Introduction
Advection–diffusion of differential forms arises in a wide range of important phenomena, including scalar advection–diffusion, magnetohydrodynamics (MHD), and Fokker–Planck equations [7 , 25 , 26 ] . These models admit the unified form
∂ ω ∂ t + L β ω + η Δ H ω = 0 , \frac{\partial\omega}{\partial t} + L_\beta\omega+ \eta\Delta_{\mathrm{H}}\omega= 0, ∂ t ∂ ω + L β ω + η Δ H ω = 0 ,
where ω \omega ω is a differential k k k -form, L β ω L_\beta\omega L β ω denotes the Lie derivative of ω \omega ω along a vector field β \beta β modeling transport, Δ H ω : = ( d k − 1 δ k + δ k + 1 d k ) ω \Delta_{\mathrm{H}}\omega:= (d^{k-1}\delta^k + \delta^{k+1}d^k)\omega Δ H ω := ( d k − 1 δ k + δ k + 1 d k ) ω is the Hodge Laplacian modeling diffusion, and η \eta η is a constant coefficient measuring the relative strength between advection and diffusion.
Understanding the long-time behavior of advection–diffusion of differential forms (referred to as Lie advection-diffusion below) is of fundamental importance. For densities, this is closely related to topological and Morse-theoretic aspects of Fokker–Planck and other nonlinear equations [7 Section V.3] . For MHD problems, asymptotics plays a central role in understanding force-free fields in solar physics and in the design of equilibrium fields for fusion devices [12 , 41 ] , while long-time behavior is crucial for understanding dynamo action. Numerical computation has become indispensable for investigating these phenomena. Finite element methods have been developed for advection–diffusion of differential forms and full MHD problems [25 , 26 , 27 , 28 , 31 , 32 , 33 , 37 , 49 , 54 , 55 ] . Rigorous numerical analysis is therefore required to understand the reliability and interpretation of computational results.
Most existing works (such as the error analysis in [9 ] ) focus on local numerical analysis , namely the convergence of numerical solutions as the mesh size h → 0 h \to0 h → 0 and/or the polynomial degree p → ∞ p \to\infty p → ∞ for p p p /hp methods, either at a fixed time or over a finite time interval. However, many problems arising in the applications above concern long-time asymptotics rather than finite-time behavior. An important example is dynamo theory, which studies the mechanism by which magnetic fields are generated and sustained by the motion of conducting fluids, with fundamental applications to geophysical and astrophysical systems such as planets and stars. Mathematically, the dynamo problem asks whether the magnetic field B \mathbf{B} B , viewed as a differential 2-form in 3D, can exhibit exponential growth under the evolution
∂ t B − curl ( β × B ) + R m − 1 curl curl B = 0 , div B = 0. (1) \partial_t \mathbf{B} - \operatorname{curl}(\boldsymbol{\beta} \times\mathbf{B}) + R_m^{-1} \operatorname{curl}\operatorname{curl}\mathbf{B} = 0,\qquad\operatorname{div}\mathbf{B} = 0. \tag*{(1)} ∂ t B − curl ( β × B ) + R m − 1 curl curl B = 0 , div B = 0. (1)
Here β \boldsymbol{\beta} β is the prescribed velocity field and R m R_m R m is the magnetic Reynolds number. A central question in dynamo theory is whether the transport term can overcome magnetic diffusion and generate exponentially growing magnetic fields, particularly in the large-R m R_m R m regime [7 ] . Dynamo theory remains an active area of research, and many fundamental questions are still open. Nevertheless, several antidynamo theorems identify situations in which dynamo action cannot occur [7 ] . Antidynamo theory addresses the complementary question: which geometric, topological, or structural conditions rule out such growth?
Numerical methods are widely used to investigate dynamo and antidynamo phenomena. However, local numerical analysis is insufficient for addressing many questions of practical and theoretical interest in this setting. For example, convergence as h → 0 h \to0 h → 0 over a finite time interval does not determine whether numerical dynamo action (exponential growth of numerical solutions) will occur for a particular implementation with fixed discretization parameters. This calls for numerical analysis of global and long-time behavior. Similar global questions have already attracted attention in geometric numerical integration, for example in the study of KAM-type behavior and long-time preservation properties of symplectic integrators [46 ] .
Lie advection–diffusion is closely related to global, topological, and dynamical-system properties. In the spirit of structure-preserving numerical methods, we aim to formulate and analyze discrete analogues of such questions arising in topological hydrodynamics. One example of this paradigm is the long-time relaxation of MHD systems under topological constraints [20 , 24 ] (see [7 Chapter III] ). In the context of dynamo theory, several aspects interact. Spectral properties of the operators provide information about the evolution, while pseudospectra reveal additional dynamical features [47 , 48 ] . The convergence of eigenvalue problems and pseudospectra for Lie advection–diffusion as h → 0 h \to0 h → 0 has been studied in another work [11 ] . However, spectra and pseudospectra alone do not fully characterize long-time dynamics, and that analysis remains local in the sense that it studies convergence of (pseudo)spectra as h → 0 h \to0 h → 0 . That analysis does not address the practically relevant situation of fixed discretization parameters (h h h or p p p ), where one wishes to determine whether a numerical scheme faithfully predicts dynamo or antidynamo behavior. For example, a computation at fixed discretization parameters may produce positive eigenvalues that later disappear under mesh refinement, leading to potentially misleading conclusions about long-time growth. This motivates the development of global numerical analysis for advection–diffusion of differential forms, which is the primary focus of this paper.
In this paper, we study properties of finite element methods for Lie advection–diffusion, particularly properties that do not require asymptotic restrictions on discretization parameters such as sufficiently small mesh size. In other words, in addition to asymptotic results as h → 0 h \to0 h → 0 , we focus on the behavior of numerical schemes at finite discretization parameters, which is the regime relevant in practical computation. We show that certain classes of numerical methods faithfully inherit structural properties of the underlying continuous dynamical systems arising in topological hydrodynamics. In particular, we establish antidynamo theorems for schemes based on finite element exterior calculus (FEEC) [4 , 5 , 32 , 51 ] . Consequently, these results help exclude undesirable situations in which the continuous system does not exhibit dynamo action while the numerical discretization spuriously predicts exponential growth.
More broadly, the discussion extends beyond antidynamo theorems and addresses advection–diffusion from a Hodge-theoretic viewpoint. Classical Hodge theory establishes that harmonic forms, namely the kernel of the Hodge Laplacian, have dimension equal to the Betti numbers of the underlying manifold. FEEC preserves this property and therefore provides a faithful discretization of Hodge theory. When advection is introduced, a natural question is how this correspondence changes. Topological hydrodynamics provides a partial answer [6 , 7 ] : when diffusion dominates, the dimension of the generalized harmonic forms, namely the kernel of the Lie advection–diffusion operator, still coincides with the Betti numbers, while in general the dimension is bounded below by the Betti numbers. Numerical evidence for this phenomenon was presented in [11 ] . In this paper, we show that FEEC discretizations faithfully preserve these structural properties.
Although our theory does not cover all cases of Lie advection–diffusion, paralleling the limitations of the continuous theory itself, a particularly transparent situation occurs when the drift is a potential flow. If there exists a smooth function θ \theta θ such that β ♭ = η d θ \beta^\flat= \eta\mathrm{d}\theta β ♭ = η d θ , then the generalized Hodge Laplacian can be written as a weighted Hodge Laplacian, which is self-adjoint and positive semidefinite. This is closely related to Witten’s deformation of the de Rham complex [50 ] . The weighted Hodge decomposition then implies that the space of weighted harmonic k k k -forms is naturally isomorphic to the k k k -th de Rham cohomology group. Since the governing operator has non-negative real spectrum, the evolution preserves the weighted harmonic component contained in the kernel, while all orthogonal components decay to zero. Consequently, the long-time limit is completely determined by the weighted harmonic component of the initial data. This yields a differential-form version of an antidynamo mechanism for gradient-type drifts.
The paper is organized as follows. § 2 reviews notation from Hodge theory and FEEC and introduces the generalized Hodge Laplacian. § 3 develops the weighted Hodge theory for potential flows, while § 4 presents the FEEC discretization. The finite element antidynamo theorems are established in § 5 . § 6 verifies the structural assumptions for sufficiently fine meshes and for lowest-order exponential-fitting discretizations on arbitrary mesh size. Numerical experiments are presented in § 7 to show the long-time evolution of the proposed discretizations. Finally, § 8 concludes the paper.
Preliminaries
This section introduces the necessary mathematical tools and notation. We begin with an overview of Hodge theory and its generalization through incorporating Lie advection. Following this, we analyze the spectral properties of the generalized operator to characterize its a priori steady-state behavior. To conclude, we present finite element exterior calculus (FEEC), which serves as the discrete framework for our subsequent numerical analysis.
Hodge theory and its generalization
Unless otherwise stated, all metric dependent constructions and results are understood with respect to a fixed Riemannian metric. Euclidean settings are specified explicitly when required. We first recall the basic notation from Hodge theory and the de Rham complex. Given a compact, connected, oriented manifold M M M , possibly with boundary, let L 2 Λ k ( M ) L^2\Lambda^k(M) L 2 Λ k ( M ) denote the Hilbert space of square-integrable differential k k k -forms. The L 2 L^2 L 2 inner product is
( u , v ) : = ∫ M u ∧ ⋆ v , u , v ∈ L 2 Λ k ( M ) , (2) (u,v) := \int_M u \wedge\star v,\qquad u,v \in L^2\Lambda^k(M), \tag*{(2)} ( u , v ) := ∫ M u ∧ ⋆ v , u , v ∈ L 2 Λ k ( M ) , (2)
where ⋆ : Λ k ( M ) → Λ n − k ( M ) \star:\Lambda^k(M)\to\Lambda^{n-k}(M) ⋆ : Λ k ( M ) → Λ n − k ( M ) is the Hodge star. The associated norm is denoted by
∥ ω ∥ 2 = ∥ ω ∥ L 2 Λ k 2 : = ( ω , ω ) . \|\omega\|^2=\|\omega\|^2_{L^2\Lambda^k}:=(\omega,\omega). ∥ ω ∥ 2 = ∥ ω ∥ L 2 Λ k 2 := ( ω , ω ) .
For s ∈ R s\in\mathbb{R} s ∈ R , we denote by H s Λ k ( M ) H^s\Lambda^k(M) H s Λ k ( M ) the Sobolev space of k k k -forms whose coefficients belong to H s ( M ) H^s(M) H s ( M ) , and we write its norm as ∥ ⋅ ∥ H s \|\cdot\|_{H^s} ∥ ⋅ ∥ H s . In particular, given exterior derivatives d k d^k d k satisfying d k ∘ d k − 1 = 0 d^k\circ d^{k-1}=0 d k ∘ d k − 1 = 0 , we set
V k : = H Λ k ( M ) : = { ω ∈ L 2 Λ k ( M ) : d k ω ∈ L 2 Λ k + 1 ( M ) } . (3) V^k:=H\Lambda^k(M):=\{\omega\in L^2\Lambda^k(M):d^k\omega\in L^2\Lambda^{k+1}(M)\}. \tag*{(3)} V k := H Λ k ( M ) := { ω ∈ L 2 Λ k ( M ) : d k ω ∈ L 2 Λ k + 1 ( M )} . (3)
equipped with the norm
∥ ω ∥ V k 2 = ∥ ω ∥ H Λ k 2 : = ∥ ω ∥ 2 + ∥ d k ω ∥ 2 . (4) \|\omega\|^2_{V^k}=\|\omega\|^2_{H\Lambda^k}:=\|\omega\|^2+\|d^k\omega\|^2. \tag*{(4)} ∥ ω ∥ V k 2 = ∥ ω ∥ H Λ k 2 := ∥ ω ∥ 2 + ∥ d k ω ∥ 2 . (4)
We then have the L 2 L^2 L 2 -based de Rham domain complex
0 ⟶ H Λ 0 ( M ) → d H Λ 1 ( M ) → d ⋯ → d H Λ n ( M ) ⟶ 0. (5) 0\longrightarrow H\Lambda^0(M)\xrightarrow{d}H\Lambda^1(M)\xrightarrow{d}\cdots\xrightarrow{d}H\Lambda^n(M)\longrightarrow0. \tag*{(5)} 0 ⟶ H Λ 0 ( M ) d H Λ 1 ( M ) d ⋯ d H Λ n ( M ) ⟶ 0. (5)
The coderivative δ k : Λ k ( M ) → Λ k − 1 ( M ) \delta^k:\Lambda^k(M)\to\Lambda^{k-1}(M) δ k : Λ k ( M ) → Λ k − 1 ( M ) is given by
⋆ δ k = ( − 1 ) k d n − k ⋆ . (6) \star\delta^k=(-1)^k d^{n-k}\star. \tag*{(6)} ⋆ δ k = ( − 1 ) k d n − k ⋆ . (6)
In analogy with H Λ k ( M ) H\Lambda^k(M) H Λ k ( M ) , we define the space
H ∗ Λ k ( M ) : = { ω ∈ L 2 Λ k ( M ) : δ k ω ∈ L 2 Λ k − 1 ( M ) } . (7) H^*\Lambda^k(M):=\{\omega\in L^2\Lambda^k(M):\delta^k\omega\in L^2\Lambda^{k-1}(M)\}. \tag*{(7)} H ∗ Λ k ( M ) := { ω ∈ L 2 Λ k ( M ) : δ k ω ∈ L 2 Λ k − 1 ( M )} . (7)
For smooth forms u ∈ Λ k − 1 ( M ) u\in\Lambda^{k-1}(M) u ∈ Λ k − 1 ( M ) and v ∈ Λ k ( M ) v\in\Lambda^k(M) v ∈ Λ k ( M ) , Stokes’ theorem gives
( d k − 1 u , v ) = ( u , δ k v ) + ∫ ∂ M tr ( u ) ∧ tr ( ⋆ v ) . (8) (d^{k-1}u,v)=(u,\delta^kv)+\int_{\partial M}\operatorname{tr}(u)\wedge\operatorname{tr}(\star v). \tag*{(8)} ( d k − 1 u , v ) = ( u , δ k v ) + ∫ ∂ M tr ( u ) ∧ tr ( ⋆ v ) . (8)
The spaces H Λ k H\Lambda^k H Λ k and H ∗ Λ k H^*\Lambda^k H ∗ Λ k with homogeneous boundary conditions are defined as
H ˚ Λ k ( M ) : = { ω ∈ H Λ k ( M ) : tr ω = 0 } , (9) \mathring{H}\Lambda^k(M):=\{\omega\in H\Lambda^k(M):\operatorname{tr}\omega=0\}, \tag*{(9)} H ˚ Λ k ( M ) := { ω ∈ H Λ k ( M ) : tr ω = 0 } , (9)
H ˚ ∗ Λ k ( M ) : = { ω ∈ H ∗ Λ k ( M ) : tr ( ⋆ ω ) = 0 } . (10) \mathring{H}^*\Lambda^k(M):=\{\omega\in H^*\Lambda^k(M):\operatorname{tr}(\star\omega)=0\}. \tag*{(10)} H ˚ ∗ Λ k ( M ) := { ω ∈ H ∗ Λ k ( M ) : tr ( ⋆ ω ) = 0 } . (10)
Throughout this paper, we work in settings in which the exterior derivatives d k d^k d k in the underlying de Rham Hilbert complex have closed ranges. The exterior derivative is regarded as an unbounded operator with domain D ( d k ) = H Λ k ( M ) D(d^k)=H\Lambda^k(M) D ( d k ) = H Λ k ( M ) . The domain of the coderivative depends on the existence of the boundary. In the presence of boundary, we take D ( δ k ) = H ˚ ∗ Λ k ( M ) D(\delta^k)=\mathring{H}^*\Lambda^k(M) D ( δ k ) = H ˚ ∗ Λ k ( M ) , whereas in the boundaryless case, D ( δ k ) = H ∗ Λ k ( M ) D(\delta^k)=H^*\Lambda^k(M) D ( δ k ) = H ∗ Λ k ( M ) . The space of harmonic k k k -forms is defined by
H k = { ω ∈ D ( d k ) : d k ω = 0 , ( ω , d k − 1 ζ ) = 0 , ∀ ζ ∈ D ( d k − 1 ) } = { ω ∈ D ( d k ) ∩ D ( δ k ) : d k ω = 0 , δ k ω = 0 } . (11) \begin{aligned}
\mathcal{H}^k=\{\omega\in D(d^k):d^k\omega=0,\;(\omega,d^{k-1}\zeta)=0,\ \forall\zeta\in D(d^{k-1})\} \\
&=\{\omega\in D(d^k)\cap D(\delta^k):d^k\omega=0,\ \delta^k\omega=0\}. \tag*{(11)}
\end{aligned} H k = { ω ∈ D ( d k ) : d k ω = 0 , ( ω , d k − 1 ζ ) = 0 , ∀ ζ ∈ D ( d k − 1 )} = { ω ∈ D ( d k ) ∩ D ( δ k ) : d k ω = 0 , δ k ω = 0 } . (11)
The Hodge Laplacian on k k k -forms is
Δ H L k : = δ k + 1 d k + d k − 1 δ k , (12) \Delta_{\mathrm{HL}}^k:=\delta^{k+1}d^k+d^{k-1}\delta^k, \tag*{(12)} Δ HL k := δ k + 1 d k + d k − 1 δ k , (12)
with domain
D ( Δ H L k ) = { ω ∈ D ( d k ) ∩ D ( δ k ) : d k ω ∈ D ( δ k + 1 ) , δ k ω ∈ D ( d k − 1 ) } . D(\Delta_{\mathrm{HL}}^k)=\{\omega\in D(d^k)\cap D(\delta^k):d^k\omega\in D(\delta^{k+1}),\ \delta^k\omega\in D(d^{k-1})\}. D ( Δ HL k ) = { ω ∈ D ( d k ) ∩ D ( δ k ) : d k ω ∈ D ( δ k + 1 ) , δ k ω ∈ D ( d k − 1 )} .
The L 2 L^2 L 2 -Hodge decomposition is the orthogonal direct sum
L 2 Λ k ( M ) = R ( d k − 1 ) ⊕ R ( δ k + 1 ) ⊕ H k , (13) L^2\Lambda^k(M)=\mathcal{R}(d^{k-1})\oplus\mathcal{R}(\delta^{k+1})\oplus\mathcal{H}^k, \tag*{(13)} L 2 Λ k ( M ) = R ( d k − 1 ) ⊕ R ( δ k + 1 ) ⊕ H k , (13)
where H k = N ( Δ H L k ) \mathcal{H}^k=\mathcal{N}(\Delta_{\mathrm{HL}}^k) H k = N ( Δ HL k ) . Here, N ( T ) \mathcal{N}(T) N ( T ) and R ( T ) \mathcal{R}(T) R ( T ) denote the kernel and range of an operator T T T , respectively. The de Rham cohomology group associated with the complex is
H d R k ( M ) : = N ( d k ) / R ( d k − 1 ) . (14) H_{\mathrm{dR}}^k(M):=\mathcal{N}(d^k)/\mathcal{R}(d^{k-1}). \tag*{(14)} H dR k ( M ) := N ( d k ) / R ( d k − 1 ) . (14)
Hodge theory identifies each de Rham cohomology class with a unique harmonic representative. Equivalently, H k ≅ H d R k ( M ) \mathcal{H}^k\cong H_{\mathrm{dR}}^k(M) H k ≅ H dR k ( M ) . Consequently, the k k k -th Betti number satisfies
b k ( M ) = dim H d R k ( M ) = dim H k . (15) b_k(M)=\dim H_{\mathrm{dR}}^k(M)=\dim\mathcal{H}^k. \tag*{(15)} b k ( M ) = dim H dR k ( M ) = dim H k . (15)
We now consider the transport of differential forms by a given velocity field β \beta β . The advection-diffusion evolution of ω ∈ V k \omega\in V^k ω ∈ V k with velocity field β \beta β and diffusion coefficient η > 0 \eta>0 η > 0 is given by
∂ ω ∂ t + L β ω + η Δ H L ω = 0 , (16) \frac{\partial\omega}{\partial t}+\mathcal{L}_{\beta}\omega+\eta\Delta_{\mathrm{HL}}\omega=0, \tag*{(16)} ∂ t ∂ ω + L β ω + η Δ HL ω = 0 , (16)
where L β \mathcal{L}_{\beta} L β denotes the Lie derivative along β \beta β . By Cartan’s formula,
L β ω = d i β ω + i β d ω , (17) \mathcal{L}_{\beta}\omega=d\mathsf{i}_{\beta}\omega+\mathsf{i}_{\beta}d\omega, \tag*{(17)} L β ω = d i β ω + i β d ω , (17)
where i β \mathsf{i}_{\beta} i β is the contraction with β \beta β .
