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
104 changes: 104 additions & 0 deletions research/fixed-basis-idle-coalescing/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,104 @@
# A fixed independent idle family reduces to coordinates

## Contract

Let F be any field, V=F^n with named coordinate basis e_1,...,e_n,
and let stages s=1,...,m have coordinate masks M_s contained in {1,...,n}.
Write U_s=span{e_i : i in M_s}. Fix integers 0<=q_s<=|M_s|.

Choose one basis v_1,...,v_n of V for the entire schedule. A feasible
allocation chooses J_s contained in {1,...,n}, with |J_s|=q_s and
v_j in U_s for every j in J_s. Put I_j={s:j in J_s} and d_j=|I_j|.
Directions are independent globally, not merely separately at each stage.
A direction's scalar multiple represents the same line. No direction changes
between its uses. Choosing a different basis for a different allocation is
allowed. All basis setup, scalar mixing and transport costs are optimistically
zero. Eligibility uses only the output direction's support, so imposing
additional right-vector or physical restrictions can only shrink this model.

The coordinate model is the special case v_j=e_j. Its schedules have exactly
q_s distinct idle coordinates in M_s at each stage. Neither model preserves
a prescribed HOST action or imposes chronological barriers on coalescing.

## Theorem

For every feasible fixed-basis allocation there is a coordinate allocation
and a permutation pi of {1,...,n} with

I'_{pi(j)} = I_j for every j.

In particular, the two models have exactly the same feasible multisets of
idle stage sets, and exactly the same feasible idle-degree multisets.
They have the same minimum sum_j f(d_j) for every real-valued function
f on {0,...,m}. No concavity or monotonicity assumption is needed for
this equality. More generally any objective invariant under relabeling the
directions has the same attainable values in both models.

The same reduction holds for a globally independent family of at most n
directions. Extend the family to a basis and give every added direction
the empty stage set. The permutation argument then applies to the original
family as an injection into the coordinate indices.

## Proof

Let B have columns v_j in the coordinate basis. Since det B is nonzero,
the determinant expansion has a nonzero summand. There is a permutation pi
such that B_{pi(j),j} is nonzero for every j. A nonzero sum must have a
nonzero summand over every field, including characteristic two. Thus pi(j)
belongs to the support of v_j.

If j belongs to J_s, all coordinates in the support of v_j belong to M_s,
because v_j belongs to U_s. Hence pi(j) belongs to M_s. Set
J'_s={pi(j):j in J_s}. The permutation is injective, so |J'_s|=q_s,
and J'_s is contained in M_s. This is a feasible coordinate allocation.
Moreover pi(j) belongs to J'_s exactly when j belongs to J_s, proving
the equality of stage sets and of degrees.

This constructs one inclusion of feasible multiset collections. The reverse
inclusion holds because the coordinate basis is itself an allowed fixed
basis. Objective equality follows by relabeling. All allocation domains are
finite at the stage-set level, even when F is infinite, so every nonempty
set of attainable objective values has a minimum. Coordinate feasibility
also follows directly from 0<=q_s<=|M_s|. This proves the theorem.

The support-matching fact is classical. Equivalently, every subset of the
independent columns has support union of size at least its cardinality,
and Hall's marriage theorem supplies the injection. The new contribution
is the degree-preserving reduction for this idle-coalescing question.

## Application to the connected eighteen-stage root

CubeJerry's [#7002](https://github.com/CrocSwap/integer-mult-bounds/pull/7002)
defines a coordinate-role relaxation in
[IDLE-COALESCING.md](../connected-l18-root-caller/IDLE-COALESCING.md).
It optimistically allows any two eligible coordinates to be hosted at a stage,
leaves the other mask coordinates idle, and coalesces all occurrences of a
coordinate for free. Its costs depend on idle degrees alone. Set
q_s=|M_s|-2 in the theorem.

