Rigorous methods for computing magnetotelluric response function estimates and their variances from electromagnetic field measurements using the singular value decomposition (SVD) are presented. The relative noise level between the electric and magnetic field components is treated as a free variable, and is shown profoundly to affect both the response function estimates and their variances. In general, the variances decrease to a minimum as the field components share noise more equally. A statistical parameter for checking the validity of the usual magnetotelluric assumptions (two-dimensional conductivity structure and homogeneous source fields) as a function of frequency is developed and applied to the estimation of the ratio of electric and magnetic field noise in data.

The method is illustrated using both synthetic and real magnetotelluric data. Results from a seafloor experiment in the NW Pacific show that the response functions obtained primarily from the east electric and north magnetic fields are better determined than the conjugate function, reflecting higher field coherences, but yields no significant statistical difference between the principal components of the response tensor. The ratio of magnetic to electric field noise appears to increase at frequencies below 0.5 cph, and evidence for source field complications, as suggested by high vertical to horizontal magnetic field coherence, is widespread and may reflect ocean-generated electromagnetic fields.