Technologies are generally described for multi-resolution based power spectral density estimation. According to some examples, a first power spectral density (PSD) of a signal may be computed to determine a first frequency resolution and a first error variance (EV) value associated with the PSD. The first PSD may be computed using fewer segments of the signal to achieve a higher resolution. The higher resolution is associated with a higher EV because of the fewer segments used for the computation of the first PSD. Next, a second PSD may be computed using the signal while sampling more segments to determine second resolution and EV. The second PSD may be associated with a lower EV because of the additional segments used for the computing of the second PSD. Missing frequency variables of the second PSD may be determined using the first resolution and the second EV through an expectation maximization technique.