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
17 changes: 17 additions & 0 deletions research/p112-farkas-kernel/NOTICE
Original file line number Diff line number Diff line change
@@ -0,0 +1,17 @@
New finite-kernel generator, certificate checker, proof and package metadata:
ToikY, with substantial OpenAI Codex assistance. Apache-2.0.

The two source files are retained verbatim from Nick Petalas's PR6425, exact
head092dcbd13d1263e691ad29f34fbf286fea1fac2c. Their original headings and credit
remain. SOURCE-PINS.json records exact Git blob identities, byte lengths and
SHA256 digests; no predecessor executable is imported or replayed.

Retain predecessor credit to Jeremy Diviney/PR1422 for the complete-clique guard,
geometry/caps and smaller-family question; Cinna Davis/PR1347/1323 for the support
model and geometry; Navinor, David Arias, Hanson Sun, DreamingOfClouds,
themilkman216 and the retained E8/label/decoder lineage.

PR5767 is related exact-certificate tooling, not the source of an optimization
solver here. Its shared limits and code are unchanged. Linear programming,
nonnegative implication/Farkas certificates, simplex and Bland's rule are
classical methods; no literature priority is claimed.
1 change: 1 addition & 0 deletions research/p112-farkas-kernel/ORACLE.json

Large diffs are not rendered by default.

105 changes: 105 additions & 0 deletions research/p112-farkas-kernel/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,105 @@
# Exact implications in the fixed P112 demand system

This is a finite, source-bound inequality result, quantified over all nonnegative
real multiplicity vectors. No predecessor executable or clique census is replayed.
The source matrix is the literal 280-row kernel in Nick Petalas's PR6425 at
`092dcbd13d1263e691ad29f34fbf286fea1fac2c`, authenticated by `SOURCE-PINS.json`.
Its 14 ordered cap rectangles, source clique witnesses and earlier completeness
claims retain their original meaning and attribution.

## Matrix and domain

For each saved pair `(H_i,d_i)`, let `a_i` be the 14-dimensional 0/1 indicator of
hit mask `H_i`. The row is `a_i x >= d_i`, with the common baseline `x >= 0`.
The source rows have distinct nonempty 14-bit masks and positive integer demands
at most 8. These conditions are checked, not inferred from a row count.

PR6425's [proof](https://github.com/petalas/integer-mult-bounds/blob/092dcbd13d1263e691ad29f34fbf286fea1fac2c/research/p112-demand-kernel/PROOF.md)
defines the hit predicate and proves equivalence of this kernel to its full
262,928-clique demand screen. That earlier theorem is a credited premise when
composing the present result with the full geometry. The present standalone
certificate establishes equivalence of the literal 280-row matrix and its new
retained subfamily; it does not independently certify the source clique census,
the witnesses' geometric validity, decoder coefficients or physical admission.

## Checkable deletion implication

For each omitted row `t`, the certificate supplies nonnegative rational weights
`lambda_j`, with every donor `j` in the **final retained family** `R`. The exact
checks are

sum_j lambda_j a_j <= a_t (coordinatewise),
sum_j lambda_j d_j >= d_t.

For any nonnegative `x` satisfying the retained rows,

a_t x >= (sum_j lambda_j a_j) x
>= sum_j lambda_j d_j >= d_t.

Thus the retained system implies every omitted row. Conversely a solution of
all original rows satisfies their retained subfamily. This proves equivalence
for all nonnegative real vectors, including unbounded integer vectors. No LP
solver status or floating tolerance enters this acceptance argument.

## Generation and simultaneous deletion

A positive donor weight cannot use a row having a 1 outside `H_t`, because the
target coefficient there is zero and all donor coefficients/weights are
nonnegative. Restrict candidate donors to strict subsets of `H_t`. Process
targets by increasing support size, retaining a row whenever it is not certified
implied by the already retained strict subsets. Retained rows are never deleted
later. Consequently every deletion refers directly to rows in final `R`;
simultaneous deletion is justified without circular premises.

The bounded generator uses exact rational packing simplex: maximize
`sum_j d_j lambda_j`, subject to coefficient capacities at most 1 on `H_t` and
`lambda >= 0`. Zero weights plus slacks form an initial feasible basis. Bland
entering/leaving rules include nonbasic slack columns. A feasible packing reaching
`d_t` already certifies deletion; full optimality is unnecessary for that case.
Explicit bounds reject exhausted work instead of truncating coverage.

The generator is a discovery mechanism. The elementary coefficient checks above
and the explicit points below independently establish the result even if the
simplex implementation is not trusted.

## Separating points and the restricted minimum claim

For every retained row `t`, the certificate supplies a full 14-dimensional
rational vector `x^(t) >= 0` which satisfies **all 279 other original rows** and
strictly violates row `t`. This is checked literally against the original matrix,
not just against `R`, a restricted LP, or an objective value.

