Spike sorting refers to the detection and classification of electric potentials (spikes) from multi-neuron recordings. It is a difficult but essential pre-processing step before neural data can be analyzed for information content. While several spike sorting algorithms have been proposed, our goal is to determine the ultimate limits of spike classification, and to characterize this error regardless of sorting algorithm. We have identified and incorporated three important factors that affect the sorting procedure - SNR, spike amplitude ratio and inter-spike correlation - into a signal constellation model to derive error probability bounds on any sorting procedure. We consider the cases of known and unknown time-of-occurrence of the spike(s) in question. We calculate spike timing error estimates in the case of unknown delay. Additionally, we derive a theoretical amplitude distribution for spike amplitudes at the electrode. Finally we introduce the idea of a non-gaussian "corruption" noise component that affects spike waveform.