Generalized spectrum of second order differential operators

Contribution

• The paper extends prior scalar-coefficient analyses to second-order differential operators involving symmetric tensors.The problem is formulated as the generalized eigenvalue equation ∇·(K∇u) = λ∆u on a bounded Lipschitz domain.Source: §1 · pp. 1–2

Main result

• For continuous symmetric K on a bounded Lipschitz domain, the spectrum of L−1A equals the interval determined by the ranges of the tensor’s diagonalized entries.The result applies to definite, indefinite, and semidefinite tensors and strengthens the scalar-case inclusion to equality.Source: §1 · p. 2

Scope and caveat

• In infinite-dimensional Sobolev spaces, generalized eigenvalues generally represent only part of the spectrum, so the eigenvalue problem does not determine the whole spectrum.The paper therefore distinguishes the full operator spectrum from discrete or generalized eigenvalue observations.Source: §1 · p. 2

Numerical experiments

• Discretized experiments show spectra can fill the full interval [1, 10], even when the diagonal-entry ranges do not overlap.The experiments use diagonal tensors on the unit square and piecewise linear finite elements; the discrete problem has size 381.Source: §1 · pp. 2–3, Figure 1 · p. 3

Operator framework

• Symmetry makes the preconditioned operator self-adjoint in the Laplacian-associated inner product, so its spectrum is real and this norm supports the proofs.This structure underlies the spectral characterization.Source: §1 · p. 3

Motivation

• The analysis is motivated by understanding Conjugate Gradient behavior when Laplacian-based preconditioners are applied to discretized second-order elliptic PDEs.The spectral distribution is linked to convergence behavior, with constant-coefficient preconditioners also relevant to isogeometric analysis.Source: §1 · pp. 3–4

2.1. Function values at points of continuity belong to the spectrum.

• At every continuity point of a diagonal entry κ_i, its value belongs to the spectrum of the preconditioned operator.The proof constructs localized functions whose residuals vanish while their norms do not, ruling out a bounded inverse at that value.Source: §2.1 · pp. 4–5

2.2.1. Dirichlet problem for the wave equation.

• For the localized wave equation, any positive constant can be used to obtain infinitely many nontrivial solutions on a sufficiently small solution domain.The domain-size parameter can make the relevant eigenvalue set arbitrarily small.Source: §2.2.1 · pp. 5–6

2.2. Disjoint ranges extend the spectrum.

• When the ranges of κ_1 and κ_2 are disjoint, the spectrum also contains the closed interval between their extremal ranges.This extends beyond the individual coefficient ranges and is established through localized constructions associated with a wave equation.Source: §2.2 · p. 5, §2.2.2 · pp. 6–7

2.2.2. Tensors constant on an open subdomain.

• If the tensor is constant on an open subdomain, the interval between the extremal values of κ_1 and κ_2 lies in the spectrum without requiring global continuity.The constant subdomain may be arbitrarily small, and endpoint values are covered by the pointwise-continuity result.Source: §2.2.2 · pp. 6–7

2.2.3. Tensors continuous at least at a single point.

• Continuity at a single interior point suffices to preserve the interval inclusion for the diagonal tensor.The argument locally replaces the tensor near that point and uses a bounded-inverse contradiction with a Neumann-series estimate.Source: §2.2.3 · pp. 7–8

3.2. Main result – diagonal tensors.

• The reverse inclusion completes the result: for continuous diagonal tensors, the spectrum equals the interval determined by the coefficient extrema.The proof combines the containment result with the auxiliary interval-inclusion arguments.Source: §3 · p. 8, §3.1 · p. 8, §3.2 · p. 9

3.1. The spectrum is a subset of the extremal interval.

• The spectrum is contained in the extremal interval determined by the minimum and maximum values of κ_1 and κ_2.Self-adjointness makes the spectrum real, and the interval endpoints are themselves spectral values; no definiteness assumption on K is required.Source: §3.1 · p. 8