Since the topological properties of the Hodge Laplacian Δ H L \Delta_{\mathrm{HL}} Δ HL are well understood, it is natural to view
Δ η , β : = η Δ H L + L β , η > 0 , (18) \Delta_{\eta,\beta}:=\eta\Delta_{\mathrm{HL} }+\mathcal{L}_{\beta},\qquad\eta>0, \tag*{(18)} Δ η , β := η Δ HL + L β , η > 0 , (18)
as a generalized Hodge Laplacian. The operator Δ η , β \Delta_{\eta,\beta} Δ η , β is a second-order elliptic operator. On a compact manifold, elliptic regularity for a sufficiently large positive shift Δ η , β k + c I \Delta_{\eta,\beta}^k+cI Δ η , β k + c I , together with the compact Sobolev embedding, implies that the shifted resolvent is compact on L 2 Λ k ( M ) L^2\Lambda^k(M) L 2 Λ k ( M ) . Hence Δ η , β k \Delta_{\eta,\beta}^k Δ η , β k has a compact resolvent. In particular, its spectrum consists of isolated eigenvalues of finite algebraic multiplicity, and therefore N ( Δ η , β k ) \mathcal{N}(\Delta_{\eta,\beta}^k) N ( Δ η , β k ) is finite-dimensional.
The kernel N ( Δ η , β k ) \mathcal{N}(\Delta_{\eta,\beta}^k) N ( Δ η , β k ) therefore plays the role of a space of generalized harmonic k k k -forms. Its elements are the stationary solutions of the advection–diffusion evolution, i.e., the time-invariant part of the dynamics. However, unlike the classical Hodge Laplacian, the operator Δ η , β k \Delta_{\eta,\beta}^k Δ η , β k is, in general, non-self-adjoint and indefinite. As a result, the long-time behavior of the system may not converge to a steady state given by the kernel of Δ η , β \Delta_{\eta,\beta} Δ η , β . This raises several natural questions: how does the generalized harmonic space N ( Δ η , β k ) \mathcal{N}(\Delta_{\eta,\beta}^k) N ( Δ η , β k ) reflect the topology of the underlying domain? Does the advection–diffusion evolution converge to this generalized harmonic space as t → ∞ t\to\infty t → ∞ ? At the discrete level, the first question is answered in § 3 for potential flows and in Theorem 4.3 for general advection (a lower bound by the Betti numbers), while the second is the subject of the finite element antidynamo theorems of § 5 .
Spectral properties of Δ η , β \Delta_{\eta,\beta} Δ η , β
Beyond characterizing the kernel of the generalized Hodge Laplacian operator Δ η , β k \Delta_{\eta,\beta}^{k} Δ η , β k , we also require a priori information on its spectral structure in order to understand the long-time dynamics of the evolution. The spectrum is well-understood for the scalar cases (k = 0 k=0 k = 0 and k = n k=n k = n ). For 0-forms, the operator is a second-order elliptic operator, and obeys the maximum principle. For n n n -forms, the operator corresponds to the adjoint of the 0-form operator via Hodge duality, subject to a reversed velocity field. The spectral behavior of Δ η , β \Delta_{\eta,\beta} Δ η , β on 0-forms serves as the cornerstone for analyzing the global dissipation of the system. Rooted in the maximum and comparison principles, the operator exhibits the following fundamental spectral properties.
Theorem 2.1 (spectral properties of Δ η , β \Delta_{\eta,\beta} Δ η , β for k = 0 k=0 k = 0 ). For any diffusion coefficient η > 0 \eta>0 η > 0 and velocity field β ∈ L ∞ ( M ; T M ) \beta\in L^\infty(M;TM) β ∈ L ∞ ( M ; T M ) , let σ ( Δ η , β 0 ) \sigma(\Delta_{\eta,\beta}^{0}) σ ( Δ η , β 0 ) denote the spectrum of the advection-diffusion operator of 0-forms. Δ η , β 0 \Delta_{\eta,\beta}^{0} Δ η , β 0 satisfies the following properties:
principal eigenvalue: There exists a real, algebraically simple principal eigenvalue λ 0 ∈ R \lambda_0\in\mathbb{R} λ 0 ∈ R . In particular, λ 0 = 0 \lambda_0=0 λ 0 = 0 in the closed, periodic, and Neumann cases, while λ 0 > 0 \lambda_0>0 λ 0 > 0 under Dirichlet boundary conditions.
sectorial distribution: The operator is sectorial. Its spectrum is confined within a sectorial region in the right half of the complex plane:
σ ( Δ η , β 0 ) ⊂ Σ ϑ , λ 0 : = { z ∈ C ∣ arg ( z − λ 0 ) ≤ ϑ } \sigma(\Delta_{\eta,\beta}^{0})\subset\Sigma_{\vartheta,\lambda_0}:=\{z\in\mathbb{C}\mid\arg(z-\lambda_0)\le\vartheta\} σ ( Δ η , β 0 ) ⊂ Σ ϑ , λ 0 := { z ∈ C ∣ arg ( z − λ 0 ) ≤ ϑ } for some angle ϑ ∈ [ 0 , π / 2 ) \vartheta\in[0,\pi/2) ϑ ∈ [ 0 , π /2 ) .
spectral gap: Δ η , β 0 \Delta_{\eta,\beta}^{0} Δ η , β 0 has compact resolvent and there exists a constant μ \mu μ such that all λ ∈ σ ( Δ η , β 0 ) ∖ { λ 0 } \lambda\in\sigma(\Delta_{\eta,\beta}^{0})\setminus\{\lambda_0\} λ ∈ σ ( Δ η , β 0 ) ∖ { λ 0 } satisfy Re ( λ ) ≥ μ > λ 0 \operatorname{Re}(\lambda)\ge\mu>\lambda_0 Re ( λ ) ≥ μ > λ 0 .
The spectral properties stated in Theorem 2.1 are well established (see, e.g., [19 , 22 , 39 ] . and to clarify the boundary-condition conventions used here, we include a proof for its simplified version in Appendix § 9 .§ 1 for β ∈ C 0 , 1 ( M ; T M ) \beta\in C^{0,1}(M;TM) β ∈ C 0 , 1 ( M ; T M ) in the closed, periodic, and Neumann cases.
The spectral structure described in Theorem 2.1 is precisely the property that a structure-preserving discretization should inherit. It will be formalized at the discrete level as Assumption 5.3 in § 5 and verified for concrete schemes in § 6 .
Finite element exterior calculus
We briefly introduce the basic setup for finite element exterior calculus (FEEC). Let T h \mathcal{T}_h T h be a shape-regular triangulation of the manifold M M M , with mesh size h : = max T ∈ T h diam ( T ) h:=\max_{T\in\mathcal{T}_h}\operatorname{diam}(T) h := max T ∈ T h diam ( T ) , we have finite element subcomplexes V h k ⊂ V k V_h^k\subset V^k V h k ⊂ V k , which form the discrete de Rham complex
0 ⟶ V h 0 → d V h 1 → d ⋯ → d V h n ⟶ 0. (19) 0\longrightarrow V_h^0\xrightarrow{d}V_h^1\xrightarrow{d}\cdots\xrightarrow{d}V_h^n\longrightarrow0. \tag*{(19)} 0 ⟶ V h 0 d V h 1 d ⋯ d V h n ⟶ 0. (19)
Standard choices for these discrete spaces include polynomial differential forms, V h k = P r − Λ k ( T h ) V_h^k=\mathcal{P}_r^-\Lambda^k(\mathcal{T}_h) V h k = P r − Λ k ( T h ) and P r Λ k ( T h ) \mathcal{P}_r\Lambda^k(\mathcal{T}_h) P r Λ k ( T h ) .
Similarly, the discrete cohomology groups are defined as
H h k : = ( N ( d k ) ∩ V h k ) / ( d k − 1 V h k − 1 ) . (20) H_h^k:=(\mathcal{N}(d^k)\cap V_h^k)/(d^{k-1}V_h^{k-1}). \tag*{(20)} H h k := ( N ( d k ) ∩ V h k ) / ( d k − 1 V h k − 1 ) . (20)
A fundamental result in finite element exterior calculus (FEEC) is that the cohomology of the discrete complex is isomorphic to the cohomology of the continuous complex [5 , 15 ] . Consequently, b k = dim H k = dim H h k b_k=\dim H^k=\dim H_h^k b k = dim H k = dim H h k .
We introduce the discrete L 2 L^2 L 2 adjoint operator δ h k : V h k → V h k − 1 \delta_h^k:V_h^k\to V_h^{k-1} δ h k : V h k → V h k − 1 defined by the following relation
( δ h u h , v h ) = ( u h , d v h ) , ∀ ( u h , v h ) ∈ V h k × V h k − 1 . (21) (\delta_hu_h,v_h)=(u_h,dv_h),\qquad\forall(u_h,v_h)\in V_h^k\times V_h^{k-1}. \tag*{(21)} ( δ h u h , v h ) = ( u h , d v h ) , ∀ ( u h , v h ) ∈ V h k × V h k − 1 . (21)
Additionally, we define the L 2 L^2 L 2 projection operator Q h k : L 2 Λ k → V h k Q_h^k : L^2\Lambda^k \to V_h^k Q h k : L 2 Λ k → V h k through
( Q h k u , v h ) = ( u , v h ) , ∀ v h ∈ V h k . (22) (Q_h^k u,v_h)=(u,v_h), \qquad\forall v_h\in V_h^k. \tag*{(22)} ( Q h k u , v h ) = ( u , v h ) , ∀ v h ∈ V h k . (22)
At the discrete level, one has
B h k = d V h k − 1 , Z h k = N ( d k ) ∩ V h k . (23) \mathcal{B}_h^k=dV_h^{k-1}, \qquad\mathcal{Z}_h^k=\mathcal{N}(d^k)\cap V_h^k. \tag*{(23)} B h k = d V h k − 1 , Z h k = N ( d k ) ∩ V h k . (23)
The discrete harmonic form is defined by analogy as
H h k = { v ∈ Z h k : v ⊥ B h k } , (24) \mathcal{H}_h^k=\{v\in\mathcal{Z}_h^k:v\perp\mathcal{B}_h^k\}, \tag*{(24)} H h k = { v ∈ Z h k : v ⊥ B h k } , (24)
which is isomorphic to H h k H_h^k H h k . The discrete Hodge decomposition then takes the form
V h k = Z h k ⊕ B h k ⊕ H h k = B h k , ∗ ⊕ B h k ⊕ H h k , (25) V_h^k=\mathcal{Z}_h^k\oplus\mathcal{B}_h^k\oplus\mathcal{H}_h^k=\mathcal{B}_h^{k,*}\oplus\mathcal{B}_h^k\oplus\mathcal{H}_h^k, \tag*{(25)} V h k = Z h k ⊕ B h k ⊕ H h k = B h k , ∗ ⊕ B h k ⊕ H h k , (25)
where B h k , ∗ = δ h k + 1 V h k + 1 \mathcal{B}_h^{k,*}=\delta_h^{k+1}V_h^{k+1} B h k , ∗ = δ h k + 1 V h k + 1 .
Weighted Hodge theory and antidynamo for potential flows
Building on the general framework of § 2 , we now examine a case in which the generalized Hodge Laplacian recovers a self-adjoint structure. The advection field is assumed to be a potential field at the scale determined by the diffusion coefficient: there exists a scalar function θ \theta θ such that
β ♭ = η d θ . (26) \beta^\flat=\eta d\theta. \tag*{(26)} β ♭ = η d θ . (26)
For a vector field β \beta β , let j β : = β ♭ ∧ : Λ k ( M ) → Λ k + 1 ( M ) j_\beta:=\beta^\flat\wedge:\Lambda^k(M)\to\Lambda^{k+1}(M) j β := β ♭ ∧ : Λ k ( M ) → Λ k + 1 ( M ) . Under the standard L 2 L^2 L 2 inner product, j β j_\beta j β is the formal adjoint of the contraction i β i_\beta i β . Define the deformed exterior derivative
d θ ω : = e − θ d ( e θ ω ) = d ω + d θ ∧ ω . (27) d_\theta\omega:=e^{-\theta}d(e^\theta\omega)=d\omega+d\theta\wedge\omega. \tag*{(27)} d θ ω := e − θ d ( e θ ω ) = d ω + d θ ∧ ω . (27)
Then (3.1) gives
η d θ = η d + j β , (28) \eta d_\theta=\eta d+j_\beta, \tag*{(28)} η d θ = η d + j β , (28)
which identifies the advection–diffusion flux η d k + j β k \eta d^k+j_\beta^k η d k + j β k and motivates the exponential-fitting flux used in the finite element discretization introduced later.
Let θ ∈ W 1 , ∞ ( M ) \theta\in W^{1,\infty}(M) θ ∈ W 1 , ∞ ( M ) . Then there exists constants 0 < E 0 ≤ E 0 < ∞ 0<E_0\le E^0<\infty 0 < E 0 ≤ E 0 < ∞ such that
0 < E 0 ≤ e − θ ( x ) ≤ E 0 < ∞ , for a.e. x ∈ M . (29) 0<E_0\le e^{-\theta(x)}\le E^0<\infty,\qquad\text{for a.e. }x\in M. \tag*{(29)} 0 < E 0 ≤ e − θ ( x ) ≤ E 0 < ∞ , for a.e. x ∈ M . (29)
Indeed, one may take E 0 = e − ∥ θ ∥ L ∞ ( M ) E_0=e^{-\|\theta\|_{L^\infty(M)}} E 0 = e − ∥ θ ∥ L ∞ ( M ) and E 0 = e ∥ θ ∥ L ∞ ( M ) E^0=e^{\|\theta\|_{L^\infty(M)}} E 0 = e ∥ θ ∥ L ∞ ( M ) . We then define the corresponding weighted L 2 L^2 L 2 inner product.
Definition 3.1 (weighted inner product). The weighted L 2 L^2 L 2 inner product on k k k -forms is
( α , γ ) θ : = ∫ M e − θ α ∧ ⋆ γ . (30) (\alpha,\gamma)_\theta:=\int_M e^{-\theta}\alpha\wedge\star\gamma. \tag*{(30)} ( α , γ ) θ := ∫ M e − θ α ∧ ⋆ γ . (30) The induced norm is denoted by ∥ ω ∥ θ 2 = ( ω , ω ) θ \|\omega\|_\theta^2=(\omega,\omega)_\theta ∥ ω ∥ θ 2 = ( ω , ω ) θ , and the completion of C ∞ Λ k ( M ) C^\infty\Lambda^k(M) C ∞ Λ k ( M ) is denoted by L θ 2 Λ k ( M ) L_\theta^2\Lambda^k(M) L θ 2 Λ k ( M ) . By (3.4), the weighted and unweighted L 2 L^2 L 2 -norms are equivalent.
This weighted inner product determines the corresponding adjoint of the exterior derivative.
Lemma 3.2 (weighted adjoint). The adjoint of d k d^k d k with respect to ( ⋅ , ⋅ ) θ (\cdot,\cdot)_\theta ( ⋅ , ⋅ ) θ is
δ θ k + 1 : = e θ δ k + 1 e − θ , D ( δ θ k + 1 ) = e θ D ( δ k + 1 ) . (31) \delta_\theta^{k+1}:=e^\theta\delta^{k+1}e^{-\theta},\qquad D(\delta_\theta^{k+1})=e^\theta D(\delta^{k+1}). \tag*{(31)} δ θ k + 1 := e θ δ k + 1 e − θ , D ( δ θ k + 1 ) = e θ D ( δ k + 1 ) . (31) Equivalently, δ θ = δ + i ∇ θ \delta_\theta=\delta+i\nabla\theta δ θ = δ + i ∇ θ .
Proof. Let ( d k ) θ ∗ (d^k)^*_\theta ( d k ) θ ∗ denote the Hilbert adjoint of d k d^k d k with respect to the weighted inner product. By definition, γ ∈ D ( ( d k ) θ ∗ ) \gamma\in D((d^k)^*_\theta) γ ∈ D (( d k ) θ ∗ ) if and only if there exists z ∈ L θ 2 Λ k ( M ) z\in L^2_\theta\Lambda^k(M) z ∈ L θ 2 Λ k ( M ) such that
( d k α , γ ) θ = ( α , z ) θ , ∀ α ∈ D ( d k ) . (32) (d^k\alpha,\gamma)_\theta=(\alpha,z)_\theta,\qquad\forall\alpha\in D(d^k). \tag*{(32)} ( d k α , γ ) θ = ( α , z ) θ , ∀ α ∈ D ( d k ) . (32)
This condition is equivalent to
( d k α , e − θ γ ) = ( α , e − θ z ) , ∀ α ∈ D ( d k ) . (33) (d^k\alpha,e^{-\theta}\gamma)=(\alpha,e^{-\theta}z),\qquad\forall\alpha\in D(d^k). \tag*{(33)} ( d k α , e − θ γ ) = ( α , e − θ z ) , ∀ α ∈ D ( d k ) . (33)
Since δ k + 1 = ( d k ) ∗ \delta^{k+1}=(d^k)^* δ k + 1 = ( d k ) ∗ in the standard L 2 L^2 L 2 inner product, the latter is equivalent to
e − θ γ ∈ D ( δ k + 1 ) , δ k + 1 ( e − θ γ ) = e − θ z . (34) e^{-\theta}\gamma\in D(\delta^{k+1}),\qquad\delta^{k+1}(e^{-\theta}\gamma)=e^{-\theta}z. \tag*{(34)} e − θ γ ∈ D ( δ k + 1 ) , δ k + 1 ( e − θ γ ) = e − θ z . (34)
Therefore,
γ ∈ e θ D ( δ k + 1 ) , z = e θ δ k + 1 ( e − θ γ ) . (35) \gamma\in e^\theta D(\delta^{k+1}),\qquad z=e^\theta\delta^{k+1}(e^{-\theta}\gamma). \tag*{(35)} γ ∈ e θ D ( δ k + 1 ) , z = e θ δ k + 1 ( e − θ γ ) . (35)
This proves both the action and the exact domain of the weighted Hilbert adjoint. Using the identity i X ω = ( − 1 ) k ( n − k ) ⋆ ( ⋆ ω ∧ X b ) = ( − 1 ) n ( k + 1 ) ⋆ ( X b ∧ ⋆ ω ) i_X\omega=(-1)^{k(n-k)}\star(\star\omega\wedge X^b)=(-1)^{n(k+1)}\star(X^b\wedge\star\omega) i X ω = ( − 1 ) k ( n − k ) ⋆ ( ⋆ ω ∧ X b ) = ( − 1 ) n ( k + 1 ) ⋆ ( X b ∧ ⋆ ω ) (see [29 Lemma 8.2.1] ), we have, for γ ∈ Λ k \gamma\in\Lambda^k γ ∈ Λ k :
δ ( e − θ γ ) = ( − 1 ) n ( k + 1 ) + 1 ⋆ d ( e − θ ⋆ γ ) = e − θ δ γ + ( − 1 ) n ( k + 1 ) + 1 ⋆ ( d e − θ ∧ ⋆ γ ) = e − θ δ γ + e − θ ( − 1 ) n ( k + 1 ) ⋆ ( d θ ∧ ⋆ γ ) = e − θ ( δ γ + i ∇ θ γ ) . \begin{aligned}
\delta(e^{-\theta}\gamma)&=(-1)^{n(k+1)+1}\star d(e^{-\theta}\star\gamma)\\
&=e^{-\theta}\delta\gamma+(-1)^{n(k+1)+1}\star(de^{-\theta}\wedge\star\gamma)\\
&=e^{-\theta}\delta\gamma+e^{-\theta}(-1)^{n(k+1)}\star(d\theta\wedge\star\gamma)\\
&=e^{-\theta}(\delta\gamma+i\nabla\theta\gamma).
\end{aligned} δ ( e − θ γ ) = ( − 1 ) n ( k + 1 ) + 1 ⋆ d ( e − θ ⋆ γ ) = e − θ δ γ + ( − 1 ) n ( k + 1 ) + 1 ⋆ ( d e − θ ∧ ⋆ γ ) = e − θ δ γ + e − θ ( − 1 ) n ( k + 1 ) ⋆ ( d θ ∧ ⋆ γ ) = e − θ ( δ γ + i ∇ θ γ ) .
Therefore,
δ θ : = e θ δ e − θ = δ + i ∇ θ . \delta_\theta:=e^\theta\delta e^{-\theta}=\delta+i\nabla\theta. δ θ := e θ δ e − θ = δ + i ∇ θ .
Definition 3.4 (weighted Laplacian and harmonic forms). The weighted Hodge Laplacian on k k k -forms is
Δ θ k : = d k − 1 δ θ k + δ θ k + 1 d k . (36) \Delta_\theta^k := d^{k-1}\delta_\theta^k+\delta_\theta^{k+1}d^k. \tag*{(36)} Δ θ k := d k − 1 δ θ k + δ θ k + 1 d k . (36) The space of weighted harmonic k k k -forms is
H θ k : = N ( Δ θ k ) . (37) \mathcal{H}_\theta^k := \mathcal{N}(\Delta_\theta^k). \tag*{(37)} H θ k := N ( Δ θ k ) . (37)
Because β ♭ = η d θ \beta^\flat=\eta d\theta β ♭ = η d θ , Lemma 3.2 implies
η Δ θ k = η Δ H L k + d k − 1 ι β k + ι β k + 1 d k = η Δ H L k + L β k . (38) \eta\Delta_\theta^k=\eta\Delta_{\mathrm{HL}}^k+d^{k-1}\iota_\beta^k+\iota_\beta^{k+1}d^k=\eta\Delta_{\mathrm{HL}}^k+L_\beta^k. \tag*{(38)} η Δ θ k = η Δ HL k + d k − 1 ι β k + ι β k + 1 d k = η Δ HL k + L β k . (38)
Thus the advection–diffusion operator with a potential-flow is exactly a weighted Hodge Laplacian multiplied by η \eta η .
