Here we study the variability in extracellular records of action potentials. Our work is motivated, in part, by the need to construct effective algorithms to classify single-unit waveforms from multiunit recordings. 2. We used microwire electrode pairs (stereotrodes) to record from primary somatosensory cortex of awake, behaving rat. Our data consist of continuous records of extracellular activity and segmented records of extracellular spikes. Spectral and principal component techniques are used to analyze mean single-unit wave-forms, the variability between different instances of a single-unit waveform, and the underlying background activity. 3. The spectrum of the variability between different instances of a single-unit waveforms is not white, and falls off above 1 kHz with a frequency dependence of roughly f-2. This spectrum is different from that of the mean spike waveforms, which falls off roughly as f-4, but is essentially identical with the spectrum of background activity. The spatial coherence of the variability on the 10-micron scale also falls off at high frequencies. 4. The variability between different instances of a single-unit waveform is dominated by a relatively small number of principal components. As a consequence, there is a large anisotropy in the cluster of the spike waveforms. 5. The background noise cannot be represented as a stationary Gaussian random process. In particular, we observed that the spectrum changes significantly between successive 20-ms intervals. Furthermore, the total power in the background activity exhibits larger fluctuations than is consistent with a stationary Gaussian random process. 6. Roughly half of the single-unit spike waveforms exhibit systematic changes as a function of the interspike interval. Although this results in a non-Gaussian distribution in the space of waveforms, the distribution can be modeled by a scalar function of the interspike interval. 7. We use a set of 44 mean single-unit waveforms to define the space of differences between spike waveforms. This characterization, together with that of the background activity, is used to construct a filter that optimizes the detection of differences between single-unit waveforms. Further, an information theoretic measure is defined that characterizes the detectability.