• At continuity points of K, its eigenvalue functions κ1 and κ2 are continuous, enabling local spectral constructions.The argument uses the spectral decomposition of K at a point and continuity of the tensor.Source: §4 · pp. 9–10
• The local test-function contradiction argument forces the relevant eigenvalue values into the spectrum of the preconditioned operator.The constructed functions yield a nonvanishing limiting norm, contradicting exclusion from the spectrum.Source: §4 · pp. 10–11
• The proof shows that intervals determined by the ranges of the eigenvalue functions belong to the spectrum, with the analogous statement obtained by interchanging κ1 and κ2.These interval-inclusion lemmas cover tensors that may be constant on subdomains or continuous at selected points.Source: §4 · p. 11
• Local changes of variables transform a constant symmetric tensor into a diagonal tensor, reducing the nondiagonal analysis to the established diagonal case.Orthogonality of the decomposition matrix preserves the relevant norm structure in this reduction.Source: §4 · pp. 11–12
• Theorem 1.1 is completed by combining Lemmas 4.1 to 4.4, with a proof structure fully analogous to that of Theorem 3.2 for diagonal tensors.The remaining part is a straightforward extension of the analysis in Section 3.Source: §4 · pp. 9, 11, 12

• Theorem 1.1 also holds for generalized eigenvalue problems with homogeneous Neumann boundary conditions.The Neumann formulation uses the boundary condition involving the outward normal and analogous definitions of the operators.Source: §5 · p. 12
• For the Neumann problem, the auxiliary functions used in the proof must be modified, after which the remaining argument proceeds analogously.The inverse-boundedness condition for L is noted in the Neumann setting.Source: §5 · p. 12

Experimental setup

• The experiments use finite-element discretization on uniform triangular meshes, with matrices assembled in FEniCS and eigenvalues computed in Matlab.Unless otherwise specified, the domain is the unit square and piecewise linear basis functions are used.Source: §6 · p. 13

Nondiagonal indefinite tensors

• The spectra for nondiagonal indefinite problems spread across the predicted intervals, and mesh refinement improves endpoint approximations without changing the qualitative behavior.For (P4)–(P6), the spectra cover the intervals determined by the eigenvalue ranges; the interval [−5, 5] is better approximated on finer meshes.Source: §6 · p. 13

Boundary conditions

• Neumann and Dirichlet discretizations show comparable spectral behavior, with Neumann conditions better approximating endpoints attained on the boundary.For (P7) and (P8), the relevant interval is [4, 10], while the coefficient ranges do not overlap.Source: §6 · pp. 13–14, Figure 3 · p. 14

Nonconvex domains

• For nonconvex domains, the discretized spectra differ according to the ranges of κ1 and κ2 actually attained within each domain.In (P9), extrema of the coefficient functions lie outside one of the L-shaped domains, changing the observed spectral ranges.Source: §6 · p. 14, Figure 4 · p. 15

Three-dimensional experiments

• Three-dimensional experiments show spectra spread over [1, 10] regardless of whether the associated coefficient ranges overlap, but rigorous theory is unavailable there.The 3D tests use diagonal tensors on the unit cube with zero Dirichlet conditions.Source: §6 · pp. 14–15

Open problems

• Discretized versions of the tensor results remain outside this paper's scope, although numerical experiments suggest eigenvalues may be evenly distributed across the predicted interval.The authors defer a full analysis of discretized operators and eigenvalue distributions to future work.Source: §7 · p. 15

Appendix A

• Appendix A constructs the localized function vr on Rr, including its support, boundary behavior, and nondifferentiability sets needed for the proof inequalities.The function is centered at (0, 0), vanishes outside Rr, and reaches its maximum along an explicitly described set.Source: §7 · pp. 15–16

Numerical illustrations

• The 3D experiments show spectra spanning [1, 10], with selected diagonal-entry ranges κ1(Ω) = [4, 6] and κ3(Ω) = [2, 10] for (P12).Source: Figure 5 · p. 16

Appendix B

• Appendix B establishes the perturbation bound by controlling the product of the resolvent and coefficient difference, using continuity of K to choose sufficiently small l.Source: §7 · pp. 17–18