Weighted Hodge decomposition
With these preparations, we now establish the weighted Hodge decomposition.
Lemma 3.5 (equivalence and orthogonality). A form ω ∈ H θ k \omega\in\mathcal{H}_\theta^k ω ∈ H θ k if and only if d ω = 0 d\omega=0 d ω = 0 and δ θ ω = 0 \delta_\theta\omega=0 δ θ ω = 0 . Moreover, R ( d k − 1 ) \mathcal{R}(d^{k-1}) R ( d k − 1 ) , R ( δ θ k + 1 ) \mathcal{R}(\delta_\theta^{k+1}) R ( δ θ k + 1 ) , and H θ k \mathcal{H}_\theta^k H θ k are mutually orthogonal with respect to ( ⋅ , ⋅ ) θ (\cdot,\cdot)_\theta ( ⋅ , ⋅ ) θ .
Proof. If ω ∈ H θ k \omega\in\mathcal{H}_\theta^k ω ∈ H θ k , then
0 = ( Δ θ k ω , ω ) θ = ∥ δ θ ω ∥ θ 2 + ∥ d ω ∥ θ 2 . 0=(\Delta_\theta^k\omega,\omega)_\theta=\|\delta_\theta\omega\|_\theta^2+\|d\omega\|_\theta^2. 0 = ( Δ θ k ω , ω ) θ = ∥ δ θ ω ∥ θ 2 + ∥ d ω ∥ θ 2 .
Hence d ω = 0 d\omega=0 d ω = 0 and δ θ ω = 0 \delta_\theta\omega=0 δ θ ω = 0 . The converse is immediate from the definition of Δ θ k \Delta_\theta^k Δ θ k . Orthogonality follows from the weighted adjoint relation and from d ∘ d = 0 d\circ d=0 d ∘ d = 0 : for example, ( d α , δ θ u ) θ = ( d ∘ d α , u ) θ = 0 (d\alpha,\delta_\theta u)_\theta=(d\circ d\alpha,u)_\theta=0 ( d α , δ θ u ) θ = ( d ∘ d α , u ) θ = 0 . The orthogonality with H θ k \mathcal{H}_\theta^k H θ k follows similarly.□
Remark 3.6 (implications of Witten deformation). Via the unitary transformation U ω = e − θ / 2 ω U\omega=e^{-\theta/2}\omega U ω = e − θ /2 ω , the weighted Hodge Laplacian Δ θ \Delta_\theta Δ θ satisfies
U Δ θ U − 1 = e − θ / 2 ( δ θ d + d δ θ ) e θ / 2 = ( e θ / 2 δ e − θ / 2 ) ( e − θ / 2 d e θ / 2 ) + ( e − θ / 2 d e θ / 2 ) ( e θ / 2 δ e − θ / 2 ) = δ θ / 2 d θ / 2 + d θ / 2 δ θ / 2 = ( δ θ / 2 + d θ / 2 ) 2 (39) \begin{aligned}
U\Delta_\theta U^{-1}&=e^{-\theta/2}(\delta_\theta d+d\delta_\theta)e^{\theta/2}\\
&=(e^{\theta/2}\delta e^{-\theta/2})(e^{-\theta/2}de^{\theta/2})+(e^{-\theta/2}de^{\theta/2})(e^{\theta/2}\delta e^{-\theta/2})\\
&=\delta_{\theta/2}d_{\theta/2}+d_{\theta/2}\delta_{\theta/2}\\
&=(\delta_{\theta/2}+d_{\theta/2})^2 \tag*{(39)}
\end{aligned} U Δ θ U − 1 = e − θ /2 ( δ θ d + d δ θ ) e θ /2 = ( e θ /2 δ e − θ /2 ) ( e − θ /2 d e θ /2 ) + ( e − θ /2 d e θ /2 ) ( e θ /2 δ e − θ /2 ) = δ θ /2 d θ /2 + d θ /2 δ θ /2 = ( δ θ /2 + d θ /2 ) 2 (39)
Hence, Δ θ \Delta_\theta Δ θ is unitarily equivalent to the Witten Laplacian ( d θ / 2 + δ θ / 2 ) 2 (d_{\theta/2}+\delta_{\theta/2})^2 ( d θ /2 + δ θ /2 ) 2 , which is self-adjoint and non-negative.
Remark 3.7 (closedness of the ranges). By the standard closed-range property of the de Rham complex, R ( d k ) \mathcal{R}(d^k) R ( d k ) and R ( δ k + 1 ) \mathcal{R}(\delta^{k+1}) R ( δ k + 1 ) are closed. Because the weighted codifferential satisfies D ( δ θ k + 1 ) = e θ D ( δ k + 1 ) D(\delta_\theta^{k+1})=e^\theta D(\delta^{k+1}) D ( δ θ k + 1 ) = e θ D ( δ k + 1 ) , it immediately follows that R ( δ θ k + 1 ) = e θ R ( δ k + 1 ) \mathcal{R}(\delta_\theta^{k+1})=e^\theta\mathcal{R}(\delta^{k+1}) R ( δ θ k + 1 ) = e θ R ( δ k + 1 ) . Since multiplication by e θ e^\theta e θ defines a bounded invertible operator on L 2 Λ L^2\Lambda L 2 Λ , it preserves closed subspaces. Thus, R ( δ θ k + 1 ) \mathcal{R}(\delta_\theta^{k+1}) R ( δ θ k + 1 ) is closed in L 2 Λ k L^2\Lambda^k L 2 Λ k . By the equivalence of the L 2 L^2 L 2 - and L θ 2 L_\theta^2 L θ 2 -norms, these ranges remain closed in the weighted space L θ 2 Λ k L_\theta^2\Lambda^k L θ 2 Λ k .
Theorem 3.8 (weighted Hodge decomposition). The weighted Hilbert space admits the weighted orthogonal decomposition
L θ 2 Λ k ( M ) = R ( d k − 1 ) ⊕ R ( δ θ k + 1 ) ⊕ H θ k . (40) L_\theta^2\Lambda^k(M)=\mathcal{R}(d^{k-1})\oplus\mathcal{R}(\delta_\theta^{k+1})\oplus\mathcal{H}_\theta^k. \tag*{(40)} L θ 2 Λ k ( M ) = R ( d k − 1 ) ⊕ R ( δ θ k + 1 ) ⊕ H θ k . (40)
Proof. We write ⊥ θ \perp_\theta ⊥ θ for the orthogonal complement with respect to the weighted inner product ( ⋅ , ⋅ ) θ (\cdot,\cdot)_\theta ( ⋅ , ⋅ ) θ . By Lemma 3.2, δ θ k + 1 \delta_\theta^{k+1} δ θ k + 1 is the L θ 2 L^2_\theta L θ 2 -adjoint of d k d^k d k . Hence, by the closed range theorem and the closedness of the ranges,
R ( δ θ k + 1 ) = N ( d k ) ⊥ θ . \mathcal{R}(\delta_\theta^{k+1})=\mathcal{N}(d^k)^{\perp_\theta}. R ( δ θ k + 1 ) = N ( d k ) ⊥ θ .
Therefore,
L θ 2 Λ k ( M ) = N ( d k ) ⊕ R ( δ θ k + 1 ) . (41) L^2_\theta\Lambda^k(M)=\mathcal{N}(d^k)\oplus\mathcal{R}(\delta_\theta^{k+1}). \tag*{(41)} L θ 2 Λ k ( M ) = N ( d k ) ⊕ R ( δ θ k + 1 ) . (41)
It remains to decompose the closed subspace N ( d k ) \mathcal{N}(d^k) N ( d k ) . Since
R ( d k − 1 ) ⊂ N ( d k ) \mathcal{R}(d^{k-1})\subset\mathcal{N}(d^k) R ( d k − 1 ) ⊂ N ( d k )
and R ( d k − 1 ) \mathcal{R}(d^{k-1}) R ( d k − 1 ) is closed, we have the orthogonal decomposition inside N ( d k ) \mathcal{N}(d^k) N ( d k )
N ( d k ) = R ( d k − 1 ) ⊕ ( N ( d k ) ∩ R ( d k − 1 ) ⊥ θ ) . \mathcal{N}(d^k)=\mathcal{R}(d^{k-1})\oplus\left(\mathcal{N}(d^k)\cap\mathcal{R}(d^{k-1})^{\perp_\theta}\right). N ( d k ) = R ( d k − 1 ) ⊕ ( N ( d k ) ∩ R ( d k − 1 ) ⊥ θ ) .
Again by Lemma 3.2, we have R ( d k − 1 ) ⊥ θ = N ( δ θ k ) \mathcal{R}(d^{k-1})^{\perp_\theta}=\mathcal{N}(\delta_\theta^k) R ( d k − 1 ) ⊥ θ = N ( δ θ k ) . Consequently,
N ( d k ) ∩ R ( d k − 1 ) ⊥ θ = N ( d k ) ∩ N ( δ θ k ) = H θ k . \mathcal{N}(d^k)\cap\mathcal{R}(d^{k-1})^{\perp_\theta}=\mathcal{N}(d^k)\cap\mathcal{N}(\delta_\theta^k)=\mathcal{H}_\theta^k. N ( d k ) ∩ R ( d k − 1 ) ⊥ θ = N ( d k ) ∩ N ( δ θ k ) = H θ k .
Substituting this decomposition of N ( d k ) \mathcal{N}(d^k) N ( d k ) into (41) gives
L θ 2 Λ k ( M ) = R ( d k − 1 ) ⊕ R ( δ θ k + 1 ) ⊕ H θ k . L^2_\theta\Lambda^k(M)=\mathcal{R}(d^{k-1})\oplus\mathcal{R}(\delta_\theta^{k+1})\oplus\mathcal{H}_\theta^k. L θ 2 Λ k ( M ) = R ( d k − 1 ) ⊕ R ( δ θ k + 1 ) ⊕ H θ k .
This proves the weighted Hodge decomposition.
Corollary 3.9 (cohomology isomorphism). The natural map
Φ : H θ k → H d R k ( M ) , Φ ( ω ) = [ ω ] , \Phi:\mathcal{H}_\theta^k\to H_{\mathrm{dR}}^k(M),\qquad\Phi(\omega)=[\omega], Φ : H θ k → H dR k ( M ) , Φ ( ω ) = [ ω ] , is an isomorphism. In particular, dim H θ k = b k \dim\mathcal{H}_\theta^k=b_k dim H θ k = b k .
Proof. Let [ ω ] ∈ H d R k ( M ) [\omega]\in H_{\mathrm{dR}}^k(M) [ ω ] ∈ H dR k ( M ) and choose a closed representative ω \omega ω . By Theorem 3.8,
ω = d α + δ θ γ + h θ , h θ ∈ H θ k . \omega=d\alpha+\delta_\theta\gamma+h_\theta,\qquad h_\theta\in\mathcal{H}_\theta^k. ω = d α + δ θ γ + h θ , h θ ∈ H θ k .
Since d ω = 0 d\omega=0 d ω = 0 , we have d δ θ γ = 0 d\delta_\theta\gamma=0 d δ θ γ = 0 . Therefore
∥ δ θ γ ∥ θ 2 = ( δ θ γ , δ θ γ ) θ = ( d δ θ γ , γ ) θ = 0 , \|\delta_\theta\gamma\|_\theta^2=(\delta_\theta\gamma,\delta_\theta\gamma)_\theta=(d\delta_\theta\gamma,\gamma)_\theta=0, ∥ δ θ γ ∥ θ 2 = ( δ θ γ , δ θ γ ) θ = ( d δ θ γ , γ ) θ = 0 ,
so δ θ γ = 0 \delta_\theta\gamma=0 δ θ γ = 0 . Hence ω − h θ = d α \omega-h_\theta=d\alpha ω − h θ = d α , and [ h θ ] = [ ω ] [h_\theta]=[\omega] [ h θ ] = [ ω ] . This proves existence.
If h θ , h θ ′ ∈ H θ k h_\theta,h'_\theta\in\mathcal{H}_\theta^k h θ , h θ ′ ∈ H θ k represent the same cohomology class, then h θ − h θ ′ = d α h_\theta-h'_\theta=d\alpha h θ − h θ ′ = d α . Since H θ k ⊥ θ R ( d ) \mathcal{H}_\theta^k\perp_\theta\mathcal{R}(d) H θ k ⊥ θ R ( d ) , this means d α = 0 d\alpha=0 d α = 0 . Thus h θ = h θ ′ h_\theta=h'_\theta h θ = h θ ′ , proving uniqueness and the isomorphism.
Remark 3.10. This weighted Hodge isomorphism was previously established in [36 ] , Eq. (2.13) for closed manifolds with a smooth positive measure, which allows one to identify weighted harmonic one-forms with parallel one-forms under nonnegative Bakry–Emery Ricci curvature.
Long-time behavior and the antidynamo mechanism
Beyond the cohomology isomorphism, the weighted formulation also yields the long-time behavior. Since Δ k \Delta_k Δ k is similar to a self-adjoint non-negative operator with compact resolvent, it also has compact resolvent, and its spectrum is real and non-negative. Decomposing the solution into its weighted harmonic and orthogonal components, the weighted harmonic component is time-invariant, while the remaining components dissipate. This gives the following corollary.
Corollary 3.11 (antidynamo for potential flows). Assume that β ♭ = η d θ \beta^\flat= \eta d\theta β ♭ = η d θ . Then the evolution
∂ t ω + ( η Δ H L + L β ) ω = 0 \partial_t\omega+ (\eta\Delta_{\mathrm{HL}} + L_\beta)\omega= 0 ∂ t ω + ( η Δ HL + L β ) ω = 0 does not exhibit exponential growth. Moreover, the solution converges to its weighted harmonic component as t → ∞ t \to\infty t → ∞ .
Discrete weighted Hodge theory
On the compact manifold M M M , we show that the weighted Hodge decomposition admits an analogue on the discrete complex ( V h ∙ , d ) (V_h^\bullet,d) ( V h ∙ , d ) for every mesh size h h h and introduce a weighted Galerkin discretization of the advection–diffusion equation (16) . In this case, we prove that this scheme satisfies the antidynamo mechanism without any restriction on the mesh.
Definition 3.13 (discrete weighted codifferential and harmonic forms). The discrete weighted codifferential δ θ , h k : V h k → V h k − 1 \delta_{\theta,h}^k:V_h^k\to V_h^{k-1} δ θ , h k : V h k → V h k − 1 is defined by
( δ θ , h k u h , τ h ) θ = ( u h , d k − 1 τ h ) θ , ∀ τ h ∈ V h k − 1 , (42) (\delta_{\theta,h}^k u_h,\tau_h)_\theta=(u_h,d^{k-1}\tau_h)_\theta,\qquad\forall\tau_h\in V_h^{k-1}, \tag*{(42)} ( δ θ , h k u h , τ h ) θ = ( u h , d k − 1 τ h ) θ , ∀ τ h ∈ V h k − 1 , (42) i.e., δ θ , h k \delta_{\theta,h}^k δ θ , h k is the adjoint of d k − 1 ∣ V h k − 1 d^{k-1}|_{V_h^{k-1}} d k − 1 ∣ V h k − 1 with respect to the weighted inner product (30) . The space of discrete weighted harmonic k k k -forms is
H ˚ θ , h k : = { v h ∈ V h k : d k v h = 0 , δ θ , h k v h = 0 } . (43) \mathring{\mathcal{H}}_{\theta,h}^k:=\{v_h\in V_h^k:d^k v_h=0,\ \delta_{\theta,h}^k v_h=0\}. \tag*{(43)} H ˚ θ , h k := { v h ∈ V h k : d k v h = 0 , δ θ , h k v h = 0 } . (43)
By (42) , the condition δ θ , h k v h = 0 \delta_{\theta,h}^k v_h=0 δ θ , h k v h = 0 is equivalent to ( v h , d k − 1 τ h ) θ = 0 (v_h,d^{k-1}\tau_h)_\theta=0 ( v h , d k − 1 τ h ) θ = 0 for all τ h ∈ V h k − 1 \tau_h\in V_h^{k-1} τ h ∈ V h k − 1 . Therefore
H ˚ θ , h k = Z h k ∩ ( B h k ) ⊥ θ . (44) \mathring{\mathcal{H}}_{\theta,h}^k=\mathcal{Z}_h^k\cap(\mathcal{B}_h^k)^{\perp_\theta}. \tag*{(44)} H ˚ θ , h k = Z h k ∩ ( B h k ) ⊥ θ . (44)
Since (29) guarantees that ( ⋅ , ⋅ ) θ (\cdot,\cdot)_\theta ( ⋅ , ⋅ ) θ is an inner product on each finite-dimensional space V h k V_h^k V h k , the operator δ θ , h k \delta_{\theta,h}^k δ θ , h k is well defined on V h k V_h^k V h k .
Theorem 3.14 (discrete weighted Hodge decomposition). For every h > 0 h>0 h > 0 and 0 ≤ k ≤ n 0\leq k\leq n 0 ≤ k ≤ n , the discrete space admits the weighted orthogonal decomposition
V h k = B h k ⊕ H ˚ θ , h k ⊕ δ θ , h k + 1 V h k + 1 , (45) V_h^k=\mathcal{B}_h^k\oplus\mathring{\mathcal{H}}_{\theta,h}^k\oplus\delta_{\theta,h}^{k+1}V_h^{k+1}, \tag*{(45)} V h k = B h k ⊕ H ˚ θ , h k ⊕ δ θ , h k + 1 V h k + 1 , (45) where the sums are orthogonal with respect to ( ⋅ , ⋅ ) θ (\cdot,\cdot)_\theta ( ⋅ , ⋅ ) θ . Moreover, the natural map
Φ h : H ˚ θ , h k ⟶ H h k , Φ h ( v h ) = [ v h ] , \Phi_h:\mathring{\mathcal{H}}_{\theta,h}^k\longrightarrow H_h^k,\qquad\Phi_h(v_h)=[v_h], Φ h : H ˚ θ , h k ⟶ H h k , Φ h ( v h ) = [ v h ] , is an isomorphism. In particular,
dim H ˚ θ , h k = dim H h k = b k . (46) \dim\mathring{\mathcal{H}}_{\theta,h}^k=\dim H_h^k=b_k. \tag*{(46)} dim H ˚ θ , h k = dim H h k = b k . (46)
Proof. Since δ θ , h k + 1 \delta_{\theta,h}^{k+1} δ θ , h k + 1 is the ( ⋅ , ⋅ ) θ (\cdot,\cdot)_{\theta} ( ⋅ , ⋅ ) θ -adjoint of the linear map d k : V h k → V h k + 1 d^k:V_h^k\to V_h^{k+1} d k : V h k → V h k + 1 between finite-dimensional spaces, the standard duality gives
R ( δ θ , h k + 1 ) = ( N ( d k ) ∩ V h k ) ⊥ θ = ( Z h k ) ⊥ θ , \mathcal{R}(\delta_{\theta,h}^{k+1})=\left(\mathcal{N}(d^k)\cap V_h^k\right)^{\perp_\theta}=(\mathcal{Z}_h^k)^{\perp_\theta}, R ( δ θ , h k + 1 ) = ( N ( d k ) ∩ V h k ) ⊥ θ = ( Z h k ) ⊥ θ ,
whence
V h k = Z h k ⊕ θ R ( δ θ , h k + 1 ) . (47) V_h^k=\mathcal{Z}_h^k\oplus_\theta\mathcal{R}(\delta_{\theta,h}^{k+1}). \tag*{(47)} V h k = Z h k ⊕ θ R ( δ θ , h k + 1 ) . (47)
For Z h k \mathcal{Z}_h^k Z h k , we split θ \theta θ -orthogonally along the subspace B h k ⊂ Z h k \mathcal{B}_h^k\subset\mathcal{Z}_h^k B h k ⊂ Z h k and use (44) :
Z h k = B h k ⊕ ( Z h k ∩ ( B h k ) ⊥ θ ) = B h k ⊕ H θ , h k . (48) \mathcal{Z}_h^k=\mathcal{B}_h^k\oplus\left(\mathcal{Z}_h^k\cap(\mathcal{B}_h^k)^{\perp_\theta}\right)=\mathcal{B}_h^k\oplus\mathcal{H}_{\theta,h}^k. \tag*{(48)} Z h k = B h k ⊕ ( Z h k ∩ ( B h k ) ⊥ θ ) = B h k ⊕ H θ , h k . (48)
Substituting (48) into (47) proves (45) .
For the dimension count, (48) yields
dim H θ , h k = dim Z h k − dim B h k = dim H h k = b k . \dim\mathcal{H}_{\theta,h}^k=\dim\mathcal{Z}_h^k-\dim\mathcal{B}_h^k=\dim H_h^k=b_k. dim H θ , h k = dim Z h k − dim B h k = dim H h k = b k .
