S U M M A R YObtaining new seismic responses from existing recordings is generally referred to as seismic interferometry (SI). Conventionally, the SI responses are retrieved by simple crosscorrelation of recordings made by separate receivers: one of the receivers acts as a 'virtual source' whose response is retrieved at the other receivers. When SI is applied to recordings of ambient seismic noise, mostly surface waves are retrieved. The newly retrieved surface wave responses can be used to extract receiver-receiver phase velocities. These phase velocities often serve as input parameters for tomographic inverse problems. Another application of SI exploits the temporal stability of the multiply scattered arrivals of the newly retrieved surface wave responses. Temporal variations in the stability and/or arrival time of these multiply scattered arrivals can often be linked to temporally varying parameters such as hydrocarbon production and precipitation. For all applications, however, the accuracy of the retrieved responses is paramount. Correct response retrieval relies on a uniform illumination of the receivers: irregularities in the illumination pattern degrade the accuracy of the newly retrieved responses. In practice, the illumination pattern is often far from uniform. In that case, simple crosscorrelation of separate receiver recordings only yields an estimate of the actual, correct virtual-source response. Reformulating the theory underlying SI by crosscorrelation as a multidimensional deconvolution (MDD) process, allows this estimate to be improved. SI by MDD corrects for the non-uniform illumination pattern by means of a so-called point-spread function (PSF), which captures the irregularities in the illumination pattern. Deconvolution by this PSF removes the imprint of the irregularities on the responses obtained through simple crosscorrelation. We apply SI by MDD to surface wave data recorded by the Malargüe seismic array in western Argentina. The aperture of the array is approximately 60 km and it is located on a plateau just east of the Andean mountain range. The array has a T-shape: the receivers along one of the two lines act as virtual sources whose responses are recorded by the receivers along the other (perpendicular) line. We select time windows dominated by surface wave noise travelling in a favourable direction, that is, traversing the line of virtual sources before arriving at the receivers at which we aim to retrieve the virtual-source responses. These time windows are selected through a frequency-dependent slowness analysis along the two receiver lines. From the selected time windows, estimates of virtual-source responses are retrieved by means of crosscorrelations. Similarly, crosscorrelations between the positions of the virtual sources are computed to build C The