Skip to content

Research: one shared call at block 3, which two-term parts finish, which remove (no new κ) - #7629

Open
jacobalansussman wants to merge 1 commit into
CrocSwap:mainfrom
jacobalansussman:research/complex-m3-shared-call-finish-or-remove
Open

jacobalansussman wants to merge 1 commit into
CrocSwap:mainfrom
jacobalansussman:research/complex-m3-shared-call-finish-or-remove

Conversation

@jacobalansussman

Copy link
Copy Markdown
Contributor

Change and mathematical scope

Adds one package, research/complex-m3-shared-call-finish-or-remove: an exact table, in the frame game at 3
address bits, of what one array can hold in the output of a call that it shares with a second array. Among the
two-term parts tested (two kinds; coefficients 1, −1, 2 and 1/2 for the first kind and 1, −1 and 2 for the
second), none can be both finished by a last call and removed from the other array's last call. Checked by
computer; not a theorem about all programs. No new κ; no record, target or frozen package changes.

Finding Game Status
Two-term parts of kinds A and B of one shared middle call at 3 address bits, coefficient 1 (10,752 free elements each): 128 finishable by the last call alone and 639 removable by a call that could be a last call (kind A), 0 and 128 (kind B), none both frame game checked by computer in exact rational arithmetic; every case was also run one at a time with no symmetry reduction
The kind A rule. Same frame: finishable by the call alone exactly when 1 + c·g is invertible, removable exactly when it is not. Other frames: neither frame game checked by computer on all 9,983 tested cases for each of c = 1, −1, 2 and 1/2, none both; no proof by hand
Kind B with c = −1 and c = 2: none finishable; the same 128 parts removable with −1, none with 2 frame game checked by computer on all 9,984 tested cases each
With a free step after the last call, 128 more kind A parts are finished at c = 1 (512 at 2 and at 1/2), none of them removable frame game checked by computer, each identity term by term; that the list is complete is local only
Kind A parts on adjacent frames are removable at c = ±1 only by four calls that cannot be a last call frame game checked by computer
One first call and one shared call are enough, by symmetry frame game argued by hand; the counts, a second shared call and 368 cases of the map are checked by computer
The same rule with one spare address bit (4 address bits) frame game checked by computer on a seeded sample of 100 cases only
A 2 in 5 at three directions with one spare address bit (1.66e-1, a saving in the paid calls of one unit, not a κ) engine game open; no evidence either way

Remaining obligations: no person has checked the table, the rule or the symmetry argument. The table is a result
about the cases it lists: parts with three or more terms, a first call that reads a combination of the raw
input, last calls repaired against each other, other arrangements of the five calls, and the engine game itself
are not covered. At 3 address bits a 2 in 5 was already excluded by #4628, so nothing new is excluded.

Task and overlap

Validation

python3 scripts/research_package.py --package research/complex-m3-shared-call-finish-or-remove --check-only
python3.12 -B scripts/research_package.py --package research/complex-m3-shared-call-finish-or-remove
# = python3 -B research/complex-m3-shared-call-finish-or-remove/verify.py --part controls       (PASS_COMPLEX_M3_SHARED_CALL_CONTROLS)
#   ... --part main            (PASS_COMPLEX_M3_SHARED_CALL_MAIN)
#   ... --part poststep        (PASS_COMPLEX_M3_SHARED_CALL_POSTSTEP)
#   ... --part symmetry        (PASS_COMPLEX_M3_SHARED_CALL_SYMMETRY)
#   ... --part coefficients    (PASS_COMPLEX_M3_SHARED_CALL_COEFFICIENTS)
#   ... --part spare           (PASS_COMPLEX_M3_SHARED_CALL_SPARE)
python3 -B scripts/check_research_commit.py --package research/complex-m3-shared-call-finish-or-remove --check-source-history

Standard library only, exact rational arithmetic, no file read or written. With Python 3.12.13 on an Apple M4
(macOS) the six registered parts take about 3, 38, 4, 42, 85 and 22 seconds and under 30 MB each. In a clone of
the hub's main at e54523517d, scripts/research_package.py passes the metadata check
(PASS_PACKAGE_METADATA_AND_PINS) and runs the six registered commands with Python 3.12.13
(PASS_REGISTERED_RESEARCH_PACKAGE), and scripts/check_research_commit.py passes on the commit that adds the
package, with --check-source-history (PASS_RESEARCH_COMMIT_BYTES).

