Skip to content
Merged
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
59 changes: 59 additions & 0 deletions .github/workflows/c7-total-bounded.yml
Original file line number Diff line number Diff line change
@@ -0,0 +1,59 @@
name: C7-total bounded-difference checks

'on':
pull_request:
paths:
- 'frontiers/c7_total_bounded_20260929/**'
- '.github/workflows/c7-total-bounded.yml'
workflow_dispatch:

permissions:
contents: read

jobs:
exact-and-probe:
runs-on: ubuntu-latest
timeout-minutes: 10
steps:
- uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262
with:
persist-credentials: false
- uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065
with:
python-version: '3.11.16'
- name: Verify sources, replay both modes, reject mutants
env:
PYTHONDONTWRITEBYTECODE: '1'
run: |
set -euo pipefail
python -B -S - <<'PY'
import hashlib, json, pathlib, subprocess, sys
root = pathlib.Path('frontiers/c7_total_bounded_20260929').resolve()
manifest = json.loads((root / 'SOURCE_FILES.json').read_text())['files']
present = sorted(p.name for p in root.iterdir())
expected_names = sorted([e['path'] for e in manifest] + ['SOURCE_FILES.json'])
if present != expected_names:
raise ValueError('packet tree differs from manifest: ' + repr(present))
for entry in manifest:
path = root / entry['path']
if pathlib.PurePosixPath(entry['path']).name != entry['path'] or path.is_symlink():
raise ValueError('flat regular source required: ' + entry['path'])
data = path.read_bytes()
if len(data) != entry['bytes'] or hashlib.sha256(data).hexdigest() != entry['sha256']:
raise ValueError('source identity mismatch: ' + entry['path'])
script = str(root / 'exponent_check.py')
expected = (root / 'RESULTS.json').read_bytes()
mutants = ['M1', 'M2', 'M3']
for flags in (['-B', '-S'], ['-B', '-O', '-S']):
out = subprocess.run([sys.executable, *flags, script],
check=True, capture_output=True, timeout=300).stdout
if out != expected:
raise RuntimeError('output differs from RESULTS.json under ' + ' '.join(flags))
for m in mutants:
rc = subprocess.run([sys.executable, *flags, script, '--mutant', m],
capture_output=True, timeout=300).returncode
if rc != 1:
raise RuntimeError('mutant ' + m + ' not rejected')
print(json.dumps({'scientific_effect': 'NONE', 'scientific_acceptance': False}))
PY
git diff --exit-code
203 changes: 203 additions & 0 deletions frontiers/c7_total_bounded_20260929/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,203 @@
# C7-total: a gap-mark-explicit failure intensity and the bounded unrestricted selection difference

**Object:** CL-C7-TOTAL-BOUNDED-20260929-v1.2 (v1.1 + the fixed-frame wording of the shifted-Gaussian step, see Revision history; v1.1 supersedes the withdrawn v1 memo CL-C7-TOTAL-K-SCALING-RECON-20260929-v1;
see `RECONNAISSANCE.md` for the retraction). **Author:** Anthropic Claude (session `session_01NMeKEismAyeqgdB4sy2NJU`).
**Disposition:** author-side proof candidate; **nonauthor analytic review required**. Scientific effect: NONE.

## 0. Statement of what is closed, and what is consumed

The C7 record (`reviews/c7_nonvanishing_openai_20260929/REVIEW.md` §5) leaves open the **order of the total
unrestricted rejected density** `ρ_rej(ℓ) := ν_cand^all(ℓ) − ν_eld^all(ℓ)` (known: `o(ℓ^{−1/3})`; far part `Θ(1)`) and
the **inverse-moment strip `q ∈ (−1, −2/3]`** of the rejected measure.

