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
178 changes: 178 additions & 0 deletions research/binary-dual-polar-cut-ceiling/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,178 @@
# The exact binary dual-polar cut ratio at every rank

This contribution was produced by Nick Petalas. Apache-2.0.

## Statement and scope

Let X be the Lagrangian subspaces of a nondegenerate alternating space of
dimension 2m over F2, where m is any positive integer. Put
d(L,M)=m-dim(L intersection M). Let A_i be the zero-one distance-i matrix,
k_i its row sum and T_i=A_i/k_i. A distance-i pair means a uniformly
chosen ordered pair at that distance. For S contained in X, define

p_i(S) = Pr[exactly one endpoint of a distance-i pair belongs to S].

For every nonempty proper S,

p_m(S) / p_1(S) <= K_m = 2 - 2^(1-m). (1)

Point cuts S_v={L : v belongs to L}, for v nonzero, attain equality. Thus
K_m is the exact Boolean-cut optimum for every m. The unique nonconstant
adjacency eigenspace maximizing the spectral quotient is the first one.
This does not classify all attaining Boolean sets.

Theorem 6 of Jacob Sussman's PR #4628, in
`author-evidence/induction-across-block-sizes.md`, explicitly leaves the
all-rank spectral comparison open after finite computations through rank 12.
The new step here is its universal proof. The graph, spectral-comparison
method, point-cut witness and coordinate-family interpretation are credited
predecessor context. No predecessor scientific executable is replayed.

This limits a stated certificate family. It is neither a word lower bound
of K_m calls per array nor an obstruction to every kept-subspace certificate.
It supplies no physical program, new kappa or supplier admission.

## 1. Intersection numbers

Every neighbor of L is obtained by choosing a hyperplane H of L and one of
the two other Lagrangians over H in the symplectic plane H-perp/H. Hence
k_1=2(2^m-1). Fix another vertex Z and put R=L intersection Z, of dimension
m-i. There are 2^i-1 hyperplanes H of L containing R. Over each such H,
one alternative gains one dimension of intersection with Z and the other
keeps the intersection dimension. Indeed the pairing from Z to H has rank
i-1, so Z intersection H-perp has dimension m-i+1 and determines exactly
one alternative line over H.

There are 2^m-2^i hyperplanes not containing R. The pairing from Z to H
then has rank i. Thus Z intersection H-perp=R, and either alternative
intersects Z in H intersection R, losing one dimension. Consequently

c_i = 2^i-1, a_i = 2^i-1,
b_i = 2^(i+1)(2^(m-i)-1). (2)

Here c_i,a_i,b_i count neighbors at distances i-1,i,i+1 from Z. The graph
is connected since c_i>0 whenever i>0. Transverse Lagrangians exist, so
its diameter is m. Double counting edges between distance spheres gives
k_i c_i=k_(i-1)b_(i-1), and in particular

k_m = 2^(m(m+1)/2). (3)

The distance matrices satisfy

A_1 A_i = b_(i-1) A_(i-1) + a_i A_i + c_(i+1) A_(i+1), (4)

with terms outside 0..m omitted. Since c_i>0 for 1<=i<=m, each A_i is
a degree-i polynomial in A_1. These m+1 matrices are linearly independent
because their supports are disjoint. At i=m, (4) gives a degree-(m+1)
polynomial annihilating A_1. Its minimal polynomial therefore has degree
m+1. Since A_1 is real symmetric, it has exactly m+1 distinct eigenvalues.

## 2. Factored distance eigenvalues

For 0<=j<=m define

theta_j = 2^(m-j+1)-2^j-1,
F_j(z) = product_(r=1..m-j)(1+2^r z)
product_(r=0..j-1)(1-2^r z).