The map Φ h \Phi_h Φ h is well defined because every element of H θ , h k \mathcal{H}_{\theta,h}^k H θ , h k is closed, and it is injective because
N ( Φ h ) = H θ , h k ∩ B h k = { 0 } \mathcal{N}(\Phi_h)=\mathcal{H}_{\theta,h}^k\cap\mathcal{B}_h^k=\{0\} N ( Φ h ) = H θ , h k ∩ B h k = { 0 }
by the θ \theta θ -orthogonality. Injectivity together with the equality of dimensions implies that Φ h \Phi_h Φ h is an isomorphism. □
Remark 3.15 (boundary conditions). If M M M has a boundary and the complex with boundary conditions is used, V ˚ h k = V h k ∩ H ˚ Λ k ( M ) \mathring{V}_h^k=V_h^k\cap\mathring{H}\Lambda^k(M) V ˚ h k = V h k ∩ H ˚ Λ k ( M ) , then the statement and proof of Theorem 3.14 are similar with the relative discrete cohomology in place of H h k H_h^k H h k . In particular, the discrete weighted harmonic space is
H ˚ θ , h k : = { v h ∈ V ˚ h k : d k v h = 0 , δ ˚ θ , h k v h = 0 } = Z ˚ h k ∩ ( B ˚ h k ) ⊥ θ . \mathring{\mathcal{H}}_{\theta,h}^k:=\{v_h\in\mathring{V}_h^k:d^k v_h=0,\ \mathring{\delta}_{\theta,h}^k v_h=0\}=\mathring{\mathcal{Z}}_h^k\cap(\mathring{\mathcal{B}}_h^k)^{\perp_\theta}. H ˚ θ , h k := { v h ∈ V ˚ h k : d k v h = 0 , δ ˚ θ , h k v h = 0 } = Z ˚ h k ∩ ( B ˚ h k ) ⊥ θ .
with δ ˚ θ , h k : V ˚ h k → V ˚ h k − 1 \mathring{\delta}_{\theta,h}^k:\mathring{V}_h^k\to\mathring{V}_h^{k-1} δ ˚ θ , h k : V ˚ h k → V ˚ h k − 1
( δ ˚ θ , h k u h , τ h ) θ = ( u h , d k − 1 τ h ) θ , ∀ τ h ∈ V ˚ h k − 1 . (\mathring{\delta}_{\theta,h}^k u_h,\tau_h)_{\theta}=(u_h,d^{k-1}\tau_h)_{\theta},\qquad\forall\ \tau_h\in\mathring{V}_h^{k-1}. ( δ ˚ θ , h k u h , τ h ) θ = ( u h , d k − 1 τ h ) θ , ∀ τ h ∈ V ˚ h k − 1 .
We can obtain dim H ˚ θ , h k = dim Z ˚ h k − dim B ˚ h k = b k ( M , ∂ M ) = b n − k ( M ) \dim\mathring{\mathcal{H}}_{\theta,h}^k=\dim\mathring{\mathcal{Z}}_h^k-\dim\mathring{\mathcal{B}}_h^k=b_k(M,\partial M)=b_{n-k}(M) dim H ˚ θ , h k = dim Z ˚ h k − dim B ˚ h k = b k ( M , ∂ M ) = b n − k ( M ) .
A weighted Galerkin discretization and its long-time behavior
By Lemma 3.2, the weighted codifferential δ θ \delta_\theta δ θ is the adjoint of d d d with respect to the weighted inner product ( ⋅ , ⋅ ) θ (\cdot,\cdot)_\theta ( ⋅ , ⋅ ) θ , and by (38) the advection–diffusion operator for potential flows factorizes as
η Δ θ k = η ( d k − 1 δ θ k + δ θ k + 1 d k ) = η Δ H L k + L β k . \eta\Delta_\theta^k=\eta(d^{k-1}\delta_\theta^k+\delta_\theta^{k+1}d^k)=\eta\Delta_{\mathrm{HL}}^k+L_\beta^k. η Δ θ k = η ( d k − 1 δ θ k + δ θ k + 1 d k ) = η Δ HL k + L β k .
This suggests discretizing (16) by a Galerkin method posed in the weighted inner product, in which no contraction operator is required: find ( σ h , u h ) ∈ V h k − 1 × V h k (\sigma_h,u_h)\in V_h^{k-1}\times V_h^k ( σ h , u h ) ∈ V h k − 1 × V h k such that, for t > 0 t>0 t > 0 ,
( σ h , τ h ) θ − η ( u h , d k − 1 τ h ) θ = 0 , ∀ τ h ∈ V h k − 1 , (\sigma_h,\tau_h)_\theta-\eta(u_h,d^{k-1}\tau_h)_\theta=0,\qquad\forall\ \tau_h\in V_h^{k-1}, ( σ h , τ h ) θ − η ( u h , d k − 1 τ h ) θ = 0 , ∀ τ h ∈ V h k − 1 ,
( ∂ t u h , v h ) θ + ( d k − 1 σ h , v h ) θ + η ( d k u h , d k v h ) θ = 0 , ∀ v h ∈ V h k . (49) (\partial_tu_h,v_h)_\theta+(d^{k-1}\sigma_h,v_h)_\theta+\eta(d^ku_h,d^kv_h)_\theta=0,\qquad\forall\ v_h\in V_h^k. \tag*{(49)} ( ∂ t u h , v h ) θ + ( d k − 1 σ h , v h ) θ + η ( d k u h , d k v h ) θ = 0 , ∀ v h ∈ V h k . (49)
The first equation of (49) states that σ h = η δ θ , h k u h \sigma_h=\eta\delta_{\theta,h}^ku_h σ h = η δ θ , h k u h . Eliminating σ h \sigma_h σ h yields the finite-dimensional evolution
∂ t u h + A θ , h k u h = 0 , A θ , h k : = η Δ θ , h k : = η ( d k − 1 δ θ , h k + δ θ , h k + 1 d k ) . (50) \partial_tu_h+A_{\theta,h}^ku_h=0,\qquad A_{\theta,h}^k:=\eta\Delta_{\theta,h}^k:=\eta(d^{k-1}\delta_{\theta,h}^k+\delta_{\theta,h}^{k+1}d^k). \tag*{(50)} ∂ t u h + A θ , h k u h = 0 , A θ , h k := η Δ θ , h k := η ( d k − 1 δ θ , h k + δ θ , h k + 1 d k ) . (50)
Since d k + 1 d k = 0 d^{k+1}d^k=0 d k + 1 d k = 0 , we have the following commutativity
d k A θ , h k = η d k δ θ , h k + 1 d k = η ( d k δ θ , h k + 1 + δ θ , h k + 2 d k + 1 ) d k = A θ , h k + 1 d k . d^kA_{\theta,h}^k=\eta d^k\delta_{\theta,h}^{k+1}d^k=\eta\left(d^k\delta_{\theta,h}^{k+1}+\delta_{\theta,h}^{k+2}d^{k+1}\right)d^k=A_{\theta,h}^{k+1}d^k. d k A θ , h k = η d k δ θ , h k + 1 d k = η ( d k δ θ , h k + 1 + δ θ , h k + 2 d k + 1 ) d k = A θ , h k + 1 d k .
The following theorem is the discrete counterpart of the antidynamo mechanism for gradient advection fields established in § 3.2 for any h , η > 0 h,\eta>0 h , η > 0 .
Theorem 3.16 (finite element antidynamo theorem for potential flows). Let M M M be a connected, compact, oriented manifold, and let β / η = ( d θ ) ♯ \beta/\eta=(d\theta)^{\sharp} β / η = ( d θ ) ♯ be a potential advection field. Then, for any h , η > 0 h,\eta>0 h , η > 0 and every 0 ≤ k ≤ n 0\leq k\leq n 0 ≤ k ≤ n , the evolution (50) has the following properties.
(no discrete dynamo modes) The operator A θ , h k A_{\theta,h}^k A θ , h k is self-adjoint and positive semidefinite on ( V h k , ( ⋅ , ⋅ ) θ ) (V_h^k,(\cdot,\cdot)_\theta) ( V h k , ( ⋅ , ⋅ ) θ ) . Consequently σ ( A θ , h k ) ⊂ [ 0 , ∞ ) \sigma(A_{\theta,h}^k)\subset[0,\infty) σ ( A θ , h k ) ⊂ [ 0 , ∞ ) , the map t ↦ ∥ u h ( t ) ∥ θ t\mapsto\lVert u_h(t)\rVert_\theta t ↦ ∥ u h ( t ) ∥ θ is non-increasing, and
∥ e − t A θ , h k ∥ L ( V h k , ∥ ⋅ ∥ θ ) ≤ 1 , t ≥ 0. \left\lVert e^{-tA_{\theta,h}^k}\right\rVert_{\mathcal{L}(V_h^k,\lVert\cdot\rVert_\theta)}\leq1,\qquad t\geq0. e − t A θ , h k L ( V h k , ∥ ⋅ ∥ θ ) ≤ 1 , t ≥ 0. In particular, no solution of (50) exhibits exponential (or any) growth in the weighted norm.
(steady states) Every stationary solution is closed, and
N ( A θ , h k ) = H θ , h k ≅ H h k , dim N ( A θ , h k ) = b k . \mathcal{N}(A_{\theta,h}^k)=\mathcal{H}_{\theta,h}^k\cong H_h^k,\qquad\dim\mathcal{N}(A_{\theta,h}^k)=b_k. N ( A θ , h k ) = H θ , h k ≅ H h k , dim N ( A θ , h k ) = b k . (exponential convergence) Let P θ , h k P_{\theta,h}^k P θ , h k denote the θ \theta θ -orthogonal projection onto H θ , h k \mathcal{H}_{\theta,h}^k H θ , h k . Then, for every initial state u h ( 0 ) u_h(0) u h ( 0 ) ,
∥ u h ( t ) − P θ , h k u h ( 0 ) ∥ θ ≤ e − λ 1 , h t ∥ u h ( 0 ) − P θ , h k u h ( 0 ) ∥ θ , t ≥ 0. (51) \lVert u_h(t)-P_{\theta,h}^ku_h(0)\rVert_\theta\leq e^{-\lambda_{1,h}t}\lVert u_h(0)-P_{\theta,h}^ku_h(0)\rVert_\theta,\qquad t\geq0. \tag*{(51)} ∥ u h ( t ) − P θ , h k u h ( 0 ) ∥ θ ≤ e − λ 1 , h t ∥ u h ( 0 ) − P θ , h k u h ( 0 ) ∥ θ , t ≥ 0. (51) where λ 1 , h > 0 \lambda_{1,h}>0 λ 1 , h > 0 is the smallest nonzero eigenvalue of A θ , h k A_{\theta,h}^k A θ , h k . Therefore u h ( t ) u_h(t) u h ( t ) converges to P θ , h k u h ( 0 ) P_{\theta,h}^ku_h(0) P θ , h k u h ( 0 ) exponentially. In particular, for k = n k=n k = n , the total mass ∫ M u h ( t ) \int_Mu_h(t) ∫ M u h ( t ) is conserved if the manifold is closed, and hence the limit is the unique element of the one-dimensional space H θ , h n \mathcal{H}_{\theta,h}^n H θ , h n with total mass ∫ M u h ( 0 ) \int_Mu_h(0) ∫ M u h ( 0 ) .
Proof. By (42) , for all u h , v h ∈ V h k u_h,v_h\in V_h^k u h , v h ∈ V h k ,
( A θ , h k u h , v h ) θ = η [ ( δ θ , h k u h , δ θ , h k v h ) θ + ( d k u h , d k v h ) θ ] . (52) (A_{\theta,h}^ku_h,v_h)_\theta=\eta\left[(\delta_{\theta,h}^ku_h,\delta_{\theta,h}^kv_h)_\theta+(d^ku_h,d^kv_h)_\theta\right]. \tag*{(52)} ( A θ , h k u h , v h ) θ = η [ ( δ θ , h k u h , δ θ , h k v h ) θ + ( d k u h , d k v h ) θ ] . (52)
so A θ , h k A_{\theta,h}^k A θ , h k is self-adjoint and positive semidefinite on ( V h k , ( ⋅ , ⋅ ) θ ) (V_h^k,(\cdot,\cdot)_\theta) ( V h k , ( ⋅ , ⋅ ) θ ) . Taking v h = u h v_h=u_h v h = u h in (52) shows that A θ , h k u h = 0 A_{\theta,h}^ku_h=0 A θ , h k u h = 0 if and only if d k u h = 0 d^ku_h=0 d k u h = 0 and δ θ , h k u h = 0 \delta_{\theta,h}^ku_h=0 δ θ , h k u h = 0 , i.e.,
N ( A θ , h k ) = H θ , h k . \mathcal{N}(A_{\theta,h}^k)=\mathcal{H}_{\theta,h}^k. N ( A θ , h k ) = H θ , h k .
by (43) . In particular, every stationary solution is closed, so N ( A θ , h k ) ∩ Z h k = N ( A θ , h k ) \mathcal{N}(A_{\theta,h}^k)\cap\mathcal{Z}_h^k=\mathcal{N}(A_{\theta,h}^k) N ( A θ , h k ) ∩ Z h k = N ( A θ , h k ) . The isomorphism H θ , h k ≅ H h k \mathcal{H}_{\theta,h}^k\cong H_h^k H θ , h k ≅ H h k and the dimension count dim H θ , h k = b k \dim\mathcal{H}_{\theta,h}^k=b_k dim H θ , h k = b k are the content of Theorem 3.14. This proves item 2.
Since A θ , h k A_{\theta,h}^k A θ , h k is self-adjoint, it admits a θ \theta θ -orthonormal eigenbasis { ψ j } j = 1 N \{\psi_j\}_{j=1}^N { ψ j } j = 1 N with eigenvalues 0 ≤ μ 1 ≤ μ 2 ≤ ⋯ ≤ μ N 0\leq\mu_1\leq\mu_2\leq\cdots\leq\mu_N 0 ≤ μ 1 ≤ μ 2 ≤ ⋯ ≤ μ N , where the zero eigenspace is H θ , h k \mathcal{H}_{\theta,h}^k H θ , h k and λ 1 , h : = min { μ j : μ j > 0 } \lambda_{1,h}:=\min\{\mu_j:\mu_j>0\} λ 1 , h := min { μ j : μ j > 0 } . The semigroup admits the spectral representation
u h ( t ) = e − t A θ , h k u h ( 0 ) = ∑ j = 1 N e − μ j t ( u h ( 0 ) , ψ j ) θ ψ j . (53) u_h(t)=e^{-tA_{\theta,h}^k}u_h(0)=\sum_{j=1}^{N}e^{-\mu_jt}(u_h(0),\psi_j)_\theta\psi_j. \tag*{(53)} u h ( t ) = e − t A θ , h k u h ( 0 ) = j = 1 ∑ N e − μ j t ( u h ( 0 ) , ψ j ) θ ψ j . (53)
Since e − μ j t ≤ 1 e^{-\mu_jt}\le1 e − μ j t ≤ 1 for all j j j and t ≥ 0 t\ge0 t ≥ 0 , we obtain ∥ e − t A θ , h k v h ∥ L ( V h k , ∥ ⋅ ∥ θ ) ≤ 1 \lVert e^{-tA_{\theta,h}^k}v_h\rVert_{L(V_h^k,\lVert\cdot\rVert_\theta)}\le1 ∥ e − t A θ , h k v h ∥ L ( V h k , ∥ ⋅ ∥ θ ) ≤ 1 , and consequently t ↦ ∥ u h ( t ) ∥ θ t\mapsto\lVert u_h(t)\rVert_\theta t ↦ ∥ u h ( t ) ∥ θ is non-increasing. This proves item 1. Moreover, P θ , h k u h ( 0 ) P_{\theta,h}^k u_h(0) P θ , h k u h ( 0 ) collects exactly the terms in (53) with μ j = 0 \mu_j=0 μ j = 0 , which are invariant in time. The remaining terms carry the factors e − μ j t ≤ e − λ 1 , h t e^{-\mu_jt}\le e^{-\lambda_{1,h}t} e − μ j t ≤ e − λ 1 , h t , whence
∥ u h ( t ) − P θ , h k u h ( 0 ) ∥ θ 2 = ∑ μ j > 0 e − 2 μ j t ∣ ( u h ( 0 ) , ψ j ) θ ∣ 2 ≤ e − 2 λ 1 , h t ∥ u h ( 0 ) − P θ , h k u h ( 0 ) ∥ θ 2 . \lVert u_h(t)-P_{\theta,h}^k u_h(0)\rVert_\theta^2=\sum_{\mu_j>0}e^{-2\mu_jt}\lvert(u_h(0),\psi_j)_\theta\rvert^2\le e^{-2\lambda_{1,h}t}\lVert u_h(0)-P_{\theta,h}^k u_h(0)\rVert_\theta^2. ∥ u h ( t ) − P θ , h k u h ( 0 ) ∥ θ 2 = μ j > 0 ∑ e − 2 μ j t ∣( u h ( 0 ) , ψ j ) θ ∣ 2 ≤ e − 2 λ 1 , h t ∥ u h ( 0 ) − P θ , h k u h ( 0 ) ∥ θ 2 .
This proves (51) and the exponential convergence in item 3.
Since d n u h = 0 d^n u_h=0 d n u h = 0 , the second equation of (49) reads ( ∂ t u h + d n − 1 σ h , v h ) θ = 0 (\partial_tu_h+d^{n-1}\sigma_h,v_h)_\theta=0 ( ∂ t u h + d n − 1 σ h , v h ) θ = 0 for all v h ∈ V h n v_h\in V_h^n v h ∈ V h n . As ∂ t u h + d n − 1 σ h ∈ V h n \partial_tu_h+d^{n-1}\sigma_h\in V_h^n ∂ t u h + d n − 1 σ h ∈ V h n , it vanishes identically, i.e., ∂ t u h = − d n − 1 σ h \partial_tu_h=-d^{n-1}\sigma_h ∂ t u h = − d n − 1 σ h . Hence, by Stokes’ theorem on the closed manifold,
d d t ∫ M u h = − ∫ M d n − 1 σ h = 0. \frac{d}{dt}\int_Mu_h=-\int_Md^{n-1}\sigma_h=0. d t d ∫ M u h = − ∫ M d n − 1 σ h = 0.
so the total mass is conserved.
It remains to identify the limit. Since d n = 0 d^n=0 d n = 0 , Theorem 3.14 reduces to
V h n = B h n ⊕ θ H θ , h n , dim H θ , h n = b n = 1 , V_h^n=\mathfrak{B}_h^n\oplus_\theta\mathfrak{H}_{\theta,h}^n,\qquad\dim\mathfrak{H}_{\theta,h}^n=b_n=1, V h n = B h n ⊕ θ H θ , h n , dim H θ , h n = b n = 1 ,
where b n = 1 b_n=1 b n = 1 because M M M is closed and oriented. By Stokes’ theorem and the natural isomorphism H h n ≅ H d R n ( M ) H_h^n\cong H_{\mathrm{dR}}^n(M) H h n ≅ H dR n ( M ) , integration induces an isomorphism
∫ M : H θ , h n ⟶ R . \int_M:\mathfrak{H}_{\theta,h}^n\longrightarrow\mathbb{R}. ∫ M : H θ , h n ⟶ R .
Hence, by (51) and mass conservation, u h ( t ) u_h(t) u h ( t ) converges to the unique element of H θ , h n \mathfrak{H}_{\theta,h}^n H θ , h n with mass ∫ M u h ( 0 ) \int_Mu_h(0) ∫ M u h ( 0 ) . This completes the proof. □
Remark 3.17. Since A θ , h k A_{\theta,h}^k A θ , h k is self-adjoint in ( ⋅ , ⋅ ) θ (\cdot,\cdot)_\theta ( ⋅ , ⋅ ) θ , the θ \theta θ -orthogonal projection P θ , h k P_{\theta,h}^k P θ , h k coincides with the spectral projection onto N ( A θ , h k ) \mathcal{N}(A_{\theta,h}^k) N ( A θ , h k ) . This is special to the potential-flow setting: for general advection fields the discrete operator is non-normal in L 2 L^2 L 2 , and the long-time limit is given by a spectral projection instead; see Remark 4.7 and Section 5.
Discretization for general advection field
In this section, we discretize the generalized Hodge Laplacian Δ η , β k \Delta_{\eta,\beta}^k Δ η , β k by a FEEC-compatible mixed formulation for general β \beta β . Let A k : = Δ η , β k A^k:=\Delta_{\eta,\beta}^k A k := Δ η , β k , so that the continuous evolution is ∂ t ω + A k ω = 0 \partial_t\omega+A^k\omega=0 ∂ t ω + A k ω = 0 . The only nonstandard term is the contraction i β i_\beta i β , which does not generally map V h k V_h^k V h k into V h k − 1 V_h^{k-1} V h k − 1 . We therefore choose a family of bounded operators
i β , h k : V h k ⟶ V h k − 1 , k = 1 , … , n , i_{\beta,h}^k:V_h^k\longrightarrow V_h^{k-1},\qquad k=1,\ldots,n, i β , h k : V h k ⟶ V h k − 1 , k = 1 , … , n ,
whose concrete realization will be discussed in Section 6.
