Liste des Groupes | Revenir à theory |
Op 30.mei.2024 om 16:43 schreef olcott:The actual behavior of DD correctly simulated by ANY pure function HOn 5/28/2024 11:16 AM, olcott wrote:Why are you referring to the 'pathological behavior of x' if your claim is that the simulator does not even reach the part of DD (below) that contradicts the result of HH? This 'pathological behavior of x' is completely irrelevant.>>
When Ĥ is applied to ⟨Ĥ⟩
Ĥ.q0 ⟨Ĥ⟩ ⊢* embedded_H ⟨Ĥ⟩ ⟨Ĥ⟩ ⊢* Ĥ.qy ∞
Ĥ.q0 ⟨Ĥ⟩ ⊢* embedded_H ⟨Ĥ⟩ ⟨Ĥ⟩ ⊢* Ĥ.qn
>
*Formalizing the Linz Proof structure*
∃H ∈ Turing_Machines
∀x ∈ Turing_Machines_Descriptions
∀y ∈ Finite_Strings
such that H(x,y) = Halts(x,y)
>
A decider computes the mapping from finite string inputs to
its own accept or reject state.
>
A decider does not and cannot compute the mapping from
Turing_Machine inputs to its own accept or reject state.
>
Halts(x,y) would report on the direct execution of x(y) thus ignores
the pathological behavior of x correctly simulated by pure function H.
This makes Halts(x,y) an incorrect measure of the correctness of H(x,y).
The problem is that a simulating decider is unable to handle the simulation of itself because it gets stuck in recursive simulation). That DD contradicts HH's result is completely irrelevant.--
>
This is easier to see when we can see every single detail of all of
the steps as an x86 execution trace of D correctly simulated by pure
function H.
>
typedef int (*ptr)(); // ptr is pointer to int function in C
00 int HH(ptr p, ptr i);
01 int DD(ptr p)
02 {
03 int Halt_Status = HH(p, p);
04 if (Halt_Status)
05 HERE: goto HERE;
06 return Halt_Status;
07 }
08
09 int main()
10 {
11 HH(DD,DD);
12 return 0;
13 }
Les messages affichés proviennent d'usenet.