Skip to content
Draft
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
321 changes: 321 additions & 0 deletions research/nonidempotent-n-constant-insertion/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,321 @@
# A constant insertion beside the non-idempotent factor N

No new kappa. The selected frontier stays 363005817/100000000000.

PR #7027 exhibits a non-idempotent 4-by-4 matrix `N` over `F_2` whose gate
`I + x N` is the leading factor of a triple multiplying to `(1 + x) I_4`.
One matrix `R` is inserted immediately beside that gate, with `e` central and
`e^2 = 0`. The coefficient of `e` is never `I`. On the exhibited stage-0
triple, the same coefficient equals `(1 + x) I_4` if and only if `R = I_4`.

The evidence is split by class:

1. A proof, for every trailing matrix, that the right-hand insertion has
`(0, 0)` entry in the principal ideal `(1 + x)`. The identity is therefore
impossible for every matrix `R` with entries in `F_2[x, y, z]`.
2. A proof, by exact cancellation in that integral domain, that the exhibited
stage-0 triple has coefficient `(1 + x) I_4` if and only if `R = I_4`, on
either side.
3. An exhaustive finite census of the `65,536` constant 4-by-4 matrices, on
both sides. The domain is finite, every bit pattern is visited once, and
the count is the whole domain.
4. A proof that the same one insertion, read in the eight-dimensional word
obtained by embedding the triple on stage 0, never has coefficient `I_8`.
This is not a census of the `2^64` constant 8-by-8 matrices.

The diagonal entry formula of PR #7039 and PR #7103 is not used. `N` is not
diagonal, and the formula is false for this gate.

## 1. Hypotheses

The scalar ring is `F_2[x, y, z]`. It is an integral domain. The homomorphism
onto `F_2` given by `x ↦ 1`, `y ↦ 0`, `z ↦ 0` will be called evaluation at
`(1, 0, 0)`. The scalar `e` is central and `e^2 = 0`. A constant matrix has
entries in `F_2`. The coefficient of `e` is the degree-1 term in the quotient
by `e^2`.

The exhibited factor, copied from the bit rows in PR #7027, is

```text
N = [[1, 0, 0, 0],
[0, 0, 1, 0],
[0, 0, 0, 0],
[0, 0, 0, 0]],
B = E_22,
C = [[0, 0, 0, 0],
[0, 0, 1, 0],
[0, 0, 1, 0],
[0, 0, 0, 1]].
```

Bit `c` of row `r` is the entry `(r, c)`, so these are the rows
`(1, 4, 0, 0)`, `(0, 2, 0, 0)` and `(0, 4, 4, 8)`. Then `N^2 = E_11`, so `N`
is not idempotent. The eight-dimensional coordinates are

```text
(core_0, core_1, leaf0_0, leaf0_1, leaf1_0, leaf1_1, leaf2_0, leaf2_1).
```

Stage `j` uses the core together with leaf `j`. Embedding a 4-by-4 matrix on
stage 0 writes it on coordinates `(0, 1, 2, 3)`.

## 2. The expansion

Row 0 of `N` is `(1, 0, 0, 0)`, so

```text
I + x N = [[1 + x, 0, 0, 0],
[0, 1, x, 0],
[0, 0, 1, 0],
[0, 0, 0, 1]].
```

For an arbitrary matrix `M` of the same size,

```text
(I + x N)(I + e R) M = (I + x N) M + e (I + x N) R M.
```

The coefficient of `e` is `K = (I + x N) R M`. Its `(0, 0)` entry uses only
row 0 of `I + x N`:

```text
K_00 = (1 + x) (R M)_00.
```

The product `(R M)_00` lies in `F_2[x, y, z]`. Evaluation at `(1, 0, 0)`
sends `1 + x` to `0` and sends `1` to `1`, so `(1 + x) f = 1` has no
solution. Therefore `K` is not the identity.

The trailing factor `M` is arbitrary. It may be `I`, the exhibited product
`(I + x B)(I + x C)`, or the product of every later gate in the embedded
word below. No choice of those gates removes the factor `1 + x`.

The same row survives embedding. On coordinates `(0, 1, 2, 3)` the embedded
gate still has row 0 equal to `(1 + x)` times the first standard basis
covector. For every 8-by-8 trailing matrix, the coefficient of a right-hand
insertion therefore has `(0, 0)` entry in the ideal `(1 + x)`, and is not
`I_8`.

This is the expansion the question asks for. It does not depend on the
diagonal identity.

## 3. The exhibited stage-0 triple

Direct multiplication gives

```text
(I + x B)(I + x C) = [[1, 0, 0, 0 ],
[0, 1+x, x+x^2, 0 ],
[0, 0, 1+x, 0 ],
[0, 0, 0, 1+x ]]
```

and `(I + x N)` times that matrix is `(1 + x) I_4`. Write `M` for the
displayed product.