Introducing the auxiliary ( k − 1 ) (k-1) ( k − 1 ) -form
σ h = η δ h k u h + i β , h k u h , \sigma_h=\eta\delta_h^ku_h+i_{\beta,h}^ku_h, σ h = η δ h k u h + i β , h k u h ,
we define the semidiscrete problem as follows: find ( σ h , u h ) ∈ V h k − 1 × V h k (\sigma_h,u_h)\in V_h^{k-1}\times V_h^k ( σ h , u h ) ∈ V h k − 1 × V h k such that, for t > 0 t>0 t > 0 ,
( σ h , τ h ) − ( i β , h k u h , τ h ) − η ( u h , d h k − 1 τ h ) = 0 , ∀ τ h ∈ V h k − 1 , (54) (\sigma_h,\tau_h)-(i_{\beta,h}^ku_h,\tau_h)-\eta(u_h,d_h^{k-1}\tau_h)=0,\qquad\forall\tau_h\in V_h^{k-1}, \tag*{(54)} ( σ h , τ h ) − ( i β , h k u h , τ h ) − η ( u h , d h k − 1 τ h ) = 0 , ∀ τ h ∈ V h k − 1 , (54)
( ∂ t u h , v h ) + ( d h k − 1 σ h , v h ) + η ( d h k u h , d h k v h ) + ( i β , h k + 1 d h k u h , v h ) = 0 , ∀ v h ∈ V h k . (\partial_tu_h,v_h)+(d_h^{k-1}\sigma_h,v_h)+\eta(d_h^ku_h,d_h^kv_h)+(i_{\beta,h}^{k+1}d_h^ku_h,v_h)=0,\qquad\forall v_h\in V_h^k. ( ∂ t u h , v h ) + ( d h k − 1 σ h , v h ) + η ( d h k u h , d h k v h ) + ( i β , h k + 1 d h k u h , v h ) = 0 , ∀ v h ∈ V h k .
The initial value is u h ( 0 ) = ω h ( 0 ) u_h(0)=\omega_h(0) u h ( 0 ) = ω h ( 0 ) , for example u h ( 0 ) = π h k ω ( 0 ) u_h(0)=\pi_h^k\omega(0) u h ( 0 ) = π h k ω ( 0 ) , where π h k \pi_h^k π h k is a cochain projection.
Definition 4.1 (discrete stationary pair and closed stationary form). A pair ( σ h , u h ) ∈ V h k − 1 × V h k (\sigma_h,u_h)\in V_h^{k-1}\times V_h^k ( σ h , u h ) ∈ V h k − 1 × V h k is called a discrete stationary pair if it satisfies
( σ h , τ h ) − ( i β , h k u h , τ h ) − η ( u h , d h k − 1 τ h ) = 0 , ∀ τ h ∈ V h k − 1 , ( d h k − 1 σ h , v h ) + η ( d h k u h , d h k v h ) + ( i β , h k + 1 d h k u h , v h ) = 0 , ∀ v h ∈ V h k . (55) \begin{aligned}
(\sigma_h,\tau_h)-(i_{\beta,h}^k u_h,\tau_h)-\eta(u_h,d_h^{k-1}\tau_h)&=0,\quad\forall\tau_h\in V_h^{k-1},\\
(d_h^{k-1}\sigma_h,v_h)+\eta(d_h^k u_h,d_h^k v_h)+(i_{\beta,h}^{k+1}d_h^k u_h,v_h)&=0,\quad\forall v_h\in V_h^k. \tag*{(55)}
\end{aligned} ( σ h , τ h ) − ( i β , h k u h , τ h ) − η ( u h , d h k − 1 τ h ) ( d h k − 1 σ h , v h ) + η ( d h k u h , d h k v h ) + ( i β , h k + 1 d h k u h , v h ) = 0 , ∀ τ h ∈ V h k − 1 , = 0 , ∀ v h ∈ V h k . (55) The form u h u_h u h is called a closed discrete stationary k k k -form if, in addition, d h k u h = 0 d_h^k u_h=0 d h k u h = 0 .
The first equation of (54) is the weak identity
σ h = i β , h k u h + η δ h k u h . \sigma_h=i_{\beta,h}^k u_h+\eta\delta_h^k u_h. σ h = i β , h k u h + η δ h k u h .
Substitution into the second equation gives the operator form
∂ t u h + A h k u h = 0 , (56) \partial_tu_h+A_h^k u_h=0, \tag*{(56)} ∂ t u h + A h k u h = 0 , (56)
where
A h k u h : = d h k − 1 σ h + ( η δ h k + 1 + i β , h k + 1 ) d h k u h = d h k − 1 ( i β , h k u h + η δ h k u h ) + ( η δ h k + 1 + i β , h k + 1 ) d h k u h . (57) \begin{aligned}
A_h^k u_h&:=d_h^{k-1}\sigma_h+(\eta\delta_h^{k+1}+i_{\beta,h}^{k+1})d_h^k u_h\\
&=d_h^{k-1}(i_{\beta,h}^k u_h+\eta\delta_h^k u_h)+(\eta\delta_h^{k+1}+i_{\beta,h}^{k+1})d_h^k u_h. \tag*{(57)}
\end{aligned} A h k u h := d h k − 1 σ h + ( η δ h k + 1 + i β , h k + 1 ) d h k u h = d h k − 1 ( i β , h k u h + η δ h k u h ) + ( η δ h k + 1 + i β , h k + 1 ) d h k u h . (57)
Thus
A h k = η Δ h k + L β , h k , Δ h k = d h k − 1 δ h k + δ h k + 1 d h k , A_h^k=\eta\Delta_h^k+L_{\beta,h}^k,\qquad\Delta_h^k=d_h^{k-1}\delta_h^k+\delta_h^{k+1}d_h^k, A h k = η Δ h k + L β , h k , Δ h k = d h k − 1 δ h k + δ h k + 1 d h k ,
with the discrete Lie advection operator
L β , h k : = d h k − 1 i β , h k + i β , h k + 1 d h k . (58) L_{\beta,h}^k:=d_h^{k-1}i_{\beta,h}^k+i_{\beta,h}^{k+1}d_h^k. \tag*{(58)} L β , h k := d h k − 1 i β , h k + i β , h k + 1 d h k . (58)
We denote the closed stationary space by
S h k : = N ( A h k ) ∩ Z h k . (59) S_h^k:=\mathcal{N}(A_h^k)\cap Z_h^k. \tag*{(59)} S h k := N ( A h k ) ∩ Z h k . (59)
Because d h k + 1 d h k = 0 d_h^{k+1}d_h^k=0 d h k + 1 d h k = 0 , (58) gives the cochain commutation relation
d h k L β , h k = d h k ( d h k − 1 i β , h k + i β , h k + 1 d h k ) = d h k i β , h k + 1 d h k = ( d h k i β , h k + 1 + i β , h k + 2 d h k + 1 ) d h k = L β , h k + 1 d h k . (60) \begin{aligned}
d_h^kL_{\beta,h}^k&=d_h^k(d_h^{k-1}i_{\beta,h}^k+i_{\beta,h}^{k+1}d_h^k)\\
&=d_h^ki_{\beta,h}^{k+1}d_h^k=(d_h^ki_{\beta,h}^{k+1}+i_{\beta,h}^{k+2}d_h^{k+1})d_h^k=L_{\beta,h}^{k+1}d_h^k. \tag*{(60)}
\end{aligned} d h k L β , h k = d h k ( d h k − 1 i β , h k + i β , h k + 1 d h k ) = d h k i β , h k + 1 d h k = ( d h k i β , h k + 1 + i β , h k + 2 d h k + 1 ) d h k = L β , h k + 1 d h k . (60)
The discrete Hodge Laplacian also commutes with d d d , hence
d h k A h k = A h k + 1 d h k . (61) d_h^kA_h^k=A_h^{k+1}d_h^k. \tag*{(61)} d h k A h k = A h k + 1 d h k . (61)
Remark 4.2. The commutation in (60) requires one compatible family { i β , h k } k = 1 n \{i_{\beta,h}^k\}_{k=1}^n { i β , h k } k = 1 n . In particular, the same algebraic contraction must be used whenever the same form degree appears in adjacent equations.
In [7 ] , a lower bound for the dimension of the space of closed stationary solutions was established at the continuous level using the index of Δ η , β \Delta_{\eta,\beta} Δ η , β . We now prove the corresponding result for the FEEC discretization. Specifically, we show that the space of closed discrete stationary k k k -forms has dimension at least the k k k -th Betti number. This result is not limited to closed manifolds, but also applies to periodic domains and, more generally, to the setting of the finite element de Rham complex without boundary conditions.
Theorem 4.3. The closed discrete stationary space satisfies
dim S h k ≥ b k . \dim S_h^k \ge b_k. dim S h k ≥ b k . Proof. Restrict A h k A_h^k A h k to Z ˚ h k \mathring{Z}_h^k Z ˚ h k . If ω h ∈ Z ˚ h k \omega_h \in\mathring{Z}_h^k ω h ∈ Z ˚ h k , then d k ω h = 0 d^k\omega_h=0 d k ω h = 0 , and (57) gives
A h k ω h = d k − 1 ( i β , h k ω h + η δ h k ω h ) ∈ B h k . A_h^k\omega_h=d^{k-1}\left(i_{\beta,h}^k\omega_h+\eta\delta_h^k\omega_h\right)\in B_h^k. A h k ω h = d k − 1 ( i β , h k ω h + η δ h k ω h ) ∈ B h k . Therefore A h k ( Z ˚ h k ) ⊂ B h k A_h^k(\mathring{Z}_h^k)\subset B_h^k A h k ( Z ˚ h k ) ⊂ B h k . By the rank–nullity theorem, we obtain
dim S h k = dim N ( A h k ∣ Z ˚ h k ) = dim Z ˚ h k − dim A h k Z ˚ h k ≥ dim Z ˚ h k − dim B h k = dim ( Z ˚ h k / B h k ) = b k . \begin{aligned}
\dim S_h^k&=\dim\mathcal{N}\left(A_h^k|_{\mathring{Z}_h^k}\right)\\
&=\dim\mathring{Z}_h^k-\dim A_h^k\mathring{Z}_h^k\\
&\ge\dim\mathring{Z}_h^k-\dim B_h^k=\dim(\mathring{Z}_h^k/B_h^k)=b_k.
\end{aligned} dim S h k = dim N ( A h k ∣ Z ˚ h k ) = dim Z ˚ h k − dim A h k Z ˚ h k ≥ dim Z ˚ h k − dim B h k = dim ( Z ˚ h k / B h k ) = b k .
This result shows that the dimension of ker A h k \ker A_h^k ker A h k is at least the dimension of the space of harmonic k k k -forms, suggesting that the kernel may contain additional elements.
We close this section with an elementary lemma on the long-time behavior of finite-dimensional semigroups, which is an immediate consequence of the spectral decomposition of finite-dimensional semigroups (see [18 Theorem V.3.1] ) and will be used repeatedly in § 5 to pass from the spectral information to convergence statements. Its proof is given in Appendix § 9 .§ 2 .
Lemma 4.5. Let X X X be a finite-dimensional subspace of a Hilbert space H H H , and A : X → X A:X\to X A : X → X a linear operator. Suppose 0 0 0 is an algebraically simple eigenvalue of A A A , and all other eigenvalues λ ∈ σ ( A ) ∖ { 0 } \lambda\in\sigma(A)\setminus\{0\} λ ∈ σ ( A ) ∖ { 0 } satisfy Re ( λ ) > 0 \operatorname{Re}(\lambda)>0 Re ( λ ) > 0 . Let P : X → X P:X\to X P : X → X denote the spectral projection onto the kernel of A A A . Then, for any exponential decay rate
0 < γ < min { Re ( λ ) ∣ λ ∈ σ ( A ) ∖ { 0 } } , 0<\gamma<\min\{\operatorname{Re}(\lambda)\mid\lambda\in\sigma(A)\setminus\{0\}\}, 0 < γ < min { Re ( λ ) ∣ λ ∈ σ ( A ) ∖ { 0 }} , there exists a constant C > 0 C>0 C > 0 such that
∥ e − t A − P ∥ L ( X ) ≤ C e − γ t , ∀ t ≥ 0. \left\|e^{-tA}-P\right\|_{\mathcal{L}(X)}\le Ce^{-\gamma t},\qquad\forall t\ge0. e − t A − P L ( X ) ≤ C e − γ t , ∀ t ≥ 0. Consequently, e − t A → P e^{-tA}\to P e − t A → P exponentially as t → ∞ t\to\infty t → ∞ .
and only if R ( A ) ⊥ N ( A ) \mathcal{R}(A) \perp\mathcal{N}(A) R ( A ) ⊥ N ( A ) , e.g., when A A A is self-adjoint. Concretely, if 0 0 0 is an algebraically simple eigenvalue of A A A , let r r r and ℓ \ell ℓ denote the associated right and left null vectors, i.e., A r = 0 Ar=0 A r = 0 and A ∗ ℓ = 0 A^{*}\ell=0 A ∗ ℓ = 0 , where A ∗ A^{*} A ∗ is the adjoint of A A A with respect to the given inner product. Then
P x = ( x , ℓ ) ( r , ℓ ) r Px=\frac{(x,\ell)}{(r,\ell)}r P x = ( r , ℓ ) ( x , ℓ ) r
since ℓ ⊥ R ( A ) \ell\perp\mathcal{R}(A) ℓ ⊥ R ( A ) .
Discrete preservation of long-time asymptotics
In this section, we investigate the long-time asymptotic behavior of the semidiscrete scheme (54) and identify conditions under which the structural properties of the continuous dynamics are preserved at the discrete level. These properties include the characterization of steady states and long-time limits for differential forms of different degrees. As a principal consequence, we establish a finite element antidynamo theorem in two dimensions.
To streamline the presentation, we first consider closed, connected, oriented manifolds and periodic domains, for which no boundary terms arise. The modifications required for manifolds with boundary, including the effects of natural and essential boundary conditions on the long-time asymptotics, are discussed separately in Subsection 5.4. We begin with a concise summary of the main results in Table 5.1 , including the characterization of steady states, the existence of long-time limits, the relevant diffusion regimes, and the assumptions under which each result holds.
Table 5.1. Summary of the long-time asymptotic results on closed manifolds. Here S h k = N ( A h k ) ∩ Z h k S_h^k=\mathcal{N}(A_h^k)\cap Z_h^k S h k = N ( A h k ) ∩ Z h k .
The two assumptions referenced in Table 5.1 are introduced at the beginning of Subsection 5.1 and Subsection 5.2, respectively, and are verified for concrete discretizations in § 6 . The modifications required in the presence of boundaries and boundary conditions are discussed in Subsection 5.4.
Long-time behavior for sufficiently large diffusion coefficients
In this subsection we show the long-time behavior of all form degrees 0 ≤ k ≤ n 0\leq k\leq n 0 ≤ k ≤ n simultaneously, requiring the diffusion coefficient to dominate the advection. The only structural condition is the following uniform L 2 L^2 L 2 -boundedness of the discrete contraction.
Assumption 5.1 (L 2 L^2 L 2 -boundedness of the discrete contraction). There exists a
constant C β C_\beta C β , independent of h h h and η \eta η , such that
∥ i β , h k ω h ∥ ≤ C β ∥ ω h ∥ , ∀ ω h ∈ V h k , k = 1 , … , n . (63) \left\|i_{\beta,h}^k\omega_h\right\| \le C_\beta\left\|\omega_h\right\|,\qquad\forall\omega_h\in V_h^k,\qquad k=1,\ldots,n. \tag*{(63)} i β , h k ω h ≤ C β ∥ ω h ∥ , ∀ ω h ∈ V h k , k = 1 , … , n . (63)
Under this assumption, we prove that, for sufficiently large η \eta η , the stationary space is characterized by the discrete cohomology and the corresponding evolution converges to a stationary solution as t → ∞ t\to\infty t → ∞ . The analogous result at the continuous level was established in [7 Chapter V, Theorem 3.6] .
Theorem 5.2. Under Assumption 5.1, there exists η 0 > 0 \eta_0>0 η 0 > 0 , independent of h h h , such that for every η > η 0 \eta>\eta_0 η > η 0 the following statements hold for all degrees k k k :
Every stationary solution is closed, i.e., S h k = N ( A h k ) S_h^k=N(A_h^k) S h k = N ( A h k ) .
dim S h k = b k \dim S_h^k=b_k dim S h k = b k , and the natural map S h k → H h k S_h^k\to H_h^k S h k → H h k is an isomorphism.
If the initial form is closed, then the solution exponentially converges to the unique stationary form in its discrete cohomology class.
For an arbitrary initial form, the solution exponentially converges to a closed stationary form.
Proof. The proof is structured into three parts based on the discrete Hodge decomposition.
Part 1: Every stationary solution of (55) is closed. Let ω h ∈ V h k \omega_h\in V_h^k ω h ∈ V h k satisfy A h ω h = 0 A_h\omega_h=0 A h ω h = 0 . Applying d d d yields
d A h ω h = d ( η δ h d ω h + i β , h d ω h ) = 0. (64) dA_h\omega_h=d(\eta\delta_hd\omega_h+i_{\beta,h}d\omega_h)=0. \tag*{(64)} d A h ω h = d ( η δ h d ω h + i β , h d ω h ) = 0. (64)
Setting ζ h = d ω h ∈ B h k + 1 \zeta_h=d\omega_h\in\mathcal{B}_h^{k+1} ζ h = d ω h ∈ B h k + 1 and taking the inner product with ζ h \zeta_h ζ h , we obtain
( ζ h , d ( η δ h ζ h + i β , h ζ h ) ) = ( δ h ζ h , η δ h ζ h + i β , h ζ h ) = 0. (65) (\zeta_h,d(\eta\delta_h\zeta_h+i_{\beta,h}\zeta_h))=(\delta_h\zeta_h,\eta\delta_h\zeta_h+i_{\beta,h}\zeta_h)=0. \tag*{(65)} ( ζ h , d ( η δ h ζ h + i β , h ζ h )) = ( δ h ζ h , η δ h ζ h + i β , h ζ h ) = 0. (65)
Under Assumption 5.1, applying the Cauchy–Schwarz inequality together with the discrete Poincaré inequality for exact forms, ∥ ζ h ∥ ≤ C P ∥ δ h ζ h ∥ \|\zeta_h\|\le C_P\|\delta_h\zeta_h\| ∥ ζ h ∥ ≤ C P ∥ δ h ζ h ∥ for some constant C P C_P C P , which follows from a standard duality argument; see also the generalized Gaffney inequality in [23 ] , we obtain
η ∥ δ h ζ h ∥ 2 = − ( δ h ζ h , i β , h ζ h ) ≤ C β ∥ δ h ζ h ∥ ∥ ζ h ∥ ≤ C β C P ∥ δ h ζ h ∥ 2 . (66) \eta\|\delta_h\zeta_h\|^2=-(\delta_h\zeta_h,i_{\beta,h}\zeta_h)\le C_\beta\|\delta_h\zeta_h\|\|\zeta_h\|\le C_\beta C_P\|\delta_h\zeta_h\|^2. \tag*{(66)} η ∥ δ h ζ h ∥ 2 = − ( δ h ζ h , i β , h ζ h ) ≤ C β ∥ δ h ζ h ∥∥ ζ h ∥ ≤ C β C P ∥ δ h ζ h ∥ 2 . (66)
Let η 0 : = C β C P \eta_0:=C_\beta C_P η 0 := C β C P (independent of h h h ). For any η > η 0 \eta>\eta_0 η > η 0 , inequality (66) forces δ h ζ h = 0 \delta_h\zeta_h=0 δ h ζ h = 0 and ζ h = 0 \zeta_h=0 ζ h = 0 due to discrete Poincaré inequality. Thus, all stationary solutions are closed forms, which proves item 1.
Part 2: Cohomology isomorphism and closed-form dynamics. Since A h Z h k ⊂ B h k ⊂ Z h k A_h\mathcal{Z}_h^k\subset\mathcal{B}_h^k\subset\mathcal{Z}_h^k A h Z h k ⊂ B h k ⊂ Z h k , the space of discrete closed forms is invariant under the evolution (54) . Let ω h ( t ) \omega_h(t) ω h ( t ) be an evolving closed form. Using the discrete Hodge decomposition, we uniquely decompose ω h ( t ) = p h ( t ) + h h ( t ) \omega_h(t)=p_h(t)+h_h(t) ω h ( t ) = p h ( t ) + h h ( t ) . The evolution equation gives
∂ t ω h = − A h ω h = − d ( i β , h ω h + η δ h ω h ) . (67) \partial_t\omega_h=-A_h\omega_h=-d(i_{\beta,h}\omega_h+\eta\delta_h\omega_h). \tag*{(67)} ∂ t ω h = − A h ω h = − d ( i β , h ω h + η δ h ω h ) . (67)
Since the right-hand side is an exact form, the harmonic component must remain invariant, i.e., ∂ t h h = 0 \partial_t h_h=0 ∂ t h h = 0 , which implies h h ( t ) ≡ h h ( 0 ) h_h(t)\equiv h_h(0) h h ( t ) ≡ h h ( 0 ) . This confirms that the evolution strictly preserves the discrete cohomology class.
