This has nothing to do with floating point error accumulation or numerical stability in the floating point sense. You can do this with arbitrary sized floating point values and you will still get the same non-convergence result.