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
159 changes: 159 additions & 0 deletions research/permutation-cycle-layout/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,159 @@
# Optimal simultaneous column order for a wire permutation

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

Claim type: scoped written proof. The finite replay corroborates bounded
instances; it is not the universal proof or a proof-kernel certificate.

## 1. Contract and result

Let n >= 1, d >= 2 and let pi be a permutation of {0,...,n-1}.
Each site is C^d. The full operator U_pi routes an input site to its
specified destination:

U_pi[y,x] = product_i [y_(pi(i)) = x_i].

Every input and output coordinate is retained. There is no side information,
restriction to a subspace, dirty workspace, scalar phase or discarded output.
Choose one order o of the sites and use it for both input and output columns.
At a cut let A be the sites preceding it, B its complement, and put

c(A) = |{i in A : pi(i) in B}|.

There are also c(A) wires from B to A, since pi is a bijection. A fixed-width
matrix-product representation may use different local tensors at different
positions, with scalar boundary bonds. Its peak bond is the largest internal
virtual dimension; define it as one when n=1.

**Theorem.** The coefficient unfolding across A|B has exactly

r(A) = d^(2 c(A))

nonzero singular values, all with square d^(n-2 c(A)). In the prescribed
order the minimum possible peak bond is max_A r(A), over its prefix cuts.
When the simultaneous order may be chosen arbitrarily,

min_order min_exact_train peak_bond = 1 if pi is identity,
= d^2 otherwise.

Consecutive directed cycles attain the optimum. These are algebraic layout
statements. They do not assert that a caller permits that layout, that a
permutation description is available cheaply, or that physical reordering
has zero cost. The fixed-width representation does not invoke the all-width
fixed-table equivalence theorem of zxv's PR #5441.

## 2. Complete cut spectrum

Unfold U_pi with rows (y_A,x_A) and columns (y_B,x_B).
There are 2c crossing wires. Index their independent d-valued wire labels
by z. For fixed z, let u_z be the left 0/1 vector imposing the crossing
labels and every equality on an internal A-to-A wire. Define v_z on B
similarly. Orient each label by its original source wire, so that

H = sum_z u_z v_z^T.

Different z have disjoint left supports and disjoint right supports: each
crossing label appears in a distinct left physical coordinate and a distinct
right physical coordinate. They therefore give orthogonal nonzero vectors.
There are |A|-c internal A-to-A wires. Each contributes one freely chosen
label after its equality constraint, independently of every other wire.
Consequently

||u_z||^2 = d^(|A|-c),
||v_z||^2 = d^(|B|-c).

Normalizing these orthogonal vectors gives a singular value decomposition
with d^(2c) equal singular values. Their squares are d^(n-2c).
In particular their sum is d^n, the squared Frobenius norm of U_pi.
This includes empty cuts and c=0. The same disjoint-support factorization
also proves the rank formula over any coefficient field.

## 3. Exact attaining tensor train

At each cut carry one d-valued state label for each directed wire whose
endpoints lie on opposite sides of that cut. Thus the virtual space has
dimension d^(2c). For a local tensor at site v, impose these equalities:

- Every wire present on both adjacent cut frontiers keeps its label.
- The wire with source v has label x_v at its source endpoint.
- The wire with destination v has label y_v at its destination endpoint.
- A self-wire imposes x_v=y_v locally.

All other tensor entries are zero. Both endpoint frontiers are empty.
Contracting the local tensors sums over every wire label. A full physical
input/output pair has exactly one satisfying assignment if it obeys every
routing equality, and has none otherwise. The contracted coefficient is
therefore precisely U_pi[y,x], with no multiplicative d factor or phase.
This constructs all cut dimensions simultaneously.

Conversely, at a bond of dimension D the unfolding factors through that
D-dimensional virtual space, so its rank is at most D. Section 2 forces
D >= d^(2c) at every cut. The explicit construction attains every such
lower bound in the prescribed order. It uses standard tensor-network
factorization; no new general factorization principle is claimed.