**Left insertion.** The coefficient in `(I + e R)(I + x N) M` is
`(1 + x) R`, because `(I + x N) M = (1 + x) I_4`. The equation
`(1 + x) R = I_4` asks `(1 + x) R_00 = 1`, which section 2 forbids. The
equation `(1 + x) R = (1 + x) I_4` is `(1 + x)(R_ij - δ_ij) = 0`. The
multiplier is nonzero, so `R = I_4` is the only solution in
`F_2[x, y, z]`.

**Right insertion.** Let `Q = (I + x N) R` and `K = Q M`. The shapes above
give

```text
Q_0j = (1 + x) R_0j,
Q_1j = R_1j + x R_2j,
Q_2j = R_2j,
Q_3j = R_3j,
```

and then

```text
K_i0 = Q_i0,
K_i1 = (1 + x) Q_i1,
K_i2 = x(1 + x) Q_i1 + (1 + x) Q_i2,
K_i3 = (1 + x) Q_i3.
```

In entries, with `π = 1 + x`,

```text
K_00 = π R_00
K_10 = R_10 + x R_20
K_20 = R_20
K_30 = R_30
K_01 = π^2 R_01
K_11 = π R_11 + x π R_21
K_21 = π R_21
K_31 = π R_31
K_02 = x π^2 R_01 + π^2 R_02
K_12 = x π R_11 + x^2 π R_21 + π R_12 + x π R_22
K_22 = x π R_21 + π R_22
K_32 = x π R_31 + π R_32
K_03 = π^2 R_03
K_13 = π R_13 + x π R_23
K_23 = π R_23
K_33 = π R_33.
```

Every identity is an equality in `F_2[x, y, z]`, not a sample of monomials.
In characteristic two, `π^2 = 1 + x^2` and `x π = x + x^2`.

The identity `K = I_4` requires `K_00 = π R_00 = 1`, which is impossible.
The remaining fifteen entries are not needed for that conclusion.

The identity `K = π I_4` cancels in the following order. Each pivot is a
nonzero element of `F_2[x]`, and the ring is an integral domain, so the
quotient is unique in `F_2[x, y, z]`.

```text
K_20 = R_20 = 0
K_30 = R_30 = 0
K_10 = R_10 + x R_20 = 0 so R_10 = 0
K_00 = π R_00 = π so R_00 = 1
K_21 = π R_21 = 0 so R_21 = 0
K_31 = π R_31 = 0 so R_31 = 0
K_11 = π R_11 + x π R_21 = π so R_11 = 1
K_23 = π R_23 = 0 so R_23 = 0
K_13 = π R_13 + x π R_23 = 0 so R_13 = 0
K_33 = π R_33 = π so R_33 = 1
K_01 = π^2 R_01 = 0 so R_01 = 0
K_03 = π^2 R_03 = 0 so R_03 = 0
K_02 = x π^2 R_01 + π^2 R_02 = 0 so R_02 = 0
K_22 = x π R_21 + π R_22 = π so R_22 = 1
K_32 = x π R_31 + π R_32 = 0 so R_32 = 0
K_12 = x π R_11 + x^2 π R_21 + π R_12 + x π R_22 = 0
```

The last equation reduces to `π R_12 = 0` after the earlier values are
substituted, because the two copies of `x π` cancel. Thus `R_12 = 0`. Every
entry has been determined, and `R = I_4`.

The same calculation shows where the diagonal identity stops applying.
That identity would require `K_11 = π R_11`. The formula here is
`K_11 = π R_11 + x π R_21`. The matrix unit `E_21` has `K_11 = x π`, which
is not zero. PR #7039 records that its diagonal argument does not cover this
factor. The `(0, 0)` entry is a different computation, using row 0 of `N`
rather than diagonality of every gate.

The left-hand and right-hand statements are the two readings of "beside".
The right-hand expansion is the one the question names first. Both use the
same factor `N` and the same two remaining stage-0 gates.

Any other middle matrix `B'` with `C' = I + N + B'` and
`(I + x N)(I + x B')(I + x C') = (1 + x) I_4` has the same two obstructions
to coefficient `I_4`: the right-hand `(0, 0)` entry is covered by section 2
for every trailing matrix, and the left-hand coefficient is `(1 + x) R`
because the pure product is `(1 + x) I_4`. This package does not recount the
128 middle matrices of PR #7027. It also does not classify, for a
non-exhibited companion, the solutions of coefficient `(1 + x) I_4`.

## 4. The constant census

A constant 4-by-4 matrix is an integer from `0` through `65,535`, with row
`r` stored in bits `4r` through `4r + 3`. Each of the `2^16` matrices occurs
once. Nothing stops early.

