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
164 changes: 164 additions & 0 deletions research/acyclic-interaction-budgets/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,164 @@
# Sharp budgets from a common acyclic interaction support

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

## Contract and notation

Let D_1,...,D_K be real or complex n by n matrices. Their union support is
the directed graph having an edge c to r whenever some (D_i)_(r,c) is
nonzero. Assume this graph is acyclic and every directed path has at most h
edges, with 0<=h<=n-1. Fix any induced matrix norm and bounds
d_i>=||D_i||. The d_i may be zero and need not be small.

For a list d=(d_1,...,d_j), define

e_r(d) = sum_{1<=i_1<...<i_r<=j} d_i1 ... d_ir,
H_r(d) = sum_{a_1+...+a_j=r, a_i>=0} d_1^a1 ... d_j^aj,
E_h(d) = sum_{r=0}^h e_r(d),
C_h(d) = sum_{r=0}^h H_r(d).

Both degree-zero polynomials are one. For the empty list, all positive
degree polynomials vanish. These are the classical elementary symmetric
and complete homogeneous symmetric polynomials, respectively.

Write R_j=(I+D_j)...(I+D_1), with R_0=I. Then

||R_j|| <= E_h(d_1,...,d_j),
||R_j-I|| <= E_h(d_1,...,d_j)-1,
||R_j^-1|| <= C_h(d_1,...,d_j),
||R_j^-1-I|| <= C_h(d_1,...,d_j)-1.

Every R_j is invertible. All four envelopes are sharp over the stated
class in the induced infinity norm, separately for forward and inverse
quantities, for every finite nonnegative list d and every h. A witness has
dimension h+1. Sharpness also holds in any larger dimension by adding zero
coordinates to the errors. A particular graph or an extra rank restriction
can have smaller optimal envelopes.

## Vanishing products and the upper bounds

An entry of a product of r supported matrices is a sum over directed walks
of length r in the union graph. A directed walk in an acyclic graph cannot
repeat a vertex, so it is a path. Thus every product of h+1 matrices chosen
from the D_i, with arbitrary repetition and order, is zero. This statement
is stronger than each individual D_i being nilpotent.

Expand R_j in its chronological order. Its terms of degree r choose r
distinct factor indices. Submultiplicativity bounds the sum of their norms
by e_r(d). Every term of degree greater than h vanishes. Keeping the
constant term proves the first bound; removing it proves the second.

Since D_i^(h+1)=0, the exact inverse of I+D_i is the finite sum

I-D_i+D_i^2-...+(-D_i)^h.

The inverse R_j^-1 is the product of these inverse factors in reverse
order, namely (I+D_1)^-1...(I+D_j)^-1. Expanding and grouping by total
degree r gives one term for each nonnegative exponent tuple summing to r.
The norm sum is at most H_r(d). Terms above degree h vanish, even when
they use different factors. This proves both inverse bounds. No infinite
series, commuting-error hypothesis or small-error limit is used.

## Exact attaining families

Let S be the (h+1)-dimensional matrix with ones just above the diagonal
and zeros elsewhere. Its norm is one when h>=1, and S^(h+1)=0. Its union
support is one directed path of length h. For h=0 use the zero 1 by 1
matrix. Take D_i=d_i S. All factors commute in this example, and direct
polynomial multiplication gives

R_j = sum_{r=0}^h e_r(d) S^r.

The first row has one entry e_r(d) in each column r+1. Every other row
has a shorter nonnegative sum. Consequently its maximum absolute row sum
is exactly E_h(d), and removing I gives exactly E_h(d)-1.

For the inverse bounds instead take D_i=-d_i S. Its inverse product is

R_j^-1 = sum_{r=0}^h H_r(d) S^r.

The same first-row argument gives C_h(d) and C_h(d)-1. These witnesses
satisfy the prescribed upper bounds on individual norms, with equality
when h>=1. The h=0 class necessarily has zero errors, and both envelopes
are one regardless of any looser prescribed d_i.

These are different sign choices. No simultaneous attainment of the
forward and inverse maxima by one error sequence is asserted. If the
positive d_i fail to use every edge, the union graph can have depth less
than h. The witness still belongs to the class of support depth at most h,
and a zero budget has the same sharp value one in either depth class.