## 4. Optimal simultaneous order

Decompose pi into disjoint directed cycles. Place each cycle consecutively,
in its directed order, and concatenate the cycles. At any prefix, completed
cycles contribute no crossing wires. A cut strictly inside one nontrivial
cycle has exactly one wire from A to B and one from B to A. A cut between
cycles has no crossing wires. Hence the peak is at most d^2, or one when
all cycles are fixed points. The algorithm visits each site once after
validating the permutation.

For the lower bound, fix any order of a nonidentity permutation and consider
its first moved site v. Every preceding site is a fixed point. The site
pi(v) is another moved site, so it is still on the right. At the cut just
after v, exactly one A-to-B wire leaves, namely the wire from v. There is
at least one further moved site, so this is an internal cut. Section 2
forces bond at least d^2. Identity routing is a product of local identities
and has peak one. This proves the optimum for all n and d in the contract.

For the half-swap pi(i)=i+k modulo 2k, natural order has c=k at its middle
cut, giving peak d^(2k). The cycle order

0,k,1,k+1,...,k-1,2k-1

alternates c=1 and c=0 and has peak exactly d^2. Thus the fixed-order
exponential obstruction from #5441 disappears under this simultaneous
layout change. Its physical legality and cost are a separate question.

## 5. Approximation consequence and its limit

Write N=d^n. At a cut of rank r, an arbitrary rank-D approximation to H
has optimal squared relative Frobenius error

max(0,1-D/r).

Indeed its squared singular values are N/r, and projection can retain at
most D of them; truncating the displayed decomposition attains the bound.
Unfolding preserves the Frobenius norm. For a physical approximation V,
||V-U_pi||_F^2 <= N ||V-U_pi||_op^2, so cut bond D and operator error
at most epsilon imply D >= (1-epsilon^2)r. This is a necessary condition,
not an attaining operator-norm construction or a joint approximation at
all cuts. The projection argument is inherited classical reasoning, also
used for #5441's two original fixed-order examples.

## 6. Attribution and scope

[zxv #5441](https://github.com/CrocSwap/integer-mult-bounds/pull/5441)
provides the repository's physical-column transfer contract, the cut-rank
argument, the half-swap example and the explicit warning that reordering
changes the obstruction. Its native fixture, equivalence algorithm and
scientific verification are cited, not replayed. Its predecessor attribution,
including Krzysztof Wójt #5262, Ali Romman #5352 and Jose Manuel #2250,
remains in that unchanged package.

Cycle decomposition, operator-Schmidt rank, projection bounds and
matrix-product factorizations are classical. The new repository result is
the complete arbitrary-permutation cut spectrum, its explicit wire-label
train, and the sharp optimum over simultaneous orders, resolving the
algebraic layout subproblem. No global literature priority is claimed.
Nick Petalas #6184 concerns arithmetic chirp phases, a different endpoint;
its formulas and construction are not used here.

No fixed finite table uniform across all widths, legal physical reordered
caller, native/tape tariff, supplier admission, new kappa or complete
multiplication theorem follows. The selected frontier is unchanged.
Independent human and formal proof-kernel review are unclaimed.
153 changes: 153 additions & 0 deletions research/permutation-cycle-layout/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,153 @@
# Every wire permutation has an exact cycle layout of peak bond d^2

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

Claim type: scoped written proof with exact finite corroboration.
For n sites of equal dimension d >= 2, route source i to destination pi(i).
Use the same column order for the operator's input and output sites.
If c wires leave a prefix, exactly c enter it, and its coefficient unfolding
has rank d^(2c). All its nonzero squared singular values equal d^(n-2c).

