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
130 changes: 130 additions & 0 deletions research/selected-phase-linear-statistic/PREDECESSOR-RECIPE.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,130 @@
{
"base_commit": "e54523517d0f03e58352e647c403a5af3e525a57",
"native_event_index": 24472,
"event": [
"MOVE",
415,
3,
99,
false
],
"side": "right",
"rank": 1,
"recipe": {
"matrix_rows": [
239,
129,
130,
132,
136,
16,
160,
192,
256
],
"offset": 1,
"quadratic_rows": [
111,
111,
111,
111,
0,
111,
111,
0,
0
],
"linear_Z4": [
3,
3,
3,
3,
0,
3,
3,
2,
0
],
"phase": 1
},
"lowered_recipe": {
"rank": 1,
"left": {
"matrix_rows": [
3,
5,
9,
17,
32,
65,
129,
1,
256
],
"offset": 0,
"quadratic_rows": [
0,
0,
0,
0,
0,
0,
0,
0,
0
],
"linear_Z4": [
2,
0,
0,
0,
0,
0,
0,
0,
0
],
"phase": 0
},
"right": {
"matrix_rows": [
239,
129,
130,
132,
136,
16,
160,
192,
256
],
"offset": 1,
"quadratic_rows": [
111,
111,
111,
111,
0,
111,
111,
0,
0
],
"linear_Z4": [
3,
3,
3,
3,
0,
3,
3,
2,
0
],
"phase": 1
}
},
"catalogue_pairs_examined": 1,
"native_compressed_sha256": "2b94953dc14a122e3ef089633f05b58ec689ec5156feb7c784ed3e137dc01883",
"native_canonical_sha256": "91429a6b96729292a8764836e00c87a6f7cd3b76f05159033f4867f69a98d097"
}
124 changes: 124 additions & 0 deletions research/selected-phase-linear-statistic/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,124 @@
# The exact linear information of one selected tensor phase

This scoped written proof was produced by Nick Petalas. Apache-2.0.

## Contract and result

Take the right monomial recipe in CubeJerry / Jerry Chen's
[#7426](https://github.com/CrocSwap/integer-mult-bounds/pull/7426), head
`172f4535ce8dcff79f626410af845f757c1563be`. Its unchanged factual recipe is
retained as PREDECESSOR-RECIPE.json. We study only its diagonal absolute
Gaussian phase, evaluated on the coordinates before the monomial's affine
address route. We cite the source's recipe binding; we do not replay that
binding or its literal tape program.

Let S={0,1,2,3,5,6}. In its strict-upper convention the source phase is

q(x) = 1 + 3 sum_(i in S) x_i
+ 2 sum_(i<k, i,k in S) x_i x_k + 2 x_7 (mod 4).

The diagonal phase is i^q(x). Coordinates 4 and 8 are spectators. For f>=1,
the tensor phase on x in F2^(9f) is

Q_f(x) = sum_(j=1..f) q(x^(j)) (mod 4).

A binary affine-linear observation is Lx+c in F2^k. An arbitrary decoder of
its output must recover Q_f(x) on every input. The observation and decoder
are fixed, with no auxiliary access to x. Observation count means the
rank of L; deleting dependent output bits always gives an equivalent
statistic. No bound on the decoder's computational cost is asserted.

**Theorem.** The exact minimum observation rank is f+1. An attaining linear
statistic is

r_j = XOR_(i in S) x_(i,j), j=1..f,
b = XOR_(j=1..f) x_(7,j).

Its decoder is

Q_f(x) = f + 3 sum_j r_j + 2b (mod 4). (1)

In particular, fifteen quadratic pair terms per column disappear from
this phase description. Compared with retaining the two natural observations
r_j and x_(7,j) separately at every column, aggregating the latter saves
f-1 independent observation bits. The sharp bound is about binary linear
information. Q_f itself has only four possible values, so two nonlinear
output bits always suffice. That is a different observation contract.

## Contraction

This contraction is inherited from #7426's PROOF.md, Section 1, which
explicitly proves q(x)=1-(sum_S x_i mod 2)+2*x_7. We recall it to fix the
phase convention. For Boolean bits, the classical identity

XOR_i x_i = sum_i x_i - 2 sum_(i<k) x_i x_k (mod 4)

follows by expanding (1-product_i(1-2x_i))/2. Every term of degree at
least three is divisible by four. Multiplying by 3 modulo four gives
coefficient 3 on each linear term and coefficient 2 on each pair. Thus

q(x) = 1 + 3 r(x) + 2 x_7 (mod 4).

Summing over columns proves (1), because twice an integer depends modulo
four only on that integer's parity. This keeps the absolute constant f.
Replacing it by one would fail already at f=2. No phase is discarded as
an irrelevant scalar.

## All translation periods and the lower bound

For a function Q on a binary vector space define its translation-period
subgroup

K = {a : Q(x+a)=Q(x) for every x}.