On the right of `N`, and on the left, the number of these matrices whose
coefficient equals `I_4` is `0`. The number whose coefficient equals
`(1 + x) I_4` is `1`, and that matrix is `I_4`, on each side.

The zero matrix has coefficient `0`. The matrix `I_4 + E_01` does not have
coefficient `(1 + x) I_4`. Those are controls. The census agrees with
sections 2 and 3 on the constant subdomain. It is not a substitute for the
polynomial argument, and the polynomial argument is not a sampling of the
census. The degree checks in `verify.py` exercise the division routine; they
are not a degree cutoff in the theorem.

## 5. The embedded three-stage word

Replace the three diagonal stage-0 gates of the nine-gate word by the
embedded factors `N`, `B` and `C`, in that order, and keep the six diagonal
gates of stages 1 and 2. Stage 1 multiplies by `1 + y` on coordinate `0`,
then on coordinate `1`, then on coordinates `4` and `5`. Stage 2 does the
same with `1 + z` on coordinates `0`, `1`, `6` and `7`.

The stage-0 product is `I + x E_0`. The full pure product `P` is therefore
the same diagonal operator as in the nine-gate word:

```text
P_00 = P_11 = (1 + x)(1 + y)(1 + z),
P_22 = P_33 = 1 + x,
P_44 = P_55 = 1 + y,
P_66 = P_77 = 1 + z.
```

**Right insertion, immediately after the embedded `I + x N`.** Section 2
already excludes coefficient `I_8`. For this particular tail, column 0 is
`(1 + y)(1 + z)` times the first basis vector, so the general formula
specialises to

```text
K_00 = (1 + x)(1 + y)(1 + z) R_00.
```

Evaluation at `(1, 0, 0)` sends the right-hand side to `0`. The same
conclusion holds for a polynomial entry `R_00`, not only for a bit.

**Left insertion, with `N` first.** The coefficient is `R P`. Since `P` is
diagonal, `K_00 = R_00 P_00`, and `P_00` is the same nonunit. The coefficient
is not `I_8`.

Both statements are bilinear, so they are checked on the 64 matrix units and
then extended by the ring action. A sum of two units is checked directly as a
control. This unit check is not an enumeration of `M_8(F_2)`.

The pure-product equation `K = P` is not solved. Its column 0 is. On the
right, that column is

```text
K_00 = (1 + x)(1 + y)(1 + z) R_00,
K_10 = (1 + y)(1 + z) (R_10 + x R_20),
K_20 = (1 + y)(1 + z) R_20,
K_30 = (1 + y)(1 + z) R_30,
K_i0 = (1 + y)(1 + z) R_i0 for i = 4, 5, 6, 7.
```

`P` has `(1 + x)(1 + y)(1 + z)` in position `(0, 0)` and zeros below it.
Cancelling the nonzero factor `1 + y + z + y z` forces `R_00 = 1`,
`R_20 = 0`, hence `R_10 = 0`, and `R_i0 = 0` for every `i > 0`. Column 0 of
`R` is the first standard basis vector. The other seven columns are open.

## 6. Limits

The selected record remains `363005817/100000000000`. No native caller,
compiler bill, or thin map is constructed. Profiles `g = 13` and `g = 31`
are not considered. PR #6956 is not rerun.

The result is not the two-gap classification of PR #7103, and it is not the
polynomial count of PR #7191. Those papers insert constants or linear
polynomials into the diagonal word. The factor here is `N`.

A matrix over the rational function field is outside the statement. The
scalar matrix `(1/(1 + x)) I_4` satisfies `(1 + x) R = I_4` in that larger
ring, which is why the obstruction is that `1 + x` is not a unit. It is not
a solution in `F_2[x, y, z]`.

`current-priorities.md` at this base can still print an older decimal. The
live selected record is `363005817/100000000000`, from
`certificates/selected-result.json`. This package uses that record and edits
neither file. PR #7027, PR #7039 and PR #7103 name the previous fraction
`781451/250000000`; that is not the frontier this package is measured
against, and this package does not move either value.

## Credit

The paired-star contract, the operator `T`, and the pure target are from PR
#5747 (Ali Romman), including PR #5673, PR #5631, and PR #5617. The
T-commuting projector obstruction is PR #6956. The factor `N`, the triple,
and the embedding convention are PR #7027, head
`1ba2ac5adfe58fbe9dba34d50d4e5d30633ce3d0`. The diagonal obstruction for the
idempotent gates, and the statement that it does not cover `N`, are PR
#7039, head `e4ddcd2fde8c46b7e528e8806ab72d9bf7690978`. The question answered
here is open question 3 of PR #7103, which is the same non-idempotent slot
already named as question 2 of PR #7039. Prepared by unpaidmoon with
substantial Cursor Grok assistance. Independent human review and a
proof-assistant proof are unclaimed. Apache-2.0.
Loading