the edge entry `d_a_i+d_b_i`. Original
diagonality forces these monomials to agree, so the endpoint is the
same diagonal before and after S.

Assign to original coordinate r the paid calls of its corresponding
old projector type. Assign the empty subset to every other coordinate.
These subsets are disjoint and give all target exponents exactly.
For rational address matrices, rank subadditivity on each subset gives

```
sum_r rank(target_r)
<= sum over nonbridge calls j rank(A_j)
<= sum over all paid calls j rank(A_j).
```

Bridge calls remain paid; their deletion is an endpoint identity, not
a claim that they cost nothing in the original invoice. No saving
in total rational target rank is possible in this class.

## 7. Exact checks and hypothesis controls

The verifier checks the star-quotient ranks for t=1 through 20. It
also checks that the fresh star is dependent. It computes degree-two
terms from ordered matrix products and substituted linear exponent
vectors for 88748 type orders in eleven stated domains. These domains
have one call of each old type, one through four bridges for t=1,2,
and one through three bridges for t=3. Exactly the orders with one
consecutive bridge block pass. Bridge positions are not separately
labelled permutations; their exponents are assigned in chronological
order. Each model has exactly t+1 saturated relations, and its listed
paid exponents generate the stated free exponent group.

Separate full Laurent products check all 20532 type orders with two
bridges for t=1,2,3. Exactly 5166 have the expected diagonal endpoint;
every other word has a forbidden off-edge entry. The t=2 case is a
prior five-core control, not a new result. The recurrence is adapted
with credit from the source's
[Laurent checker](https://github.com/CubeJerry/integer-mult-bounds/blob/685457252b51b9c7932ece363365bdf0b12c9c31/research/bit-three-relation-core/independent_five_word.py).
This adapted implementation is not a separately authored review.

Ten full ambient fixtures use nonsymmetric fresh factors `(3,5)`,
two unused coordinates, t=1,2,3,4,6 and two or three bridge calls.
All full matrix coefficients agree before and after deleting the
whole bridge block. Omitting just one bridge changes an explicit
dirty-spectator column outside V. These checks do not project away
the spectator state.

A wrong-sign bridge control has the same linear kernel and passes
the degree-two quotient, but the full Laurent endpoint is not diagonal.
Thus the exact endpoint step cannot be replaced by a quadratic test.
With just one extra relation, an exactly diagonal word can contain a
cancelling coordinate excursion inside two separated bridge calls.
This defeats literal chronology beyond the stated budget; it is not
a rank-saving counterexample. A separate even-generator control shows
why the paid-generated group is needed for the parity kernel.

Finite checks corroborate the written proof and do not prove all
lengths. Full predecessor package replays, selected cold replay,
Lean and Docker were not run. No coordinate-call split, arbitrary
other alphabet, native realization, rounding contract, complete
multiplier or improvement in kappa is claimed. The selected frontier
and all five inherited assumptions remain unchanged.

## Open next questions

Can a reducible or subextremal image give a legal saving at the same
relation budget, or can an explicitly larger relation budget support
a saving? First give one exact scalar alphabet and its complete paid
and target invoice. The one-extra-relation control only removes the
chronology obstruction; it supplies no saving or native interface.
65 changes: 65 additions & 0 deletions research/extremal-core-word-obstruction/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,65 @@
# A word obstruction for every extremal irreducible binary core

Ali Romman, with substantial OpenAI Codex assistance. Apache-2.0.

Consider binary scalar calls `I+([A_j]-1)P_j` over the paid-generated
torsion-free exponent group. Suppose the occurring projectors are
noncoordinate rank-one idempotents, their pattern dimension is `k>=2`,
their joint image is irreducible of dimension `2k-1`, their saturated
exponent relation rank is at most k, and the full ambient endpoint is
diagonal monomial.

Then all bridge calls lie in one consecutive block and their exponent
sum is zero. Deleting that block preserves the full ambient operator,
including spectator-input columns. The target exponents are sums of
disjoint paid subsets, so total rational target rank cannot exceed
paid rank.

Read [PROOF.md](PROOF.md), [RESULT.json](RESULT.json), [CREDITS.md](CREDITS.md)
and [TASK.json](TASK.json). The proof extends the prior five-core
restricted word theorem in
[PR #6782](https://github.com/CrocSwap/integer-mult-bounds/pull/6782).
The full ambient conclusion uses the unique ambient link from
[PR #7693](https://github.com/CrocSwap/integer-mult-bounds/pull/7693).

## Reproduce

Use CPython 3.12 and its standard library, from the repository root.

```sh
python3.12 -I -W error -B research/extremal-core-word-obstruction/verify.py
python3.12 -I -W error -B -O research/extremal-core-word-obstruction/verify.py
python3.12 -B scripts/research_package.py --package research/extremal-core-word-obstruction
```

Expected markers are `PASS_EXTREMAL_CORE_WORD_OBSTRUCTION` and
`PASS_REGISTERED_RESEARCH_PACKAGE`. The verifier compares its result
with the committed receipt and writes no file.

It checks 88748 degree-two type orders, 20532 full Laurent type orders,
the star quotient through 20 edge pairs, and ten ambient deletion
fixtures with nonsymmetric fresh factors and dirty spectators. Wrong
bridge signs, one extra relation and an incorrect ambient parity
kernel have explicit controls. The finite checks corroborate the
written all-length proof; they do not replace it.

## Limits

Coordinate calls, reducible or smaller images, and larger relation
budgets are outside this theorem. The one-extra-relation control
breaks the chronology claim but gives no rank saving. The normal-form
shear is a proof device, not a free physical operation.

No native supplier, rounding contract, complete multiplier invoice
or new kappa is supplied. The selected frontier and its five assumptions
remain unchanged. Independent human and formal-kernel review are
unclaimed. Full predecessor package replays, selected cold replay,
Lean and Docker were not run.

## Open next questions

Can a specified subextremal or reducible core, or an explicitly larger
relation budget, support a legal saving? Give the exact alphabet and
full paid/target invoice before searching. A scalar word would still
need native realization, dirty-state restoration and a complete cost
and precision account.
Loading
Loading