Technical Report 2017-016

Are the Eigenvalues of the B-spline IgA Approximation of -Delta u = lambda u Known in Almost Closed Form?

Sven-Erik Ekström, Isabella Furci, and Stefano Serra-Capizzano

August 2017


In this paper we consider the B-spline IgA approximation of the second-order eigenvalue problem -Delta u = lambda u on Omega =(0,1)d, with zero Dirichlet boundary conditions and with Delta = sumj=1dpartial2/partial xj2, dge 1. We use B-splines of degree p=(p1,...,pd) and maximal smoothness and we consider the natural Galerkin approach. By using elementary tensor arguments, we show that the eigenvalue-eigenvector structure of the discrete problem can be reduced to the case of d=1, pge 1, regularity Cp-1, with coefficient matrix Ln[p] having size N(n,p)=n+p-2. In previous works, it has been established that the normalized sequence {n-2Ln[p]}n has a canonical distribution in the eigenvalue sense and the so-called spectral symbol ep(theta) has been identified.

In this paper we provide numerical evidence of a precise asymptotic expansion for the eigenvalues, which obviously begins with the function ep, up to the largest nprm out=p+rm mod(p,2)-2 eigenvalues which behave as outliers. More precisely, for every integer alphage 0, every n, every pge 3 and every j=1,...,hat N=N(n,p)-nprm out=n-rm mod(p,2), the following asymptotic expansion holds: begin{align*} lambdaj(n-2 Ln[p])=ep(thetaj,n,p)+sumk=1alphack(p)(thetaj,n,p)hk+Ej,n,alpha(p), end{align*} where: • the eigenvalues of n-2Ln[p] are arranged in nondecreasing order and ep is increasing; • {ck(p)}k=1,2,... is a sequence of functions from [0,pi] to mathbb R which depends only on ep; • for any pge 3 and k, there exists bartheta (p,k)>0 such that ck(p) vanishes (at least numerically) on the whole nontrivial interval [0,bartheta (p,k)], so that the formula is exact, up to machine precision, for a large portion of the small eigenvalues; • h=1/n and thetaj,n,p=jpi/n=jpi h, j=1,..., n-rm mod(p,2); • Ej,n,alpha(p)=O(halpha +1) is the remainder (the error), which satisfies the inequality |Ej,n,alpha(p)|le C\alpha halpha +1 for some constant C\alpha depending only on alpha and ep. For the case p=1,2 the complete structure of the eigenvalues and eigenvectors is identified exactly. Indeed, for such values of p, the matrices Ln[p] belong to Toeplitz -minus- Hankel algebras and this is also the reason why there are no outliers, that is nprm out=0.

Moreover, for pge 3 and based on the eigenvalue asymptotics for n-2Ln[p], we devise an extrapolation algorithm for computing the eigenvalues of the discrete problem with a high level of accuracy and with a relatively negligible computational cost. However, the algorithm is not necessary for all the spectrum and indeed, for pge 3 and thetaj,n,p belonging to the interval [0,bartheta (p)], bartheta (p)=infkbartheta (p,k), the value ep(thetaj,n,p) coincides with lambdaj(n-2 Ln[p]), up to machine precision.

Such expansions are of the same type studied in the literature for the eigenvalues of a sequence of Toeplitz matrices {Tn(f)} and of a sequence of preconditioned Toeplitz matrices {Tn-1(g)Tn(f)}, for f trigonometric polynomial, g nonnegative, not identically zero trigonometric polynomial.

Extensive numerical experiments are discussed and further future research steps are illustrated at the end of the paper.

Available as PDF (1.7 MB, no cover)

Download BibTeX entry.