“Vector infidelity” is the failure of multicomponent ocean-bottom-seismic geophones to respond isotropically to incoming seismic energy. We model vector infidelity as an unknown linear convolutional filter that the geophone and its environment apply to the “true data” to produce the “corrupted” data that we actually record. Our goal is to determine the inverse of this filter, so that we can correct the recorded data. A perfect vector-infidelity correction filter, when applied to the data, should cause the horizontal components of the direct arrival to become linearly polarized in the radial direction for all source azimuths. We show how to calculate the linear filter that optimally accomplishes this, and the results of applying the method to a real OBS dataset.