SabaPivot's picture
download
raw
11.3 kB
#!/usr/bin/env python3
"""Independent analytic certificate for Lemmas 2.1 and 2.2.
The finite property test in the original reproduction could only establish a
toy version of Claim 5. This script records the exact measure-zero argument,
checks its finite combinatorial reduction with rational arithmetic, and runs
destructive controls for the assumptions that make the lemmas true.
"""
from __future__ import annotations
import hashlib
import itertools
import json
import platform
import time
from fractions import Fraction
from pathlib import Path
ROOT = Path(__file__).resolve().parent
OUT = ROOT / "claim5_analytic_evidence"
PAPER_TEX = ROOT / "source_arxiv_2508.09628" / "main.tex"
PAPER_TAR = ROOT / "source_arxiv_2508.09628.tar"
def sha256(path: Path) -> str:
h = hashlib.sha256()
with path.open("rb") as handle:
for block in iter(lambda: handle.read(1 << 20), b""):
h.update(block)
return h.hexdigest()
def update_matrix(leaders: tuple[int, ...], gamma: Fraction):
"""Return L with y_i=(1-gamma)x_i+gamma*x_{leaders[i]}."""
n = len(leaders)
matrix = [[Fraction(0) for _ in range(n)] for _ in range(n)]
for i, leader in enumerate(leaders):
matrix[i][i] += 1 - gamma
matrix[i][leader] += gamma
return matrix
def outer_zero_symmetric(a, b) -> bool:
"""Whether a b^T+b a^T is exactly zero."""
n = len(a)
return all(a[p] * b[q] + b[p] * a[q] == 0 for p in range(n) for q in range(n))
def outer_zero_skew(a, b) -> bool:
"""Whether a b^T-b a^T is exactly zero."""
n = len(a)
return all(a[p] * b[q] - b[p] * a[q] == 0 for p in range(n) for q in range(n))
def exact_combinatorial_audit() -> dict:
gammas = [Fraction(1, 3), Fraction(1, 2), Fraction(2, 3)]
maps = 0
row_checks = 0
candidate_pair_checks = 0
symmetric_identity_only_when_same = True
skew_identity_only_when_same = True
for n in range(2, 6):
for leaders in itertools.product(range(n), repeat=n):
for gamma in gammas:
maps += 1
matrix = update_matrix(leaders, gamma)
for row in matrix:
row_checks += 1
assert all(value >= 0 for value in row)
assert sum(row) == 1
for query in range(n):
a = matrix[query]
for j in range(n):
for k in range(j + 1, n):
b = [matrix[j][q] - matrix[k][q] for q in range(n)]
assert sum(a) == 1
assert sum(b) == 0
candidate_pair_checks += 1
same_candidate_map = all(value == 0 for value in b)
if outer_zero_symmetric(a, b) and not same_candidate_map:
symmetric_identity_only_when_same = False
if outer_zero_skew(a, b) and not same_candidate_map:
skew_identity_only_when_same = False
assert symmetric_identity_only_when_same
assert skew_identity_only_when_same
return {
"n_values": [2, 3, 4, 5],
"gamma_values": ["1/3", "1/2", "2/3"],
"functional_leader_maps_times_gamma": maps,
"row_stochasticity_checks": row_checks,
"exact_query_candidate_pair_checks": candidate_pair_checks,
"symmetric_part_identity_implies_same_candidate_map":
symmetric_identity_only_when_same,
"skew_part_identity_implies_same_candidate_map":
skew_identity_only_when_same,
"arithmetic": "fractions.Fraction; no floating-point comparisons",
}
def destructive_controls() -> dict:
# Invertibility is necessary for a generic singleton statement.
points = [-1, 2]
zero_b_scores = [[0 * x * y for y in points] for x in points]
zero_b_top_multiplicity = [
sum(score == max(row) for score in row) for row in zero_b_scores
]
# Scalar gamma outside [0,1] is no longer a convex combination.
gamma = Fraction(6, 5)
scalar_outside = (1 - gamma) * Fraction(0) + gamma * Fraction(1)
# Paper Remark 2.3: coordinate-wise convexity does not imply membership in
# a non-axis-aligned convex hull. Unit triangle is x>=0,y>=0,x+y<=1.
x = (Fraction(1), Fraction(0))
leader = (Fraction(0), Fraction(1))
coordinate_gamma = (Fraction(3, 5), Fraction(7, 10))
matrix_step = (
x[0] + coordinate_gamma[0] * (leader[0] - x[0]),
x[1] + coordinate_gamma[1] * (leader[1] - x[1]),
)
return {
"singular_B_control": {
"B": 0,
"points": points,
"top_multiplicity": zero_b_top_multiplicity,
"distinct_argmax_not_singleton": all(v == 2 for v in zero_b_top_multiplicity),
},
"gamma_outside_convex_range_control": {
"gamma": "6/5",
"old_hull": "[0,1]",
"updated_point": str(scalar_outside),
"outside_old_hull": scalar_outside > 1,
},
"matrix_valued_step_control": {
"old_hull": "conv{(0,0),(1,0),(0,1)}",
"coordinate_step": ["3/5", "7/10"],
"updated_point": [str(value) for value in matrix_step],
"coordinate_sum": str(sum(matrix_step)),
"outside_old_hull": sum(matrix_step) > 1,
},
}
def proof_certificate() -> dict:
return {
"lemma_2_1": {
"statement": (
"For invertible B^t, for Lebesgue-almost-every initial "
"configuration the hardmax set is a singleton for every "