Write P_i(j) for the coefficient of z^i in F_j. Empty products are one.
The telescoping products give the polynomial identity

(1-z)(1+2z) F_j(2z)
= (1+2^(m-j+1)z)(1-2^j z) F_j(z). (5)

This is a polynomial identity, with no division by a possibly zero factor.
Comparing coefficients and using (2) yields

theta_j P_i(j)
= b_(i-1)P_(i-1)(j) + a_i P_i(j)
+ c_(i+1)P_(i+1)(j), (6)

again omitting outside terms. This includes the terminal equation i=m.
The initial value P_0(j)=1 and (4) uniquely determine every distance
eigenvalue from an adjacency eigenvalue. Equation (6) says each theta_j
is a root of the degree-(m+1) annihilating polynomial just constructed.
They are strictly decreasing, so they give its complete list of roots.
On the theta_j eigenspace, A_i has eigenvalue P_i(j).

The leading coefficient of F_j and (3) give

u_1(j) = theta_j / [2(2^m-1)],
u_m(j) = (-1)^j 2^(-j(m-j+1)), (7)

the eigenvalues of T_1 and T_m. Indeed the leading-coefficient exponent
is `(m-j)(m-j+1)/2 + j(j-1)/2`; subtracting `m(m+1)/2`
gives `-j(m-j+1)`.

The distance-regular graph spectrum is classical. Lee and Tanaka,
[SIGMA 14 (2018), 009, section 3](https://doi.org/10.3842/SIGMA.2018.009),
record the general intersection numbers and adjacency eigenvalues. The
binary factor/recurrence argument above makes the proof self-contained;
it does not extrapolate a finite table.

## 3. Uniform maximization

For j>=1, u_1(j)<1. The proposed bound satisfies

K_m (1-u_1(j))
= 2 - 2^(1-j) + (2^j-1)2^(-m). (8)

At j=1, both (8) and 1-u_m(j) equal 1+2^(-m), so the quotient is K_m.
If j>=2, then m>=2 and

j(m-j+1)-m = (j-1)(m-j) >= 0.

Thus `1-u_m(j) <= 1+2^(-m) <= 5/4`, whereas (8) is strictly greater
than 3/2. The quotient is strictly less than K_m in every remaining
nonconstant sector. For m=1 there is only j=1. We have proved

max_(1<=j<=m) (1-u_m(j))/(1-u_1(j)) = K_m. (9)

This argument holds for every positive m, independently of the checker's
finite rank limits.

## 4. Boolean cuts and sharpness

Use the uniform inner product on functions on X. Decompose f=1_S into
mutually orthogonal adjacency eigenspaces, f=sum_j f_j. Symmetry and
regularity give

p_i(S) = 2 <f,(I-T_i)f>
= 2 sum_(j>=1)(1-u_i(j)) ||f_j||^2.

Apply (9) term by term. Connectedness gives p_1(S)>0 for every nontrivial
S, proving (1). A directed-exit convention multiplies both probabilities
by the same constant, leaving the ratio unchanged.

The credited point-cut witness works at every rank. Each Lagrangian
contains 2^m-1 nonzero points, and the symplectic group is transitive on
the 2^(2m)-1 nonzero points. Thus the fraction containing a fixed point is
alpha=1/(2^m+1). Two transverse Lagrangians cannot both contain that point,
so p_m(S_v)=2alpha. From a Lagrangian containing v, exactly 2^(m-1)
hyperplanes omit v, and both alternatives over each such hyperplane omit
v. The conditional edge-exit probability is 2^m/[2(2^m-1)]. Edge reversal
therefore gives

p_1(S_v) = 2alpha * 2^(m-1)/(2^m-1).

The ratio is `(2^m-1)/2^(m-1)=K_m`, proving sharpness. At m=1, distance
one and distance m coincide, as the formula requires.

## 5. Certificate consequence and limits

A coordinate family assigns Y_L=span{e_x : x belongs to S(L)} in one
fixed finite coordinate basis, with constant dimension. Its directed
defect `dim(Y_L)-dim(Y_L intersection Y_M)` is a sum of Boolean-cut
exits over coordinates. Averaging over each distance relation turns this
into a sum of half the disagreement probabilities above. Consequently
the opposite-to-edge average-defect ratio is at most K_m whenever the
denominator is positive. Nonnegative sums obey the same ceiling. The
coordinates indexed by nonzero points, with membership v in L, give a
constant-dimension family attaining it.

