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
44 changes: 44 additions & 0 deletions research/no-edge-total-image-bound/CREDITS.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,44 @@
# Credit and source boundary

Ali Romman prepared this package with substantial OpenAI Codex assistance.
New work is Apache-2.0. Independent human and formal-kernel review are
unclaimed. No predecessor package is needed at runtime.

CubeJerry supplied the intersecting diagonal-dependency argument and the
sharp dimension-three total bound in
[PR #6838](https://github.com/CrocSwap/integer-mult-bounds/pull/6838),
[exact proof](https://github.com/CubeJerry/integer-mult-bounds/blob/f177c4f27e63dfa2e886033bb6a4e2dcd91f40b6/research/bit-zero-edge-image-bound/PROOF.md).
The new proof extends the coordinate-column method to every diagonal
relation at once.

unpaidmoon supplied the normal/column triple classification, the sharp
dimension-four bound in
[PR #7139](https://github.com/CrocSwap/integer-mult-bounds/pull/7139),
[exact proof](https://github.com/unpaidmoon/integer-mult-bounds/blob/007869ccdc22ee591b5a88d99a35f02511bffbee/research/no-edge-dim4-image-at-most-six/PROOF.md),
and the seven-image construction in
[PR #7237](https://github.com/CrocSwap/integer-mult-bounds/pull/7237) and
[PR #7281](https://github.com/CrocSwap/integer-mult-bounds/pull/7281),
[exact source](https://github.com/unpaidmoon/integer-mult-bounds/blob/0cd5324c41490b856d0e94daa8b011683fc153f6/research/no-edge-dim5-fresh-link/PROOF.md).
The seven-image witness is rechecked as prior sharpness evidence.

Nick Petalas supplied the binary fiber lemma in
[PR #5921](https://github.com/CrocSwap/integer-mult-bounds/pull/5921),
as credited in the exact proofs of PR #6838 and PR #7139. A short proof is
included here. The new total bound does not require the fiber lemma; it is
used for the irreducible equality boundary and complete sharp pools.

Ali Romman's image-basis theorem in
[PR #7686](https://github.com/CrocSwap/integer-mult-bounds/pull/7686),
[exact proof](https://github.com/aliromman/integer-mult-bounds/blob/9da06cbe0b888293cbb249cb29eef0f37addaa4f/research/no-edge-image-basis-bound/PROOF.md),
and extraction counterexample in
[PR #7688](https://github.com/CrocSwap/integer-mult-bounds/pull/7688),
[exact proof](https://github.com/aliromman/integer-mult-bounds/blob/d9aba79eaed566640353efcb0a27b0d5649f705a/research/no-edge-basis-extraction-failure/PROOF.md),
supplied the immediate scope question. Their exact verifiers supply the
reused binary matrix, span, lift, orbit and subspace-enumeration primitives:
[PR #7686 verifier](https://github.com/aliromman/integer-mult-bounds/blob/9da06cbe0b888293cbb249cb29eef0f37addaa4f/research/no-edge-image-basis-bound/verify.py),
[PR #7688 verifier](https://github.com/aliromman/integer-mult-bounds/blob/d9aba79eaed566640353efcb0a27b0d5649f705a/research/no-edge-basis-extraction-failure/verify.py).

The disjoint-circuit structure, the sharp total bound for every k, and the
general irreducible consequence are the new result. Earlier examples,
dimension-specific bounds and normal-form classifications retain their
source credit.
312 changes: 312 additions & 0 deletions research/no-edge-total-image-bound/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,312 @@
# A sharp total-image bound for every no-edge pattern dimension

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

## 1. Statement

Work over F2 in a fixed ambient coordinate space. Let C be any collection
of rank-one idempotents `P=u v^T`, with `v^T u=1`. Assume that every
off-diagonal pattern off(P) is nonzero. Let H be their pattern span and V
their joint image. Write `k=dim H` and `n=dim V`. Assume that H contains no
off-diagonal matrix unit.

**Total-image theorem.** `2n <= 3k`. Thus `n <= floor(3k/2)`. The bound is
sharp for every positive integer k, even when C is the complete pool of H.

**Irreducible consequence.** If `n>0` and C acts irreducibly on V, then
`2n <= 3k-1`. The total-bound equality at even k is always reducible.

The theorem applies to incomplete collections and does not assume that an
image-basis subcollection is irreducible. The proof uses the whole space of
diagonal relations among any independent image subcollection. It also proves
that the pattern relations of such a subcollection are generated by disjoint
circuits, each of size at least three. A circuit is a nonempty relation
support that contains no smaller nonempty relation support.