The minimum exact matrix-product peak bond in that order is the largest
of these cut ranks. Optimizing the simultaneous column order gives exactly
one for identity routing and d^2 for every nonidentity permutation.
Consecutive directed cycles attain the optimum with an explicit wire-label
tensor train. [PROOF.md](PROOF.md) proves the statements for every n and d
in the mathematical contract. The executable caps formula/layout inputs
at 64 sites and local dimension 16; materialized trains have smaller caps.

For the half-swap of 2k sites, the natural order needs peak d^(2k).
The order 0,k,1,k+1,...,k-1,2k-1 has peak d^2. At d=2 and k=2 this
changes the exact representation requirement from 16 to 4. The full cut
spectrum also gives a necessary operator-error bound and a sharp
single-cut relative Frobenius approximation bound, with their scopes
separated in the proof.

## What this adds

[zxv #5441](https://github.com/CrocSwap/integer-mult-bounds/pull/5441)
supplies the physical-column transfer contract, cut-rank method and
fixed-order half-swap obstruction, and explicitly leaves column reordering
as an escape. This package determines the algebraic optimum under
unrestricted simultaneous reordering and constructs the attaining train.
It does not replay the predecessor's verifier or audit its result.
Classical cycle decomposition, operator-Schmidt rank and tensor-network
factorization retain their credit. Global literature priority is unclaimed.

The card `formal-interface-lemma` is explicitly changed to this algebraic
side result. It can prevent a fixed-order lower bound from being treated as
a representation-independent routing obstruction. It does not establish
that a physical caller permits this order or pays less for it. Site-dependent
fixed-width tensors do not automatically form a fixed finite transfer table
uniform across widths. No supplier admission, new kappa or complete
multiplication theorem is claimed, and the selected frontier is unchanged.

## Exact finite evidence and reproduction

From the worktree root, with the standard library of CPython 3.12:

```sh
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -W error -B research/permutation-cycle-layout/verify.py
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -W error -O -B research/permutation-cycle-layout/verify.py
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -B scripts/research_package.py --package research/permutation-cycle-layout
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -B scripts/check_research_commit.py --package research/permutation-cycle-layout --check-source-history --check-modes
```

Another CPython 3.12 path may replace the absolute interpreter path.
Expected markers are `PASS_PERMUTATION_CYCLE_LAYOUT`,
`PASS_REGISTERED_RESEARCH_PACKAGE` and `PASS_RESEARCH_COMMIT_BYTES`.
The direct verifier emits its actual counters, interpreter version and
SHA256 hashes of all seven package files. It writes no files and needs
neither Git nor predecessor files. Receipts are generated by execution.

The exact replay checks:

- 11,680 literal coefficient unfoldings containing 694,400 nonzero entries,
and 164,014 stored Gram entries. It constructs the full routed basis
entries for every binary permutation through six sites and every ternary
permutation through four sites, in natural and cycle orders when distinct.
The exact identities G^2=lambda G and trace(G)=lambda*r certify the
complete proposed spectrum, since G=H H^T is positive semidefinite.
- All 5,913 permutations through seven sites. An independent subset dynamic
program represents every column order, using 695,482 prefix subsets
and 2,406,609 transitions. Its optimum agrees with the constructed order.
This is exhaustive dynamic programming, not literal listing of every order.
- 64 explicit trains and 17,047 contracted physical entries. Complete small
matrices are compared with literal routing; two six-site half-swap trains
use every input with its correct output and one incorrect output.
- Thirteen invalid-input rejections and seven semantic controls covering
ordering, crossing directions, singular values, a missing local factor,
absolute phase and routing orientation.

Finite checks corroborate the written theorem. No numerical tolerance,
Monte Carlo, predecessor execution, selected cold replay, Lean or Docker
run is required or supplied. Independent human and formal proof-kernel
review are unclaimed.

## Sources and overlap coverage

Canonical base is e54523517d0f03e58352e647c403a5af3e525a57, priority
revision 8, selected #7000. [SOURCES.json](SOURCES.json) binds the exact
predecessor notes and orientation documents as historical provenance.
There are no live external scientific inputs or imported predecessor code.
The current selected metadata orientation passed; the catalogue's #5599
assessment predates selected #7000 and is not used as a scientific limit.

