diff --git a/research/binomial-source-floor/PROOF.md b/research/binomial-source-floor/PROOF.md new file mode 100644 index 0000000000..eb69d8150b --- /dev/null +++ b/research/binomial-source-floor/PROOF.md @@ -0,0 +1,156 @@ +# Sharp source floor from five inventory summaries + +This is a scoped written proof, produced by Nick Petalas. Apache-2.0. +It uses Ali Romman's all-height source telescope from +[PR #5318](https://github.com/CrocSwap/integer-mult-bounds/pull/5318) as a +premise. The binomial counting identity and monotone comparison are elementary; +no literature-wide priority for those methods is claimed. + +## Contract and credited premise + +Treat each productive or paid position as a separate token, including repeated +ranks. Let productive ranks have exact minimum L and maximum M, with +2 <= L < M. There are W productive tokens, exactly t of rank M, where +1 <= t < W. There are Q paid tokens, each with positive integer rank below L. +Assume Q >= W. For an integer cap c >= 1, let g be the unique nonnegative +integer satisfying L^g <= c < L^(g+1). + +For d in 0,...,g, let F_d(c) count words consisting of d ordered paid tokens +and g-d ordered productive tokens whose rank product is at most c. +The paid and productive positions are distinguished. There is no factor +binomial(g,d) choosing their interleaving. The chronological source model +and its Pascal matching are exactly those of the credited predecessor. + +Write D_d(c) = Q^d W^(g-d) - F_d(c). The predecessor proves + +```text +S(c) = W^(g+1) + (Q-W) sum_(d=0,...,g-1) D_d(c). (1) +``` + +This note neither replays that package nor claims a new source census. +The new question is the sharp lower envelope of (1) with only +L, M, W, Q, t fixed. + +## Pointwise theorem + +Define + +```text +B_d(c) = Q^d sum_(k=0,...,g-d) + binomial(g-d,k) t^k (W-t)^(g-d-k) + [L^(g-d-k) M^k > c]. + +E(c) = W^(g+1) + (Q-W) sum_(d=0,...,g-1) B_d(c). (2) +``` + +Square brackets denote an indicator. Then S(c) >= E(c). Moreover E(c) +is the exact minimum of S(c) over all inventories with these five summaries. +The empty sum gives E(c)=W when g=0. + +To prove the inequality, preserve token identities and replace every +nonmaximum productive rank by L and every paid rank by 1. Preserve all t +maximum-rank tokens at M. Every word's rank product weakly decreases. +Consequently each word still above c after replacement was above c before +replacement, so D_d(c) >= B_d(c). + +For the replaced word, choosing its k maximum-rank productive positions +and their identities gives precisely the summand in (2). All Q^d paid +token words have the same replaced rank 1, but remain distinct words. +Multiplying by the nonnegative Q-W in (1) proves the inequality. + +The replaced inventory itself has the required exact minimum, maximum, +counts and separation. Its productive histogram is +{L: W-t, M: t}; its paid histogram is {1: Q}. Every preceding inequality +is equality for that inventory, proving sharpness. This is sharpness over +the summary-data class, not attainment by a native or physical compiler. + +The sign condition matters. For example, take L=2, M=4, W=3, t=1, Q=1 +and c=4. The collapsed inventory has 17 sources, whereas productive ranks +(2,3,4) and the same single rank-one paid token have 11. Thus (2) is not +a lower bound if Q= L^(g+1). (3) +``` + +The distinction between >c and >=L^(g+1) is the integer boundary, not +a relaxation. In particular, rank product equal to c is eligible and +must not be charged as a deficit. + +## Power-1024 application + +The following are cited source facts from PR #5318 and the retained +[#4470 invoice](../joint-helper-lattice/RESULT.json), not results +reverified here. Under that unchanged historical product, completion and +split rule, + +```text +v = 1024, +L = 21^v, M = 145^v, W = 10928^v, t = 224^v, +Q >= Q0 = 64620^v > W, and every paid rank is below L. +``` + +Exactly t maximum-rank tokens remain because a product of v productive +ranks equals 145^v precisely when every factor is one of the 224 rank-145 +tokens. The completion floor is below this maximum. The inherited lower +bound Q>=Q0 includes the fact that each raw paid call yields at least one +positive split call; completion may add more. + +The exact small integer comparisons are + +```text +21 < 145 < 21^2, 145^2 > 21^3. +``` + +At g=1 and the slab's upper cap, even M is eligible, so the best uniform +summary-data floor remains W^2. At g=2, exactly the pair of maximum-rank +productive tokens is ineligible in the collapsed depth-zero word. A word +with only one maximum has product L M < L^3. At depth one, even M < L^3 +is eligible. Equations (2) and (3) therefore give the exact summary minimum + +```text +min_(L^2 <= c < L^3) S(c) = W^3 + (Q-W)t^2. (4) +``` + +The minimum in (4) ranges over inventories with the declared summaries +as well as caps. For the actual historical inventory it is a lower bound, +not an asserted exact source count. Substituting Q>=Q0 gives the new +explicit correction + +```text +S(c) >= 10928^(3072) + + (64620^1024 - 10928^1024) * 224^2048 + > W^3. +``` + +More generally, for every g>=2 and every cap in its height slab, any +depth-zero word with at least two maximum-rank tokens is ineligible: +L^(g-k)M^k >= L^(g-2)M^2 > L^(g+1) > c for k>=2. +Counting just these words in (1) proves + +```text +S(c) >= W^(g+1) + (Q0-W) + * (W^g - (W-t)^g - g*t*(W-t)^(g-1)) + > W^(g+1). (5) +``` + +Equation (2) can be stronger because it retains other depths and threshold +terms. Formula (5) is a simple explicit correction valid throughout every +height slab g>=2 of the actual power-1024 model. The all-height quantifier +is proved above; the executable does not enumerate its enormous heights. + +## Scope + +This improves a mathematical source-count lower bound for an unchanged +historical nested compiler. It supplies neither a higher accumulated K/N +coefficient nor a smaller compiler width. The selected #7000 construction, +kappa, physical stock, routing, ownership, native realization, precision/tape, +all-size admission and analytic recovery are unchanged. No complete +multiplication theorem or independent human or proof-kernel review is claimed. diff --git a/research/binomial-source-floor/README.md b/research/binomial-source-floor/README.md new file mode 100644 index 0000000000..fb4005563b --- /dev/null +++ b/research/binomial-source-floor/README.md @@ -0,0 +1,134 @@ +# Sharp binomial source floors for separated inventories + +Scoped written proof with exact finite corroboration, produced by Nick Petalas. +No new kappa. Apache-2.0. + +The all-height floor `S(c) >= W^(g+1)` from Ali Romman's +[PR #5318](https://github.com/CrocSwap/integer-mult-bounds/pull/5318) now has +a sharp summary-data refinement. Fix the productive minimum L, maximum M, +maximum-rank multiplicity t, productive count W and paid count Q>=W, with +all positive paid ranks below L. The exact minimum source count among those +inventories is a finite binomial-tail expression, attained by productive +ranks `{L:W-t,M:t}` and paid ranks `{1:Q}`. + +For `L^g <= c < L^(g+1)`, the new floor is + +```text +W^(g+1) + (Q-W) sum_(d=0,...,g-1) Q^d + * sum_(k=0,...,g-d) binomial(g-d,k) t^k (W-t)^(g-d-k) + [L^(g-d-k) M^k > c]. +``` + +[PROOF.md](PROOF.md) gives the tokenwise comparison and sharpness proof. +The entire height slab has its sharp uniform floor at `c=L^(g+1)-1`. +This is an exact extremal result for an abstract inventory class, not an +exact count or an attaining implementation for a selected supplier. + +For the unchanged historical #4470 rule at product power 1024, the new +first strict uniform correction occurs at height two: + +```text +L=21^1024, M=145^1024, W=10928^1024, t=224^1024, +Q >= Q0=64620^1024, + +S(c) >= W^3 + (Q0-W)t^2 > W^3 when L^2 <= c < L^3. +``` + +For every larger height g>=2, a simple uniform correction is + +```text +S(c) >= W^(g+1) + (Q0-W) + * (W^g - (W-t)^g - g*t*(W-t)^(g-1)) > W^(g+1). +``` + +This quantifies part of #5318's explicitly open top-slab deficit question +at the actual product power. The correction is small. No improved accumulated +K/N coefficient, reduced compiler width or physical saving is claimed. +The selected #7000 frontier remains unchanged. + +## Evidence and reproduction + +From the repository root, using CPython 3.12 and the standard library: + +```sh +/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -W error -B research/binomial-source-floor/verify.py +/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -O -W error -B research/binomial-source-floor/verify.py +/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -B scripts/research_package.py --package research/binomial-source-floor +/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -B scripts/check_research_commit.py --package research/binomial-source-floor --check-source-history +``` + +On another machine, replace the absolute interpreter path with its CPython +3.12 executable. The direct marker is `PASS_BINOMIAL_SOURCE_FLOOR no_new_kappa`. +The checker prints a generated receipt with package and cited-input hashes. +`--receipt /absolute/path/outside/the/checkout/receipt.json` also saves it. +It creates no files by default. + +The independent root/nonroot slot enumerator passes 2,250 inventory/cap cases, +including 459 strict comparisons. A product-histogram implementation passes +1,640 attaining cases through height six. Eight semantic checks cover the +strict cutoff, height change, exact first correction and the Q cap + ) + correction += paid**depth * missing + return targets**(degree + 1) + (paid - targets) * correction diff --git a/research/binomial-source-floor/verify.py b/research/binomial-source-floor/verify.py new file mode 100644 index 0000000000..39d7996378 --- /dev/null +++ b/research/binomial-source-floor/verify.py @@ -0,0 +1,197 @@ +#!/usr/bin/env python3 +"""Check the new lower envelope without running predecessor executables.""" + +import argparse +from collections import Counter +from hashlib import sha256 +from itertools import combinations_with_replacement, product +import json +from math import prod +from pathlib import Path +import sys + +from floor import height, need, source_floor + + +PACKAGE = Path(__file__).resolve().parent +ROOT = PACKAGE.parents[1] + + +def source_inputs(): + manifest = json.loads((PACKAGE / "SOURCES.json").read_text()) + hashes = {} + for row in manifest["files"]: + if row["usage"] == "input": + raw = (ROOT / row["path"]).read_bytes() + digest = sha256(raw).hexdigest() + need(digest == row["sha256"], "changed input: " + row["path"]) + hashes[row["path"]] = digest + need(len(hashes) == 2, "expected two cited source inputs") + return hashes + + +def products(tokens, length): + """Literal token words: multiplicities remain distinct positions.""" + need(len(tokens) <= 5 and 0 <= length <= 3, "literal word bound") + return tuple(prod(word) for word in product(tokens, repeat=length)) + + +def source_slots(targets, paid, cap): + """Independent root/nonroot slots, without the telescope or binomial law.""" + low = min(targets) + degree = height(low, cap) + need(degree <= 3 and len(targets) <= 5 and len(paid) <= 5, "slot domain") + need(max(paid) < low, "paid/target separation") + total = len(targets) * sum(value <= cap for value in products(targets, degree)) + for depth in range(1, degree + 1): + prefixes = products(paid, depth - 1) + suffixes = products(targets, degree - depth) + for prefix, rank, suffix, last in product(prefixes, paid, suffixes, targets): + if prefix * rank * suffix <= cap < prefix * suffix * last: + total += 1 + return total + + +def histogram_source(targets, paid, cap): + """Independent product histograms for larger attaining examples.""" + degree = height(min(targets), cap) + need(len(targets) <= 8 and len(paid) <= 8, "histogram token bound") + need(max(paid) < min(targets), "histogram separation") + target_powers, paid_powers = [Counter({1: 1})], [Counter({1: 1})] + for _ in range(degree): + for tokens, powers in ((targets, target_powers), (paid, paid_powers)): + following = Counter() + for value, count in powers[-1].items(): + for rank in tokens: + following[value * rank] += count + powers.append(following) + eligible = [ + sum(a * b for r, a in paid_powers[d].items() + for k, b in target_powers[degree - d].items() if r * k <= cap) + for d in range(degree + 1) + ] + w, q = len(targets), len(paid) + return w * eligible[0] + sum( + w * eligible[d] - q * eligible[d - 1] for d in range(1, degree + 1) + ) + + +def literal_checks(): + cases = strict = 0 + for low in (2, 3): + for maximum in (low + 1, low + 2): + for middle in range(low, maximum): + targets = (low, middle, maximum) + for q in (3, 4): + for paid in combinations_with_replacement(range(1, low), q): + for cap in range(1, low**4): + actual = source_slots(targets, paid, cap) + bound = source_floor(low, maximum, 3, q, 1, cap) + need(actual >= bound, "lower envelope failed") + if middle == low and set(paid) == {1}: + need(actual == bound, "collapsed inventory must attain") + strict += actual > bound + cases += 1 + return {"literal_inventories_and_caps": cases, "strict_comparisons": strict} + + +def attaining_checks(): + cases = 0 + for low in (2, 3, 4): + for maximum in (low + 1, 2 * low + 1): + for w in range(2, 6): + for t in range(1, w): + for q in (w, w + 1): + targets = (low,) * (w - t) + (maximum,) * t + for degree in range(7): + caps = {low**degree, low**(degree + 1) - 1} + for cap in sorted(caps): + bound = source_floor(low, maximum, w, q, t, cap) + actual = histogram_source(targets, (1,) * q, cap) + need(actual == bound, "sharpness failed") + cases += 1 + return {"attaining_inventories_and_caps": cases} + + +def controls(): + need(source_floor(3, 8, 3, 5, 1, 26) == 29, "sharp first correction") + need(source_slots((3, 7, 8), (1,) * 5, 26) == 35, "uncollapsed fixture") + need(source_floor(3, 8, 3, 3, 1, 26) == 27, "Q=W boundary") + need(source_floor(3, 8, 3, 5, 1, 8) == 9, "height-one slab") + need(source_floor(3, 8, 3, 5, 1, 9) == 37, "new slab entry") + need(source_floor(2, 3, 2, 3, 1, 3) == 4, "strict cutoff equality") + # Q low**3, "two-block threshold") + need(q0 > w and 0 < t < w, "positive correction hypotheses") + extra = (q0 - w) * t**2 + missing = sum(a * b for a, n in ((w - t, low), (t, maximum)) + for b, m in ((w - t, low), (t, maximum)) if n * m >= low**3) + need(missing == t**2, "first activation count") + need(all(rank < low**3 for rank in (low, maximum)), "depth-one deficit") + result = w**3 + extra + return {"product_power": v, "first_strict_height": 2, + "first_strict_floor_sha256": sha256(str(result).encode()).hexdigest(), + "correction_bit_length": extra.bit_length(), + "correction_sha256": sha256(str(extra).encode()).hexdigest()} + + +def run(): + hashes = source_inputs() + result = {"schema_version": 1, "status": "PASS_BINOMIAL_SOURCE_FLOOR"} + for check in (literal_checks, attaining_checks, controls, selected_application): + result.update(check()) + need(source_inputs() == hashes, "input changed during verification") + result["source_sha256"] = hashes + result["package_sha256"] = { + p.name: sha256(p.read_bytes()).hexdigest() + for p in sorted(PACKAGE.iterdir()) if p.is_file() and p.suffix in (".py", ".md", ".json") + } + return result + + +def main(): + parser = argparse.ArgumentParser(description=__doc__) + parser.add_argument("--receipt", type=Path, help="write a hash-bound receipt outside the checkout") + args = parser.parse_args() + if args.receipt: + need(not args.receipt.resolve().is_relative_to(ROOT), "receipt must be outside checkout") + sys.set_int_max_str_digits(100000) + result = run() + encoded = json.dumps(result, indent=2, sort_keys=True) + "\n" + if args.receipt: + args.receipt.write_text(encoded) + print(encoded, end="") + print("PASS_BINOMIAL_SOURCE_FLOOR no_new_kappa") + + +if __name__ == "__main__": + main()