Skip to content
Merged
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
225 changes: 225 additions & 0 deletions research/projection-recursion-limit/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,225 @@
# Exact limit of the projection recursion

This is a theorem about an explicit arithmetic recurrence. Its interpretation
as a complex-engine lower bound inherits the representation and hook premises
of Jacob Alan Sussman #4628 and Maximilian Lutz #4643. None of those premises
is proved or certified here.

## Statement

Fix an integer b >= 1 and a nonnegative real profile a_b(0),...,a_b(b),
with a_b(0)=0 and a_b(1)>0. For k>b and 0<=r<=k, define

```
a_k(r) = sum_s N_k(r,s) a_(k-1)(s),
N_k(r,r) = 2^(k-r)-1 + 4^k-2^(k+r),
N_k(r,r-1) = 2(2^k-2^(k-r)) + (2^r-1)2^(k-1),
N_k(r,r-2) = (2^r-1)(2^(r-1)-1)2^(k-r).
```

The third entry is zero at r=0,1. The first entry is zero at r=k.
Zero entries outside the previous profile's range are omitted. At r=0
only N_k(0,0)=4^k-1 remains. These are the cited predecessor counts.
No optional triangle/max strengthening is part of this recurrence.

Write [b choose r]_2 for the Gaussian binomial coefficient and put

```
Z_b = product_(i=1)^b (2^i+1),
C_b = (2^(b+1)+1)/(2^b-1),
f(r) = a_b(r)/a_b(1),
F = max_(0<=r<=b) f(r),
T_m = a_m(m)/a_m(1).
```

Then the exact limit is

```
L_b = (C_b/2) / Z_b
* sum_(r=0)^b [b choose r]_2 2^(r(r+1)/2) f(r).
```

For every pair of integers b<=K<m, the following explicit error bound holds:

```
|T_m-L_b| <= F * ((C_b/2) * (9/(2(m-K)) + 2^(2-K))
+ 3 C_b/(2(2^(m+1)+1))). (1)
```

Choosing K=floor(m/2) for m>=2b proves convergence. The formula is exact
for every seed profile in the stated domain, rather than a decimal estimate
of one seed's tail. It does not assert monotonicity or that L_b is an
all-m lower bound.

For b=4 and the cited seed (0,27456,48576,68320,86592),

```
Z_4 = 2295,
([4 choose r]_2 2^(r(r+1)/2))_(r=0)^4 = (1,30,280,960,1024),
C_4/2 = 11/10,
L_4 = 439277/149175 = 2.9447092341209986... .
```

The exact rational limit and the general limit/error theorem are the new
result. The predecessor's quoted decimal 2.944709234 agrees to its displayed
precision; that agreement is not a new hook-rank verification.

## 1. Normalize the recurrence

Each row of N_k sums to 4^k-1. Thus P_k(r,s)=N_k(r,s)/(4^k-1)
is a stochastic reduction from block k to k-1. If X starts at distance m
at block m and follows these reductions down to b, then

```
a_m(m) = a_b(1) product_(k=b+1)^m (4^k-1) * E[f(X_b)].
```

At r=1 only the s=1 term contributes, with coefficient

```
N_k(1,1) = (2^(k-1)-1)(2^(k+1)+1).
```

Both products telescope to give

```
T_m = G_(m,b) E[f(X_b)],
G_(m,b) = C_b (2^m-1)/(2^(m+1)+1),
G_(m,b) -> C_b/2,
C_b/2-G_(m,b) = 3 C_b/(2(2^(m+1)+1)). (2)
```

## 2. A coherent family of reference distributions

In a binary symplectic space of dimension 2k, fix a Lagrangian L.
There are

```
D_k(r) = [k choose r]_2 2^(r(r+1)/2)
```

Lagrangians M with distance r=k-dim(L intersect M). To see this, first
choose I=L intersect M of codimension r in L. In I-perp/I, a transverse
Lagrangian to L/I is the graph of a symmetric r-by-r binary matrix over
a chosen complementary Lagrangian. There are 2^(r(r+1)/2) such matrices.

The total number Z_k is product_(i=1)^k(2^i+1). An elementary induction
counts pairs (p,M), with 0!=p in M. There are 2^k-1 choices in each M,
and each of the 4^k-1 nonzero p belongs to Z_(k-1) Lagrangians, by
reduction in p-perp/<p>. Hence Z_k=(2^k+1)Z_(k-1).

