Eigenvalue approximation and discrete compactness in FEEC

Intuition Beginner

When you strike a drum or hum into a metal box, only certain tones come out. Those special tones are the eigenvalues of the system, and the shapes that vibrate at each tone are the eigenmodes. Maxwell's equations have the same structure: an empty metal cavity rings at a discrete list of electromagnetic frequencies, and engineers need to compute that list accurately to design antennas, lasers, and photonic crystals.

The surprise is that a finite element method can compute the wrong list. It can be a perfectly good solver for the forced problem, where you push the system at a chosen frequency and ask for the response, and still produce a spectrum riddled with fake numbers.

Visual Beginner

The left column shows the real cavity frequencies. A bad method (middle) scatters extra fake values among them; these are spurious modes. A compatible FEEC method (right) reproduces the real list with nothing extra and nothing missing.

Worked example Beginner

Picture a simple box cavity whose first three true squared-frequencies are $2$, $5$, and $5$ (the value $5$ appears twice because two different field shapes ring at the same tone).

A naive method using ordinary vertex-based vector elements might report $2$, then a fake value near $3.7$, then $5$. The $3.7$ corresponds to no real field. An engineer who trusted it would design for a resonance that does not exist.

A compatible edge-element method instead reports $2$, $5$, $5$. It even reproduces the doubled value correctly, because it respects the geometry that produced the pair in the first place.

What this tells us: getting the source problem right is not enough. Computing a spectrum correctly is a stricter test, and it needs a method that preserves the structure of the equations.

Check your understanding Beginner

Formal definition Intermediate+

Let the de Rham complex of $L^2$ differential forms on a bounded domain $\Omega$ furnish a Hilbert complex $(L^2{\Lambda }^{\bullet },d)$ with domains $H{\Lambda }^k$, and let $L=d{d}^{\ast }+d^{\ast }d$ denote the Hodge Laplacian on $k$-forms with its natural boundary conditions. The eigenvalue problem seeks pairs $(\lambda ,u)$ with $u\ne 0$ and $$ L u = \lambda u . $$ Following 24.04.01, one writes this in mixed/compatible form. Introducing the auxiliary variable $\sigma =d^{\ast }u$, the eigenproblem becomes: find $\lambda \in \mathbb{R}$, $\sigma \in H{\Lambda }^{k-1}$, $u\in H{\Lambda }^k$, not both zero, with $$ \langle \sigma, \tau \rangle - \langle u, d\tau \rangle = 0 \quad \forall, \tau \in H\Lambda^{k-1}, $$ $$ \langle d\sigma, v \rangle + \langle du, dv \rangle = \lambda \langle u, v \rangle \quad \forall, v \in H\Lambda^k . $$ The harmonic forms ${\mathfrak{H}}^k=\ker d\cap \ker d^{\ast }$ are precisely the eigenspace for $\lambda =0$; by the Hodge isomorphism $\dim {\mathfrak{H}}^k=b_k$, the $k$-th Betti number.

A FEEC discretization replaces $H{\Lambda }^{\bullet }$ by a finite element subcomplex $V_h^{\bullet }\subset H{\Lambda }^{\bullet }$ with $d{V}_h^k\subset V_h^{k+1}$, yielding discrete eigenpairs $({\lambda }_{h,i},u_{h,i})$. Let $V_h=V_h^k$ and let $Z_h=\{v\in V_h:⟨v,dw⟩=0\forall w\in V_h^{k-1}\}$ be the discrete coexact (divergence-free) fields.

Definition (discrete compactness, Kikuchi). The family $\{V_h^{\bullet }{\}}_{h>0}$ has the discrete compactness property at degree $k$ if every sequence $u_h\in Z_h$ that is bounded in the graph norm $(\Vert u_h{\Vert }^2+\Vert d{u}_h{\Vert }^2{)}^{1/2}$ admits a subsequence converging in $L^2{\Lambda }^k$ to a limit in $H{\Lambda }^k$.

This is the discrete echo of the Rellich-Kondrachov compactness 02.11.05 underlying the continuous problem and of the compactness axiom of the abstract Hilbert complex 24.03.07: the continuous solution operator is compact because divergence-free fields with bounded $d$ are compact in $L^2$, and a method is spectrally trustworthy only if its discrete divergence-free fields inherit that compactness uniformly in $h$.