## Arithmetic noise and the original precision interface

For z_0=x and z_i=(I+D_i)z_(i-1)+v_i, finite expansion gives

||z_j-x|| <= (E_h(d_1,...,d_j)-1)||x||
+ sum_{i=1}^j E_h(d_(i+1),...,d_j)||v_i||.

Each noise term meets only the later factors. The proof above applies to
that suffix, whose support depth is at most h. This bound is sharp in the
same class in the infinity norm. Use D_i=d_i S, x=M times the all-ones
vector, and v_i=eta_i times that vector, with M,eta_i>=0. Every contribution
is nonnegative and its first-row sum attains the displayed bound. This
also covers h=0 and empty suffixes.

For exact gates A_i with F_i=A_i...A_1 and perturbed gates A_i+Delta_i,
y_i=(A_i+Delta_i)y_(i-1)+u_i, the interaction identity of ywz65's PR2701 is

D_i=F_i^-1 Delta_i F_(i-1),
z_i=F_i^-1 y_i,
v_i=F_i^-1 u_i.

If these actual D_i satisfy the common acyclic hypothesis, the preceding
bounds apply. In particular, for ||F_i||,||F_i^-1||<=B,

||Q_j|| <= B E_h(d_1,...,d_j),
||Q_j^-1|| <= B C_h(d_1,...,d_j),
||Q_j-F_j|| <= B(E_h(d_1,...,d_j)-1),
||Q_j^-1-F_j^-1|| <= B(C_h(d_1,...,d_j)-1).

The physical trajectory bound is ||F_j|| times the preceding noise formula
with ||v_i||=||F_i^-1 u_i||. The B factors are sufficient; their optimality
for arbitrary prescribed physical frames is not claimed. For data-dependent
errors, apply the inequalities to the realized sequence and its stated
norm bounds. Inverting that realized linear product is not an inverse for
a nonlinear rounded execution.

## A strict improvement and its boundary

At depth h=1, both exact envelopes are

1+sum_i d_i.

All interactions are then mutually annihilating, which is stronger than
the separate square-zero condition in Ali Romman's PR5288. Whenever two
d_i are positive, this is strictly below its sufficient product
product_i(1+d_i). That earlier theorem is cited, not replayed. Twenty
errors of norm 1/10 have exact envelope 3, instead of (11/10)^20.

The higher-depth contract also allows individual errors with D_i^2!=0.
For h=2 and the same twenty norms, the exact forward envelope is 49/10
and the exact inverse envelope is 51/10. Their total norm budget is two;
the general small-total-error criterion in PR2701 is unavailable.

Common acyclicity is essential for truncation at h. In dimension two,
D_1=D_3=E_12 and D_2=E_21 are individually square-zero, but their union
has a directed two-cycle. Their product (I+D_3)(I+D_2)(I+D_1) has infinity
norm five, exceeding the invalid depth-one prediction four.

Physical gate supports do not establish the hypothesis in interaction
coordinates. For example, with physical Delta_2=E_12, exact A_1=I+E_21
and A_2=I, one gets D_2=A_1^-1 E_12 A_1, which has nonzero diagonal
entries. Thus even a triangular physical perturbation can lose the
coordinate-support certificate after conjugation.

## Scope

This is a scoped mathematical precision-interface theorem. The norm and
support of an actual selected interaction word remain to be supplied.
No actual selected caller, coefficient storage policy, bounded arithmetic
kernel, physical inverse, overflow margin, tape/storage invoice, native
operation, supplier admission or exponent improvement is proved here.
All five selected framework assumptions remain. Finite rational checks
corroborate the written universal argument; they are not its quantifier
proof or independent human or formal proof-kernel acceptance.
161 changes: 161 additions & 0 deletions research/acyclic-interaction-budgets/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,161 @@
# Sharp polynomial budgets for acyclic interaction errors

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

Claim type: scoped written mathematical proof with exact finite corroboration.
If all interaction-error matrices share an acyclic coordinate support with
paths of at most h edges, the sharp forward norm envelope is the degree-h
sum of elementary symmetric polynomials in their individual norm budgets.
The sharp inverse envelope uses complete homogeneous symmetric polynomials.
No small-total-error restriction is needed. A common shift matrix attains
each envelope exactly in the infinity norm, with separate sign choices.

For R=(I+D_K)...(I+D_1), d_i>=||D_i||, write e_r and H_r for these two
classical symmetric polynomials. [PROOF.md](PROOF.md) proves

```text
||R|| <= sum_(r=0..h) e_r(d),
||R^-1|| <= sum_(r=0..h) H_r(d).
```

Subtracting the degree-zero term gives sharp operator-defect bounds. The
proof also gives a sharp suffix-weighted arithmetic-noise budget. Products
of more than h errors vanish because every such product follows a directed
walk in the common graph. This argument permits arbitrary finite real or
complex matrices; they need not commute or be individually square-zero.

At depth one, the exact forward and inverse envelope is `1+sum_i d_i`.
Twenty errors of norm `1/10` therefore have bound **3**, strictly below the
earlier sufficient shear product `(11/10)^20` of Ali Romman's
[PR5288](https://github.com/CrocSwap/integer-mult-bounds/pull/5288).
This uses the stronger common-depth-one hypothesis. At depth two the exact
bounds for the same norm list are **49/10 forward** and **51/10 inverse**.
Those higher-depth errors may have nonzero squares, outside that earlier
shear theorem. Neither predecessor result is re-audited here.

## Reproduce

From the repository root, with CPython 3.12 and its standard library:

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

Another CPython 3.12 path may replace that executable. Expected direct marker:
`PASS_ACYCLIC_INTERACTION_BUDGETS`, with `errors=0`. Each run emits a newly
computed JSON receipt bound to every package file by SHA256. The verifier
writes nothing. No receipt is handwritten, and no solver or predecessor
executable is required. The two registered commands each have a 120-second
timeout. Provenance pins are authenticated separately by the package checks.

The finite replay covers:

- 326 coefficient comparisons against independently enumerated subsets and
multisets, plus two maximum-capacity checks.
- All 729 pairs of three-dimensional strictly upper triangular matrices with
entries in `{-1/2,0,1/2}`. A separate adjacency-power oracle checks graph
depth, and rational Gaussian elimination computes the actual inverses.
- Fifteen permuted-coordinate words and two twenty-step words. Across these
and the exhaustive pairs, 746 support graphs, 1,553 prefixes and 1,553
arithmetic-noise steps pass.
- Fifty-one attaining families, including empty words, zero depth, zero
norms and the executable's maximum dimension and step count.
- Twenty malformed-input rejections and seven mathematical controls. These
distinguish individual from common nilpotence, elementary from homogeneous
inverse coefficients, chronological inverse order, suffix noise, and
physical from interaction-coordinate support.

These finite checks corroborate the universal written proof. They do not
exhaust all matrices or supply formal proof-kernel acceptance.

## Scope, sources and limitations

The interaction identity `D_i=F_i^-1 Delta_i F_(i-1)` and the physical
prefix/noise contract come from ywz65's
[PR2701](https://github.com/CrocSwap/integer-mult-bounds/pull/2701), head
`c21f5a50b9f4682ff9d4fe622d3d4e81c3cc86fb`. Ali Romman's PR5288,
head `cd33c723408bb98875211ee3145582c507881f29`, supplies the distinct
square-zero shear product comparison. Classical nilpotent inverse formulas,
symmetric polynomials and directed-path expansions retain their mathematical
credit. No general literature-priority claim is made for those methods.

The new step is the pair of sharp common-depth budgets, their attaining
families and exact precision-interface consequence. The predecessor's
production-caller question is refined to this additional-structure question;
it is not claimed as a completed caller binding. No predecessor executable,
construction or scientific replay is used. Predecessor files and credit lines
remain unchanged. Their retained proof texts, the original card and selected
record are pinned as historical provenance in [SOURCES.json](SOURCES.json).

The `uniform-transfer-audit` card explicitly changes scope to this side
result. It does not discharge the selected `precision_tape` assumption.
Canonical base is `e54523517d0f03e58352e647c403a5af3e525a57`, priority
revision eight, selected #7000 with conditional kappa `363005817/10^11`.
The selected bindings-only baseline passed. No selected cold replay,
predecessor scientific replay, native build, Lean or Docker run is claimed.