The exact component p h ∈ B h k p_h\in\mathcal{B}_h^k p h ∈ B h k satisfies a linear dynamical system
∂ t p h = − d i β , h p h − η d δ h p h − d i β , h h h . (68) \partial_t p_h=-di_{\beta,h}p_h-\eta d\delta_h p_h-di_{\beta,h}h_h. \tag*{(68)} ∂ t p h = − d i β , h p h − η d δ h p h − d i β , h h h . (68)
The linear part A 1 p h : = − d ( i β h p h + η δ h p h ) A_{1ph}:=-d(i\beta_h p_h+\eta\delta_h p_h) A 1 p h := − d ( i β h p h + η δ h p h ) satisfies
( A 1 p h , p h ) = − η ∥ δ h p h ∥ 2 − ( i β h p h , δ h p h ) ≤ − ( η − C β C P ) ∥ δ h p h ∥ 2 ≤ − η − η 0 C P 2 ∥ p h ∥ 2 , (69) (A_{1ph},p_h)=-\eta\|\delta_h p_h\|^2-(i\beta_h p_h,\delta_h p_h)\leq-(\eta-C_\beta C_P)\|\delta_h p_h\|^2\leq-\frac{\eta-\eta_0}{C_P^2}\|p_h\|^2, \tag*{(69)} ( A 1 p h , p h ) = − η ∥ δ h p h ∥ 2 − ( i β h p h , δ h p h ) ≤ − ( η − C β C P ) ∥ δ h p h ∥ 2 ≤ − C P 2 η − η 0 ∥ p h ∥ 2 , (69)
for η > η 0 = C β C P \eta>\eta_0=C_\beta C_P η > η 0 = C β C P . This implies A 1 A_1 A 1 restricted to B h k \mathcal{B}_h^k B h k is strictly dissipative. Consequently, A 1 A_1 A 1 generates an exponentially stable semigroup and is invertible on B h k \mathcal{B}_h^k B h k .
Setting ∂ t p h = 0 \partial_t p_h=0 ∂ t p h = 0 in (68) demonstrates that each harmonic form h h ∈ H h k h_h\in\mathcal{H}_h^k h h ∈ H h k uniquely dictates a stationary exact component p ∞ = A 1 − 1 ( d i β h h h ) ∈ B h k p_\infty=A_1^{-1}(d i\beta_h h_h)\in\mathcal{B}_h^k p ∞ = A 1 − 1 ( d i β h h h ) ∈ B h k . Conversely, by the discrete Hodge decomposition, any stationary closed form in S h k \mathcal{S}_h^k S h k is uniquely decomposed into exact and harmonic components that satisfy this identical relationship. Consequently, the mapping h h ↦ p ∞ + h h h_h\mapsto p_\infty+h_h h h ↦ p ∞ + h h establishes a linear isomorphism between the harmonic space H h k \mathcal{H}_h^k H h k and the space of stationary closed forms S h k \mathcal{S}_h^k S h k .
Therefore, a one-to-one correspondence is established between the discrete cohomology classes and the stationary forms, which yields dim S h k = dim H h k = b k \dim\mathcal{S}_h^k=\dim\mathcal{H}_h^k=b_k dim S h k = dim H h k = b k (proving item 2). Furthermore, the contracting dynamics of (68) guarantee that any evolving closed form asymptotically converges to this unique stationary form ω ∞ = p ∞ + h h \omega_\infty=p_\infty+h_h ω ∞ = p ∞ + h h determined by its initial cohomology class h h h_h h h (proving item 3).
Part 3: Convergence to the closed form. Consider an initial form, and decompose the solution as ω h ( t ) = p h ( t ) + h h ( t ) + q h ( t ) \omega_h(t)=p_h(t)+h_h(t)+q_h(t) ω h ( t ) = p h ( t ) + h h ( t ) + q h ( t ) , where q h ∈ B h k , ∗ q_h\in\mathcal{B}_h^{k,*} q h ∈ B h k , ∗ . Applying d d d to the evolution equation and utilizing the commutativity d A h = A h d dA_h=A_hd d A h = A h d , we have
∂ t ( d ω h ) = − A h ( d ω h ) . (70) \partial_t(d\omega_h)=-A_h(d\omega_h). \tag*{(70)} ∂ t ( d ω h ) = − A h ( d ω h ) . (70)
Since d ω h ∈ B h k + 1 d\omega_h\in\mathcal{B}_h^{k+1} d ω h ∈ B h k + 1 is an exact form, a similar strict contraction argument from Part 2 gives that
1 2 d d t ∥ d ω h ∥ 2 ≤ − η − η 0 C P 2 ∥ d ω h ∥ 2 . (71) \frac{1}{2}\frac{d}{dt}\|d\omega_h\|^2\leq-\frac{\eta-\eta_0}{C_P^2}\|d\omega_h\|^2. \tag*{(71)} 2 1 d t d ∥ d ω h ∥ 2 ≤ − C P 2 η − η 0 ∥ d ω h ∥ 2 . (71)
By Gronwall’s inequality, we obtain
∥ d ω h ( t ) ∥ 2 ≤ e − γ t ∥ d ω h ( 0 ) ∥ 2 , \|d\omega_h(t)\|^2\leq e^{-\gamma t}\|d\omega_h(0)\|^2, ∥ d ω h ( t ) ∥ 2 ≤ e − γ t ∥ d ω h ( 0 ) ∥ 2 ,
where γ = 2 ( η − η 0 ) C P 2 > 0 \gamma=\frac{2(\eta-\eta_0)}{C_P^2}>0 γ = C P 2 2 ( η − η 0 ) > 0 . Because d q h = d ω h dq_h=d\omega_h d q h = d ω h , the Poincaré inequality ∥ q h ∥ ≤ C P ∥ d q h ∥ \|q_h\|\leq C_P\|dq_h\| ∥ q h ∥ ≤ C P ∥ d q h ∥ ensures that the co-exact component q h ( t ) q_h(t) q h ( t ) vanishes exponentially as t → ∞ t\to\infty t → ∞ , i.e., ∥ q h ( t ) ∥ 2 ≤ C P 2 e − γ t ∥ d q h ( 0 ) ∥ 2 \|q_h(t)\|^2\leq C_P^2e^{-\gamma t}\|dq_h(0)\|^2 ∥ q h ( t ) ∥ 2 ≤ C P 2 e − γ t ∥ d q h ( 0 ) ∥ 2 . Thus, the system asymptotically approaches the space of closed forms.
To prove the existence of the limit ω h ( ∞ ) \omega_h(\infty) ω h ( ∞ ) , let P H h k P_{\mathcal{H}_h^k} P H h k denote the L 2 L^2 L 2 orthogonal projection onto H h k \mathcal{H}_h^k H h k , projecting the evolution equation yields
∂ t h h = − P H h k A h ω h = − P H h k ( i β h , d q h ) . (72) \partial_t h_h=-P_{\mathcal{H}_h^k}A_h\omega_h=-P_{\mathcal{H}_h^k}(i\beta_h,dq_h). \tag*{(72)} ∂ t h h = − P H h k A h ω h = − P H h k ( i β h , d q h ) . (72)
Since P H h k i β h P_{\mathcal{H}_h^k}i\beta_h P H h k i β h is a bounded linear operator in L 2 L^2 L 2 , we get
∥ ∂ t h h ( t ) ∥ ≤ ∥ P H h k i β h ∥ ∥ d q h ( t ) ∥ ≤ C e − γ 2 t . \|\partial_t h_h(t)\|\leq\|P_{\mathcal{H}_h^k}i\beta_h\|\|dq_h(t)\|\leq Ce^{-\frac{\gamma}{2}t}. ∥ ∂ t h h ( t ) ∥ ≤ ∥ P H h k i β h ∥∥ d q h ( t ) ∥ ≤ C e − 2 γ t .
Hence
∫ 0 ∞ ∥ ∂ t h h ( t ) ∥ d t ≤ C ∫ 0 ∞ e − γ 2 t d t = C ′ γ < ∞ . \int_0^\infty\|\partial_t h_h(t)\|\,dt\leq C\int_0^\infty e^{-\frac{\gamma}{2}t}\,dt=\frac{C'}{\gamma}<\infty. ∫ 0 ∞ ∥ ∂ t h h ( t ) ∥ d t ≤ C ∫ 0 ∞ e − 2 γ t d t = γ C ′ < ∞.
Therefore ∂ t h h ∈ L 1 ( ( 0 , ∞ ) ; H h k ) \partial_t h_h\in L^1((0,\infty);\mathcal{H}_h^k) ∂ t h h ∈ L 1 (( 0 , ∞ ) ; H h k ) . For 0 ≤ s < t 0\leq s<t 0 ≤ s < t ,
h h ( t ) − h h ( s ) = ∫ s t ∂ τ h h ( τ ) d τ . h_h(t)-h_h(s)=\int_s^t\partial_\tau h_h(\tau)\,d\tau. h h ( t ) − h h ( s ) = ∫ s t ∂ τ h h ( τ ) d τ .
and
∥ h h ( t ) − h h ( s ) ∥ ≤ ∫ s t ∥ ∂ τ h h ( τ ) ∥ d τ ≤ C ∫ s t e − γ 2 τ d τ ≤ C ′ γ e − γ 2 s . \lVert h_h(t)-h_h(s)\rVert\le\int_s^t \lVert\partial_\tau h_h(\tau)\rVert\,d\tau\le C\int_s^t e^{-\frac{\gamma}{2}\tau}\,d\tau\le\frac{C'}{\gamma}e^{-\frac{\gamma}{2}s}. ∥ h h ( t ) − h h ( s )∥ ≤ ∫ s t ∥ ∂ τ h h ( τ )∥ d τ ≤ C ∫ s t e − 2 γ τ d τ ≤ γ C ′ e − 2 γ s .
The right-hand side tends to 0 0 0 as s → ∞ s\to\infty s → ∞ , uniformly for t > s t>s t > s . Thus { h h ( t ) } t ≥ 0 \{h_h(t)\}_{t\ge0} { h h ( t ) } t ≥ 0 is a Cauchy family in the finite-dimensional Hilbert space S h k \mathcal{S}_h^k S h k . Hence there exists h ∞ ∈ S h k h_\infty\in\mathcal{S}_h^k h ∞ ∈ S h k such that
h h ( t ) → h ∞ as t → ∞ . h_h(t)\to h_\infty\qquad\text{as }t\to\infty. h h ( t ) → h ∞ as t → ∞.
Equivalently, the limit can be written as the absolutely convergent integral
h ∞ = h h ( 0 ) − ∫ 0 ∞ P S h k i β , h d q h ( τ ) d τ . h_\infty=h_h(0)-\int_0^\infty P_{\mathcal{S}_h^k}i\beta,h\,d q_h(\tau)\,d\tau. h ∞ = h h ( 0 ) − ∫ 0 ∞ P S h k i β , h d q h ( τ ) d τ .
Moreover, the convergence is exponential
∥ h h ( t ) − h ∞ ∥ ≤ ∫ t ∞ ∥ ∂ τ h h ( τ ) ∥ d τ ≤ C ′ γ e − γ 2 t . \lVert h_h(t)-h_\infty\rVert\le\int_t^\infty\lVert\partial_\tau h_h(\tau)\rVert\,d\tau\le\frac{C'}{\gamma}e^{-\frac{\gamma}{2}t}. ∥ h h ( t ) − h ∞ ∥ ≤ ∫ t ∞ ∥ ∂ τ h h ( τ )∥ d τ ≤ γ C ′ e − 2 γ t .
We next prove that the exact component p h ( t ) p_h(t) p h ( t ) also has a finite limit. Recall that p h ( t ) ∈ B h k p_h(t)\in\mathcal{B}_h^k p h ( t ) ∈ B h k satisfies
∂ t p h = A 1 p h + F ( t ) , \partial_t p_h=A_1p_h+F(t), ∂ t p h = A 1 p h + F ( t ) ,
where A 1 : B h k → B h k A_1:\mathcal{B}_h^k\to\mathcal{B}_h^k A 1 : B h k → B h k is strictly dissipative, and
F ( t ) = − d i β , h h h ( t ) − d i β , h q h ( t ) − P B h k i β , h d q h ( t ) . F(t)=-d_{i\beta,h}h_h(t)-d_{i\beta,h}q_h(t)-P_{\mathcal{B}_h^k}i\beta,h\,d q_h(t). F ( t ) = − d i β , h h h ( t ) − d i β , h q h ( t ) − P B h k i β , h d q h ( t ) .
Since the mesh is fixed, all discrete operators involved are bounded. Moreover, from the previous steps,
h h ( t ) → h ∞ , q h ( t ) → 0 , d q h ( t ) → 0 h_h(t)\to h_\infty,\qquad q_h(t)\to0,\qquad d q_h(t)\to0 h h ( t ) → h ∞ , q h ( t ) → 0 , d q h ( t ) → 0
exponentially as t → ∞ t\to\infty t → ∞ . Hence
F ( t ) → F ∞ : = − d i β , h h ∞ ∈ B h k . F(t)\to F_\infty:=-d_{i\beta,h}h_\infty\in\mathcal{B}_h^k. F ( t ) → F ∞ := − d i β , h h ∞ ∈ B h k .
More precisely, there exist constants C F > 0 C_F>0 C F > 0 and κ > 0 \kappa>0 κ > 0 such that
∥ F ( t ) − F ∞ ∥ ≤ C F e − κ t . \lVert F(t)-F_\infty\rVert\le C_F e^{-\kappa t}. ∥ F ( t ) − F ∞ ∥ ≤ C F e − κ t .
Indeed,
F ( t ) − F ∞ = − d i β , h ( h h ( t ) − h ∞ ) − d i β , h q h ( t ) − P B h k i β , h d q h ( t ) , F(t)-F_\infty=-d_{i\beta,h}(h_h(t)-h_\infty)-d_{i\beta,h}q_h(t)-P_{\mathcal{B}_h^k}i\beta,h\,d q_h(t), F ( t ) − F ∞ = − d i β , h ( h h ( t ) − h ∞ ) − d i β , h q h ( t ) − P B h k i β , h d q h ( t ) ,
and each term on the right-hand side decays exponentially by boundedness of the operators. Since A 1 A_1 A 1 is strictly contracting on the finite-dimensional space B h k \mathcal{B}_h^k B h k , its semigroup satisfies
∥ e t A 1 ∥ ≤ M e − α t \lVert e^{tA_1}\rVert\le M e^{-\alpha t} ∥ e t A 1 ∥ ≤ M e − α t
for some M ≥ 1 M\ge1 M ≥ 1 and α > 0 \alpha>0 α > 0 . In particular, A 1 A_1 A 1 is invertible on B h k \mathcal{B}_h^k B h k .
Define
p ∞ : = − A 1 − 1 F ∞ . p_\infty:=-A_1^{-1}F_\infty. p ∞ := − A 1 − 1 F ∞ .
Then p ∞ ∈ B h k p_\infty\in\mathcal{B}_h^k p ∞ ∈ B h k and it solves the limiting steady equation
A 1 p ∞ + F ∞ = 0. A_1p_\infty+F_\infty=0. A 1 p ∞ + F ∞ = 0.
Set
r h ( t ) : = p h ( t ) − p ∞ . r_h(t) := p_h(t) - p_\infty. r h ( t ) := p h ( t ) − p ∞ .
Using the equations for p h ( t ) p_h(t) p h ( t ) and p ∞ p_\infty p ∞ , we obtain
∂ t r h = A 1 r h + ( F ( t ) − F ∞ ) . \partial_t r_h = A_1 r_h + (F(t) - F_\infty). ∂ t r h = A 1 r h + ( F ( t ) − F ∞ ) .
By Duhamel’s formula,
r h ( t ) = e t A 1 r h ( 0 ) + ∫ 0 t e ( t − s ) A 1 ( F ( s ) − F ∞ ) d s . r_h(t) = e^{tA_1}r_h(0) + \int_0^t e^{(t-s)A_1}(F(s) - F_\infty)\,ds. r h ( t ) = e t A 1 r h ( 0 ) + ∫ 0 t e ( t − s ) A 1 ( F ( s ) − F ∞ ) d s .
Therefore
∥ r h ( t ) ∥ ≤ M e − α t ∥ r h ( 0 ) ∥ + M ∫ 0 t e − α ( t − s ) ∥ F ( s ) − F ∞ ∥ d s ≤ M e − α t ∥ r h ( 0 ) ∥ + M C F ∫ 0 t e − α ( t − s ) e − κ s d s . \begin{aligned}
\|r_h(t)\| &\leq Me^{-\alpha t}\|r_h(0)\| + M\int_0^t e^{-\alpha(t-s)}\|F(s)-F_\infty\|\,ds \\
&\leq Me^{-\alpha t}\|r_h(0)\| + M C_F\int_0^t e^{-\alpha(t-s)}e^{-\kappa s}\,ds.
\end{aligned} ∥ r h ( t ) ∥ ≤ M e − α t ∥ r h ( 0 ) ∥ + M ∫ 0 t e − α ( t − s ) ∥ F ( s ) − F ∞ ∥ d s ≤ M e − α t ∥ r h ( 0 ) ∥ + M C F ∫ 0 t e − α ( t − s ) e − κ s d s .
More explicitly, if α ≠ κ \alpha\ne\kappa α = κ , then
∫ 0 t e − α ( t − s ) e − κ s d s = e − κ t − e − α t α − κ , \int_0^t e^{-\alpha(t-s)}e^{-\kappa s}\,ds
= \frac{e^{-\kappa t}-e^{-\alpha t}}{\alpha-\kappa}, ∫ 0 t e − α ( t − s ) e − κ s d s = α − κ e − κ t − e − α t ,
and if α = κ \alpha=\kappa α = κ , then it equals
t e − α t . te^{-\alpha t}. t e − α t .
Thus, for every 0 < ν < min { α , κ } 0<\nu<\min\{\alpha,\kappa\} 0 < ν < min { α , κ } , there exists C ν > 0 C_\nu>0 C ν > 0 such that
∥ p h ( t ) − p ∞ ∥ = ∥ r h ( t ) ∥ ≤ C ν e − ν t . \|p_h(t)-p_\infty\|=\|r_h(t)\|\leq C_\nu e^{-\nu t}. ∥ p h ( t ) − p ∞ ∥ = ∥ r h ( t ) ∥ ≤ C ν e − ν t .
Thus
p h ( t ) → p ∞ = − A 1 − 1 F ∞ = A 1 − 1 d i β , h h ∞ . p_h(t)\to p_\infty=-A_1^{-1}F_\infty=A_1^{-1}d i_{\beta,h}h_\infty. p h ( t ) → p ∞ = − A 1 − 1 F ∞ = A 1 − 1 d i β , h h ∞ .
Combining this with h h ( t ) → h ∞ h_h(t)\to h_\infty h h ( t ) → h ∞ and q h ( t ) → 0 q_h(t)\to0 q h ( t ) → 0 , we conclude that
ω h ( t ) = p h ( t ) + h h ( t ) + q h ( t ) → p ∞ + h ∞ = : ω ∞ . \omega_h(t)=p_h(t)+h_h(t)+q_h(t)\to p_\infty+h_\infty=:\omega_\infty. ω h ( t ) = p h ( t ) + h h ( t ) + q h ( t ) → p ∞ + h ∞ =: ω ∞ .
Finally, the limit exists and is closed and stationary: A h ω ∞ = A h p ∞ + A h h ∞ = − A 1 p ∞ + d i β , h h ∞ = F ∞ + d i β , h h ∞ = 0 A_h\omega_\infty=A_hp_\infty+A_hh_\infty=-A_1p_\infty+d i_{\beta,h}h_\infty=F_\infty+d i_{\beta,h}h_\infty=0 A h ω ∞ = A h p ∞ + A h h ∞ = − A 1 p ∞ + d i β , h h ∞ = F ∞ + d i β , h h ∞ = 0 , using δ h h ∞ = 0 \delta_hh_\infty=0 δ h h ∞ = 0 , d h h ∞ = 0 d_hh_\infty=0 d h h ∞ = 0 , and the definition of p ∞ p_\infty p ∞ . This proves item 4 and completes the proof. □
Long-time behavior for arbitrary diffusion coefficients
When the diffusion coefficient is arbitrary, the discretization requires additional care. To this end, we assume the following discrete spectral preservation property for the cases k = 0 k=0 k = 0 and k = n k=n k = n .
Assumption 5.3 (discrete spectral preservation for 0- and n n n -forms). For k = 0 k=0 k = 0 and k = n k=n k = n , assume that 0 is an algebraically simple eigenvalue of A h k A_h^k A h k , and the nonzero eigenvalues lie in the open right half-plane, i.e.,
Re λ > 0 , ∀ λ ∈ σ ( A h k ) ∖ { 0 } . \operatorname{Re}\lambda>0,\qquad\forall\lambda\in\sigma(A_h^k)\setminus\{0\}. Re λ > 0 , ∀ λ ∈ σ ( A h k ) ∖ { 0 } .
We now study the long-time asymptotic behavior of the FEEC discretization of the advection–diffusion equation and derive the corresponding finite element antidynamo theorems. The results at the continuous level were established in [7 Chapter V, Theorems 3.7–3.12] . At the discrete level, the arguments follow a similar strategy, but the finite-dimensional setting provides additional simplifications: the discrete operators are bounded, and finite element inverse inequalities imply, for instance, that asymptotic convergence of u h u_h u h also yields asymptotic convergence of d u h du_h d u h .
We first establish the long-time asymptotic behavior for the cases k = 0 k=0 k = 0 and k = n k=n k = n under Assumption 5.3.
Theorem 5.5 (k = 0 k=0 k = 0 ). Under Assumption 5.3, for every η > 0 \eta>0 η > 0 the stationary 0 0 0 -forms are exactly the constants, N ( A h 0 ) = S h 0 = H h 0 \mathcal{N}(A_h^0)=S_h^0=\mathcal{H}_h^0 N ( A h 0 ) = S h 0 = H h 0 . Moreover, every solution of (56) with initial data in V h 0 V_h^0 V h 0 converges exponentially to a constant.
