|
| 1 | +# Mode-fiber rounding radius: a finite reduction and 1 ≤ R ≤ 6 |
| 2 | + |
| 3 | +Prepared for apoursh with substantial Anthropic Claude assistance. Apache-2.0. |
| 4 | + |
| 5 | +## Setting |
| 6 | + |
| 7 | +M is the 112×8 mode matrix of Dillon Amburgey's |
| 8 | +[PR4933](https://github.com/CrocSwap/integer-mult-bounds/pull/4933), read from |
| 9 | +`research/coded-native-caller/NORMALIZERS.json` (`mode_columns`), with rows |
| 10 | +m_0, …, m_111. Nick Petalas's #6312 (`research/mode-preserving-rounding`, open |
| 11 | +at head `46593234bfe658b63fb829f0d22b7eb216e392a2`) studies rounding x ∈ ℝ^112 to |
| 12 | +z ∈ ℤ^112 with Mᵀz = Mᵀx, and asks in its first open question for the least |
| 13 | +uniform radius |
| 14 | + |
| 15 | + R = sup { ρ(x) : the fiber F(x) = {z ∈ ℤ^112 : Mᵀz = Mᵀx} is nonempty }, |
| 16 | + ρ(x) = min { ‖z − x‖∞ : z ∈ F(x) }. |
| 17 | + |
| 18 | +The minimum exists because F(x) is discrete. #6312 proves R ≥ 1, using |
| 19 | +x = (e_0+e_7+e_20+e_111)/2. This note gives a finite characterisation of R = 1 |
| 20 | +and a first explicit upper bound. |
| 21 | + |
| 22 | +Write L = Mᵀℤ^112 ⊂ ℤ^8, V = ker Mᵀ ⊂ ℝ^112, Z = Mᵀ[0,1]^112 (a zonotope) and |
| 23 | +S = {Mᵀz : z ∈ {0,1}^112} (subset sums of rows). |
| 24 | + |
| 25 | +## 1. Structure of M (exact, verify.py) |
| 26 | + |
| 27 | +* Every row has four nonzero entries, all ±1. The rows are distinct, and MᵀM = 56·I. |
| 28 | +* The 14 row supports are exactly the 14 weight-four words of the extended |
| 29 | + Hamming code H8 = [8,4,4]. Each support carries 8 rows, so 8 of its 16 sign |
| 30 | + patterns occur. The row set is **not** closed under negation. |
| 31 | +* L = {x ∈ ℤ^8 : x mod 2 ∈ H8}, the Hamming construction of √2·E8, with |
| 32 | + [ℤ^8 : L] = 16. verify.py proves this by comparing Hermite normal forms |
| 33 | + of the rows and of H8 ∪ 2ℤ^8. |
| 34 | +* For every i, the other 111 rows have rank 8. |
| 35 | + |
| 36 | +## 2. Theorem A (finite reduction) |
| 37 | + |
| 38 | +**R ≤ 1 if and only if L ∩ int Z ⊆ S.** If some t ∈ L ∩ int Z is not a 0/1 |
| 39 | +subset sum of the rows, then every y ∈ (0,1)^112 with Mᵀy = t has ρ(y) > 1, |
| 40 | +so R > 1. With #6312's lower bound: |
| 41 | + |
| 42 | + R = 1 ⇔ every lattice point of L in the open zonotope int Z is a subset sum of rows. |
| 43 | + |
| 44 | +*Proof.* |
| 45 | + |
| 46 | +(a) *Invariance and continuity.* F(x+n) = F(x)+n for n ∈ ℤ^112, so ρ is |
| 47 | +ℤ^112-periodic. For v ∈ V, F(x+v) = F(x), so |ρ(x+v) − ρ(x)| ≤ ‖v‖∞. |
| 48 | + |
| 49 | +(b) *Generic points suffice.* Fix x with F(x) ≠ ∅. The coordinate functional |
| 50 | +v ↦ v_i is not identically zero on V. Otherwise e_i ∈ V^⊥ = col span(M), say |
| 51 | +e_i = Ma. Then m_k·a = 0 for all k ≠ i, and since those rows have rank 8 |
| 52 | +(§1), a = 0, a contradiction. So {v ∈ V : x_i + v_i ∈ ℤ} is a countable union |
| 53 | +of proper affine hyperplanes of V for each i. The union over i is a null set, |
| 54 | +so there are v ∈ V with arbitrarily small ‖v‖∞ such that x+v has no integral |
| 55 | +coordinate. By (a), R is the supremum of ρ over points with no integral coordinate. |
| 56 | + |
| 57 | +(c) *Reduction to the open cube.* For such x, put y = x − ⌊x⌋ ∈ (0,1)^112. |
| 58 | +Then ρ(y) = ρ(x), and t := Mᵀy ∈ L, because Mᵀx ∈ L (the fiber is nonempty) |
| 59 | +and Mᵀ⌊x⌋ ∈ L. For y_i ∈ (0,1), an integer z_i satisfies |z_i − y_i| ≤ 1 if |
| 60 | +and only if z_i ∈ {0,1}. Hence ρ(y) ≤ 1 ⇔ t ∈ S, and otherwise ρ(y) > 1. |
| 61 | + |
| 62 | +(d) *Which t occur.* Mᵀ is a surjective linear map ℝ^112 → ℝ^8. The image of the |
| 63 | +open cube is the interior of the image of the closed cube, so |
| 64 | +{Mᵀy : y ∈ (0,1)^112} = int Z. |
| 65 | + |
| 66 | +Combining (b)–(d): R ≤ 1 ⇔ L ∩ int Z ⊆ S. If t ∈ (L ∩ int Z) \ S, every y in |
| 67 | +the open cube over t has ρ(y) > 1. ∎ |
| 68 | + |
| 69 | +Consequences. Deciding R = 1 is a **finite** problem, since Z is bounded. A |
| 70 | +counterexample has a finite exact certificate: a lattice point t, a rational |