A packing dual optimum below `d_t` can suggest the coordinates on `H_t`.
Setting each coordinate outside `H_t` to 8 covers any source row touching an
outside bit, since every demand is at most 8. Original rows wholly inside `H_t`
are covered by retained rows and their certified implications. This explains a
construction of the points; their literal checks are the acceptance evidence.

Given the full-original checks, any subfamily `S` of the original 280 rows
equivalent to the full system must contain every `t` in `R`: otherwise `x^(t)`
would satisfy `S` but fail the full system. Since `R` is itself equivalent, it is
the unique minimum-cardinality and unique inclusion-minimal **subfamily of these
original rows**, with the baseline `x >= 0` fixed. This claim is conditional on
every retained point passing the stated full-original checks.

No minimum over arbitrary new inequalities or alternative representations is
claimed. Fractional separating points do not establish integer-only
indispensability. Changing caps, adding rectangles or trimming actual supports
changes the matrix and requires new certificates.

## Credit and limits

Credit Nick Petalas/PR6425 for the fixed matrix, pairwise kernel and explicit
combined-implication handoff; Jeremy Diviney/PR1422 for the complete-clique guard,
source geometry, caps and smaller-family question; Cinna Davis/PR1347/1323 and
the retained E8/label/decoder lineage. Exact LP certificate semantics also have
repository precedent in PR5767; no shared solver or 128-variable cap is changed.
Nonnegative implication, packing/covering duality, simplex and Bland's rule are
classical. No general mathematical priority is asserted for them.

New work uses substantial OpenAI Codex assistance. A separately implemented
certificate checker is independent code verification of a frozen certificate,
not blind rediscovery or independent human/formal proof-kernel acceptance.
No coefficient synthesis, complete COPY floor, native/physical supplier,
restoration/access/precision/stock/compiler closure, signed invoice, outer
integration, κ improvement or selected-frontier change follows.
99 changes: 99 additions & 0 deletions research/p112-farkas-kernel/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,99 @@
# A 237-row exact P112 demand system

For the fixed 14-variable, 280-row cap-demand matrix of Nick Petalas's
[PR6425](https://github.com/CrocSwap/integer-mult-bounds/pull/6425), exact
nonnegative rational combinations remove 43 rows. The retained 237 rows are
equivalent for every nonnegative real multiplicity vector. Each retained row has
a rational point satisfying all 279 other original rows and strictly violating
that row. Together these certificates establish the **unique minimum subfamily
of the original 280 rows**, with `x >= 0` fixed.

This answers PR6425's first Open next question. Its pairwise reduction, source
clique census and geometry-to-demand equivalence remain credited predecessor
results; no predecessor executable or census is replayed. The present result is
an exact reduction of the literal source matrix. It supplies no coefficients,
decoder, physical supplier, COPY floor or κ improvement.

## Concrete implication

The first deleted mask is `0x1d1a`, demand 3. One half of each of the retained
rows `0x1412`, `0x1818` and `0xd0a`, each demand 2, gives RHS 3. The resulting
coefficient is 1 on target coordinates 1, 3, 4, 10, 11 and 12, and one half on
coordinate 8 (zero-based); all are at most the target's coefficients. The
missing half of coordinate 8 is supplied by `x_8 >= 0`. No single retained row
dominates this target: the source family was already pairwise nondominated.
`RESULT.json` retains all 43 exact combinations and all 237 full separating
points; `PROOF.md` gives the universal implication and restricted minimum proof.

## Replay

CPython 3.12 and the standard library suffice. No Git executable, network, solver
package or predecessor executable is needed for the package's two replays.

```sh
python3.12 -I -W error -B research/p112-farkas-kernel/verify.py
python3.12 -I -W error -B research/p112-farkas-kernel/oracle.py
```

Also run with `-O`. Default commands require strict complete retained JSON;
`--emit` prints regenerated output without overwriting files. The primary
regenerates the certificate with bounded exact `Fraction` packing simplex and
checks every implication and separator literally. The separate checker treats
the sealed primary result as untrusted certificate data and computes all
coefficient/RHS budgets and full-original-row separations independently. It does
not import the generator, solve an LP or claim blind rediscovery.

The generator uses 2,272 pivots, at most 49 per target, on this exact instance.
Explicit time/pivot/coefficient bounds fail closed. These are generation counts,
not a general runtime improvement. `VERIFY.json` registers the two bounded checks.
`SOURCE-PINS.json` authenticates both inert source files before parsing them.
`SOURCES.json` pins the originating current-main card/selected record as
provenance; all runtime bytes are package-local.