Proof. For k = 0 k=0 k = 0 , the operator reduces to A h 0 = ( i β h + η δ h ) d A_h^0=(\mathrm{i}\beta_h+\eta\delta_h)d A h 0 = ( i β h + η δ h ) d . Since the exterior derivative d d d annihilates constant functions, all constant discrete forms belong to S h 0 S_h^0 S h 0 . According to Assumption 5.3, the eigenvalue λ = 0 \lambda=0 λ = 0 is algebraically simple. Therefore, any stationary solution ω ˉ h ∈ N ( A h ) \bar{\omega}_h\in\mathcal{N}(A_h) ω ˉ h ∈ N ( A h ) must be a constant, establishing N ( A h 0 ) = S h 0 = H h 0 \mathcal{N}(A_h^0)=S_h^0=\mathcal{H}_h^0 N ( A h 0 ) = S h 0 = H h 0 .
For an arbitrary initial condition ω h ( 0 ) ∈ V h 0 \omega_h(0)\in V_h^0 ω h ( 0 ) ∈ V h 0 , Lemma 4.5 yields the exponential convergence
lim t → ∞ ω h ( t ) = lim t → ∞ e − t A h 0 ω h ( 0 ) = P 0 s p ω h ( 0 ) , (73) \lim_{t\to\infty}\omega_h(t)=\lim_{t\to\infty}e^{-tA_h^0}\omega_h(0)=P_0^{\mathrm{sp}}\omega_h(0), \tag*{(73)} t → ∞ lim ω h ( t ) = t → ∞ lim e − t A h 0 ω h ( 0 ) = P 0 sp ω h ( 0 ) , (73)
where P 0 s p P_0^{\mathrm{sp}} P 0 sp is the spectral projection onto N ( A h 0 ) = H h 0 \mathcal{N}(A_h^0)=\mathcal{H}_h^0 N ( A h 0 ) = H h 0 along R ( A h 0 ) \mathcal{R}(A_h^0) R ( A h 0 ) (cf. Remark 4.7). In particular, the limit is a global constant. We emphasize that P 0 s p P_0^{\mathrm{sp}} P 0 sp differs in general from the L 2 L^2 L 2 -orthogonal projection onto constants. This completes the proof. □
Theorem 5.6 (k = n k=n k = n ). Under Assumption 5.3, for every η > 0 \eta>0 η > 0 the stationary n n n -form space is one-dimensional: dim N ( A h n ) = dim S h n = 1 \dim\mathcal{N}(A_h^n)=\dim S_h^n=1 dim N ( A h n ) = dim S h n = 1 . Moreover:
Every solution with initial data in V h n V_h^n V h n converges exponentially to the unique stationary n n n -form with the same total mass.
If the initial condition is exact, ω h ( 0 ) = d n − 1 B h \omega_h(0)=d^{n-1}B_h ω h ( 0 ) = d n − 1 B h with B h ∈ V h n − 1 B_h\in V_h^{n-1} B h ∈ V h n − 1 , then the solution converges exponentially to zero.
Proof. For k = n k=n k = n the operator reduces to A h n = d n − 1 ( i β h + η δ h n ) A_h^n=d^{n-1}(\mathrm{i}\beta_h+\eta\delta_h^n) A h n = d n − 1 ( i β h + η δ h n ) . Since every n n n -form is closed, S h n = N ( A h n ) ∩ Z h n = N ( A h n ) S_h^n=\mathcal{N}(A_h^n)\cap Z_h^n=\mathcal{N}(A_h^n) S h n = N ( A h n ) ∩ Z h n = N ( A h n ) , and Assumption 5.3 gives dim S h n = 1 \dim S_h^n=1 dim S h n = 1 . Lemma 4.5 then yields the exponential convergence
lim t → ∞ ω h ( t ) = P n s p ω h ( 0 ) , (74) \lim_{t\to\infty}\omega_h(t)=P_n^{\mathrm{sp}}\omega_h(0), \tag*{(74)} t → ∞ lim ω h ( t ) = P n sp ω h ( 0 ) , (74)
where P n s p P_n^{\mathrm{sp}} P n sp is the spectral projection onto N ( A h n ) \mathcal{N}(A_h^n) N ( A h n ) along R ( A h n ) \mathcal{R}(A_h^n) R ( A h n ) (cf. Remark 4.7).
First, Stokes’ theorem implies strict mass conservation:
d d t ∫ M ω h = − ∫ M d ( i β , h ω h + η δ h ω h ) = 0. (75) \frac{d}{dt}\int_M\omega_h=-\int_M d(\mathrm{i}_{\beta,h}\omega_h+\eta\delta_h\omega_h)=0. \tag*{(75)} d t d ∫ M ω h = − ∫ M d ( i β , h ω h + η δ h ω h ) = 0. (75)
Thus, ∫ M ω h ( t ) = ∫ M ω h ( 0 ) \int_M\omega_h(t)=\int_M\omega_h(0) ∫ M ω h ( t ) = ∫ M ω h ( 0 ) for all t ≥ 0 t\geq0 t ≥ 0 . Because initial conditions can carry arbitrary non-zero mass, the basis of the one-dimensional stationary space S h n S_h^n S h n must inherently have a non-zero mass to preserve it.
Finally, consider an initially exact n n n -form ω h ( 0 ) = d B h \omega_h(0)=dB_h ω h ( 0 ) = d B h with B h ∈ V h n − 1 B_h\in V_h^{n-1} B h ∈ V h n − 1 . By Stokes’ theorem, its initial mass is zero (∫ M d B h = 0 \int_M dB_h=0 ∫ M d B h = 0 ). By mass conservation, it must exponentially converge to the unique stationary state carrying zero mass. Since any non-trivial stationary form carries non-zero mass, this limit must be strictly 0. This completes the proof.
The result for 0-forms implies a corresponding statement for closed 1-forms.
Theorem 5.7 (k = 1 k = 1 k = 1 ). Under Assumption 5.3 , for every η > 0 \eta> 0 η > 0 the closed stationary space satisfies
dim S h 1 = b 1 ( M ) , S h 1 ∩ B h 1 = { 0 } . \dim S_h^1 = b_1(M), \qquad S_h^1 \cap\mathcal{B}_h^1 = \{0\}. dim S h 1 = b 1 ( M ) , S h 1 ∩ B h 1 = { 0 } . Consequently, each class in H h 1 H_h^1 H h 1 contains a unique closed stationary 1-form, and the evolution of any closed initial 1-form converges exponentially to that representative.
Proof. First, consider an initially exact 1-form ω h ( 0 ) = d u h ( 0 ) \omega_h(0) = du_h(0) ω h ( 0 ) = d u h ( 0 ) with u h ( 0 ) ∈ V h 0 u_h(0) \in V_h^0 u h ( 0 ) ∈ V h 0 . By the commutativity A h 1 d = d A h 0 A_h^1d = dA_h^0 A h 1 d = d A h 0 , its evolution satisfies
ω h ( t ) = e − t A h 1 d u h ( 0 ) = d ( e − t A h 0 u h ( 0 ) ) . (76) \omega_h(t) = e^{-tA_h^1}du_h(0) = d\left(e^{-tA_h^0}u_h(0)\right). \tag*{(76)} ω h ( t ) = e − t A h 1 d u h ( 0 ) = d ( e − t A h 0 u h ( 0 ) ) . (76)
By Theorem 5.5 , the 0-form evolution e − t A h 0 u h ( 0 ) e^{-tA_h^0}u_h(0) e − t A h 0 u h ( 0 ) converges exponentially to a global constant c c c as t → ∞ t \to\infty t → ∞ . Since ω h ( t ) = d ( e − t A h 0 u h ( 0 ) ) \omega_h(t) = d(e^{-tA_h^0}u_h(0)) ω h ( t ) = d ( e − t A h 0 u h ( 0 )) and the exact subspace is invariant, this implies lim t → ∞ ω h ( t ) = d c = 0 \lim_{t\to\infty}\omega_h(t) = dc = 0 lim t → ∞ ω h ( t ) = d c = 0 . This guarantees that any exact 1-form dissipates to zero. Consequently, the only exact stationary 1-form is the trivial zero field, i.e., S h 1 ∩ B h 1 = { 0 } S_h^1 \cap\mathcal{B}_h^1 = \{0\} S h 1 ∩ B h 1 = { 0 } .
By Theorem 4.3 and its proof, we have dim S h 1 ≥ b 1 ( M ) = dim H h 1 \dim S_h^1 \ge b_1(M) = \dim H_h^1 dim S h 1 ≥ b 1 ( M ) = dim H h 1 . Since S h 1 ∩ B h 1 = { 0 } S_h^1 \cap\mathcal{B}_h^1 = \{0\} S h 1 ∩ B h 1 = { 0 } , the natural projection from S h 1 S_h^1 S h 1 onto the cohomology space H h 1 = Z h 1 / B h 1 H_h^1 = \mathcal{Z}_h^1/\mathcal{B}_h^1 H h 1 = Z h 1 / B h 1 is injective. This forces dim S h 1 ≤ dim H h 1 \dim S_h^1 \le\dim H_h^1 dim S h 1 ≤ dim H h 1 . Hence, dim S h 1 = dim H h 1 = b 1 ( M ) \dim S_h^1 = \dim H_h^1 = b_1(M) dim S h 1 = dim H h 1 = b 1 ( M ) . This isomorphism dictates that each cohomology class uniquely corresponds to a closed stationary form, proving the dimension and uniqueness statements.
Finally, recall from the proof of Theorem 5.2 that the evolution of an initially closed 1-form strictly preserves its discrete cohomology class. Since the natural projection from S h 1 S_h^1 S h 1 to the cohomology space H h 1 = Z h 1 / B h 1 H_h^1 = \mathcal{Z}_h^1/\mathcal{B}_h^1 H h 1 = Z h 1 / B h 1 is an isomorphism, the initial condition ω h ( 0 ) \omega_h(0) ω h ( 0 ) shares its cohomology class with the unique stationary form ω ˉ h ∈ S h 1 \bar{\omega}_h \in S_h^1 ω ˉ h ∈ S h 1 . By the definition of cohomology classes, their difference must be an exact form, allowing us to uniquely decompose the initial condition as ω h ( 0 ) = ω ˉ h + ω h e x ( 0 ) \omega_h(0) = \bar{\omega}_h + \omega_h^{\mathrm{ex}}(0) ω h ( 0 ) = ω ˉ h + ω h ex ( 0 ) , where ω h e x ( 0 ) ∈ B h 1 \omega_h^{\mathrm{ex}}(0) \in\mathcal{B}_h^1 ω h ex ( 0 ) ∈ B h 1 . Since stationary forms are time-invariant and exact forms dissipate to zero, we obtain
lim t → ∞ ω h ( t ) = lim t → ∞ e − t A h 1 ω ˉ h + lim t → ∞ e − t A h 1 ω h e x ( 0 ) = ω ˉ h + 0 = ω ˉ h . (77) \lim_{t\to\infty}\omega_h(t) = \lim_{t\to\infty}e^{-tA_h^1}\bar{\omega}_h + \lim_{t\to\infty}e^{-tA_h^1}\omega_h^{\mathrm{ex}}(0) = \bar{\omega}_h + 0 = \bar{\omega}_h. \tag*{(77)} t → ∞ lim ω h ( t ) = t → ∞ lim e − t A h 1 ω ˉ h + t → ∞ lim e − t A h 1 ω h ex ( 0 ) = ω ˉ h + 0 = ω ˉ h . (77)
Thus, the solution asymptotically converges to the unique stationary form ω ˉ h \bar{\omega}_h ω ˉ h within the cohomology class of the initial condition. This completes the proof. □ \square □
Corollary 5.8. For k = 1 k = 1 k = 1 , we have the decomposition
V h 1 = S h 1 ⊕ B h 1 ⊕ B h 1 , ∗ . V_h^1 = S_h^1 \oplus\mathcal{B}_h^1 \oplus\mathcal{B}_h^{1,*}. V h 1 = S h 1 ⊕ B h 1 ⊕ B h 1 , ∗ .
Proof. The discrete Hodge decomposition implies
V h 1 = B h 1 , ∗ ⊕ B h 1 ⊕ H h 1 . V_h^1 = \mathcal{B}_h^{1,*} \oplus\mathcal{B}_h^1 \oplus\mathcal{H}_h^1. V h 1 = B h 1 , ∗ ⊕ B h 1 ⊕ H h 1 .
Theorem 5.7 shows that the natural projection S h 1 → H h 1 = Z h 1 / B h 1 S_h^1 \to H_h^1 = \mathcal{Z}_h^1/\mathcal{B}_h^1 S h 1 → H h 1 = Z h 1 / B h 1 is an isomorphism. Hence every closed form can be written uniquely as a sum of a stationary closed form and an exact form, that is, Z h 1 = S h 1 ⊕ B h 1 \mathcal{Z}_h^1 = S_h^1 \oplus\mathcal{B}_h^1 Z h 1 = S h 1 ⊕ B h 1 . Since B h 1 , ∗ \mathcal{B}_h^{1,*} B h 1 , ∗ is the orthogonal complement of Z h 1 \mathcal{Z}_h^1 Z h 1 , the stated direct decomposition follows. □ \square □
The result for n n n -forms gives the complementary statement in degree n − 1 n - 1 n − 1 .
Theorem 5.9 (k = n − 1 k=n-1 k = n − 1 ). Under Assumption 5.3, for every η > 0 \eta>0 η > 0 the exterior derivative of every evolving ( n − 1 ) (n-1) ( n − 1 ) -form decays exponentially:
d n − 1 ω h ( t ) ⟶ 0 ( t → ∞ ) . d^{n-1}\omega_h(t)\longrightarrow0 \qquad(t\to\infty). d n − 1 ω h ( t ) ⟶ 0 ( t → ∞ ) . In particular, every stationary ( n − 1 ) (n-1) ( n − 1 ) -form is closed, i.e., N ( A h n − 1 ) ⊂ Z h n − 1 \mathcal{N}(A_h^{n-1})\subset\mathfrak{Z}_h^{n-1} N ( A h n − 1 ) ⊂ Z h n − 1 .
Proof. Since d d d and A h A_h A h commute, the exact n n n -form d ω h d\omega_h d ω h satisfies the n n n -form equation (56) . Theorem 5.6 implies that it dissipates to zero. Thus, we have
lim t → ∞ d ω h ( t ) = 0 , (78) \lim_{t\to\infty}d\omega_h(t)=0, \tag*{(78)} t → ∞ lim d ω h ( t ) = 0 , (78)
which proves that the exterior derivative of the evolving ( n − 1 ) (n-1) ( n − 1 ) -form asymptotically vanishes.
If ω ˉ h \bar{\omega}_h ω ˉ h is stationary for the ( n − 1 ) (n-1) ( n − 1 ) -form problem, then A h n d ω ˉ h = d A h n − 1 ω ˉ h = 0 A_h^n d\bar{\omega}_h=dA_h^{n-1}\bar{\omega}_h=0 A h n d ω ˉ h = d A h n − 1 ω ˉ h = 0 . Moreover, d ω ˉ h d\bar{\omega}_h d ω ˉ h has zero total mass (∫ M d ω ˉ h = 0 \int_M d\bar{\omega}_h=0 ∫ M d ω ˉ h = 0 ). By Theorem 5.6, the only exact stationary n n n -form with zero mass is zero. Therefore d ω ˉ h = 0 d\bar{\omega}_h=0 d ω ˉ h = 0 , so every stationary ( n − 1 ) (n-1) ( n − 1 ) -form is closed. ▫
The two-dimensional finite element antidynamo theorem
In dynamo theory, the central question is whether the evolution semigroup e − t A e^{-tA} e − t A admits exponentially growing modes. Cowling’s antidynamo theorem [16 ] rules out the maintenance of a self-sustained axisymmetric magnetic field. It inspired a broad family of antidynamo results [43 , 53 ] , often summarized by the principle that genuinely two-dimensional flows cannot sustain fast kinematic dynamo action. Arnold placed this principle in a geometric framework based on differential forms [7 Chapter V] . Our objective is to determine whether FEEC discretizations [3 , 5 ] inherit this topological obstruction and thereby exclude spurious discrete dynamo growth.
We now specialize the results of subsection 5.2 to n = 2 n=2 n = 2 . In this case, the degrees k = 1 k=1 k = 1 and k = n − 1 k=n-1 k = n − 1 coincide, allowing the two corresponding conclusions to be combined. Consequently, for every η > 0 \eta>0 η > 0 and every fixed discretization satisfying Assumption 5.3, the discrete 1-form solution converges exponentially to a closed stationary 1-form. This provides a discrete FEEC counterpart of Arnold’s two-dimensional antidynamo theorem: no exponentially growing dynamo modes can be sustained by the discrete evolution.
Theorem 5.10 (discrete antidynamo). Under Assumption 5.3 and n = 2 n=2 n = 2 , for every η > 0 \eta>0 η > 0 and every initial discrete 1-form, the solution of (56) converges exponentially to a closed stationary 1-form. Thus, there are no exponentially growing discrete dynamo modes. In particular, if M M M is simply connected, the limiting stationary form is zero.
Proof. Using the discrete Hodge decomposition, we express the solution as
ω h ( t ) = p h ( t ) + h h ( t ) + q h ( t ) , (79) \omega_h(t)=p_h(t)+h_h(t)+q_h(t), \tag*{(79)} ω h ( t ) = p h ( t ) + h h ( t ) + q h ( t ) , (79)
where p h ∈ B h p_h\in\mathfrak{B}_h p h ∈ B h , h h ∈ H h h_h\in\mathfrak{H}_h h h ∈ H h and q h ∈ B h ∗ q_h\in\mathfrak{B}_h^* q h ∈ B h ∗ . We analyze these components successively. First, the coexact component q h q_h q h is controlled through the identity d ω h = d q h d\omega_h=dq_h d ω h = d q h . Its decay then determines the evolution of the harmonic component h h h_h h h , whose time derivative is driven by d q h dq_h d q h . Finally, the exact component p h p_h p h satisfies a forced evolution equation whose forcing is determined by h h h_h h h , q h q_h q h , and d q h dq_h d q h .
For n = 2 n=2 n = 2 , a 1-form is also an ( n − 1 ) (n-1) ( n − 1 ) -form. By Theorem 5.9, d ω h ( t ) = d q h ( t ) → 0 d\omega_h(t)=dq_h(t)\to0 d ω h ( t ) = d q h ( t ) → 0 exponentially for any η > 0 \eta>0 η > 0 . The discrete Poincaré inequality then ensures the coexact component q h ( t ) ∈ B h 1 , ∗ q_h(t)\in\mathfrak{B}_h^{1,*} q h ( t ) ∈ B h 1 , ∗ also vanishes exponentially. The argument proceeds similarly to the proof of Theorem 5.2 . Projecting the evolution onto S h 1 \mathfrak{S}_h^1 S h 1 gives
∂ t h h = − P S h 1 i β , h 2 d q h . \partial_t h_h=-P_{\mathfrak{S}_h^1}i_{\beta,h}^2\,dq_h. ∂ t h h = − P S h 1 i β , h 2 d q h .
Since d q h → 0 dq_h\to0 d q h → 0 exponentially, the integral argument in the proof of Theorem 5.2 (Part 3) yields h h ( t ) → h ∞ ∈ S h 1 h_h(t)\to h_\infty\in\mathfrak{S}_h^1 h h ( t ) → h ∞ ∈ S h 1 exponentially.
For the exact component p h ( t ) ∈ B h 1 p_h(t)\in\mathcal{B}_h^1 p h ( t ) ∈ B h 1 , since the discrete Hodge decomposition is time-independent, we have
p h ( t ) = P B h 1 ω h ( t ) , ∂ t p h ( t ) = P B h 1 ∂ t ω h ( t ) . p_h(t)=P_{\mathcal{B}_h^1}\omega_h(t),\qquad\partial_t p_h(t)=P_{\mathcal{B}_h^1}\partial_t\omega_h(t). p h ( t ) = P B h 1 ω h ( t ) , ∂ t p h ( t ) = P B h 1 ∂ t ω h ( t ) .
Projecting the evolution equation onto B h 1 \mathcal{B}_h^1 B h 1 , and using d p h = 0 dp_h=0 d p h = 0 , d h h = 0 dh_h=0 d h h = 0 , δ h h h = 0 \delta_hh_h=0 δ h h h = 0 , and d ω h = d q h d\omega_h=dq_h d ω h = d q h , gives
∂ t p h = A 1 p h + F ( t ) , (80) \partial_t p_h=A_1p_h+F(t), \tag*{(80)} ∂ t p h = A 1 p h + F ( t ) , (80)
where
A 1 = − A h 1 ∣ B h 1 = − d ( i β , h + η δ h ) , A_1=-A_h^1\big|_{\mathcal{B}_h^1}=-d(i_{\beta,h}+\eta\delta_h), A 1 = − A h 1 B h 1 = − d ( i β , h + η δ h ) ,
and
F ( t ) = − d i β , h h h ( t ) − d i β , h q h ( t ) − P B h 1 i β , h d q h ( t ) . F(t)=-di_{\beta,h}h_h(t)-di_{\beta,h}q_h(t)-P_{\mathcal{B}_h^1}i_{\beta,h}dq_h(t). F ( t ) = − d i β , h h h ( t ) − d i β , h q h ( t ) − P B h 1 i β , h d q h ( t ) .