Counterexamples to common slips

• Spectral correctness is not implied by source-problem convergence: the Galerkin solution operator $T_h$ can converge pointwise to $T$ without converging in operator norm, and only norm (or collective-compact) convergence controls the spectrum.
• Discrete compactness is not automatic from $V_h\subset H{\Lambda }^k$ being conforming; vector nodal Lagrange spaces are conforming in $L^2$ yet fail it.
• Raising the polynomial degree of an incompatible family does not remove spurious modes; the cure is structural (a subcomplex with a bounded cochain projection), not a matter of accuracy order.
• The zero eigenvalue is not a numerical artifact to be discarded: its multiplicity is the Betti number $b_k$, a topological invariant the method must reproduce.

Key theorem with proof Intermediate+

The forced problem defines a solution operator $T:L^2{\Lambda }^k\to L^2{\Lambda }^k$, $Tf=u$, where $u$ solves $Lu=f$ on the $L^2$-orthogonal complement of ${\mathfrak{H}}^k$ (and $T$ kills harmonic fields). Then $\lambda \ne 0$ is a Hodge-Laplacian eigenvalue iff $\mu =1/\lambda$ is a nonzero eigenvalue of $T$, with the same eigenspaces. The FEEC method produces $T_h:L^2\to V_h$, the Galerkin solution operator. Spectral correctness is the statement that the spectrum of $T_h$ converges to that of $T$ with no extra or missing points.

Theorem (spectral correctness from a bounded cochain projection). Suppose the FEEC subcomplex $V_h^{\bullet }$ admits cochain projections ${\pi }_h^{\bullet }:H{\Lambda }^{\bullet }\to V_h^{\bullet }$ that commute with $d$ and are bounded uniformly in $h$, $\Vert {\pi }_h^{k}v\Vert \le C\Vert v\Vert$ with $C$ independent of $h$ 24.03.05. Then:

1. the family has the discrete compactness property at every degree $k$;
2. $T_h\to T$ in $L(L^2,L^2)$ operator norm;
3. consequently every nonzero eigenvalue $\lambda$ of $L$ is approximated by exactly $m(\lambda )$ discrete eigenvalues counted with multiplicity (where $m(\lambda )$ is the multiplicity of $\lambda$), with no spurious values accumulating in any compact set of ${\mathbb{R}}_{>0}$ avoiding the spectrum, and the harmonic eigenvalue $\lambda =0$ is reproduced with exact multiplicity $b_k$.

Proof. Step 1 (discrete compactness). Let $u_h\in Z_h$ with $\Vert u_h{\Vert }^2+\Vert d{u}_h{\Vert }^2\le 1$. The continuous discrete Poincaré-Friedrichs inequality from 24.03.06, itself a consequence of the bounded cochain projection, gives $\Vert u_h\Vert \le c\Vert d{u}_h\Vert$ on the discrete coexact space, so $\{u_h\}$ is bounded in $H{\Lambda }^k$-graph norm. Extract a weak limit $u_h\rightharpoonup u$ in the graph norm. The continuous embedding of bounded-$d$ coexact fields into $L^2$ is compact by Rellich-Kondrachov 02.11.05; the bounded commuting projection lets us compare $u_h$ with ${\pi }_{h}u$ and transfer this compactness to the discrete space, giving a strongly $L^2$-convergent subsequence. This is exactly the discrete compactness property.

Step 2 (norm convergence of $T_h$). Write $T-T_h=(T-{\pi }_{h}T)+({\pi }_{h}T-T_h)$. The first term is bounded by the best-approximation error of $T$ in $V_h$, which tends to zero; the second is controlled by the discrete inf-sup/Poincaré stability of the mixed system 24.04.01. Discrete compactness upgrades pointwise convergence to uniform convergence on the unit ball: if it failed, a sequence $f_h$ with $\Vert f_h\Vert =1$ and $\Vert (T-T_h)f_h\Vert \ge \delta$ would, after passing to the $L^2$-precompact images, yield a contradiction with the established pointwise convergence. Hence $\Vert T-T_h{\Vert }_{L(L^2,L^2)}\to 0$.

Step 3 (spectral conclusion). $T$ is compact and self-adjoint, so its nonzero spectrum is discrete with finite multiplicities. The Babuška-Osborn spectral-approximation theorem states that for $T_h\to T$ in operator norm, each eigenvalue $\mu =1/\lambda$ of $T$ is approximated by exactly $m(\lambda )$ eigenvalues of $T_h$ inside any small circle isolating $\mu$, and no eigenvalue of $T_h$ converges to a point outside $\mathrm{spec}(T)$. Translating back through $\mu =1/\lambda$ gives the count for $L$; the harmonic part is reproduced exactly because $\ker T_h={\mathfrak{H}}_h^k$ has dimension $b_k$ by the discrete Hodge decomposition 24.04.01. $\square$

