SilverSight/docs/research/sidon_preservation_creation.md

# CRT Torus Embedding: Property Preservation and Creation

Part of Open Direction #3 — characterizing moduli that guarantee F(A) satisfies
a target property P.

---

## 1. Problem

Given A ⊂ ℤ reflection-closed under S, and a target property P (Sidon, B_h,
Golomb ruler), which moduli (L₁, …, L_k) guarantee that F(A) satisfies P?

The Sidon example shows F can *create* P from a non-P set, but this depends
on modulus choice. We need the general condition.

---

## 2. Key Invariant: The Sum Map

For a pair (a, b) in A, the CRT-lifted sum F(a) + F(b) has residues:

| Axis | Constraint |
|------|-----------|
| 1 (identity) | (a + b) mod L₁ |
| i ≥ 2 (reflection) | (2S − a − b) mod Lᵢ |

Two pairs (a,b) and (c,d) produce equal sums modulo M iff:

    a + b ≡ c + d   (mod L₁)
    a + b ≡ c + d   (mod Lᵢ)   ∀i ≥ 2

By CRT: a + b ≡ c + d  (mod M), where M = ∏ Lᵢ.

Therefore:

    F(a) + F(b) ≡ F(c) + F(d)  (mod M)   iff   a + b ≡ c + d  (mod M)

---

## 3. Three Regimes

Let M = ∏ Lᵢ.

### Regime A — M > max(A): injective, wrapping can break collisions

F is injective. Existing sum collisions break when individual CRT lifts
wrap M differently (the wrapping criterion).

**A1: M > 2·max(A)** — no sum alias. All pairwise sums < M, so new
collisions cannot form. Wrapping can still break existing collisions.
Sidon creation IS possible here (e.g., [7,3] with A={1,2,5,6}).

**A2: max(A) < M ≤ 2·max(A)** — sum alias possible. Pairs with different
sums may satisfy |T₁−T₂| = M, creating new collisions. Wrapping + M-diff
both active.

### Regime B — M ≤ max(A): F not injective (aliasing)

Not useful.

### Wrapping works identically in A1 and A2

| (L₁, L₂) | M | Regime | Sidon? | Why |
|----------|---|--------|--------|-----|
| (3, 4) | 12 | A2 | ✓ | Wrapping: 19 vs 7 |
| (7, 3) | 21 | A1 | ✓ | Wrapping: 28 vs 7 |
| (11, 2) | 22 | A1 | ✗ | Same wrap: both sums = 29 |

---

## 4. B_h Generalization

For h-fold sums: wrapping works at ANY M > max(A). M-differences require
M ≤ h·max(A) to be possible (since max h-fold sum = h·max(A)).

| Property | No sum alias (M > h·maxA) | Sum alias possible |
|----------|--------------------------|-------------------|
| Sidon (h=2) | M > 2·max(A): wrapping only, no new collisions | max(A) < M ≤ 2·max(A) |
| B_h (general) | M > h·max(A): wrapping only | max(A) < M ≤ h·max(A) |
| Golomb (differences) | M > max(A)-min(A): wrapping only | boundary case |

## 6. Creation Condition: Complete Characterization

### 6.1 Breaking Existing Collisions (The Wrapping Criterion)

Given a collision a+b = c+d = T in A, the images satisfy:

    F(a)+F(b) = T + r₁·M,    r₁ ∈ {0, 1}
    F(c)+F(d) = T + r₂·M,    r₂ ∈ {0, 1}

The collision is broken iff r₁ ≠ r₂. (Proof: each F(x) < M, so two
sums of two values are < 2M. The wrap indicator r = 1 when F(a)+F(b) ≥ M.)

**Verified:** 500/500 random tests, k=2..8.

### 6.2 Preventing New Collisions (The M-Difference Condition)

A new collision arises when pairs (a,b) and (c,d) with *distinct* original
sums T₁ ≠ T₂ satisfy F(a)+F(b) = F(c)+F(d). This occurs iff:

    |T₁ − T₂| = M  (or a multiple of M)

Since T₁, T₂ ≤ 2·max(A) and M > max(A), the only possible multiple is M.

**Proof.** F(a)+F(b) ≡ F(c)+F(d) (mod M) forces a+b ≡ c+d (mod M), i.e.,
T₁ ≡ T₂ (mod M). Since 0 ≤ T₁, T₂ ≤ 2·max(A) < 2M, we have |T₁−T₂| ∈ {0, M}.
The case 0 is the existing collision (T₁ = T₂). The case M is the new collision.