**Theorem K (gap-mark-explicit failure intensity).** In the setting of P §1 (fixed `d ≥ 2`, `L > 0`; law `Q_r =
Q_{r,b,k,R}`; typed weight `W_r`; cap region `G_r` of P §7), there are `r_0 > 0`, `N < ∞`, `C < ∞` depending on
`d, L` only, such that for every `b ∈ ℝ`, every `k > 0`, every frame, and `0 < r ≤ r_0`:

(K1) E_Q[W_r / r²] ≤ C (k + r)² (1 + |b| + k)^N ;
(K2) E_Q[(W_r / r²) 1_{G_r^c}] ≤ C (r³ / k) (1 + |b| + k)^N whenever r ≤ min{k, r_0}.

No lower bound on the normalizer `Z_r` is used anywhere.

**Corollary T (C7-total).** For every fixed `d ≥ 2`, `L > 0`, as `ℓ ↓ 0`:

(T1) ν_cand^all(ℓ) − ν_eld^all(ℓ) = O(1) (a version bounded on (0, ℓ_0]);
(T2) with Theorem U (far rejected density ≥ c_* > 0): ν_cand^all − ν_eld^all = Θ(1);
(T3) the rejected measure ρ_rej(ℓ) dℓ satisfies ∫_0^t ℓ^q ρ_rej(ℓ) dℓ < ∞ ⇔ q > −1.

So the unrestricted selection difference is bounded (the "no unrestricted `O(1)` conclusion" caveat of the C7
record is discharged), and the strip `(−1, −2/3]` lies entirely on the finite side for the rejected measure; the
candidate and elder measures individually keep the parent's threshold `−2/3`.

Consumed unchanged: P = `imports/lifetime_parent_20260925/UNIFORM_MATRIX_CAP_AND_LIFETIME.md` (blob `dfed3b8d…`)
with the reconciliation reading rule — §§1–5 (in particular (4.1), (4.3), (5.1), (5.2), (5.3)), §§6–7 verbatim
except where a bound is replaced as stated, §8, §§9–10 (the intensity identity), (13.2), (13.3), (14.1);
`frontiers/unrestricted_selection_difference_20260929/PROOF.md` (Theorem U). Nothing in those sources is
re-proved.

## 1. Where the parent loses `k`, and why it does not matter to the parent

P (6.1) bounds the axial entry by `|α_M| ≤ h/2 = M_3/2` (from (5.1)), although P (5.2) gives the sharper
`|α_M + 6k| ≤ rM_4/2`, i.e. `|α_M| ≤ 6k + rM_4/2`. On a compact gap window `k ≥ k_- > 0` the two are equivalent up to
constants, so the parent loses nothing. As `k ↓ 0` the difference is a factor `k²` in the weight — the same `k²`
that appears in the normalizer `z_0 = (6k)² E[det(A_0)² 1{A_0 < 0}]` (P (5.4)) and produces the `k^{4/3}` of the (15.2)
gamma integral. Theorem K is P §§6–7 with that factor kept and with the constant `D = 4K²/(3k)` of (7.1) tracked.

## 2. Proof of (K1)

By the pin identity P (5.3), for `i ∈ {M, S}`,
`|det H_i| / r ≤ |α_i| |det A_i| + r ‖β_i‖² ‖adj A_i‖`. Use (5.2) for `α_i`, (5.1) for `β_i`, and `‖A_i‖ ≤ M_2`,
`|det A_i| ≤ M_2^m`, `‖adj A_i‖ ≤ M_2^{m−1}` (`m = d − 1`, `M_j` the local `C^j` seminorms of P §4):

|det H_i| / r ≤ (6k + rM_4/2) M_2^m + (r M_3²/4) M_2^{m−1} ≤ C (k + r) (1 + ‖f‖_{C⁴})^{m+2}.