Bridge. This theorem builds toward the recognition that a single structural hypothesis — a bounded commuting cochain projection 24.03.05 — controls source accuracy, spectral correctness, and topology simultaneously, and the mechanism appears again in the Maxwell cavity computation of 24.04.02, which is exactly the $k=1$ instance on a three-manifold. The foundational reason the proof works is that the discrete solution operator inherits compactness from Rellich 02.11.05; this is exactly the discrete analogue of the abstract compactness axiom of 24.03.07, and the bridge is that operator-norm convergence of compact operators forces convergence of their entire discrete spectra. Putting these together, the spurious-mode pathology generalises into a clean spectral-approximation statement that the central insight of FEEC — discretise the whole complex, not one space — resolves once and for all.

Exercises Intermediate+

Advanced results Master

The sharp form of the FEEC eigenvalue theorem is quantitative. Under the bounded-cochain-projection hypothesis, for an isolated eigenvalue $\lambda$ of multiplicity $m$ with eigenspace $E_{\lambda }\subset H{\Lambda }^k$, the discrete eigenvalues ${\lambda }_{h,1},\dots ,{\lambda }_{h,m}$ converging to it satisfy a double-order estimate of Babuška-Osborn type, $$ |\lambda - \lambda_{h,i}| \le C,\big(,\sup_{u \in E_\lambda,,|u|=1} \operatorname{dist}(u, V_h^k),\big)^2 + \text{(higher order)}, $$ so eigenvalues converge at twice the rate of the eigenfunction best-approximation error, while eigenfunctions converge at the single rate, with the usual loss in a gap-dependent constant near clusters. The estimate is robust under clustering: the constant degrades with the separation of $\lambda$ from the rest of the spectrum, but the count never does — exactly $m$ discrete values track a multiplicity-$m$ eigenvalue however tightly it is approached by neighbours. This cluster-robustness is the practical payoff of norm convergence over pointwise convergence.

The phenomenon admits a clean operator-theoretic reading through the Anselone notion of collectively compact convergence, which is equivalent here to discrete compactness plus pointwise convergence and is what Boffi's survey identifies as the unifying hypothesis. Two classical sufficient conditions appear in the literature: the discrete compactness property (Kikuchi, for edge elements) and the completeness of discrete divergence-free fields plus a discrete Friedrichs inequality (Boffi-Brezzi-Gastaldi). FEEC subsumes both: a bounded commuting cochain projection implies the discrete Poincaré-Friedrichs inequality of 24.03.06 and discrete compactness in one stroke, so the abstract Hilbert-complex hypotheses of 24.03.07 are precisely the right level of generality at which spectral correctness is a theorem rather than a per-family verification.

The Maxwell cavity is the headline instance at $k=1$ in ${\mathbb{R}}^3$. There the curl-curl operator $\mathrm{curl}\mathrm{curl}$ on $H_0(\mathrm{curl})$ has an infinite-dimensional kernel (the gradients) carrying $\lambda =0$, and the nonzero spectrum is the resonant frequencies. Nédélec edge elements 24.04.02 form the degree-one slot of a discrete de Rham complex, possess discrete compactness, and compute the resonances with no spurious modes; vector nodal Lagrange elements do not form such a complex, their discrete gradients do not sit inside the curl kernel correctly, and they famously produce spectra peppered with non-physical eigenvalues. The same mechanism governs the $k=2$ ($\mathrm{div}$-$\mathrm{curl}$) and scalar $k=0$ cases, where conforming Lagrange elements already form a subcomplex and the classical Galerkin spectral theory of Strang-Fix and Babuška-Osborn applies without incident — which is why the eigenvalue pathology was historically invisible until vector problems forced it into view.

Synthesis. Spectral correctness is the foundational reason FEEC is more than a convergence theory: this is exactly the point where a method that is adequate for the source problem can still fail, and the discrete compactness property is the central insight that separates the two regimes. The bridge is that the bounded cochain projection of 24.03.05 generalises from a source-stability device into a spectral-stability device, putting these together so that the harmonic eigenvalue at zero, the nonzero resonances, and the Betti-number multiplicities are all reproduced by one structural hypothesis. The Maxwell case 24.04.02 is dual to the scalar case in the de Rham complex, and the abstract Hilbert-complex framing of 24.03.07 is the foundational reason a single proof covers every degree $k$ at once; what fails for naive vector elements is not accuracy but compactness, and that is the structural fact the whole chapter has been building toward.