Consequently, allowing an arbitrary fixed common basis, or a fixed globally
independent subset, cannot lower that relaxation's optimum. The predecessor's
stated coordinate optimum and fixed-profile failure therefore apply to this
larger subclass as a corollary. This is a new applicability step. The
predecessor's numerical optimum, transcendental bounds and complete invoice
are cited premises; this package does not recompute or certify them.

The proof maps optimistic allocations, not physical operators. It need not
map the HOSTs or retain a source's original endpoint. This causes no gap in
the lower-bound use: every admitted fixed-basis allocation maps to a feasible
member of the already optimistic coordinate relaxation. A map in the other
direction with physical HOSTs would require a separate construction.

## Boundary

Three directions e_1,e_2,e_1+e_2 in F_2^2 are dependent and cannot receive
distinct coordinate anchors. The matching proof therefore cannot cover
arbitrary dependent families, even if every individual stage uses independent
directions. This observation does not prove that dependent directions save
cost. A basis that changes between stages, changing a direction's identity,
also falls outside the contract. No theorem about either extension is claimed.

Noncoordinate common bases are thus insufficient on their own. A route
around this obstruction must change at least one premise, for example use
dependent directions across the complete schedule, rotate the basis, or
change masks, idle counts, native profile or the degree-only tariff.
127 changes: 127 additions & 0 deletions research/fixed-basis-idle-coalescing/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,127 @@
# Fixed common bases cannot improve idle coalescing

This contribution is produced by Nick Petalas. Claim type: scoped written
proof with exact finite corroboration. No new kappa.

For coordinate stage masks over every field, allowing one arbitrary fixed
common basis gives exactly the same feasible idle-degree multisets as the
coordinate basis. Match each basis vector to a distinct coordinate in its
support. Replacing its idle uses by that coordinate preserves every stage
and every degree. The result also covers fixed globally independent
subfamilies. [PROOF.md](PROOF.md) gives the contract and proof.

This settles a subclass of the noncoordinate-direction gap in CubeJerry's
[#7002](https://github.com/CrocSwap/integer-mult-bounds/pull/7002). Its
coordinate-role idle-coalescing obstruction now applies to arbitrary fixed
common bases under the same degree-only tariff. The predecessor's coordinate
optimum and numerical shortfall are cited inputs, not replayed results.

The equality holds for every common function of idle degree, including its
concave rank-power tariff. It does not require concavity. The theorem does
not cover dependent direction families, a basis that changes between stages,
changed masks or idle counts, or a changed native/caller contract. Those are
possible routes beyond the obstruction, rather than proved improvements.

## Evidence and reproduction

From the repository root, use standard-library CPython 3.12:

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

Any CPython 3.12 interpreter may replace that absolute interpreter path.
The direct marker is `PASS_FIXED_BASIS_IDLE_COALESCING`; the runner also
reports `PASS_REGISTERED_RESEARCH_PACKAGE`. The verifier writes nothing.
Each execution emits a generated JSON receipt containing the exact census
counts and SHA256 hashes of all six package files.

The checker enumerates 85,833 square matrices and checks 31,617 bases and
1,251,947 eligible stage/subset pairs. Its domains are F2 in dimensions one
through four and F3 in dimensions one through three. For each basis it compares
an augmenting-path matching with all supported permutations. It then tests
every coordinate mask and every eligible subset of directions. A separate
five-stage noncoordinate fixture checks the complete degree vector. Six
invalid input controls reject, and two boundary controls expose dependence
and the need to match within support. Explicit exceptions keep every check
active under optimized execution. The universal theorem comes from the
written determinant argument, not extrapolation from these finite domains.

## Scope and provenance

Base is canonical main `e54523517d0f03e58352e647c403a5af3e525a57`, priority
revision 8. The selected record still names conditional #7000. SOURCES.json
pins #7002's README and IDLE-COALESCING.md as historical mathematical context.
They are not executable inputs. The standalone verifier imports no predecessor
code or data and does not repeat a predecessor census or invoice check.
TASK.json declares a changed joint-query-controller scope, a local allocation
obstruction rather than a completed producer/consumer program.

