In their study of upper mantle structure beneath the Paran¿ Basin of SE Brazil, Snoke and James <1997> concluded, on the basis of a linearized least squares inversion (LLSI) of surface wave dispersion data, that a strong (5% contrast) low-velocity zone (LVZ) beginning at a depth less than ~150 km was not required to fit the data. They were unable to establish a quantitative estimate, however, on the maximum depth at which such a LVZ could be resolved by their data. Sambridge <1999a, 1999b> has introduced the Neighbourhood Algorithm (NA), a direct search method for nonlinear inversion which can be tuned to extract information from an ensemble of models in addition to finding a single best fit model. Applying NA to the Brazilian dispersion data quantifies the statistics of the ensemble of models classified as acceptable based on a data misfit criterion and a smoothness constraint. The NA best fit model is not significantly different from the LLSI best fit model, but the analysis of the ensemble of models provides new insights regarding how well constrained the model is. Synthetics runs show that for this data set, our modeling procedures could resolve a strong LVZ that began at a depth of 120 km but could not rule out such an LVZ beginning at a depth of 180 km.