Full proof set Master

Proposition 1 (zero eigenvalue has Betti-number multiplicity, preserved discretely). The eigenspace of $L$ for $\lambda =0$ on $k$-forms is the harmonic space ${\mathfrak{H}}^k$ with $\dim {\mathfrak{H}}^k=b_k$, and a FEEC subcomplex with a bounded cochain projection reproduces this exactly: $\dim {\mathfrak{H}}_h^k=b_k$ for all $h$.

Proof. $Lu=0$ gives $0=⟨Lu,u⟩=\Vert du{\Vert }^2+\Vert d^{\ast }u{\Vert }^2$, so $u\in \ker d\cap \ker d^{\ast }={\mathfrak{H}}^k$; conversely harmonic forms are annihilated by $L$. The Hodge isomorphism ${\mathfrak{H}}^k\cong H_{\mathrm{dR}}^k$ gives $\dim {\mathfrak{H}}^k=b_k$. Discretely, the bounded commuting projection makes ${\pi }_h$ a cochain map inducing an isomorphism on cohomology of the subcomplex (the FEEC cohomology theorem of 24.03.06), so $\dim {\mathfrak{H}}_h^k=\dim H^k(V_h^{\bullet })=b_k$. $\square$

Proposition 2 (discrete compactness implies no spectral pollution). Assume the discrete coexact spaces $Z_h$ have discrete compactness and that $T_h\to T$ pointwise on $L^2$. Then no sequence of discrete eigenvalues ${\lambda }_h\to {\lambda }^{\ast }$ can have ${\lambda }^{\ast }\notin \mathrm{spec}(L)\cup \{+\infty \}$ a nonzero finite limit outside the true spectrum.

Proof. Suppose ${\lambda }_h\to {\lambda }^{\ast }$ finite, ${\lambda }^{\ast }\ne 0$, with eigenforms $u_h\in Z_h$, $\Vert u_h\Vert =1$. Then $\Vert d{u}_h{\Vert }^2={\lambda }_h\to {\lambda }^{\ast }$, so $\{u_h\}$ is graph-norm bounded; discrete compactness extracts $u_h\to u$ in $L^2$ with $\Vert u\Vert =1$. For fixed $v\in H{\Lambda }^k$ and its projections ${\pi }_{h}v$, passing to the limit in $⟨d{u}_h,d{\pi }_{h}v⟩={\lambda }_h⟨u_h,{\pi }_{h}v⟩$ using the commuting property $d{\pi }_h={\pi }_{h}d$ and pointwise convergence yields $⟨du,dv⟩={\lambda }^{\ast }⟨u,v⟩$ for all $v$, i.e. $u$ is a genuine eigenform of $L$ with eigenvalue ${\lambda }^{\ast }$. Hence ${\lambda }^{\ast }\in \mathrm{spec}(L)$, contradicting the assumption. $\square$

Proposition 3 (failure of conforming-but-incompatible spaces). Let $W_h\subset H_0(\mathrm{curl};\Omega )=H{\Lambda }^1$ be a conforming vector finite element family that is not part of a discrete de Rham subcomplex with bounded cochain projection. Then $W_h$ need not have the discrete compactness property, and its curl-curl eigenvalue approximation can exhibit spurious eigenvalues.

Proof. Discrete compactness at $k=1$ requires that the discrete curl-free part and the discrete coexact part split $W_h$ compatibly with the continuous Helmholtz decomposition. Without a subcomplex $V_h^0\stackrel{\nabla }{\to }{W}_h\stackrel{\mathrm{curl}}{\to }{V}_h^2$ and a commuting projection, $\nabla (\text{discrete scalars})⊈\ker (\mathrm{curl}{|}_{W_h})$ in general, so the discrete kernel of the curl is the wrong dimension and the discrete coexact space is not uniformly precompact in $L^2$. One then constructs a graph-bounded sequence in the discrete coexact space with no $L^2$-convergent subsequence; the associated Rayleigh quotients accumulate at a value carrying no continuous eigenform, producing a spurious branch in the computed spectrum. (Vector nodal Lagrange elements on standard meshes realise this failure; this is the classical Maxwell spurious-mode phenomenon.) $\square$

Connections Master

• Abstract Hilbert complexes and the abstract Hodge decomposition 24.03.07. The general-degree spectral-correctness theorem lives most naturally at the abstract Hilbert-complex level: the compactness axiom there is precisely the hypothesis that makes the continuous solution operator compact, and discrete compactness is its discrete shadow. Proving spectral correctness once abstractly specialises to every concrete complex — de Rham, elasticity, Stokes — at no extra cost.