Hence `W_r / r² ≤ C (k + r)² (1 + ‖f‖_{C⁴})^{2m+4}`. P (4.1) holds for every derivative order `q` and every finite
`p`; its target-growth version P (13.3) is stated for `C³` but its proof ("the same summability/Minkowski proof as
Section 4") uses only the summable Fourier weights and the regression-mean bound `C|v_r|`, hence applies verbatim
to `C⁴`: `E_Q (1 + ‖f‖_{C⁴})^p ≤ C_p (1 + |b| + k)^p` uniformly for `r ≤ r_0` and all frames. Taking
`p = 2m + 4` gives (K1). ∎

## 3. Proof of (K2)

Work on the typed support with `B = −A_M ≻ 0`, eigenvalues `0 < λ_1 ≤ … ≤ λ_m = Λ`, `h = M_3`. Two of the parent's
§7 inputs are stated there for compact `B × K` only and must be replaced by their polynomial-growth versions before
`b` and `k` are unrestricted (P §13 itself warns against "the compact-target bound (4.1) applied outside its scope"):

- **(4.3) with growth.** In the regression decomposition P (4.2), `f = μ_r + B_r·(A_r − E_Q A_r) + g_r`, the mean
`μ_r` is linear in the target, so `‖μ_r‖_{C⁴} ≤ C |v_r| ≤ C(|b| + k)`, and likewise `‖E_Q A_r‖ ≤ C(|b| + k)`;
`B_r` and `g_r` are as in P §4 (target-independent). Hence `M_3, M_4 ≤ K_0 (J_r + ‖A_r‖_F) + C(|b| + k)`. Put
`U := J_r + ‖A_r‖_F + |b| + k ≥ 1`; then `h, M_4 ≤ K U` with `K` depending on `d, L` only.
- **(3.5)/(7.2) with a shifted center.** The `Q`-density of `A_r` is Gaussian with covariance uniformly bounded
above and below (P §3) and mean `m_A = E_Q A_r`, `‖m_A‖_F ≤ C(|b| + k)`: `density_Q(A) ≤ C_0 exp(−c_0 ‖A − m_A‖_F²)`.
The eigenvalue integration is performed **after fixing the orthogonal frame**: write `B = −A = O diag(λ) O^T`
and, for fixed `O`, `‖A − m_A‖_F² = ‖diag(λ) + O^T m_A O‖_F² = Σ_i (λ_i + q_i)² + (fixed off-diagonal squares)`,
`q = diag(O^T m_A O)`, `|q_i| ≤ ‖m_A‖`. The Gaussian factor in `λ_1` is dropped on the layer (as in P §7); the
remaining `λ_2, …, λ_m` integrals are one-dimensional *shifted* Gaussian polynomial moments, `∫ λ^p e^{−c_0(λ+q)²}dλ
≤ C_p (1 + |q|)^p`, hence polynomial in `‖m_A‖ ≤ C(|b| + k)`, uniformly over the compact `O`; the Vandermonde has
already been replaced by the nonnegative polynomial `Λ^{m(m−1)/2}`, and the Haar integral over `O` is then taken.
This replaces P's centered `e^{−c Σ λ_j²}` without any `exp(C‖m_A‖²)` loss, and it does **not** assert that the
shifted matrix law is orthogonally invariant — only the centered covariance bound is.

With these two replacements, P §7 runs verbatim.

**(6.2′) — the weight bound with the axial entries at their true size.** Write `a := 6k + rM_4/2 ≤ 6k + rKU/2`, so
that `|α_M|, |α_S| ≤ a` by P (5.2). P's canceled-pivot argument for (6.1) reads `|det H_M| ≤ r |α_M| det B`
(the Schur complement `B − (r/|α_M|) β β^T ≻ 0` and determinant monotonicity); it does not depend on how `|α_M|` is
bounded afterwards, so `|det H_M| ≤ r a λ_1 ∏_{j≥2} λ_j`. For the saddle there is **no** (5.2)-type improvement of
the mixed term: P's derivation of the second line of (6.2) uses `|α_S| ≤ h/2` *and* `‖β_S‖² ≤ h²/4` (from (5.1)),
and only the first of these sharpens. Keeping the second as it is,

|det H_S| ≤ r [ a (λ_1 + rh) + r h²/4 ] ∏_{j≥2} (λ_j + rh),

and therefore

W_r ≤ r² a · λ_1 [ a λ_1 + a r K U + r K² U² / 4 ] · ∏_{j≥2} λ_j (λ_j + r K U). (6.2′)

(At `a = h/2` this is P's (6.2) exactly: `(h/2)[λ_1 + rh] + rh²/4 = (h/2)[λ_1 + (3/2) rh]`.)