Live open PR bodies and Discussions were searched with the overlap helper,
plus the latest handoffs, all-state recent account bodies, merged PR searches
and all-main README searches for permutation bonds, wire crossings, cycle
layouts and reordering. The helper's sparse local note inventory was
incomplete. Its separate closed-PR bulk refresh timed out, so focused
merged searches and the direct all-state account query supplement it.
No exhaustive historical-diff or unpublished-work coverage is claimed.
The [advance claim](https://github.com/CrocSwap/integer-mult-bounds/discussions/1166#discussioncomment-19595275)
records the bounded scope. Nick Petalas #6184 is an older different-endpoint
chirp-spectrum result; its construction and formulas are not used here.
The unchanged #5441 source retains its full predecessor attribution.

## Open next questions

Structured questions, relevance and blockers are in [TASK.json](TASK.json).

1. Which actual caller permits a cycle order on both input and output sites?
Bind its requested wire permutation and complete bank layout, then price
preparation and restoration of that layout. The layout permission and
physical movement cost are currently missing.
2. Can one useful restricted permutation family be represented by a fixed
transfer table in its legal order, uniformly across widths? Start with
adjacent paired swaps after interleaving and declare the schedule state.
The present position-dependent trains do not establish that contract.
3. How does the optimum change with retained side information or unequal
site dimensions? First specify which input and output coordinates remain
demanded; the disjoint crossing-label argument then needs a new endpoint
and dimension calculation.

## Verifier resource contract

New code uses no recursion, subprocesses, network, dynamic evaluation or
optional dependencies. All functions are shorter than 60 lines. Formula
inputs are capped at 64 sites and dimension 16. Materialization is capped
at six sites and 64 virtual labels per cut. Literal spectra use at most
81 routed basis inputs and a smaller-side unfolding with at most 81 rows.
The subset verifier caps n at seven. Boundary validation rejects malformed
permutations, orders and unsupported dimensions. Checks use explicit
exceptions and remain active under optimized Python.

Power of Ten rule 3 has a managed-runtime deviation in routing.py and
verify.py: Python allocates integers and containers during exact arithmetic.
The explicit train has at most six local tables, each with at most 256
physical pairs and 64*64 scalar entries. One train is retained at a time.
The verifier's matrices have at most 81 nonzero endpoint entries, and the
subset program has at most 128 states per permutation. Layout arithmetic
has integers no larger than 16^64. File hashing reads only seven fixed local
files. These bounds and rejection controls constrain live allocations;
platform allocator behavior is not instrumented. Any interface expansion
must set new caps and recheck this resource contract first.

Power of Ten rule 9 has a bounded-dispatch deviation in verify.py: main
calls four explicitly listed verification functions, and the rejection helper
accepts only three listed boundary-check targets in this closed verifier.
These seven fixed targets have no recursive call paths. The registered warning-strict
normal and optimized runs exercise every target. No generic Python static
analyzer is configured for this isolated package; repository metadata,
publication preflight and these two executions are the recurring checks.
30 changes: 30 additions & 0 deletions research/permutation-cycle-layout/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/finite-memory-lifting/README.md",
"usage": "provenance",
"git_blob": "5ce4fa86af1d5487f9a639acdc9fa0177c76c192",
"sha256": "ca0be421d5cb1835a8814685907db78a7a3ea24b7e1e474f3038c87b0e46498e"
},
{
"path": "research/finite-memory-lifting/PROOF.md",
"usage": "provenance",
"git_blob": "36152f9e31555a88ce3636ae59abf07a8058d78d",
"sha256": "de4a6b79d480d7718b3d157d2be73b4f45723ccac6719160dd815f8cef40356b"
}
]
}
Loading
Loading