"particle and every finite integer time."
),
"independent_steps": [
(
"Fix a finite leader history through time t. Each current "
"token is A_i X^0, where A_i is a nonnegative coefficient "
"row summing to one, because every update is a scalar "
"convex combination."
),
(
"A tie for query i between distinct candidate maps j,k is "
"Q(X^0)=<B^t A_i X^0,(A_j-A_k)X^0>=0. Let a=A_i and "
"b=A_j-A_k; then sum(a)=1 and sum(b)=0."
),
(
"If the symmetric part of B^t is nonzero and Q were the "
"zero polynomial, diagonal coefficients force a_p b_p=0; "
"cross coefficients then force b=0. If B^t is "
"skew-symmetric, zero polynomial coefficients force "
"a_p b_q=a_q b_p, so b=c a; the row sums give c=0. Thus "
"Q is nonzero whenever the candidate maps are distinct."
),
(
"The zero set of a nonzero real polynomial has Lebesgue "
"measure zero. There are finitely many leader histories, "
"queries and candidate pairs at fixed t, and countably "
"many integer times. Finite and countable unions preserve "
"measure zero. If b=0, the candidate positions coincide "
"identically and represent one element of the set, not a "
"distinct tie."
),
],
"logical_dependencies": [
"finite token count",
"invertible (hence nonzero) B^t",
"scalar gamma^t in (0,1)",
"discrete times t in nonnegative integers",
],
"decision": "verified_by_independent_measure_zero_proof",
},
"lemma_2_2": {
"statement": "K^{t+1} is a subset of K^t for gamma^t in (0,1).",
"independent_steps": [
(
"The selected leader y_i^t belongs to K^t. Therefore "
"x_i^{t+1}=(1-gamma^t)x_i^t+gamma^t y_i^t belongs to K^t "
"by convexity."
),
(
"K^{t+1} is the convex hull of points x_i^{t+1}, all of "
"which lie in K^t. Since K^t is convex, their entire convex "
"hull is contained in K^t."
),
],
"decision": "verified_by_direct_convexity_proof",
},
}
def markdown_report(payload: dict) -> str:
audit = payload["exact_combinatorial_audit"]
controls = payload["destructive_controls"]
return f"""# Claim 5 analytic certificate
Decision: **verified** by an independent measure-zero proof and a direct
convexity proof.
The all-time singleton statement is reduced, on each finite leader history, to
the zero set of a nonzero quadratic polynomial. The coefficient argument uses
only `sum(a)=1`, `sum(b)=0`, and invertibility of the key-query matrix. A
finite union covers each time and a countable union covers all integer times.
Lemma 2.2 follows immediately because every new token is a scalar convex
combination of two points in the old hull.
The rational-arithmetic audit checked {audit['functional_leader_maps_times_gamma']:,}
leader-map/step-size combinations, {audit['row_stochasticity_checks']:,}
row invariants, and {audit['exact_query_candidate_pair_checks']:,} exact
query/candidate coefficient systems. It found no case where the tie
polynomial vanished identically for distinct candidate maps.
Destructive controls passed:
- Singular `B=0` produced multiplicity
{controls['singular_B_control']['top_multiplicity']}.
- `gamma=6/5` moved 0 toward 1 to
{controls['gamma_outside_convex_range_control']['updated_point']}, outside
`[0,1]`.
- The coordinate-wise matrix step produced
{controls['matrix_valued_step_control']['updated_point']} with coordinate
sum {controls['matrix_valued_step_control']['coordinate_sum']}, outside the
unit triangle.
"""
def main() -> None:
started = time.perf_counter()
OUT.mkdir(exist_ok=True)
payload = {
"paper": {
"submission_number": 8097,
"openreview_id": "zrn7rRuvhW",
"arxiv": "2508.09628",
"source_tar_sha256": sha256(PAPER_TAR),
"main_tex_sha256": sha256(PAPER_TEX),
"source_lines": {
"lemma_2_1": "main.tex:285-306",
"lemma_2_2": "main.tex:346-355",
},
},
"proof_certificate": proof_certificate(),
"exact_combinatorial_audit": exact_combinatorial_audit(),
"destructive_controls": destructive_controls(),
"environment": {
"python": platform.python_version(),
"platform": platform.platform(),
},
}
payload["runtime_seconds"] = time.perf_counter() - started
certificate = OUT / "certificate.json"
report = OUT / "PROOF.md"
certificate.write_text(json.dumps(payload, indent=2, sort_keys=True) + "\n")
report.write_text(markdown_report(payload))
files = [certificate, report, ROOT / "claim5_analytic_certificate.py"]
(OUT / "SHA256SUMS").write_text(
"".join(f"{sha256(path)} {path.relative_to(ROOT)}\n" for path in files)
)
print(json.dumps(payload, indent=2, sort_keys=True))
if __name__ == "__main__":
main()

Xet Storage Details

Size:
11.3 kB
·
Xet hash:
c883455bf9b71e8e094aae0f2f91c20f4d9a833e3e3d1b89ef2401b8c3b14509

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.