**The failure set.** P §7: `G_r^c ⊂ {λ_1 ≤ D r U²} ∪ {r M_4 > 3k/10}` with `D = 4K²/(3k)` (P (7.1)), and for
`m = 1` the additional far branch `{λ > 1/(4Dr)}` of P (7.5).

**Depth failure, `m ≥ 2`.** The layer integral of the `λ_1`-bracket of (6.2′) is exact (checker):

∫_0^{DrU²} λ_1 [ a λ_1 + a r K U + r K² U²/4 ] dλ_1 = r³ U⁵ [ a D³ U / 3 + a K D² / 2 + K² D² U / 8 ].

With `D = 4K²/(3k)`: `D³ ∝ k^{−3}`, `D² ∝ k^{−2}`. Insert (6.2′) into the eigenvalue measure (P (7.2), shifted
version above) and integrate `λ_1` first, as in P (7.4); every other factor is bounded as there. Collecting the
`a`-dependence, and using `a ≤ 6k + rKU/2`, the numerator is

E_Q[W_r 1_{depth}] ≤ C r⁵ E∫ [ a² k^{−3} + a² k^{−2} + a k^{−2} ] U^{2m+6} Λ^{m(m−1)/2} (shifted Gaussian) dλ_2 … dλ_m .

Now `a² ≤ 72k² + r²K²U²/2` gives `a² k^{−3} ≤ C k^{−3}(k² + r²) U²`, `a² k^{−2} ≤ C k^{−3}(k² + r²) U²` for `k ≤ 1`,
and `a k^{−2} ≤ C (k + rU) k^{−2} ≤ C k^{−3}(k² + r²) U` since `r k^{−2} ≤ (k^{−1} + r² k^{−3})/2` (AM–GM). Moments
are polynomial in `(1 + |b| + k)` by the fixed-frame shifted-Gaussian step above and the `C⁴` moment of §2. So

E_Q[W_r 1_{depth}] ≤ C r⁵ k^{−3} (k² + r²) (1 + |b| + k)^N (k ≤ 1),

and for `r ≤ k`: `E_Q[(W_r/r²) 1_{depth}] ≤ C (r³/k)(1 + |b| + k)^N`. For `k ≥ 1`, `D ≤ 4K²/3` and `a ≤ C(k + r)U`,
so `E_Q[W_r 1_{depth}] ≤ C r⁵ (k + r)² (1 + |b| + k)^N ≤ C r⁵ k^{-1} (1 + |b| + k)^{N+3}` — the same form with a
larger `N`; henceforth `N` denotes the larger exponent.

**`m = 1`.** P (7.5): depth failure is contained in `{λ ≤ 4DrJ_r²} ∪ {λ > 1/(4Dr)}` (with `U` as redefined above). The
near branch gives the same `r⁵ k^{−3}(k² + r²)` numerator by direct integration of the bracket of (6.2′) in `λ`
(P's `C r⁵ E J_r^{10}` with the axial factor kept and the `r h²/4` mixed term retained, as above). The far branch, by P (7.6) with the weight bounded as in §2,
`E_Q[(W_r/r²) 1_{far}] ≤ (4Dr)⁴ E_Q[(W_r/r²) λ⁴] ≤ C (r/k)⁴ (k + r)² (1 + |b| + k)^N`, and for `r ≤ k` this is
`≤ 4C (r/k)⁴ k² = 4C (r³/k)(r/k) ≤ 4C r³/k` (times the polynomial).

**The fourth-derivative exception.** P (7.7) with the weight of §2:
`E_Q[(W_r/r²) 1{rM_4 > 3k/10}] ≤ (10r/(3k))⁴ E_Q[(W_r/r²) M_4⁴] ≤ C (r/k)⁴ (k + r)² (1 + |b| + k)^N ≤ 4C r³/k`
for `r ≤ k`, as above.

