Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Subtraction-preserving observation systems and an integer instability family

Definition
NewMathDiscovery_LinearBlindness

by Sodelin · Sep 8, 2026 · Mathlib 0df444a (Lean v4.33.1)

algebrainformationkernelsodelin-newmath

Let I,X,VI,X,VI,X,V be arbitrary types with distinguished zeros in X,VX,VX,V, subtraction operations, observation maps ri:X→Vr_i:X\to Vri​:X→V and target q:X→Vq:X\to Vq:X→V. Assume the maps preserve zero and subtraction, and subtraction in VVV is zero exactly when its arguments are equal. For a selected set S⊆IS\subseteq IS⊆I, let KS={d∈X:∀i∈S, ri(d)=0}K_S=\{d\in X:\forall i\in S,\ r_i(d)=0\}KS​={d∈X:∀i∈S, ri​(d)=0}.

Blind⁡(S)  ⟺  ∃x,y, q(x)≠q(y) ∧ ∀i∈S, ri(x)=ri(y),Witness⁡(S)  ⟺  ∃d∈KS, q(d)≠0.\operatorname{Blind}(S)\iff\exists x,y,\ q(x)\ne q(y)\ \land\ \forall i\in S,\ r_i(x)=r_i(y),\qquad\operatorname{Witness}(S)\iff\exists d\in K_S,\ q(d)\ne0.Blind(S)⟺∃x,y, q(x)=q(y) ∧ ∀i∈S, ri​(x)=ri​(y),Witness(S)⟺∃d∈KS​, q(d)=0.

Target determination means equal selected observations force equal targets. A separate concrete interface on Z2\mathbb Z^2Z2 sets q(x)=x1q(x)=x_1q(x)=x1​, rn(x)=x1+(n+1)x2r_n(x)=x_1+(n+1)x_2rn​(x)=x1​+(n+1)x2​, wn=(n+1,−1)w_n=(n+1,-1)wn​=(n+1,−1), and the usual 2-by-2 determinant.

These definitions support the exact kernel criteria and integer witness theorems. The abstract interface is weaker than a vector space; no scalar operations, normalized metric, or condition-number assumptions are introduced.

Definition code
import Std
import Init.Grind.Ordered.Module
/-!
# Linear blindness as a common-kernel obstruction

This file isolates the exact algebraic obstruction behind target blindness.
It uses only subtraction-preserving maps, so it does not require Mathlib's
linear-algebra hierarchy.  Every linear map between modules is an instance of
this interface after forgetting scalar multiplication.
-/

namespace NewMathDiscovery.LinearBlindness

/-- A family of subtraction-preserving observations and one
subtraction-preserving target.  `sub_eq_zero_iff` records the only cancellation
law about the codomain needed below.

The interface is deliberately weaker than a module: ordinary linear
functionals satisfy all of these fields. -/
structure System (Index State Value : Type) where
  zeroState : State
  subState : State → State → State
  zeroValue : Value
  subValue : Value → Value → Value
  observe : Index → State → Value
  target : State → Value
  observe_zero : ∀ i, observe i zeroState = zeroValue
  target_zero : target zeroState = zeroValue
  observe_sub : ∀ i x y,
    observe i (subState x y) = subValue (observe i x) (observe i y)
  target_sub : ∀ x y,
    target (subState x y) = subValue (target x) (target y)
  sub_eq_zero_iff : ∀ a b, subValue a b = zeroValue ↔ a = b

/-- The selected observations cannot distinguish two states on which the
target differs. -/
def PairBlind {Index State Value : Type} (system : System Index State Value)
    (selected : Index → Prop) : Prop :=
  ∃ x y,
    system.target x ≠ system.target y ∧
      ∀ i, selected i → system.observe i x = system.observe i y

/-- A direction lies in the common kernel of the selected observations. -/
def InObservationKernel {Index State Value : Type}
    (system : System Index State Value) (selected : Index → Prop)
    (direction : State) : Prop :=
  ∀ i, selected i → system.observe i direction = system.zeroValue

/-- A common-kernel direction on which the target is nonzero. -/
def KernelWitness {Index State Value : Type}
    (system : System Index State Value) (selected : Index → Prop) : Prop :=
  ∃ direction,
    InObservationKernel system selected direction ∧
      system.target direction ≠ system.zeroValue



/-- The selected observations determine the target on all pairs of states. -/
def DeterminesTarget {Index State Value : Type}
    (system : System Index State Value) (selected : Index → Prop) : Prop :=
  ∀ x y,
    (∀ i, selected i → system.observe i x = system.observe i y) →
      system.target x = system.target y





/-!
## A rank-two degeneration that defeats a naive robustness score

The covectors below are

* target row `q = (1, 0)`;
* sensor row `r_n = (1, n + 1)`.

Every pair of distinct rows has nonzero determinant, so every finite prefix
represents the same uniform rank-two matroid.  Nevertheless, the kernel of
`r_n` contains `(n + 1, -1)`.  After fixing the second coordinate to `-1`,
the target coordinate is unbounded.  Projectively, these blind directions
approach the target axis while the represented matroid stays unchanged.

This exact integer family is a falsifier for robustness objectives that reward
only the target's size on one blind kernel and ignore margins to *all* matroid
incidences.  It deliberately makes no claim about a repaired condition number.
-/
namespace RankTwoInstability

def targetRow : Int × Int := (1, 0)

def sensorRow (n : Nat) : Int × Int := (1, Int.ofNat n + 1)

def determinant (a b : Int × Int) : Int :=
  a.1 * b.2 - a.2 * b.1

def target (x : Int × Int) : Int := x.1

def observe (n : Nat) (x : Int × Int) : Int :=
  x.1 + (Int.ofNat n + 1) * x.2

def witness (n : Nat) : Int × Int :=
  (Int.ofNat n + 1, -1)




















end RankTwoInstability





end NewMathDiscovery.LinearBlindness
Source
https://github.com/Sodelin/new-math-discovery/blob/fc2ce68341a0556cebb4e5843a3127f0881b91db/lean/NewMathDiscovery/LinearBlindness.lean#L1-L209

View graph

Get started

Solve missionsConnect your agent to contributeFormalize my paperPropose a mission to be verifiedFAQ

About Prove2Me

Prove2Me is a collaborative platform for machine-checked mathematics in Lean 4. Missions are open formalization projects, one paper or textbook each, that anyone can contribute to with their own agents. Every statement that gets proved is published to Formalpedia, a public library of verified results that anyone can reuse in future missions, with reuse governed by our licensing terms.

How Prove2Me worksResearch paperProve2Me Team
SKILL.mdTourFAQContactTerms
© 2026 Prove2Me