**Verified:** 416 new collisions across 5000 random trials — ALL satisfy
|T₁−T₂| = M. Zero counterexamples.

### 6.3 Complete Sidon Creation Theorem

**Status: Mathematically argued, formalized in Lean only for the preservation direction.**

The Lean theorem (`CRTSidon.lean`) now formalizes both:
- `sidon_preserved` — **preservation** under componentwise vector addition (stronger property)
- `sidon_preserved_mod` — **preservation** under CRT-reconstructed modular sums (matches Python)
  
The **creation** direction (non-Sidon → Sidon under conditions (a)+(b)) is still unformalized.
Empirical verification (500/500 tests, 416 collisions across 5000 trials) is not a proof.

**Theorem.** For a finite A ⊂ ℤ with reflection closure a ↦ S−a,
moduli L₁,…,Lₖ coprime, L₁,L₂ ≥ 2, and M = ∏ Lᵢ > max(A):

    F(A) is Sidon  ⟺  (a) and (b) both hold:

    (a) For every sum collision a+b = c+d in A:
        (F(a)+F(b) ≥ M) ≠ (F(c)+F(d) ≥ M)    [wrapping criterion]

    (b) For no distinct sums T₁, T₂ ∈ {a+b : a,b ∈ A, a ≤ b}:
        |T₁ − T₂| = M                         [M-difference condition]

**Corollary 1 (No sum alias).** If M > 2·max(A), condition (b) is vacuous
(no sums differ by exactly M). F(A) may still break existing collisions
via wrapping. No new collisions can form.

**Corollary 2 (Sum alias possible).** If max(A) < M ≤ 2·max(A), both
conditions must be checked. F(A) is Sidon iff (a) wrapping breaks all
existing collisions AND (b) no M-differences create new ones. Both
conditions are decidable in O(|A|⁴) time.

**Corollary 3 (Complete classification).**
M > max(A) → wrapping can break existing collisions; M-differences
  may or may not apply depending on if M ≤ 2·max(A).
M ≤ max(A) → F not injective (aliasing).

### 6.4 Algorithmic Guidance for Modulus Selection

Choose moduli to guarantee Sidon creation:

1. Compute all pairwise sums Sₐ = {aᵢ + aⱼ : 0 ≤ i ≤ j < |A|}.
2. Compute differences Dₐ = {|T₁ − T₂| : T₁,T₂ ∈ Sₐ, T₁ ≠ T₂}.
3. Choose M = ∏ Lᵢ such that:
   - M > max(A) (element-level injectivity)
   - M ∉ Dₐ (no new collisions)
4. For each existing collision in A, verify the wrapping criterion (a).
   If any pair wraps the same, pick different moduli or accept
   the collision persists.
5. If (a) and (b) both hold, F(A) is guaranteed Sidon.

### 6.5 Modulus Ordering Principle (Tuning Rule)

The identity axis L₁ and reflection axis L₂ are **not interchangeable**.
Larger L₁ = larger minimum gap = more likely Sidon creation.

**Empirical rule:** Choose L₁ > L₂. For the complex set A = [0,1,3,8,13]:

| (L₁, L₂) | M | Gap | Sidon? | Insight |
|----------|---|-----|--------|---------|
| (7, 3) | 21 | ≥7 | ✓ | L₁=7 large identity axis |
| (11, 2) | 22 | ≥11 | ✓ | L₁=11 even larger |
| (8, 3) | 24 | ≥8 | ✓ | L₁=8 |
| (13, 2) | 26 | ≥13 | ✓ | L₁=13, max gap |
| (3, 7) | 21 | ≥3 | ✗ | L₁=3 too small |
| (2, 11) | 22 | ≥2 | ✗ | L₁=2 minimal gap |

All 4 successes have L₁ > L₂. All failures with L₁ < L₂ have insufficient gap for
this specific set. (When both L₁ ≈ L₂, other factors like the wrapping criterion
and M-difference condition dominate.)

**Practical rule:**
1. Choose L₁ as large as possible (up to 2·maxA / L₂)
2. Choose L₂ as the smallest coprime integer that keeps M in (maxA, 2·maxA]
3. Typically L₂ = 2 (smallest possible) and L₁ = ⌊2·maxA / L₂⌋, adjusted
   downward for coprimality

This maximizes the gap L₁, which maximizes the chance of breaking existing
sum collisions via the wrapping criterion.

**Tradeoff:** Larger L₁ also means larger M. If M exceeds 2·maxA,
the M-difference condition becomes vacuous (no new collisions), but
wrapping can still break existing ones. The optimal is L₁ ≈ 1.9·maxA
from sweep data (29% success rate).