Summing the pieces gives (K2) for `r ≤ min{k, r_0}`. No division by `Z_r` occurred, and no compact-window
constant was used outside its scope. ∎

**Remark (normalized form).** On compact `b`-sets, P (5.4)–(5.5) made `k`-explicit — `z_0 ∝ k²` and the
convergence `Z_r/r² → z_0` uniform for `r ≤ ck` because the relative error in (5.2) is `rM_4/(12k)` — turn (K2) into
`1 − p_r ≤ Q^W(G_r^c) ≤ C_B (r/k)³` for `r ≤ ck`. Corollary T does not use this.

## 4. Proof of Corollary T

**The rejected intensity identity.** P §§9–10 disintegrate the marked Kac–Rice measure into the pin density
times `E_Q[W_r e]` for the elder mark `e` and `E_Q[W_r]` for candidates. Applying the same disintegration to the
mark `1 − e` (Borel, bounded), the per-volume rejected intensity is `r A_r (1 − p_r) dr db dk dσ` with

A_r (1 − p_r) = 12 π_r(v_r) · E_Q[(W_r/r²)(1 − e)],

and on the generic locus `1 − e ≤ 1_{G_r^c}` because on `G_r` the global elder partner of `M` is `S` (P §8). With
the change of variables P (11.1) and the near/far split of P §§13–14,

ρ_rej^near(ℓ) = ℓ^{−1/3} ∫∫∫ 1{k ≥ ℓ/r_0³} · 12 π_r(v_r) E_Q[(W_r/r²)(1 − e)] / (3 k^{2/3}) db dk dσ(u), r = (ℓ/k)^{1/3},

with `π_r(v_r) ≤ C e^{−c(b² + k²)}` by P (13.2) (target-free covariance band, `r ≤ r_0`).

**Split at `k_* = ℓ^{1/4}`** (equivalently `r = k`, since `r³ = ℓ/k`).

- `k ≥ k_*` (so `r ≤ k`): by (K2), the integrand is at most
`C (ℓ/k²) k^{−2/3} (1 + |b| + k)^N e^{−c(b² + k²)}`; the `b`-integral is finite and the `k`-integral is
`≤ C ℓ ∫_{k_*}^∞ k^{−8/3} dk = (3/5) C ℓ k_*^{−5/3} = C′ ℓ^{7/12}`.
- `ℓ/r_0³ ≤ k < k_*` (so `r > k`): by (K1), `(k + r)² ≤ 2k² + 2r² = 2k² + 2(ℓ/k)^{2/3}`; the `k`-integrals are
`∫_0^{k_*} k^{4/3} dk = (3/7) k_*^{7/3} = C ℓ^{7/12}` and
`ℓ^{2/3} ∫_{ℓ/r_0³}^{k_*} k^{−4/3} dk ≤ 3 ℓ^{2/3} (ℓ/r_0³)^{−1/3} = 3 r_0 ℓ^{1/3}`.

Hence `ρ_rej^near(ℓ) ≤ ℓ^{−1/3} [ C ℓ^{7/12} + C r_0 ℓ^{1/3} ] = C ℓ^{1/4} + C r_0`. The far part is
`ρ_rej^far ≤ ν_cand^far ≤ C` by P (14.1). This proves (T1). (T2) is Theorem U's lower constant on the far part.
(T3): a density that is bounded above near `0` (T1) and bounded below by a positive constant (T2) has
`∫_0^t ℓ^q dℓ`-type moments finite exactly for `q > −1`. ∎

**Remark (what the `O(1)` is made of).** The `C r_0` term comes from separations `r ∈ [k, r_0]` with tiny gap
marks `k ~ ℓ/r³` — pairs at macroscopic separation whose height gap happens to be `ℓ`. These behave like far
pairs (Theorem U's mechanism), the near/far cutoff `r_0` is arbitrary, and (T1)'s constant is not a rate. The
genuinely near part (`r ≤ k`) is `O(ℓ^{1/4})`, i.e. rejection is *rarer* than `ℓ^{2/3}`-scale only in the
compact-window sense; on the unrestricted population the far pairs dominate the rejected measure.

