Subtraction-preserving observation systems and an integer instability family
DefinitionNewMathDiscovery_LinearBlindnessalgebrainformationkernelsodelin-newmath
Let be arbitrary types with distinguished zeros in , subtraction operations, observation maps and target . Assume the maps preserve zero and subtraction, and subtraction in is zero exactly when its arguments are equal. For a selected set , let .
Target determination means equal selected observations force equal targets. A separate concrete interface on sets , , , 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