For uniform M and independently uniform nonzero p, the reduced M is
uniform among Lagrangians of p-perp/<p>. Indeed each symplectic map of
that quotient lifts, by extending a symplectic basis, to a map fixing p;
the lifted map permutes the original uniform M's and commutes with
reduction. The quotient action is transitive. The reduced L is fixed
conditional on p. Consequently the distance law D_k/Z_k reduces to
D_(k-1)/Z_(k-1). This also follows by substituting the three displayed
N_k coefficients into the Gaussian-binomial formula.

Use the deficit j=k-r. Let pi_k(j)=D_k(k-j)/Z_k for 0<=j<=k.
In deficit coordinates the coherence identity is pi_k P_k=pi_(k-1).
For fixed j,

```
pi_k(j) = w_j * product_(i=k-j+1)^k(1-2^-i)
/ product_(i=1)^k(1+2^-i),
w_0=1, w_j=1/product_(i=1)^j(2^i-1).
```

Since pi_k(j)<=w_j, and w_j<=3^(-(j-1)) for j>=1, dominated
convergence on this countable set proves total-variation convergence
pi_k -> pi, where pi(j)=w_j/sum_h w_h. It also proves that
sum_h w_h=product_(i=1)^infinity(1+2^-i). No infinite numerical product
is needed to evaluate L_b.

## 3. A homogeneous comparison chain

Put M=2^k and z=2^j. The deficit transition from j has probabilities

```
u_k(j) = (M^2/(2z)-3M/2+z)/(M^2-1), to j+1,
d_k(j) = (z-1+M^2-M^2/z)/(M^2-1), to j-1,
1-u_k(j)-d_k(j), to j.
```

At j=k the transition is deterministically to k-1; the same formulas
apply. Define the homogeneous chain Q on nonnegative integers by

```
u(j)=2^(-j-1), d(j)=1-2^-j, stay(j)=2^(-j-1).
```

Detailed balance holds because w_(j+1)/w_j=1/(2^(j+1)-1), so pi is
stationary for Q. Its expected deficit is at most 9/8: the unnormalized
first moment is at most sum_(j>=1) j/3^(j-1)=9/4 and the normalizer
is at least w_0+w_1=2.

Q is monotone. Adjacent states can cross in an ordered coupling only
if u(j)+d(j+1)>1; here their sum equals 1. Use inverse distribution
functions driven by one uniform random number to preserve order.
Couple a chain starting at 0 below a stationary chain starting with pi.
When the upper chain first reaches 0 they coalesce.

From j>=1 the upper chain's mean increment is
3*2^(-j-1)-1<=-1/4. Stopping at 0 and at a deterministic time n gives
E[min(tau,n)]<=4j by summing this drift and using nonnegativity of the
stopped state. Letting n increase gives E[tau|j]<=4j. Thus the coupling
time has mean at most 9/2, and Markov's inequality proves, for n>=1,

```
TV(delta_0 Q^n, pi) <= 9/(2n). (3)
```

This argument supplies recurrence and convergence directly; it invokes
no unproved mixing assumption.

For 0<=j<=k, direct subtraction gives

```
u_k(j)-u(j) = (-3M/2+z+1/(2z))/(M^2-1) <= 0,
d_k(j)-d(j) = (z-1/z)/(M^2-1) >= 0.
```

The total variation of the three-entry rows is the maximum of these
two absolute differences. The first is <=2/M and the second <=4/(3M)
for M>=2. Extend P_k to j>k by using Q there. Therefore

```
sup_j TV(P_k(j,.),Q(j,.)) <= 2^(1-k). (4)
```

## 4. Convergence down to the fixed seed block

Let A_(m,K)=delta_0 P_m ... P_(K+1). The starting deficit is 0.
Stochastic kernels contract total variation. Telescoping (4), with
sum_(k=K+1)^m 2^(1-k)<=2^(1-K), and (3) give

```
TV(A_(m,K),pi) <= 9/(2(m-K)) + 2^(1-K).
```

Apply the same comparison with starting distribution pi_m. Coherence
gives pi_m P_m ... P_(K+1)=pi_K, while stationarity gives
pi Q^(m-K)=pi. Since pi_m -> pi, taking m to infinity yields
TV(pi_K,pi)<=2^(1-K). Hence

```
TV(A_(m,K),pi_K) <= 9/(2(m-K)) + 2^(2-K).
```

Propagate both laws from K to b. Coherence and contraction show that
the same upper bound holds for TV(A_(m,b),pi_b). In distance coordinates
pi_b has probabilities D_b(r)/Z_b. Since 0<=f<=F, the expectation
difference is at most F times this total variation bound. Combine this
with (2) and G_(m,b)<=C_b/2 to obtain (1) and the asserted limit.