## 5. What is new relative to the parent, and what is not claimed

New: (K1) is a sharper majorant than P (13.4) by the factor `(k + r)²`, obtained from (5.2) instead of (5.1); (K2)
is P §7 with `k` tracked. Everything else is the parent's own machinery. Not claimed: any numerical constant; any
rate for `ν_cand` or `ν_eld` individually; any change to the parent's existential scope or to Theorem A's compact
formulation; any statement about the *selected* measure's threshold (still `−2/3`); nonauthor acceptance.

## Revision history

- v1 (withdrawn): `RECONNAISSANCE.md` retraction; the `ℓ^{−1/4}` conjecture is not revived.
- v1.1 (2026-09-29, blob `1fe08e2a…`): Theorem K / Corollary T with the saddle mixed term retained.
- v1.2 (2026-09-29): the shifted-Gaussian eigenvalue integration in §3 is now stated explicitly as performed after
fixing the orthogonal frame `O` (separable shifted one-dimensional moments, uniform over compact `O`, Haar integral
last), as recommended in the OpenAI reviews
[5357773376](https://github.com/d6g8k5htny-coder/Math-/pull/150#pullrequestreview-5357773376) and
[5358141048](https://github.com/d6g8k5htny-coder/Math-/pull/150#pullrequestreview-5358141048). Wording only; no
theorem, constant or checker change; `RESULTS.json` bytes unchanged.

## 6. Review requested (non-Claude lane)

(i) that P's canceled-pivot step (6.1) admits the `k`-dependent axial bound (5.2), and that the saddle bound keeps
the `r h²/4` mixed term — the v1.1 draft had wrongly sharpened it, caught in an author-side adversarial pass;
(ii) the `C³ → C⁴` extension of (13.3) and the polynomial-growth versions of (4.3) and (7.2) used for unrestricted
`b`; (iii) the `k`-tracked three-term layer integral and (7.5)–(7.7) arithmetic (exact in the checker);
(iv) the rejected-intensity identity with mark `1 − e ≤ 1_{G_r^c}`; (v) the piecewise integrals of §4 (checker).
**OpenAI** (parent author) is the natural lane for (i)–(iii); any lane for (iv)–(v).

## Reproduce

python -B -S exponent_check.py # prints RESULTS.json byte for byte; -O identical
python -B -S exponent_check.py --mutant M # exit 1 for M in {M1, M2, M3}
20 changes: 20 additions & 0 deletions frontiers/c7_total_bounded_20260929/RESULTS.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,20 @@
{
"object": "CL-C7-TOTAL-BOUNDED-20260929-v1.1-CHECKS",
"passed": true,
"results": {
"(6.2')": "layer integral r^3 U^5 [a D^3 U/3 + a K D^2/2 + K^2 D^2 U/8]; equals P (6.2) at a = h/2",
"(7.3)": "int_0^{D r U^2} lam(lam+E r U) dlam = r^3 U^5 (D^3 U/3 + E D^2/2), exact",
"(K2)_reductions": "(r/k)^4 (k+r)^2 <= 4 r^3/k for r <= k; depth term r^5 k^-3 (k^2 + r^2) <= 2 r^5/k for r <= k",
"Corollary_T": {
"far": "O(1) by (14.1)",
"near, r<=k": "O(l^{1/4})",
"near, r>k, k^2-part": "O(l^{1/4})",
"near, r>k, r^2-part": "O(1) (cutoff shell)",
"rejected_moment_threshold": "q > -1 (sharp)",
"total": "O(1); Theta(1) with Theorem U"
},
"k_reductions": "a^2 k^-3, a^2 k^-2, a k^-2 all <= C k^-3 (k^2 + r^2) poly(U) for k <= 1; a k^-2 via r k^-2 <= (k^-1 + r^2 k^-3)/2",
"withdrawn_v1": "crude majorant route gives l^{-1/4}; wrong because A_r ~ k^2 (P (5.4)); kept as mutant M1"
},
"scope": "exact exponent/integral bookkeeping for Theorem K and Corollary T; Gaussian steps not verified"
}
Loading
Loading