Direction selectivity of synaptic potentials in simple cells of the cat visual cortex. J. Neurophysiol. 78: 2772-2789, 1997. The direction selectivity of simple cells in the visual cortex is generated at least in part by nonlinear mechanisms. If a neuron were spatially linear, its responses to moving stimuli could be predicted accurately from linear combinations of its responses to stationary stimuli presented at different positions within the receptive field. In extracellular recordings, this has not been found to be the case. Although the extracellular experiments demonstrate the presence of a nonlinearity, the cellular process underlying the nonlinearity, whether an early synaptic mechanism such as a shunting inhibition or simply the spike threshold at the output, is not known. To differentiate between these possibilities, we have recorded intracellularly from simple cells of the intact cat with the whole cell patch technique. A linear model of direction selectivity was used to analyze the synaptic potentials evoked by stationary sine-wave gratings. The model predicted the responses of cells to moving gratings with considerable accuracy. The degree of direction selectivity and the time course of the responses to moving gratings were both well matched by the model. The direction selectivity of the synaptic potentials was considerably smaller than that of the intracellularly recorded action potential, indicating that a nonlinear mechanism such as threshold enhances the direction selectivity of the cell's output over that of its synaptic inputs. At the input stage, however, the cells apparently sum their synaptic inputs in a highly linear fashion. A more constrained test of linearity of synaptic summation based on principal component analysis was applied to the responses of direction-selective cells to stationary gratings. The analysis confirms that the summation in these cells is highly linear. The principal component analysis is consistent with a model in which direction selectivity in cortical simple cells is generated by only two subunits, each with a different receptive-field position and response time course. The response time course for each of the two subunits is derived for four analyzed cells. Each derived subunit is linear in spatial summation, suggesting that the neurons that comprise each subunit are either geniculate X-cells or receive their primary synaptic input from X-cells. The amplitude of the response of each subunit is linearly related to the contrast of the stimulus. The subunits are nonlinear in the time domain, however: the response to a stationary stimulus whose contrast is modulated sinusoidally in time is nonsinusoidal. The principal component analysis does not exclude models of direction selectivity based on more than two subunits, but such higher-order models would have to include the constraint that the extra subunits form a smooth continuum of interpolation between the properties derived from the two subunit solution.