## Limits of the conclusion

The hook values are premises, including the two lower bounds whose stored
rank certificates #4643 leaves open. The result is unconditional as a theorem
about the stated recurrence, and conditional when interpreted as an engine
bound. It certifies no engine implementation, new paid construction, physical
cost, selected frontier, or complete multiplication theorem. The optional
triangle strengthening is outside the theorem. No infimum over finite m,
monotonicity theorem, or formal proof-kernel review is asserted.
122 changes: 122 additions & 0 deletions research/projection-recursion-limit/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,122 @@
# Exact rational limit of the Lagrangian projection recurrence

**Scoped mathematical proof with finite corroboration; no new kappa.**
The projection recurrence of Jacob Alan Sussman [#4628](https://github.com/CrocSwap/integer-mult-bounds/pull/4628)
and Maximilian Lutz [#4643](https://github.com/CrocSwap/integer-mult-bounds/pull/4643)
has the exact asymptotic ratio

```
lim_(m->infinity) a_m(m)/a_m(1) = 439277/149175.
```

The displayed decimal is 2.9447092341209986... . The new result identifies
this limit exactly, proves convergence with an explicit error bound, and
gives the limit for every nonnegative seed profile at any fixed seed block.
It extends the predecessors' numerical tail description. It does not
re-check their hook ranks, projection geometry or engine implementation.

## The result

[PROOF.md](PROOF.md) defines the recurrence and proves that a seed profile
`a_b(0)=0`, `a_b(1)>0`, `a_b(r)>=0` has limit

```
(2^(b+1)+1) / (2(2^b-1) product_(i=1)^b(2^i+1))
* sum_(r=0)^b [b choose r]_2 2^(r(r+1)/2) a_b(r)/a_b(1).
```

Here `[b choose r]_2` is the Gaussian binomial coefficient. For the cited
block-four seed `(0,27456,48576,68320,86592)`, the distance weights are
`(1,30,280,960,1024)/2295` and the remaining factor is `11/10`.
The result is a finite rational expression, despite the infinite products
and stationary laws that appear in its proof.

The proof normalizes the integer recurrence into a Markov reduction,
uses coherent uniform-Lagrangian distance laws, and compares the deficit
with a homogeneous birth-and-death chain. A monotone coupling has expected
coalescence time at most `9/2`. A summable row-error bound proves convergence
to the coherent seed-block law. Equation (1) of PROOF.md gives the explicit
all-m error estimate. The proof invokes no numerical mixing assumption.

This theorem is unconditional for the displayed arithmetic recurrence.
Its engine-game interpretation is conditional on the predecessor's hook
profile and representation argument. In particular, the two stored-rank
certificate obligations identified in #4643 remain unresolved here.
The result does not assert monotonicity, a finite-m infimum, the optional
triangle/max strengthened recurrence, physical costs, native admission,
a new supplier, or a complete multiplication theorem. The selected
frontier is unchanged. Independent human and formal proof-kernel review
are unclaimed.

## Reproduce

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

```sh
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -B scripts/research_package.py --package research/projection-recursion-limit
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -I -X dev -W error -B research/projection-recursion-limit/verify.py
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -I -O -X dev -W error -B research/projection-recursion-limit/verify.py
```

Another CPython 3.12 interpreter may replace the absolute local path.
Success markers are `PASS_PROJECTION_RECURSION_LIMIT` and
`PASS_REGISTERED_RESEARCH_PACKAGE`. Both registered commands run with a
120-second timeout. Default verification is read-only and writes no receipt.
Explicit `--emit` prints newly generated hash-bound evidence to standard
output; [RESULT.json](RESULT.json) was generated through that command.

The standalone checker imports no predecessor code. It checks 3,320 new
kernel-coordinate and row-error identities, 63 coherence reductions,
80 detailed-balance/monotonicity indices, 73 seed-profile limit enclosures,
54 enclosures for six other seed blocks, and 38 independent comparisons
between probability-law propagation and the integer recurrence. Six
controls distinguish wrong normalization, a wrong exact limit, a changed
seed, a corrupt transition and a wrong coherent law. These finite checks
corroborate the written proof; they do not replace its infinite-limit
argument. The generated receipt binds the verifier, proof and cited source
README by SHA256. The registered runner also authenticates source metadata.