The card's first binding-only command passed. Live open PRs, active
Discussions, recent account contributions, local claims, merged searches and
all-main README searches were checked before the
[CLAIM](https://github.com/CrocSwap/integer-mult-bounds/discussions/1166#discussioncomment-19540656).
The cached overlap helper reports incomplete sparse-local coverage; absence
of a search hit is not a novelty certificate. The related open packages below
construct normalized schedules; this theorem concerns all fixed globally
independent families and imports neither schedule nor numerical comparison.

Native implementation, HOST compatibility, paid basis changes, chronological
transport, dirty restoration, ordinary costs, precision/tape, stock, compiler,
supplier admission and the full multiplication theorem remain outside scope.
No numerical frontier, selected files or predecessor files are changed. The
selected cold replay, Lean, Docker, independent human review and formal proof
kernel are unrun. The application to #7002 uses its published coordinate
obstruction as a premise without independently validating that result.

## Related open work

- Rohan Arun's [#7098](https://github.com/CrocSwap/integer-mult-bounds/pull/7098)
constructs and prices a noncoordinate projector relocation for #7002.
This proof gives a fixed globally independent family obstruction, without
recomputing that schedule or its price.
- Rohan Arun's [#7106](https://github.com/CrocSwap/integer-mult-bounds/pull/7106)
extends the normalized projector scheduling work. Its construction enforces
compatibility on overlapping intervals. This theorem requires global
independence across the whole schedule, a stricter premise. No numerical
agreement or disagreement is claimed.

## Open next questions

- Can a globally dependent family of directions lower a fully specified idle
bill while each stage still has independent idle directions? Start with
three lines in a two-dimensional coordinate space and a short mask sequence.
A favorable allocation and the actual paid preparation/restoration are
missing. The matching theorem does not decide this case.
- Can changing the basis between stages improve a complete connected-root
caller after paying conversion and HOST compatibility costs? Start with
two explicit bases and one transition. No legal conversion or favorable
complete bill is supplied here.

Structured relevance, first steps and blockers appear in TASK.json.

## Credit and engineering

Credit CubeJerry, Jerry Chen, #7002 for the connected-root question and
coordinate coalescing relaxation; DreamingOfClouds #6937 for the connected
mask lineage, and the retained #6936/#6912 native/root authors. The support
matching fact is classical determinant expansion or Hall matching. The new
step is its stage-set-preserving application, not the predecessor optimum.
Independent human and formal proof-kernel review are unclaimed. Apache-2.0.

The investigation playbook guided the work. Build the Lever supplied a
standalone checker; Prove It Works required all eligible stage subsets and
the chronological degree fixture. Power of Ten bounds dimensions by four,
fields by three, matching paths by dimension, and census domains explicitly.
All functions have fewer than sixty lines and there is no recursion, dynamic
code execution, subprocess, network or shared mutable state. Python allocation
is a managed-runtime deviation: matrices have at most sixteen entries,
permutation lists at most twenty-four entries and vector tables at most
eighty-one entries under the enforced entry-point caps. Runtime allocator
internals are not inspected. Receipt files are limited to 128 KiB each.
Warning-strict normal and optimized commands are registered diagnostics.
Ruff and mypy were unavailable on PATH. No repository-wide diagnostic policy
was changed for this isolated package.
30 changes: 30 additions & 0 deletions research/fixed-basis-idle-coalescing/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/connected-l18-root-caller/IDLE-COALESCING.md",
"usage": "provenance",
"git_blob": "91df672c3d5f4b56321e7b07ad98f011949484e3",
"sha256": "ba9d8e2caf41b18352e01edc69a3a47c1e1385f6b4ae3b2ab8c681b3bd44f5ad"
},
{
"path": "research/connected-l18-root-caller/README.md",
"usage": "provenance",
"git_blob": "6ef3bae7d8fb6081f7c6441d04a5d929cbd7016b",
"sha256": "349341634d97c3f105bdc5003de1a2e7252a31fc52e25dc7bdfd39550586ccbd"
}
]
}
Loading
Loading