Common acyclicity must hold for the actual interaction errors. Triangular
physical perturbations can acquire diagonal entries under the required
prefix conjugation. A written two-dimensional example and its exact control
show this distinction. The optimal bounds concern all support graphs of
depth at most h; a prescribed graph can have smaller optima. The physical
frame-norm factors are sufficient, with no sharpness claim for them.

Arithmetic rounding may be nonlinear. An inverse of its realized perturbed
linear product is not a reversible finite-grid program. Actual coefficient
storage, a bounded-error arithmetic kernel, overflow, physical representation,
ownership/restoration, native operations, tape/storage costs, a complete
signed invoice and supplier admission remain open. All five framework
assumptions remain. No new kappa or selected-frontier change is claimed.

The [advance claim](https://github.com/CrocSwap/integer-mult-bounds/discussions/1166#discussioncomment-19591516)
and local reservation precede the result. Live coverage included 800 open PRs
and 7,400 Discussion records with complete pagination, followed by distinctive
all-main README and merged-PR searches. Sparse local notes were incomplete;
the separate full closed-PR refresh timed out. Arbitrary historical diffs and
unpublished work remain outside coverage. These checks are not a novelty
certificate. [TASK.json](TASK.json) records the source and scope decision.

## Open next questions

1. Bind one actual emitted tile and coefficient-storage policy to the common
acyclic interaction hypothesis. Compute the exact prefixes, their inverses
and transformed error supports in one fixed basis. Include basis-change
norms, arithmetic error and physical costs before judging any precision
or complete-invoice gain. The supplied caller contract is the blocker.
2. Determine the sharper envelope for one particular acyclic graph that
excludes the shift witness. Start with its chronological paths and the
within-factor paths appearing in inverse expansions. A specific error
budget and consumer decision are needed before a new census is useful.

No further construction or search is supplied here.

## Verifier constraints

The bounded rational API accepts dimensions at most 16, words of at most 32
steps and input numerator/denominator heights of at most 256 bits. The proof
has no such finite restrictions. Polynomial updates use O(K(h+1)) arithmetic
operations and O(h+1) rational entries; the trajectory helper uses O(K^2(h+1)).
The support checker uses a bounded topological elimination on at most 256
edges. All new functions have fewer than sixty physical lines and use no
recursion. Explicit exceptions keep every scientific check enabled under
optimized Python. No subprocess, network access or persistent write occurs.

Power of Ten rule 3 deviation: CPython dynamically allocates rational numbers
and containers. The dimensions, step counts, bit heights and serial fixture
domains bound live data. The pair inventory has 27 matrices and tests one pair
at a time. An attaining fixture holds two lists of at most 32 matrices.
Internal runtime allocation is not instrumented. Receipt reads are limited to seven regular
files, each at most 128 KiB. A fixed-allocation implementation would be needed
before using this checker as an embedded runtime service.

The isolated verifier explicitly loads its one adjacent implementation module
because isolated Python omits the script directory from the import path.
This is the only dynamic module target, with no generated code or arbitrary
callback dispatch. Package-file hashes bind its source in each generated
receipt. Warning-strict normal and optimized replay and in-memory syntax
compilation supply diagnostics; no separate static analyzer is claimed.
30 changes: 30 additions & 0 deletions research/acyclic-interaction-budgets/SOURCES.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,30 @@
{
"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/simultaneous-gate-rounding/README.md",
"usage": "provenance",
"git_blob": "eaad013627f7ded744f032eea77fed6f121fc01c",
"sha256": "4bd564aa762f07d67127cba668b370bab3c1407c733f873c756cf2f62e906f2a"
},
{
"path": "research/shear-structure-perturbation/README.md",
"usage": "provenance",
"git_blob": "badc068ab14b3c80ca659799a2c8dc4c563379a2",
"sha256": "2f6eee473e3b1bccdd926ebb055d32ba299770d8037f9e0c120cbfc98e4f4a94"
}
]
}
Loading
Loading