**At pattern dimension five, the maximum total image is seven.** The existing
seven-dimensional irreducible example also attains the irreducible maximum.
This answers the open upper-bound question in
[PR #7281](https://github.com/CrocSwap/integer-mult-bounds/pull/7281),
[exact source](https://github.com/unpaidmoon/integer-mult-bounds/blob/0cd5324c41490b856d0e94daa8b011683fc153f6/research/no-edge-dim5-fresh-link/PROOF.md).
It does not classify every attaining pool up to coordinate permutation.

The dimension-four bound in
[PR #7139](https://github.com/CrocSwap/integer-mult-bounds/pull/7139),
[exact proof](https://github.com/unpaidmoon/integer-mult-bounds/blob/007869ccdc22ee591b5a88d99a35f02511bffbee/research/no-edge-dim4-image-at-most-six/PROOF.md),
uses finite pattern cases and support arguments. The present proof extends
the column-support argument to all diagonal relations at once. It uses no
finite classification step. The two-block construction, the small normal
triple and the earlier dimension-three/four bounds are prior work.

CubeJerry's intersecting-dependency argument in
[PR #6838](https://github.com/CrocSwap/integer-mult-bounds/pull/6838),
[exact proof](https://github.com/CubeJerry/integer-mult-bounds/blob/f177c4f27e63dfa2e886033bb6a4e2dcd91f40b6/research/bit-zero-edge-image-bound/PROOF.md),
is the earlier column-support method. The present argument applies it to
the full diagonal-relation space rather than two selected triples.

## 2. Select an image basis, with no irreducibility premise

Choose n members `P_i=u_i v_i^T` whose images are a basis of V. The matrices
P_i are independent: in a relation among them, compare each column and use
independence of the u_i and the fact that each v_i is nonzero.

Let A be their n-dimensional span, let `k0=dim off(A)`, and let K be the
space of diagonal matrices in A. Then

```text
d = dim K = n-k0, k0 <= k.
```

Choose a basis `D_1,...,D_d` of K. Write each D_j in the P_i basis. For each
projector i, let `r_i in F2^d` contain its coefficients in these d expressions.
For `s in F2^d`, the corresponding diagonal matrix is

```text
D(s) = sum_i (s dot r_i) P_i.
```

The vectors r_i span F2^d; otherwise a nonzero combination of the D_j would
vanish. Group the projectors by equal nonzero r_i. Put those with r_i zero
in a separate free group. No covector independence is assumed.

## 3. Each nonzero group has its own diagonal sum

Fix an ambient column b. Its value in D(s) is

```text
sum_i (s dot r_i) v_i(b) u_i.
```

Since D(s) is diagonal, all these vectors lie in the one-dimensional space
spanned by e_b. The u_i are independent. Thus the map from s to the
coefficient vector `((s dot r_i) v_i(b))_i` has rank at most one. It follows
that all nonzero r_i with `v_i(b)=1` are equal: over F2, two different
nonzero linear functionals are independent and would give rank at least two.

For a nonzero group label lambda, define

```text
Q_lambda = sum_(r_i=lambda) P_i.
```

If this sum has a nonzero column b, that column receives no contribution
from any other nonzero group. Choose s with `s dot lambda=1`. Column b of
Q_lambda is then column b of D(s), so it is diagonal. All columns of
Q_lambda are therefore diagonal. The sum is nonzero because the P_i are
independent.

The indicator vector of each group is now a relation among the patterns.
These indicators have disjoint supports and are independent. If there are
t nonzero groups, this gives `t<=d`. The r_i span F2^d, so `t>=d`.
Consequently **there are exactly d groups**, their labels are a basis of
F2^d, and their indicators generate the entire relation code. Every circuit
is one of these disjoint groups. The zero group consists of points in no
relation.

Let S_lambda be the coordinate support of the diagonal Q_lambda. Every
covector in that group is supported on S_lambda: a zero column of the sum
would force all its covector entries to be zero by independence of the
images. Conversely a nonzero covector entry gives a nonzero sum column.
Distinct S_lambda are disjoint, by the column argument above.

If a group has m members, then

```text
1 <= rank(Q_lambda) = |S_lambda| <= m,
|S_lambda| = m modulo 2.
```

The rank bound follows from its m-dimensional image span. The parity identity
is its trace: each rank-one idempotent has trace one over F2.

## 4. Every relation group has at least three members

A group of size one would make its sole projector diagonal, contrary to the
nonzero-pattern assumption.

