Subspace Dimension.

In many cases it is useful to choose $p$ the dimension of the subspace larger than $\nev$ (the number of eigenvalues sought). In that case the slowest eigenvalue converges at a rate $\vert\lambda_{p+1}/\lambda_{\nev}\vert$ which may be significantly smaller than $\vert\lambda_{\nev+1}/\lambda_{\nev}\vert$. Practitioners in structural engineering have used choices like $p = \mbox{min} (2\nev,\nev+8)$; see [42].