| 71 | +y ∈ (0,1)^112 with Mᵀy = t, and infeasibility of the 0/1 system Mᵀz = t. |
| 72 | +Conversely, R = 1 needs every one of the finitely many interior lattice points |
| 73 | +to be a subset sum. Boundary lattice points of Z impose no condition. |
| 74 | + |
| 75 | +## 3. Theorem B (explicit upper bound): R ≤ 6 |
| 76 | + |
| 77 | +Let B be the eight rows `[7, 35, 57, 73, 87, 90, 92, 99]` from CERTIFICATE.json. |
| 78 | +verify.py checks |det M_B| = 16 = covol(L). Since the rows lie in L, B is a basis |
| 79 | +of L. Each other row m_i (i ∈ N, |N| = 104) has integer coordinates w_i ∈ ℤ^8 |
| 80 | +in this basis, and max_i ‖w_i‖₁ = τ = 6. Let W be the 8×104 matrix with columns w_i. |
| 81 | + |
| 82 | +Take y ∈ (0,1)^112 with Mᵀy = t ∈ L, as in Theorem A(b,c). |
| 83 | + |
| 84 | +*Step 1 (Beck–Fiala rounding of y_N against W).* Keep x ∈ [0,1]^N, starting |
| 85 | +from x = y_N, and let F be its fractional coordinates. Call row b of W |
| 86 | +*active* if Σ_{i∈F}|W_{bi}| > τ. Since Σ_b Σ_{i∈F}|W_{bi}| ≤ τ|F|, there are |
| 87 | +fewer than |F| active rows. So some nonzero d supported on F satisfies |
| 88 | +W_active d = 0. Move x along d until one more coordinate is integral, and |
| 89 | +repeat until F = ∅. Each active row value is preserved. Once a row becomes |
| 90 | +inactive, only coordinates of F change afterwards, each by less than 1. Its |
| 91 | +total later change is therefore less than Σ_{i∈F}|W_{bi}| ≤ τ, or zero. The |
| 92 | +result z_N ∈ {0,1}^N has ‖W(y_N − z_N)‖∞ < τ. |
| 93 | + |
| 94 | +*Step 2 (absorb into the basis).* Set z_B = y_B + W(y_N − z_N). Then |
| 95 | + |
| 96 | + Σ_{b∈B} z_b m_b = Σ_B y_b m_b + Σ_N (y_i − z_i) m_i = t − Σ_N z_i m_i ∈ L, |
| 97 | + |
| 98 | +and B is a basis of L, so z_B ∈ ℤ^8. Hence z ∈ F(y), with |z_i − y_i| < 1 on N |
| 99 | +and |z_b − y_b| < τ on B. |
| 100 | + |
| 101 | +So ρ(y) < 6 for every generic y, and R ≤ 6 by Theorem A(a,b). ∎ |
| 102 | + |
| 103 | +Local search over row bases (`search/basis_search.py`, three seeds × 200 restarts) |
| 104 | +found no basis with τ < 6. This method therefore stops at 6; a better bound |
| 105 | +needs a different argument (see the open questions). |
| 106 | + |
| 107 | +## 4. Evidence for R = 1 (not a proof) |
| 108 | + |
| 109 | +`search/evidence.py` (scipy/HiGHS) samples lattice points t ∈ L near the boundary |
| 110 | +of Z, where Theorem A's condition is hardest. It does this in two ways: |
| 111 | + |
| 112 | +- near a vertex (face rank 0): the vertex minus a small inward lattice vector; |
| 113 | +- near faces spanned by 1–7 random rows: rows off the face are pinned by the side |
| 114 | + of the supporting functional, rows on the face get fractional weights with |
| 115 | + denominator 2, 3, 4 or 6, and a small inward lattice vector is subtracted. |
| 116 | + |
| 117 | +It keeps t with a positive LP interior margin and solves the 0/1 system. All |
| 118 | +1,864 interior points found (by face rank 0–7: 401, 350, 402, 372, 175, 60, 72, 32) |
| 119 | +are subset sums. verify.py re-checks each witness exactly, t ∈ L and Σ_{i∈z} m_i = t. |
| 120 | +Interiority comes only from a floating-point LP and is **not** certified. |
| 121 | +This sample supports, but does not prove, the conjecture R = 1. |
| 122 | + |
| 123 | +## Scope |
| 124 | + |
| 125 | +R is a property of the stated exact-mode contract for this M, inherited from |
| 126 | +#6312. Nothing here shows that an actual caller must preserve these modes, or |
| 127 | +supplies a physical rounder, error ledger, tape/precision bill, supplier |
| 128 | +admission, new κ or frontier change. The five framework assumptions keep |
| 129 | +their status. Theorems A and B are written proofs whose finite inputs are |
| 130 | +checked exactly; no proof-assistant or independent human review is claimed. |
| 131 | +Beck–Fiala's iterative rounding (1981) and the Hamming construction of E8 are |
| 132 | +classical and credited as such. |
0 commit comments