## Scope, evidence and limitations

The equivalence covers nonnegative real vectors and therefore preserves all
nonnegative integer tests too. Fractional separating points do **not** establish
integer-only indispensability or the smallest integer-only screen. The minimum
claim is over subfamilies of these original inequalities; it does not minimize
arbitrary newly introduced normals or alternative descriptions.

The independent checker exercises actual coefficient, demand, premise and
separator corruptions, along with source/schema controls. `VALIDATION.md`
records the actual run evidence, corrections and remaining obligations. Passing
this finite necessary-condition screen does not construct coefficient vectors
or pay for access, dirty restoration, precision, tape, stock, compiler or outer
integration. No native, Lean, Docker, selected cold supplier or complete
multiplication replay was run.

Orientation was frozen at canonical main
`7c49fe32c55de02c2c873655c804b16e4c3fef0a`, priority revision 8,
selected PR6030, selected-record SHA256
`9b8d0eea95b3bf0dfcb7938a732a6439b87a2685dec2a5ab9e525b3b34debbfa`.
The broader `complex-rank-shape-decoder` card is explicitly changed to this fixed
support-screen implication. Main/overlap are rechecked before submission;
sparse/history/hidden/unpublished coverage limits are not novelty evidence.

## Open next questions

1. Can integer multiplicity implications remove further rows? Specify the
integer domain and produce integer implication/separation certificates;
fractional witnesses alone do not answer that question.
2. For changed support caps, what is the new exact demand screen? Rebuild the
complete changed quotient before applying or attempting new combinations.
3. Can a passing vector admit actual coefficients and a fully charged decoder?
Specify one vector and its full interface before synthesis and pricing.

## Credit

Credit Nick Petalas/PR6425 for the fixed matrix, pairwise reduction and explicit
combined-implication question; Jeremy Diviney/PR1422 for the complete-clique
guard, pinned geometry/caps and smaller-family question; Cinna Davis/PR1347/1323
and Navinor, David Arias, Hanson Sun, DreamingOfClouds, themilkman216 and the
retained E8/label/decoder lineage. PR5767 is related exact-certificate tooling;
its code and 128-variable cap are unchanged. Nonnegative implication,
packing/covering duality, simplex and Bland's rule are classical.

New work is ToikY with substantial OpenAI Codex assistance. Separate AI
certificate checking is not independent human review or formal proof-kernel
acceptance. `NOTICE` preserves attribution; the two source files remain verbatim.
1 change: 1 addition & 0 deletions research/p112-farkas-kernel/RESULT.json

Large diffs are not rendered by default.

20 changes: 20 additions & 0 deletions research/p112-farkas-kernel/SOURCE-PINS.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,20 @@
{
"files": {
"README-6425.md": {
"git_blob": "d9ce3b644cff65dadd88663189638a4bfd52bac3",
"sha256": "f35f0a30796e94ced74686bc3614b4478f5460bac0adb7c8d80827991f70ed0c",
"bytes": 7410,
"repository": "petalas/integer-mult-bounds",
"commit": "092dcbd13d1263e691ad29f34fbf286fea1fac2c",
"path": "research/p112-demand-kernel/README.md"
},
"RESULT-6425.json": {
"git_blob": "e18f92e001b8ce9d44287760516bf6ea002bcbcd",
"sha256": "f8d15306112914af299acca8720779fbc9f6f07bb1336c51e99304fdea5caf89",
"bytes": 22234,
"repository": "petalas/integer-mult-bounds",
"commit": "092dcbd13d1263e691ad29f34fbf286fea1fac2c",
"path": "research/p112-demand-kernel/RESULT.json"
}
}
}
19 changes: 19 additions & 0 deletions research/p112-farkas-kernel/SOURCES.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,19 @@
{
"schema_version": 1,
"base_commit": "7c49fe32c55de02c2c873655c804b16e4c3fef0a",
"files": [
{
"path": "docs/research/agent-task-catalog.json",
"usage": "provenance",
"git_blob": "3404fbdd5e7529a0674eef26114b5071e4586cec",
"sha256": "ead6cdc9d0af6d58e8a49744d8ec8c7c15e13d2ba93c276c434bcf9b51b93a81"
},
{
"path": "certificates/selected-result.json",
"usage": "provenance",
"git_blob": "d59b5035e7b4cbaa9d688a1d32dd0d6eb79c466a",
"sha256": "9b8d0eea95b3bf0dfcb7938a732a6439b87a2685dec2a5ab9e525b3b34debbfa"
}
],
"note": "Canonical catalogue/selected are provenance only. Runtime inputs are the package-local source pair and result certificate. SOURCE-PINS.json authenticates the external source data; no predecessor executable is imported or replayed."
}
Loading