This is PR4628's coordinate-family application with the missing all-rank
premise supplied. It does not bound arbitrary noncoordinate subspaces,
temporal cuts, changed free-letter models or general physical circuits.
Supplier realization, ownership/restoration, compiler, precision/tape and
analytic recovery remain separate. No predecessor scientific replay,
selected cold replay, Lean build or Docker run is claimed.
131 changes: 131 additions & 0 deletions research/binary-dual-polar-cut-ceiling/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,131 @@
# The sharp binary dual-polar cut ceiling at every rank

This contribution was produced by Nick Petalas. Apache-2.0.
Independent human and formal proof-kernel review are unclaimed.

Claim type: scoped written proof with exact finite corroboration.
For the graph of Lagrangian subspaces of a binary symplectic space of
dimension 2m, the exact ratio of opposite-pair cut disagreement to
edge-pair cut disagreement is

max_S p_m(S)/p_1(S) = 2 - 2^(1-m), for every integer m>=1.

Point cuts attain it. The first nonconstant eigenspace uniquely maximizes
the corresponding spectral quotient. A factored distance-eigenvalue
polynomial proves the missing uniform maximization in Jacob Sussman's
[PR4628](https://github.com/CrocSwap/integer-mult-bounds/pull/4628),
Theorem 6 of its induction note. That note gives finite evidence through
rank 12 and explicitly leaves the all-rank comparison open. Its finite
results, point-cut witness and coordinate-family interpretation retain
predecessor credit. This contribution supplies the new all-rank proof,
without running predecessor scientific programs.

The resulting ceiling for constant-dimension coordinate certificates and
their nonnegative sums holds at every rank. It does not exclude general
noncoordinate certificates or prove a call-count bound for arbitrary
physical circuits. No new kappa or selected-frontier change is claimed.
The complete argument is in [PROOF.md](PROOF.md).

## Evidence and reproduction

The standalone standard-library checker generates all its scientific
fixtures. It compares 1,224 factored polynomials with 40,424 independent
intersection-recurrence equations through rank 48, then checks all 8,256
nonconstant spectral quotients through rank 128. Its separate geometric
oracle constructs the 3, 15 and 135 Lagrangians at ranks one through three
from literal binary symplectic spans and checks every pair's intersection
numbers. It enumerates all 6 and 32,766 nontrivial Boolean cuts at ranks
one and two, and checks 81 attaining point cuts across the three graphs.
The rank-two census has 170 attaining nontrivial subsets; no all-rank
classification is inferred from that count.

Fifteen controls reject altered polynomial coefficients, a changed edge
eigenvalue, smaller false ceilings and invalid checker domains. Explicit
exceptions keep the same checks active under optimized Python. Finite
checks corroborate the written proof; they are not its universal argument.

From the repository root with the local CPython 3.12.15 interpreter:

```sh
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -I -W error -B research/binary-dual-polar-cut-ceiling/verify.py
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -I -O -W error -B research/binary-dual-polar-cut-ceiling/verify.py
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -B scripts/research_package.py --package research/binary-dual-polar-cut-ceiling
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -B scripts/check_research_commit.py --package research/binary-dual-polar-cut-ceiling --check-source-history --check-modes
```

Any CPython 3.12 executable may replace that path. The direct marker is
`PASS_BINARY_DUAL_POLAR_CUT_CEILING`. Each run prints its computed counts
and a generated receipt binding all six package files by SHA256. It writes
no files. Both registered checks have a 120-second limit. Direct replay
also works from a Git-free package export with no predecessor files.

## Scope, sources and credit

The `complex-rank-shape-decoder` card explicitly changes to a certificate
limitation, rather than a physical supplier optimization. This helps delimit
what stronger lower-bound searches must change. It does not improve the
complete signed invoice. Canonical base is
`e54523517d0f03e58352e647c403a5af3e525a57`, priority revision eight. The
catalogue's #5599 context is historical; selected #7000 remains unchanged.
[TASK.json](TASK.json) records the scope change, evidence and next question.
[SOURCES.json](SOURCES.json) pins the exact predecessor induction note,
catalogue and selected record as historical provenance. None is a runtime
scientific input, and no predecessor file is copied or modified.