• Maxwell equations and FEEC edge elements 24.04.02. This unit is the general $k$-form theory of which the Maxwell cavity is the $k=1$, $n=3$ special case. The edge-element discrete compactness stated there for curl problems is the instance that historically motivated the whole investigation; here it is one slot of a uniform statement.

• Mixed FEM for the Hodge Laplacian 24.04.01. The mixed weak form and the discrete Hodge decomposition supplied there are exactly the apparatus the eigenproblem is built on; the eigenproblem is the spectral lift of that source problem, and the harmonic side condition reappears as the zero eigenvalue.

• Compact operators 02.11.05. The continuous solution operator is compact by Rellich-Kondrachov, which is why the spectrum is discrete with finite multiplicities; discrete compactness is the requirement that this property survive discretisation uniformly in the mesh.

• FEEC convergence theorem (Arnold-Falk-Winther) 24.03.06. The bounded cochain projection and discrete Poincaré-Friedrichs inequality proved there are the source-problem inputs that, once upgraded by discrete compactness to operator-norm convergence, deliver spectral correctness.

Historical & philosophical context Master

The spurious-mode problem was a practical scandal of computational electromagnetism in the 1970s and 1980s: nodal vector discretisations of cavity resonators returned eigenvalues that no measurement supported. Kikuchi isolated the missing ingredient in 1989, naming and using the discrete compactness property for Nédélec's edge elements to prove convergence of the Maxwell eigenproblem ^{[Kikuchi 1989]}. The general spectral-approximation machinery had been assembled earlier by Babuška and Osborn, whose handbook article codified the double-order eigenvalue estimate and the role of operator-norm (collectively compact) convergence in excluding pollution ^{[Babuska 1991]}.

Boffi's 2010 Acta Numerica survey consolidated three decades of eigenvalue analysis and identified discrete compactness, completeness of discrete divergence-free fields, and the commuting-diagram property as the recurring sufficient conditions across edge, face, and mixed elements ^{[Boffi 2010]}. Arnold, Falk, and Winther placed the result inside finite element exterior calculus, deriving spectral correctness from a bounded cochain projection on a discrete subcomplex and presenting it as a corollary of the same Hilbert-complex structure that governs the source problem ^{[arnold-falk-winther-2006-feec] [Arnold 2010]}. Kikuchi's discrete compactness and the FEEC commuting projection are two descriptions of one structural fact, and the edge element appears in both accounts as the degree-one slot of a discrete de Rham complex.

Bibliography Master

@article{Boffi2010Eigenvalue,
  author  = {Boffi, Daniele},
  title   = {Finite element approximation of eigenvalue problems},
  journal = {Acta Numerica},
  volume  = {19},
  pages   = {1--120},
  year    = {2010}
}

@article{Kikuchi1989DiscreteCompactness,
  author  = {Kikuchi, Fumio},
  title   = {On a discrete compactness property for the {N}{\'e}d{\'e}lec finite elements},
  journal = {Journal of the Faculty of Science, University of Tokyo, Section IA},
  volume  = {36},
  pages   = {479--490},
  year    = {1989}
}

@article{ArnoldFalkWinther2010Bams,
  author  = {Arnold, Douglas N. and Falk, Richard S. and Winther, Ragnar},
  title   = {Finite element exterior calculus: from {H}odge theory to numerical stability},
  journal = {Bulletin of the American Mathematical Society},
  volume  = {47},
  number  = {2},
  pages   = {281--354},
  year    = {2010}
}

@incollection{BabuskaOsborn1991Eigenvalue,
  author    = {Babu{\v{s}}ka, Ivo and Osborn, John E.},
  title     = {Eigenvalue problems},
  booktitle = {Handbook of Numerical Analysis, Vol. II: Finite Element Methods (Part 1)},
  editor    = {Ciarlet, P. G. and Lions, J. L.},
  publisher = {North-Holland},
  pages     = {641--787},
  year      = {1991}
}

@article{ArnoldFalkWinther2006FeecActa,
  author  = {Arnold, Douglas N. and Falk, Richard S. and Winther, Ragnar},
  title   = {Finite element exterior calculus, homological techniques, and applications},
  journal = {Acta Numerica},
  volume  = {15},
  pages   = {1--155},
  year    = {2006}
}

@article{BoffiBrezziGastaldi2000Spectral,
  author  = {Boffi, Daniele and Brezzi, Franco and Gastaldi, Lucia},
  title   = {On the problem of spurious eigenvalues in the approximation of linear elliptic problems in mixed form},
  journal = {Mathematics of Computation},
  volume  = {69},
  pages   = {121--140},
  year    = {2000}
}