New code is bounded to block/index 100 and uses explicit exceptions under
normal and optimized Python. Each function is below 60 lines. Python's
managed integers and fractions allocate during execution, a deliberate
Power of Ten deviation; the finite index limits bound live collections
and integer sizes. No recursion, dynamic code execution, external solver,
network call, subprocess or background process occurs in the verifier.
Warning-strict development-mode replay supplies interpreter diagnostics;
no static type checker or formal kernel is claimed.

## Sources and overlap

Scientific base is `a5df48c55ea806d7952b65acc9c76d878ff09b13`, priority
revision 7. [SOURCES.json](SOURCES.json) pins the predecessor README,
task catalogue and selected record. Only the predecessor README is a live
input; its source bytes remain unchanged. Arbitrary code diffs and unpublished
work are outside the overlap search. The sparse local-note inventory is
incomplete and was supplemented with targeted main-tree and merged-PR searches.
The local card changes from independent ledger checking to this original
closed-form recurrence theorem; it is a complex small-block side result.

## Open next questions

1. Does this recurrence admit a finite-m closed form, or a faster convergence
rate? The present proof gives an explicit error estimate and an exact limit.
Begin with the coherent-law comparison rather than recomputing hook ranks.
2. Which seed profiles make the top ratios monotone, and when is their
infimum the limit? A limit alone does not settle these questions. Start by
examining adjacent-block differences as linear functionals of the seed.

Neither question requires another physical construction round. The separate
engine, hook-rank, native, precision, hosting and complete-cost obligations
remain with their existing investigations. Structured questions are in
[TASK.json](TASK.json).

## Credit

This result was produced by Nick Petalas. Jacob Alan Sussman #4628 supplies
the projection recurrence and block-four hook profile. Maximilian Lutz #4643
supplies its public count formulas, coherent geometry context and prior tail
analysis, including the explicit unresolved rank obligations. Eumemic #288
and Swapnil Jain retain the underlying certificate-method credit. Classical
Gaussian-binomial counting, symplectic reduction, birth-and-death chains and
coupling arguments retain their mathematical authorship. No literature-priority
claim is made for those methods. New work is Apache-2.0.
14 changes: 14 additions & 0 deletions research/projection-recursion-limit/RESULT.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,14 @@
{
"claim_type": "scoped recurrence proof with finite corroboration",
"coherence_transitions": 63,
"failure_controls": 6,
"general_seed_enclosures": 54,
"kernel_rows": 3320,
"limit": "439277/149175",
"markov_raw_comparisons": 38,
"proof_sha256": "83843a1460d8d5ccca4c312268e3a138f5dddc8197986ba63955392a3da0da33",
"seed_limit_enclosures": 73,
"source_sha256": "59204f11aa3aae59df44fc69228cd017660139587b2dc18ece1999a98d097996",
"stationary_indices": 80,
"verifier_sha256": "5014e6aef2eb9fb3aa77d460a4a4946bbad3dbc470f4bccc28de6b905920f505"
}
24 changes: 24 additions & 0 deletions research/projection-recursion-limit/SOURCES.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,24 @@
{
"schema_version": 1,
"base_commit": "a5df48c55ea806d7952b65acc9c76d878ff09b13",
"files": [
{
"path": "certificates/selected-result.json",
"usage": "provenance",
"git_blob": "c35482cd9035843b15a634cc8276a7e1e40e8395",
"sha256": "6fe0a099b89aa2aed796a13c3a21dafc63f3388ebf397464e44e0451893b9375"
},
{
"path": "docs/research/agent-task-catalog.json",
"usage": "provenance",
"git_blob": "f126dd2e2dad2b669e75d1d0d0bd05c600725602",
"sha256": "f7a48bfa2659b88378282a8eee7d9fe52cdf505594124c4d1686b58510fffd72"
},
{
"path": "research/projection-recursion-check/README.md",
"usage": "input",
"git_blob": "1d8c9f7b1b96cd098a9a00a90022143bbced8b64",
"sha256": "59204f11aa3aae59df44fc69228cd017660139587b2dc18ece1999a98d097996"
}
]
}
Loading