It is a vector subspace, since successive translations by periods preserve
Q. A statistic Lx+c determines Q if and only if ker L is contained in K.
For necessity, x and x+a have the same observation whenever a is in ker L.
For sufficiency, every observation fiber is one such coset and Q is
constant on it. Therefore every sufficient statistic has rank at least
codim K. Taking any linear quotient by K attains this bound. The affine
offset c does not affect the argument.

We compute K for Q_f, without appealing to a finite census. Write
t_j=r_j(a) and d_j=a_(7,j). For bits u,t,

(u XOR t)-u = t-2ut as integers.

The contracted formula gives

Q_f(x+a)-Q_f(x)
= 3 sum_j t_j + 2 sum_j d_j + 2 sum_j r_j(x)t_j (mod 4). (2)

The sign in the last term is positive since -6=2 modulo four. If t_j=1,
toggle only coordinate 0 in column j of x. The last term changes by two,
while every other term stays fixed. Hence (2) cannot be zero for every x.
Every period must have t_j=0 for each j. Under those conditions (2) is
zero exactly when sum_j d_j is even. Conversely these same conditions
make (2) zero on every x. Thus the complete period subgroup is

K_f = {a : r_j(a)=0 for all j, sum_j a_(7,j)=0 in F2}. (3)

The f+1 equations are linearly independent. Each r_j has its own
coordinate 0, and the last equation uses coordinate 7, absent from S.
Consequently dim K_f=8f-1, and codim K_f=f+1. The displayed statistic has
exactly this kernel and attains the lower bound. This proves the theorem
for every positive f, including arbitrary binary-linear mixing of columns.

## Scope and inherited work

The complete tensor period subgroup and sharp linear-observation rank are
the contribution. The one-column contraction is credited to #7426.
Parity expansion and the
quotient-by-periods principle are classical. Their general principles are
not claimed as new. The phase recipe, native source binding and physical
implementation belong to CubeJerry #7426, retaining Ali Romman #5375 and
the #7000 native/Gaussian lineage. Dennys / antunishdPursuit #6500 also
records parity expansion for a different contact-family verification.

No physical tape reduction follows from rank alone. A producer must acquire
the parities, retain or regenerate the f column bits, decode their weight
modulo four, restore any dirty workspace and satisfy the source's packed
layout and return contracts. Its complete time, historical backing and
recursive invoice remain unpriced. This proof changes neither the affine
address map nor the rank-one C tensor. It supplies no new kappa, supplier
admission, complete multiplication theorem or formal proof-kernel acceptance.
134 changes: 134 additions & 0 deletions research/selected-phase-linear-statistic/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,134 @@
# The selected tensor phase needs exactly f+1 linear observations

Scoped written proof with exact finite corroboration, produced by Nick Petalas.
No new kappa is claimed. For every positive tensor column count f, the right
diagonal phase of selected MOVE24472 has exact binary linear observation
rank **f+1**, even when observations mix columns arbitrarily.

CubeJerry / Jerry Chen's [#7426](https://github.com/CrocSwap/integer-mult-bounds/pull/7426)
supplies the nine-coordinate recipe. Its fifteen quadratic pairs form the
clique on S={0,1,2,3,5,6}. That predecessor's PROOF.md, Section 1, already
proves the following contraction using the classical parity identity:

q(x) = 1 + 3 XOR_(i in S) x_i + 2 x_7 (mod 4).

For f columns, retain their f separate S-parities r_j and the single parity
b of all coordinate-7 bits. The complete absolute phase exponent is

Q_f(x) = f + 3 sum_j r_j + 2b (mod 4).

The inherited contraction removes the fifteen pair terms. The new joint
statistic saves f-1 observation bits compared with retaining the two natural
parities per column. It does not assert that #7426 allocates 2f physical bits
to them. The new lower bound follows from the complete translation-period subgroup,
of dimension 8f-1. Any affine-linear statistic determining the phase must
have its kernel inside that subgroup. [PROOF.md](PROOF.md) proves this for
all f; finite checks corroborate the stated source-specific identities.

The observation restriction matters. Computing Q_f directly is nonlinear
and always gives a two-bit output. The f+1 bound concerns linear measurements
before arbitrary decoding, rather than output size or streaming workspace.

## Reproduce

From the repository root, with the supplied CPython 3.12 interpreter:

```sh
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -I -W error -B research/selected-phase-linear-statistic/verify.py
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -I -W error -O -B research/selected-phase-linear-statistic/verify.py
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -B scripts/research_package.py --package research/selected-phase-linear-statistic
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -B scripts/check_research_commit.py --package research/selected-phase-linear-statistic --check-source-history
```

Another machine may substitute its CPython 3.12 path. Only the standard library
is required. Direct replay is read-only and works in a Git-free export. It
prints `PASS_SELECTED_PHASE_LINEAR_STATISTIC` inside a generated JSON receipt
bound to all seven package-file hashes. No receipt is handwritten or stored
as an expected answer. Normal and optimized commands use explicit exceptions.