Controls:

  • the solver finds the plain third call and refuses a target two calls away;
  • arithmetic modulo 2^61 − 1 gives the same answers on 92 seeded cases;
  • three wrong variants of the rule are each rejected;
  • the symmetry map is tested case by case with a symmetry that does not commute with the first call;
  • the program refuses to run under -O, where its assert statements would be skipped.

Local only: two AI reviewers rebuilt the counts of an earlier form of the table from the README's definitions
with their own programs, which are not in this package. Between them they found the two corrections that are
now in section 5 (a free step after the last call; removing calls that cannot be a last call).

Not registered: --part full, every case one at a time with no symmetry reduction; I ran it once in four
quarters of about a minute each, with the same totals. Not run: the 20,643,840 cases at 4 address bits (a sample
of 100 only); any other coefficient; a free step after the last call for kind B; the hub's find_overlap.py
helper (I used my own search, described above); Lean.

Hosted CI: at this base the plan job fails on one of the hub's own tests, so the package replay is skipped for
new pull requests. I did not retrigger it. The local replay above is the evidence.

A remark on method, not evidence (local only: the runs and the files are not public and not in this package).
The table came out of nine runs in which AI agents carried ideas for sharing work into this game, eight of them
from other fields. None gave a program that beats the plain way at three directions, and six arrived separately
at this arrangement. Before looking at which cases were removable, the AI agent I work with wrote a guess for
the rule in a dated file, with a confidence of 75 %, and the comparison matched. Review by other AI agents then
asked two things the guess had not asked, and both changed the table. The README has the paragraph.

Credit

On the hub: eumemic (#288); Jose Manuel (#2108); lee-phys (#4688); Yiyang, whose phase repair (#4710) the October
thread built on; Dugongue (#4743, #4885, #5803); JoyboyBrian (#4751, #4797, #5491); Nick Petalas (#4825, #7551);
Chris R (#4856); alephnan (#6714); zxv (#5283). OpenAI for the cost model, the call and its 2-point kernel.

Fields the local runs drew on, none of which gave a saving: coded caching (Maddah-Ali and Niesen) and
interference alignment (in Cadambe and Jafar's form); catalytic computing (Buhrman, Cleve, Koucký, Loff and
Speelman; Cook and Mertz) with Bennett's uncomputation; Strassen's and Pan's fast matrix multiplication; the
transform splittings of Winograd, Rader and Bluestein; branch and sector bookkeeping from quantum circuits;
storage codes (Dimakis and co-authors) and secret sharing (Shamir); Maschke's averaging over a group; addition
chains (Yao; Pippenger) with Borodin and Moenck's remainder trees.

Substantial AI assistance: I did this work with the help of AI agents (Anthropic's Claude). The programs, the
searches, the arguments and the text were written by the agents under my direction. No independent human review
of this package is claimed. Apache-2.0.

Open next questions

In README (## Open next questions) and TASK.json. The smallest useful follow-up: the same arrangement at 4
address bits with every coefficient left free and the two last outputs read together. It matters because a
2 in 5 with one spare address bit is the smallest shape at three directions that no bound I know excludes (a
saving of 1.66e-1 in the paid calls of one unit, not a κ). Blocked by nothing but size: 20,643,840 free elements
before symmetry. Two smaller ones: a proof that a kind A part on two different frames is never finishable by the
call alone; a proof that a call which exchanges the frames of the two terms removes a kind B part when c = ±1.

Workflow feedback (optional)

At this base the plan job fails for new pull requests on the hub's own test
test_current_ci_section_matches_selected_record (as far as I can tell from the job logs of #7551 and #7627:
docs/ci-verification.md still describes #6030 while the selected record is #7000), so no package replay runs.

Sources

🤖 Generated with Claude Code

…ich remove (no new κ)

Adds research/complex-m3-shared-call-finish-or-remove: an exact table, in the
frame game at 3 address bits, of what one array can hold in the output of a
call that it shares with a second array. Among the two-term parts tested, none
can be both finished by a last call and removed from the other array's last
call. Checked by computer; not a theorem about all programs. No new κ.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant