Algorithm whose running time depends on P vs. NP

Question

Is there a known, explicit example of an algorithm with the property such that if $P\neq NP$ then this algorithm doesn't run in polynomial time and if $P=NP$ then it does run in polynomial time?

Answer 1

If you assume that $P=^?NP$ is provable in PA (or ZFC), a trivial example is the following:

Input: N   (integer in binary format)
For I = 1 to N do
begin
  if I is a valid encoding of a proof of P = NP in PA (or ZFC)
    then halt and accept
End
Reject

Another - less trivial - example that relies on no assumption is the following:

Input: x   (boolean formula)
Find the minimum i such that
  1) |M_i| < log(log(|x|))  [ M_1,M_2,... is a standard fixed TM enumeration] 
  2) and  M_i solves SAT correctly 
       on all formulas |y| < log(log(|x|))
          halting in no more than |y|^|M_i| steps
          [ checkable in polynomial time w.r.t. |x| ]
  if such i exists simulate M_i on input x 
      until it stops and accept/reject according to its output
      or until it reaches 2^|x| steps and in this case reject;
  if such i doesn't exist loop for 2^|x| steps and reject.

If $P =NP$ the algorithm will soon or later - suppose on input $x_0$ - find the index of the polynomial time Turing machine (or a padded version of it) $M_{SAT}$ that solves SAT in $O( |x| ^ { |M_{SAT}| })$ and for all inputs greater than $x_0$ will continue to simulate it and halt in polynomial time (note that step 2 can also be checked in polynomial time). In other words if $P = NP$ the algorithm solves SAT in polynomial time on all but a finite number of instances.

If $P \neq NP$ the algorithm runs in exponential time.