An algorithm is presented which uses the surface waveforms originating from one event recorded in two stations which are located close to each other. The waveforms are inverted for local laterally homogeneous S velocity structure or for path-averaged horizontal S velocity gradients. In the first case the stations are situated on the same great circle to the source. In the second case the two stations lie at approximately the same epicentral distance, but with a slightly different azimuth. In both cases, the frequency-dependent phase differences between the two recordings are used as data for an inversion for S velocity values. No a priori knowledge about the source parameters and the medium between the source and the nearest receiver is needed. These factors are estimated in the first step of the inversion (linear). In the second step of the inversion (nonlinear) the S velocity information is obtained. Synthetic tests show that the algorithm is a powerful tool to image structure with only two fundamental mode surface wave recordings as data. In addition, they show that for the case where a horizontal gradient is determined, higher modes can be included in the analysis to display gradients at larger depths, without the need to identify the individual modes separately. The power of the method is illustrated with some real data examples that are all related to a pronounced tectonic boundary in western Eurasia: the Tornquist-Teisseyre zone. The results of the experiments with real data coincide well with previously obtained models for this area. ¿ American Geophysical Union 1995