The deconvolution of three-component teleseismic P waves is investigated using the autocorrelation of the P to SV scattered waves, which is important for improved imaging of crustal and upper mantle structure. The SV component of the P waveform is first estimated by transforming the three-component seismic data to the P–SV–SH frame of Kennett (1991) and also taking into account the free surface. This removes the direct P wave from the SV component leaving only the scattered P to SV waves. Assuming that the P to SV scattering coefficients are random and white, then the autocorrelation of the SV component provides an estimate of the autocorrelation of the source wavelet. This is analogous to the use of the autocorrelation of a reflection seismogram in exploration seismology to estimate the source pulse where the P-wave reflectivity is assumed to be random and white. A minimum- phase source wavelet estimated from the autocorrelation of the SV component can be used to deconvolve the unrotated radial and Z components that have been processed to be minimum phase. A minimum-phase source wavelet is not required, but the direct P wave must be larger than the scattered waves on the unrotated components. To enhance the direct wave, we use rotated coordinates about the direct P- arrival direction. This procedure is first tested on synthetic data and then applied to observed data from the 1993 Cascadia experiment where both P to P and P to SV scattered waves are estimated from the data that have been deconvolved using the autocorrelation of the SV component.