Credit Jacob Sussman, `jacobalansussman`, for PR4628's question, finite
evidence, point-cut count and coordinate-certificate application, and eumemic,
Swapnil Jain and the native/kept-subspace contributors credited there.
Dual-polar distance-regular spectra and spectral Dirichlet-form comparison
are classical. [Lee and Tanaka, SIGMA 14 (2018), 009](https://doi.org/10.3842/SIGMA.2018.009)
provide standard background; the proof here derives the binary formulas
directly. No global literature-priority claim is made.

The [advance claim](https://github.com/CrocSwap/integer-mult-bounds/discussions/1166#discussioncomment-19601539)
followed live PR/Discussion, account-history, local-reservation and
main/merged distinctive-term searches. The overlap cache covers 846 open
PRs and 7,482 Discussion records at its recorded timestamps, with incomplete
sparse local-note coverage. Distinctive searches also inspect main's README
and PROOF files. Unpublished work and arbitrary historical code remain
coverage limits.

## Limitations and unrun obligations

The theorem covers binary symplectic dual-polar graphs and their stated
cut/coordinate-certificate model. It does not classify every equality cut,
extend to other fields, construct a word or supply a physical consumer.
Native realization, legal access, ownership/restoration, stock, compiler,
precision/tape, analytic recovery and a complete supplier invoice remain
unproved here. All five selected framework assumptions remain. Selected
cold replay, Lean, Docker, independent human and formal proof-kernel review
are unrun or unclaimed.

## Open next questions

Which Boolean cuts attain equality at rank at least three? The proof
restricts their indicators to the constant and first adjacency eigenspaces,
but does not classify the resulting zero-one functions. Start with the
135-vertex rank-three graph. A complete census or a structural classification
would identify every sharp coordinate certificate; neither is supplied here.

## Code constraints

Every function is shorter than sixty physical lines. Loops have fixed
fixture bounds: rank at most 128 for scalar spectra, rank at most 48 for
polynomial checks, dimension at most six for literal geometry, at most
315 intermediate isotropic spaces and at most 32,766 cuts per graph. No
recursion, subprocess, network operation, callback registration or dynamic
code is used. Immutable tuples hold geometric spans; no shared mutable
cache exists. Explicit exceptions enforce all scientific checks.

Power of Ten rule 3 deviation: CPython and Fraction allocate during this
bounded exact offline check. At most 135 vertices and 18,225 pair cells
are retained for one graph; the largest enumerated subset family is
streamed. Polynomial degrees are bounded by 48 in the replay, and spectral
powers by 4,160 bits. Six regular package files of at most 128 KiB each
are read only for receipt hashes. These bounds constrain live mathematical
state, not interpreter internals. Rule 9's managed references and standard
iterator dispatch have fixed known targets and finite fixture sizes. Both
warning-strict registered modes and in-memory syntax checks supply the
available diagnostics. A fixed-allocation native implementation or external
static analysis is not claimed; revisit it only if this checker becomes
a resource bottleneck.
24 changes: 24 additions & 0 deletions research/binary-dual-polar-cut-ceiling/SOURCES.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,24 @@
{
"schema_version": 1,
"base_commit": "e54523517d0f03e58352e647c403a5af3e525a57",
"files": [
{
"path": "certificates/selected-result.json",
"usage": "provenance",
"git_blob": "75d77b328ea02ded535e7f23aee5b1e55d0bb1a8",
"sha256": "154c3d30455d778fd3343b6b4851f0fa4665f0f7625b0adb3675e2bafbbf6bf1"
},
{
"path": "docs/research/agent-task-catalog.json",
"usage": "provenance",
"git_blob": "3404fbdd5e7529a0674eef26114b5071e4586cec",
"sha256": "ead6cdc9d0af6d58e8a49744d8ec8c7c15e13d2ba93c276c434bcf9b51b93a81"
},
{
"path": "research/complex-small-m-lower-bounds/author-evidence/induction-across-block-sizes.md",
"usage": "provenance",
"git_blob": "9343bbd87a8e2d2b5b2b8fbca34da04033d95ce7",
"sha256": "0c2f6f5a07f1deccfa2b838c6d29bc6d86567069569890c37c820b03f7d888ef"
}
]
}
Loading
Loading