The literal polynomial oracle uses the source's strict-upper quadratic rows
and all nine linear coefficients. The contracted evaluator uses only parity
and the displayed decoder. They agree on all 512 one-column and 262,144
two-column addresses. A complete 262,144-comparison one-column truth-table
derivative calculation finds 128 periods. Its constant-derivative table
classifies all 262,144 two-column shifts, finding 32,768 periods. This last
classification uses independence of the two input columns, not 2^36 pairwise
comparisons. All 512 possible one-bit linear maps fail on their actual
one-column fibers. These are new phase-information checks; no predecessor
binding, tape program or mathematical suite is imported or replayed.

An additional 1,024 deterministic fixtures cover f=1..64, with 3,072 Gaussian
forward/inverse pairs. Seven semantic negative controls, one positive
joint-column cancellation control and five invalid-domain controls pass.
The larger fixtures are not exhaustive. The unbounded theorem rests on the
written proof, with no human or proof-kernel acceptance claimed.

## Sources, scope and credit

Credit CubeJerry #7426 at `172f4535ce8dcff79f626410af845f757c1563be`
for the supplied MOVE24472 recipe, its binding and literal phase implementation.
[PREDECESSOR-RECIPE.json](PREDECESSOR-RECIPE.json) retains that factual file
byte for byte; its original native matrices and event metadata are unchanged.
The SHA256 and exact download URL are in [TASK.json](TASK.json). Its scientific
source-binding theorem is cited. Checking this exported recipe's hash does
not independently establish its connection to the selected native catalogue.
Publication requires no pending predecessor commit in local Git history.

Retain Ali Romman #5375's multicounter phase-production lineage and #7000's
DreamingOfClouds, Ali Romman, Cinna Davis, zxv, unpaidmoon, icekylinx and earlier
native/Gaussian contributors. Dennys / antunishdPursuit #6500 also records the
classical parity expansion for a different phase-contact family. Parity
expansion and quotienting by translation periods are credited methods. The
contribution is the complete tensor period subgroup and sharp information
bound. The one-column contraction belongs to #7426, rather than this result.
No new general identity or source audit is claimed. Apache-2.0.

Base main is `e54523517d0f03e58352e647c403a5af3e525a57`, priority revision 8.
The catalogue assessment predates selected #7000; its frozen bytes remain
provenance in [SOURCES.json](SOURCES.json). The selected-binding first control
passed. The `formal-interface-lemma` card explicitly changes scope to a
written phase-information side result, with no proof-assistant build.
Live PR bodies, recent account work, current claims, local reservations,
merged searches and main README searches were checked. Search absence does
not certify novelty; arbitrary code diffs and unpublished work remain outside
coverage. The [advance claim](https://github.com/CrocSwap/integer-mult-bounds/discussions/1166#discussioncomment-19591467)
states the proposed f+1 bound. A progress clarification credits the
one-column contraction already present in the predecessor's full proof;
the delivered original result is only the new tensor information theorem.

Physical acquisition, decoder work, dirty restoration, layout, owner return,
precision, historical backing and the complete recursive invoice are unpriced.
The result alone reduces neither tapes nor runtime and admits no new supplier.
It changes neither the affine address map nor the central C tensor. All five
selected framework assumptions remain. No selected cold replay, Lean, Docker
or complete multiplication proof was run.

## Open next questions

- Can a literal producer exploit this contraction with a lower complete phase
bill? Start with the supplied #7426 packed interface, replacing quadratic
updates while charging every parity read, carry and cleanup. Its full
time/backing invoice is the blocker; information rank alone supplies none.
- Can a producer compute the nonlinear two-bit phase without retaining the
f linear observations? Start with one explicitly ordered sequential scan
and its reversible state/return contract. The theorem permits arbitrary
decoding but does not price that different streaming interface.

Structured relevance, first steps and blockers are in [TASK.json](TASK.json).
No follow-on scope is reserved.

## Verifier constraints

Build the Lever supplies the retained deterministic verifier. Prove It Works
requires the literal phase and truth-table checks above. Functions are below
60 lines, loops have fixed bounds, and no recursion, subprocess, network,
dynamic code loading or output-file write is used. Package reads are capped
at one MiB per file; the recipe is capped at 64 KiB. The public column input
is restricted to 1..64 nine-bit integers, rejecting Booleans.

Power-of-ten rule 3 deviation: CPython allocates managed integers and containers.
Live mathematical state is bounded by the fixed domains, including 512-entry
phase/derivative tables and sets of at most four phase values. The 262,144
two-column cases are streamed. File hashes cover seven bounded files.
Warning-strict ordinary/optimized replay and an AST/function-length inspection
are the available diagnostics; no project-wide tooling change is made.
Loading
Loading