We develop a new method for estimating and removing the spectrum of the sky
from deep spectroscopic observations. Our method does not rely on simultan
eous measurement of the sky spectrum with the object spectrum. The techniqu
e is based on the iterative subtraction of continuum estimates and eigenvec
tor sky models derived from singular value decompositions of sky spectra an
d sky spectra residuals. Using simulated data derived from small-telescope
observations, we demonstrate that the method is effective for faint objects
on large telescopes. We discuss simple methods to combine our new techniqu
e with the simultaneous measurement of sky to obtain sky subtraction very n
ear the Poisson limit.