It will help if I know why you think your input vector is 4096 points(this can't be arbitrary, I mean you are in frequency domain and frequency content is dependant on the grid). So make sure what you are after before getting sucked into further complications.
For example a vector indexed 1:100 in the frequency domain implies:
frequencies [0: .5Fs] are represented by values indexed 1:50(positive frequenies)
frequencies from [.5Fs:Fs] are represented by values indexed 51:100(also means negative frequencies).
For real signals, the values indexed 1:50 must be a reflection copy of values indexed 51:100(symmetrical)
In other words your original 4096 vector must make sense. And your 8192 can't be constructed by zero padding(your original vector must show reflection around midpoint of 4096/2, then you must reflect around new midpoint for new resolution at 8192/2).