9.4 The Rayleigh quotient
Theorem 9.4.1 (Rayleigh). For the Dirichlet problem (\(y(a) = y(b) = 0\)), define for continuously differentiable \(u \neq 0\) with \(u(a) = u(b) = 0\) the Rayleigh quotient \[ R[u] = \frac {\int _a^b\bigl (pu'^2 + qu^2\bigr )dt}{\int _a^bwu^2\,dt} . \] Then \(\lambda _1 = \min _uR[u]\), attained exactly at \(u = \phi _1\). More generally \(\lambda _n\) is the minimum of \(R[u]\) over \(u\) orthogonal (in \(\langle \cdot ,\cdot \rangle _w\)) to \(\phi _1, \dots , \phi _{n-1}\).
Proof. Integrating by parts, with \(u(a) = u(b) = 0\), \(\int pu'^2 = -\int u(pu')' \), so the numerator is \(\langle u, Lu\rangle \). For \(u = \phi _n\) this is \(\lambda _n\langle \phi _n, \phi _n\rangle _w\), so \(R[\phi _n] = \lambda _n\). For general \(u\), normalise the eigenfunctions so that \(\langle \phi _m, \phi _n\rangle _w = \delta _{mn}\) and expand \(u = \sum c_n\phi _n\) by Theorem 9.3.3(iii); then (with a limiting argument that we omit) the numerator is \(\sum \lambda _nc_n^2\) and the denominator \(\sum c_n^2\), so \[ R[u] = \frac {\sum _n\lambda _nc_n^2}{\sum _nc_n^2} \geq \lambda _1 , \] with equality only if \(c_n = 0\) whenever \(\lambda _n > \lambda _1\), that is \(u\) is a multiple of \(\phi _1\). If \(u\) is orthogonal to \(\phi _1, \dots , \phi _{n-1}\), the first \(n - 1\) coefficients vanish and the same argument gives \(R[u] \geq \lambda _n\). □
The Rayleigh quotient gives upper bounds for \(\lambda _1\) from any trial function, and a good trial function gives a good bound, because the error in \(R\) is quadratic in the error in \(u\). For \(-y'' = \lambda y\) on \([0, 1]\) the trial function \(t(1 - t)\) gives \(R = \frac {1/3}{1/30} = 10\), against the exact \(\pi ^2 = 9.870\), an error of \(1.3\%\). This is the basis of the Rayleigh–Ritz and finite element methods, which minimise \(R\) over a finite-dimensional space of trial functions.
Questions on this section
Stuck on something here? Ask below and it stays attached to this topic.