For a group of size two, its diagonal sum has nonzero even rank at most two,
so it has rank two. Its two-dimensional image span is exactly the coordinate
space S_lambda. Both covectors are also supported there. A noncoordinate
rank-one idempotent on two binary coordinates has exactly one off-diagonal
one: if both off-diagonal entries were one, both factors would be `(1,1)`
and their pairing would be zero. Its pattern would therefore be a matrix
unit, contrary to the hypothesis.

There are d disjoint groups of size at least three. Hence

```text
n >= 3d = 3(n-k0),
2n <= 3k0 <= 3k.
```

This proves the total-image theorem. For an empty collection the same bound
is `0<=0`. Completeness and irreducibility were not used.

## 5. Equality at even pattern dimension is reducible

Suppose `2n=3k` and `n>0`. The preceding inequalities force `k0=k`, no free
group, and exactly d groups of size three, where `n=3d` and `k=2d`. The
chosen image basis therefore spans the full pattern space H. The complete
pool of H has the same joint image V: it contains the basis, while the
total-image bound prevents it from adding a dimension. It is enough to show
that this complete pool is reducible on V.

For a size-three group, the active support S has odd size at most three.
There are two possibilities.

If `|S|=1`, every covector equals the same coordinate covector e_c. Each
image has the form `e_c+t_i`, and the diagonal-sum identity gives
`t_1+t_2+t_3=0`. The three t_i are distinct and nonzero; they are the three
nonzero vectors of a two-dimensional space T. No vector of T has weight
one, because its column-c pattern would be a matrix unit. Call this a
column triple.

If `|S|=3`, the independent images span that coordinate three-space. Their
sum is its identity projection. Expanding this identity in the image basis
shows that the covectors are its dual basis. The three projectors are
therefore mutually orthogonal and supported entirely within S. No choice
of a normal form is required here.

Let S_all be the union of these pairwise disjoint active supports. Every
pattern in H has zero columns outside S_all. Suppose a complete-pool
projector has a covector bit at a coordinate b outside S_all. Its zero
off-diagonal column forces its image to be e_b. Its pattern would be a
single row outside S_all. In the columns of a three-coordinate group,
patterns are supported within that group's own block, so this row is zero.
At a column-triple coordinate c, its column would be zero or e_b. The latter
is forbidden because T contains no weight-one vector. Thus the whole
pattern would be zero, a contradiction. Every complete-pool covector is
supported on S_all.

If at least one group is a column triple, then `|S_all|<3d=n`. The covectors
have a nonzero common kernel on V. The kernel is proper, since each
projector fixes its image. It is an invariant subspace, so the pool is
reducible.

If every group has three active coordinates, H is supported on disjoint
coordinate blocks. A noncoordinate rank-one idempotent whose pattern lies
in H must be supported in one block. Indeed, idempotence gives a coordinate
where both factors are one; the forbidden cross-block entries then force
both factors into that block. Each block has only three nonzero patterns,
and each has just one idempotent lift by the lemma in Section 6. Thus the
complete pool is precisely the original `3d` orthogonal projectors. Every
image line is invariant. It is reducible again.

Every member of the original collection belongs to this complete pool, and
both have image V. Its proper invariant subspace is therefore invariant for
the original collection as well. This rules out total-bound equality for
irreducible action. For odd k, the integer total bound already gives
`2n<=3k-1`; for even k the strict inequality does so. This proves the
irreducible consequence.

## 6. Nonunit patterns have a unique idempotent lift

This is the binary fiber lemma due to Nick Petalas, cited in the exact
source of PR #7139. A short proof is included to make the argument
self-contained.

Choose an off-diagonal one at `(a,b)` in a realized pattern p. In a lift
`x y^T`, it forces `x_a=y_b=1`. Column b fixes every x bit except x_b;
row a fixes every y bit except y_a. Call the two unknown bits alpha and beta.
Idempotence gives

```text
alpha + beta + sum_(j outside {a,b}) x_j y_j = 1.
```

If some known x bit outside `{a,b}` is one, its column-a entry fixes beta.
If some known y bit outside `{a,b}` is one, its row-b entry fixes alpha.
In either case, idempotence fixes the other bit. If there are no such bits,
both factors are supported on `{a,b}`. Idempotence then gives
`alpha+beta=1`, so the reverse off-diagonal entry `alpha beta` is zero.
The pattern is exactly the matrix unit at `(a,b)`. Thus every realized
nonzero pattern other than a matrix unit has exactly one idempotent lift.

## 7. Sharpness and the dimension-five endpoint

The standard normal triple on three coordinates is

```text
(e_1+e_2)(e_0+e_2)^T,
(e_0+e_2)(e_1+e_2)^T,
(e_0+e_1+e_2)(e_0+e_1+e_2)^T.
```

Its images are independent, its pattern space has dimension two, its three
nonzero patterns have weights three, three and six, and its complete pool
is the stated triple. This construction and its classification are prior
work, recorded in the exact source of PR #7139.

For `k=2d`, take d disjoint normal triples. For `k=2d+1`, add on three fresh
coordinates the single projector `e_0(e_0+e_1+e_2)^T`. This last pattern has
weight two and dimension one. Disjoint blocks have independent pattern
spaces and independent images. The complete pool is the union of the block
pools: a rank-one idempotent with nonzero pattern cannot cross the blocks,
by the argument in Section 5. The total image is respectively `3d` or
`3d+1`, which is `floor(3k/2)`. These prove sharpness for every k.

For the irreducible dimension-five endpoint, the prior two-block family in
[PR #7237](https://github.com/CrocSwap/integer-mult-bounds/pull/7237) and
[PR #7281](https://github.com/CrocSwap/integer-mult-bounds/pull/7281) adds the
link `x x^T`, with `x=e_2+e_5+e_6`, to normal triples on coordinates 0..2
and 3..5. Its complete pool has seven projectors, pattern dimension five
and irreducible image dimension seven. The verifier rechecks all 31 nonzero
patterns and all 127 nonzero vector orbits of this exact witness. It is
used as prior sharpness evidence, not claimed as a new construction.

The new upper bound is seven, so both the total and irreducible maxima at
pattern dimension five are now seven. The coordinate classification of all
attaining pools remains open.

## 8. Verification and limitations

The verifier checks every image-basis collection with at most three
projectors at widths three and four: 224036 collections in total. Exactly
95046 have nonzero unit-free pattern spans. Each passes the complete
diagonal-relation certificate, including the disjoint active supports and
the full relation code. Larger odd column circuits, even orthogonal circuits
and a mixed circuit collection check that the proof is not specific to
triples or to independent covectors.

It also enumerates all 2825 off-diagonal pattern subspaces at width three.
Of these, 1330 are unit-free, 336 have empty pools and 595 have irreducible
nonempty pools. Every complete pool passes both applicable bounds. All three
nonempty total-bound equality spaces occur at k=2 and are reducible.

Complete-pool checks cover the sharp families for `1<=k<=12`, with 8178
nonzero patterns in total. Five even-equality fixtures include column,
orthogonal, mixed and overlapping-support cases. The prior seven-image
irreducible witness supplies the dimension-five endpoint.

Three controls violate the total bound after a hypothesis is removed.
The idempotents with factors `(1,3)` and `(3,2)` at width two have image
dimension two and pattern dimension one, but their pattern is a matrix unit.
The factors `(6,1)` and `(7,1)` at width three also have image dimension two
and unit-free pattern dimension one, but the first matrix is nilpotent.
A coordinate projector has image dimension one and zero pattern dimension.
Integers encode binary supports with coordinate zero in the low bit.

The all-width proof does not depend on the finite checks or an unproved
basis-extraction step. In particular, it is consistent with the extraction
counterexamples in
[PR #7688](https://github.com/CrocSwap/integer-mult-bounds/pull/7688),
[exact proof](https://github.com/aliromman/integer-mult-bounds/blob/d9aba79eaed566640353efcb0a27b0d5649f705a/research/no-edge-basis-extraction-failure/PROOF.md).
The independent-image theorem and equality classification of
[PR #7686](https://github.com/CrocSwap/integer-mult-bounds/pull/7686),
[exact proof](https://github.com/aliromman/integer-mult-bounds/blob/9da06cbe0b888293cbb249cb29eef0f37addaa4f/research/no-edge-image-basis-bound/PROOF.md),
retain credit; the present bound applies without that extra irreducibility
premise. The small exact matrix and lift routines are reused from those
packages, as recorded in CREDITS.md.

No new kappa, paid BIT word, native conversion or full invoice is supplied.
The selected frontier and all five assumptions remain unchanged. Independent
human and formal-kernel review are unclaimed.

## Open next questions

Classify the complete pattern-dimension-five pools that act irreducibly on
a seven-dimensional image. An arbitrary image basis has two disjoint
diagonal-relation groups and one remaining projector, or one group of size
four and one of size three. Analyze these two structures without assuming
that the selected basis is irreducible. The total-image bound is no longer
the missing step; the attaining geometry is.
Loading
Loading