Since N ( A h 0 ) = R \mathcal{N}(A_h^0)=\mathbb{R} N ( A h 0 ) = R is invariant, A h 0 A_h^0 A h 0 induces an operator on V h 0 / R V_h^0/\mathbb{R} V h 0 / R whose spectrum is σ ( A h 0 ) ∖ { 0 } \sigma(A_h^0)\setminus\{0\} σ ( A h 0 ) ∖ { 0 } (0 being algebraically simple). The map d 0 : V h 0 / R → B h 1 d^0:V_h^0/\mathbb{R}\to\mathcal{B}_h^1 d 0 : V h 0 / R → B h 1 is a bijection and, by (4.8), the operator A h 1 ∣ B h 1 A_h^1\big|_{\mathcal{B}_h^1} A h 1 B h 1 is similar to A h 0 ∣ V h 0 / R A_h^0\big|_{V_h^0/\mathbb{R}} A h 0 V h 0 / R . Hence σ ( A h 1 ∣ B h 1 ) = σ ( A h 0 ) ∖ { 0 } ⊂ { Re z > 0 } \sigma(A_h^1\big|_{\mathcal{B}_h^1})=\sigma(A_h^0)\setminus\{0\}\subset\{\operatorname{Re}z>0\} σ ( A h 1 B h 1 ) = σ ( A h 0 ) ∖ { 0 } ⊂ { Re z > 0 } , so A 1 = − A h 1 ∣ B h 1 A_1=-A_h^1\big|_{\mathcal{B}_h^1} A 1 = − A h 1 B h 1 generates an exponentially stable semigroup and is invertible on B h 1 \mathcal{B}_h^1 B h 1 .
By Duhamel’s formula and the exponential stability of A 1 A_1 A 1 , the same argument as in the proof of Theorem 5.2 (Part 3) shows that p h ( t ) → p ∞ : = − A 1 − 1 F ∞ ∈ B h 1 p_h(t)\to p_\infty:=-A_1^{-1}F_\infty\in\mathcal{B}_h^1 p h ( t ) → p ∞ := − A 1 − 1 F ∞ ∈ B h 1 exponentially for F ∞ : = lim t → ∞ F ( t ) = − d i β , h 1 h ∞ F_\infty:=\lim_{t\to\infty}F(t)=-di_{\beta,h}^1h_\infty F ∞ := lim t → ∞ F ( t ) = − d i β , h 1 h ∞ .
We conclude that
ω h ( t ) = p h ( t ) + h h ( t ) + q h ( t ) → p ∞ + h ∞ = : ω ∞ . \omega_h(t)=p_h(t)+h_h(t)+q_h(t)\to p_\infty+h_\infty=:\omega_\infty. ω h ( t ) = p h ( t ) + h h ( t ) + q h ( t ) → p ∞ + h ∞ =: ω ∞ .
Moreover, ω ∞ \omega_\infty ω ∞ is stationary. Since q h ( t ) q_h(t) q h ( t ) , d q h ( t ) → 0 dq_h(t)\to0 d q h ( t ) → 0 , the forcing converges to F ∞ F_\infty F ∞ , and p ∞ p_\infty p ∞ solves A 1 p ∞ + F ∞ = 0 A_1p_\infty+F_\infty=0 A 1 p ∞ + F ∞ = 0 . Using d p ∞ = 0 dp_\infty=0 d p ∞ = 0 and d h ∞ = 0 dh_\infty=0 d h ∞ = 0 , δ h h ∞ = 0 \delta_hh_\infty=0 δ h h ∞ = 0 , we compute
A h 1 ω ∞ = A h 1 p ∞ + A h 1 h ∞ = − A 1 p ∞ + d 0 i β , h 1 h ∞ = F ∞ − F ∞ = 0. A_h^1\omega_\infty=A_h^1p_\infty+A_h^1h_\infty=-A_1p_\infty+d^0i_{\beta,h}^1h_\infty=F_\infty-F_\infty=0. A h 1 ω ∞ = A h 1 p ∞ + A h 1 h ∞ = − A 1 p ∞ + d 0 i β , h 1 h ∞ = F ∞ − F ∞ = 0.
Thus ω h ( t ) \omega_h(t) ω h ( t ) converges exponentially to the closed stationary 1-form ω ∞ \omega_\infty ω ∞ .
If the manifold is simply connected, then b 1 ( M ) = 0 b_1(M)=0 b 1 ( M ) = 0 and S h 1 \mathfrak{S}_h^1 S h 1 is trivial, forcing h ∞ = 0 h_\infty=0 h ∞ = 0 and F ∞ = 0 F_\infty=0 F ∞ = 0 . Consequently, ω ∞ = p ∞ = 0 \omega_\infty=p_\infty=0 ω ∞ = p ∞ = 0 , completing the proof.
Long-time asymptotics under boundary conditions
The preceding results were stated for closed manifolds. We now describe how they are modified when M M M has a boundary. The arguments based only on the Hilbert-complex structure extend to de Rham complex with absolute or relative cohomology after replacing the corresponding cohomology spaces accordingly. The results Theorems 5.5 and 5.6 , however, require separate consideration, since the boundary conditions may eliminate the zero eigenvalue in degree 0 or n n n .
Since both the continuous formulation and its FEEC discretization are built upon the de Rham complex, it is natural to consider boundary conditions that are compatible with this structure. We therefore consider two standard types of boundary conditions: absolute and relative boundary conditions. In the mixed formulation, the absolute boundary conditions arise naturally from the variational formulation, whereas the relative boundary conditions are imposed strongly by restricting the function spaces to the subcomplex with vanishing traces. More precisely, we consider
Absolute boundary conditions:
tr ( ⋆ u ) = 0 , tr ( ⋆ d u ) = 0 , \operatorname{tr}(\star u)=0,\qquad\operatorname{tr}(\star du)=0, tr ( ⋆ u ) = 0 , tr ( ⋆ d u ) = 0 ,
with the mixed variables satisfying
( σ , u ) ∈ H Λ k − 1 ( M ) × H Λ k ( M ) . (\sigma,u)\in H\Lambda^{k-1}(M)\times H\Lambda^k(M). ( σ , u ) ∈ H Λ k − 1 ( M ) × H Λ k ( M ) .
Relative boundary conditions:
tr ( u ) = 0 , tr ( σ ) = 0 , \operatorname{tr}(u)=0,\qquad\operatorname{tr}(\sigma)=0, tr ( u ) = 0 , tr ( σ ) = 0 ,
with the mixed variables satisfying
( σ , u ) ∈ H ˚ Λ k − 1 ( M ) × H ˚ Λ k ( M ) . (\sigma,u)\in\mathring{H}\Lambda^{k-1}(M)\times\mathring{H}\Lambda^k(M). ( σ , u ) ∈ H ˚ Λ k − 1 ( M ) × H ˚ Λ k ( M ) .
These choices are consistent with the classical absolute and relative boundary conditions for the Hodge Laplacian and preserve the corresponding de Rham complex structures.
Remark 5.11 (Consistency with the cases k = 0 k=0 k = 0 and k = n k=n k = n ). This is consistent with the classical scalar cases and the standard boundary conditions used to establish well-posedness. Assume that ∂ M \partial M ∂ M is smooth and β ∈ W 1 , ∞ ( M ; T M ) \beta\in W^{1,\infty}(M;TM) β ∈ W 1 , ∞ ( M ; T M ) . In degree 0 0 0 , the relative and absolute conditions reduce respectively to u ∣ ∂ M = 0 u|_{\partial M}=0 u ∣ ∂ M = 0 , g ( ∇ g u , n ) = 0 g(\nabla_g u,n)=0 g ( ∇ g u , n ) = 0 where n n n denotes the outward unit normal and g g g is the Riemannian metric. Writing an n n n -form as v = ρ vol g v=\rho\operatorname{vol}_g v = ρ vol g , where vol g = ⋆ 1 \operatorname{vol}_g=\star1 vol g = ⋆ 1 , the complementary absolute condition is ρ ∣ ∂ M = 0 \rho|_{\partial M}=0 ρ ∣ ∂ M = 0 . The identity δ v = − ι ∇ g ρ vol g \delta v=-\iota_{\nabla_g\rho}\operatorname{vol}_g δ v = − ι ∇ g ρ vol g gives σ = η δ v + ι β v = − ι η ∇ g ρ − ρ β vol g \sigma=\eta\delta v+\iota_\beta v=-\iota_{\eta\nabla_g\rho-\rho\beta}\operatorname{vol}_g σ = η δ v + ι β v = − ι η ∇ g ρ − ρβ vol g . Since tr ( ι X vol g ) = g ( X , n ) vol ∂ M , g \operatorname{tr}(\iota_X\operatorname{vol}_g)=g(X,n)\operatorname{vol}_{\partial M,g} tr ( ι X vol g ) = g ( X , n ) vol ∂ M , g (see [38 Proposition 4.1.53] ), the relative conditions reduce to g ( η ∇ g ρ − ρ β , n ) = 0 g(\eta\nabla_g\rho-\rho\beta,n)=0 g ( η ∇ g ρ − ρβ , n ) = 0 , expressing vanishing normal flux.
Long-time asymptotics under absolute boundary conditions
We first consider the finite element de Rham complex, where no essential trace constraint is imposed:
V k = H Λ k ( M ) , V h k ⊂ H Λ k ( M ) . V^k=H\Lambda^k(M),\qquad V_h^k\subset H\Lambda^k(M). V k = H Λ k ( M ) , V h k ⊂ H Λ k ( M ) .
The associated cohomology is the absolute de Rham cohomology with dim H h k = b k ( M ) \dim\mathcal{H}_h^k=b_k(M) dim H h k = b k ( M ) . Accordingly, the general long-time result in Theorem 5.2 for sufficiently large diffusion coefficient η \eta η remains valid with b k ( M ) b_k(M) b k ( M ) interpreted as the absolute Betti number.
The main difference from the closed manifold case occurs in top degree. If M M M is connected and ∂ M ≠ ∅ \partial M\ne\varnothing ∂ M = ∅ , then b n ( M ) = 0 b_n(M)=0 b n ( M ) = 0 . Thus the top-degree zero mode is absent. Assumption 5.3 for k = n k=n k = n should be Re ( λ ) > 0 , ∀ λ ∈ σ ( A h n ) \operatorname{Re}(\lambda)>0,\ \forall\lambda\in\sigma(A_h^n) Re ( λ ) > 0 , ∀ λ ∈ σ ( A h n ) . In particular, the conclusion of Theorem 5.6 should be replaced by S h n = N ( A h n ) = { 0 } S_h^n=\mathcal{N}(A_h^n)=\{0\} S h n = N ( A h n ) = { 0 } , and every solution of the n n n -form evolution decays exponentially to zero. Equivalently, there is no nontrivial stationary n n n -form and total mass is not conserved in general under the natural boundary condition. The remaining structural conclusions are unchanged after replacing the cohomology spaces by the corresponding absolute cohomology spaces.
Long-time asymptotics under relative boundary conditions
The essential boundary conditions require tr ( u ) = tr ( σ ) = 0 \operatorname{tr}(u)=\operatorname{tr}(\sigma)=0 tr ( u ) = tr ( σ ) = 0 , therefore we should consider the finite element complex with homogeneous boundary conditions
V ˚ k = H ˚ Λ k ( M ) , V ˚ h k ⊂ H ˚ Λ k ( M ) . \mathring{V}^{k}=\mathring{H}\Lambda^{k}(M), \qquad\mathring{V}^{k}_{h}\subset\mathring{H}\Lambda^{k}(M). V ˚ k = H ˚ Λ k ( M ) , V ˚ h k ⊂ H ˚ Λ k ( M ) .
The relevant cohomology is the relative cohomology H k ( M , ∂ M ) H^{k}(M,\partial M) H k ( M , ∂ M ) . We denote the corresponding discrete harmonic space by H ˚ h k \mathring{\mathcal{H}}^{k}_{h} H ˚ h k satisfying dim H ˚ h k = b k ( M , ∂ M ) = b n − k ( M ) \dim\mathring{\mathcal{H}}^{k}_{h}=b_{k}(M,\partial M)=b_{n-k}(M) dim H ˚ h k = b k ( M , ∂ M ) = b n − k ( M ) . Thus the dimension statement in Theorem 5.2 becomes
dim S h k = b k ( M , ∂ M ) = b n − k ( M ) , \dim S^{k}_{h}=b_{k}(M,\partial M)=b_{n-k}(M), dim S h k = b k ( M , ∂ M ) = b n − k ( M ) ,
and the natural map S h k ⟶ H h k ( M , ∂ M ) : = Z ˚ h k / B ˚ h k S^{k}_{h}\longrightarrow H^{k}_{h}(M,\partial M):=\mathring{Z}^{k}_{h}/\mathring{B}^{k}_{h} S h k ⟶ H h k ( M , ∂ M ) := Z ˚ h k / B ˚ h k is an isomorphism.
In degree 0 0 0 , the essential boundary condition removes the constant zero mode. Indeed, if M M M is connected and ∂ M ≠ ∅ \partial M\ne\varnothing ∂ M = ∅ , then b 0 ( M , ∂ M ) = 0 b_{0}(M,\partial M)=0 b 0 ( M , ∂ M ) = 0 . The Assumption 5.3 for k = 0 k=0 k = 0 should be Re ( λ ) > 0 \operatorname{Re}(\lambda)>0 Re ( λ ) > 0 , ∀ λ ∈ σ ( A h 0 ) \forall\lambda\in\sigma(A^{0}_{h}) ∀ λ ∈ σ ( A h 0 ) . Consequently, S h 0 = N ( A h 0 ) = { 0 } S^{0}_{h}=\mathcal{N}(A^{0}_{h})=\{0\} S h 0 = N ( A h 0 ) = { 0 } , and every solution of the 0 0 0 -form evolution decays exponentially to zero. The conclusion of Theorem 5.5 should be modified accordingly under essential boundary conditions. Theorem 5.6 and Theorem 5.9 remain valid. The corresponding result for closed stationary 1 1 1 -forms should also be interpreted in relative cohomology. More precisely,
S h 1 ≃ H ˚ h 1 , dim S h 1 = b 1 ( M , ∂ M ) = b n − 1 ( M ) , S^{1}_{h}\simeq\mathring{\mathcal{H}}^{1}_{h}, \qquad\dim S^{1}_{h}=b_{1}(M,\partial M)=b_{n-1}(M), S h 1 ≃ H ˚ h 1 , dim S h 1 = b 1 ( M , ∂ M ) = b n − 1 ( M ) ,
and
S h 1 ∩ B ˚ h 1 = { 0 } . S^{1}_{h}\cap\mathring{B}^{1}_{h}=\{0\}. S h 1 ∩ B ˚ h 1 = { 0 } .
Hence each relative cohomology class contains a unique closed stationary 1 1 1 -form, and the evolution of any closed initial 1 1 1 -form converges exponentially to that representative.
Finally, in two dimensions the degrees 1 1 1 and n − 1 n-1 n − 1 coincide. Hence the finite element antidynamo conclusion remains valid under either natural or essential boundary conditions, with the corresponding absolute or relative cohomology space: every discrete 1 1 1 -form solution converges exponentially to a closed stationary 1 1 1 -form.
Verification of the assumptions
In this section, we investigate the conditions under which the critical assumptions for the finite element antidynamo theorem are satisfied for closed manifolds. The first assumption is the L 2 L^{2} L 2 boundedness of the discrete contraction operator. The second assumption involves the spectral preservation of the discretization. For sufficiently small h h h , this follows from spectral convergence under the usual approximation and regularity assumptions. We also explore whether an M M M -matrix discretization can be derived to guarantee the desired spectral distribution for arbitrary mesh size h h h . The results are summarized in Table 6.1 .
Table 6.1. Choices used to verify the assumptions in Table 5.1 .
Existence of the zero eigenvalue
First, we establish the existence of the zero eigenvalue at the discrete level for any h > 0 h > 0 h > 0 and closed manifold M M M .
For k = 0 k = 0 k = 0 , the discrete space V h 0 V_h^0 V h 0 contains globally constant functions. Since the discrete exterior derivative satisfies d 0 1 = 0 d^0 1 = 0 d 0 1 = 0 , it follows that A h 0 1 = 0 A_h^0 1 = 0 A h 0 1 = 0 . Thus, the zero eigenvalue is present for 0-forms.
For k = n k = n k = n , let m h ∈ V h n m_h \in V_h^n m h ∈ V h n be the Riesz representation of the integration functional, i.e.,
( m h , v h ) = ∫ M v h , ∀ v h ∈ V h n . (81) (m_h,v_h) = \int_M v_h,\qquad\forall v_h \in V_h^n. \tag*{(81)} ( m h , v h ) = ∫ M v h , ∀ v h ∈ V h n . (81)
Since the integration functional is not zero, we have m h ≠ 0 m_h \ne0 m h = 0 . The mass conservation property in (5.13) gives ∫ M A h n ω h = 0 \int_M A_h^n\omega_h = 0 ∫ M A h n ω h = 0 for all ω h ∈ V h n \omega_h \in V_h^n ω h ∈ V h n . Equivalently,
( m h , A h n ω h ) = ∫ M A h n ω h = 0 , ∀ ω h ∈ V h n . (82) (m_h,A_h^n\omega_h) = \int_M A_h^n\omega_h = 0,\qquad\forall\omega_h \in V_h^n. \tag*{(82)} ( m h , A h n ω h ) = ∫ M A h n ω h = 0 , ∀ ω h ∈ V h n . (82)
Thus m h ∈ N ( ( A h n ) ∗ ) m_h \in\mathcal{N}((A_h^n)^*) m h ∈ N (( A h n ) ∗ ) , where ( A h n ) ∗ (A_h^n)^* ( A h n ) ∗ denotes the discrete adjoint of A h n A_h^n A h n . Hence, A h n A_h^n A h n is not surjective. Since A h n A_h^n A h n acts on the finite-dimensional space V h n V_h^n V h n , it is not invertible. Therefore λ = 0 \lambda= 0 λ = 0 is an eigenvalue of A h n A_h^n A h n for every mesh size h h h .
Regarding Assumption 5.3, while we have confirmed the existence of the zero eigenvalue for k ∈ { 0 , n } k \in\{0,n\} k ∈ { 0 , n } for any h > 0 h > 0 h > 0 , it remains to be verified that this eigenvalue is algebraically simple and that the remaining spectrum resides strictly in the right half of the complex plane.
Choice 1: spectrum convergence for sufficiently small h h h
In this subsection, we verify Assumption 5.1 for every h > 0 h > 0 h > 0 and Assumption 5.3 for sufficiently small h h h . We assume that β ∈ L ∞ ( M ) \beta\in L^\infty(M) β ∈ L ∞ ( M ) and that H 1 ( M ) ↪ ↪ L 2 ( M ) H^1(M) \hookrightarrow\hookrightarrow L^2(M) H 1 ( M ) ↪↪ L 2 ( M ) . The Rellich compact embedding holds for closed manifolds and, more generally, for compact manifolds with Lipschitz boundary. For the discrete contraction, we define i β , h k = Q h k − 1 i β k i_{\beta,h}^k = Q_h^{k-1}i_\beta^k i β , h k = Q h k − 1 i β k , where Q h k − 1 Q_h^{k-1} Q h k − 1 is the L 2 L^2 L 2 -projection onto V h k − 1 V_h^{k-1} V h k − 1 .
This is equivalent to applying the contraction operator i β i_\beta i β directly within the discrete variational formulation, since
( i β , h k u h , v h ) = ( Q h k − 1 i β k u h , v h ) = ( i β k u h , v h ) , ∀ u h ∈ V h k , v h ∈ V h k − 1 . (83) (i_{\beta,h}^k u_h,v_h) = (Q_h^{k-1}i_\beta^k u_h,v_h) = (i_\beta^k u_h,v_h),\qquad\forall u_h \in V_h^k,\ v_h \in V_h^{k-1}. \tag*{(83)} ( i β , h k u h , v h ) = ( Q h k − 1 i β k u h , v h ) = ( i β k u h , v h ) , ∀ u h ∈ V h k , v h ∈ V h k − 1 . (83)
This choice naturally inherits L 2 L^2 L 2 boundedness, ensuring that Assumption 5.1 holds through the stability of the L 2 L^2 L 2 -projection
∥ i β , h k u h ∥ = ∥ Q h k − 1 i β k u h ∥ ≤ ∥ i β k u h ∥ ≤ C β ∥ u h ∥ , (84) \|i_{\beta,h}^k u_h\| = \|Q_h^{k-1}i_\beta^k u_h\| \le\|i_\beta^k u_h\| \le C_\beta\|u_h\|, \tag*{(84)} ∥ i β , h k u h ∥ = ∥ Q h k − 1 i β k u h ∥ ≤ ∥ i β k u h ∥ ≤ C β ∥ u h ∥ , (84)
for β ∈ L ∞ \beta\in L^\infty β ∈ L ∞ .
Regarding Assumption 5.3, we have the following result.
Lemma 6.1. Let T h \mathcal{T}_h T h be a shape regular family, and define i β , h k = Q h k − 1 i β k i_{\beta,h}^k = Q_h^{k-1}i_\beta^k i β , h k = Q h k − 1 i β k . Assume that β ∈ L ∞ ( M ) \beta\in L^\infty(M) β ∈ L ∞ ( M ) . There exists a threshold h 0 > 0 h_0 > 0 h 0 > 0 depending on η \eta η and β \beta β such that for all h < h 0 h < h_0 h < h 0 , Assumption 5.3 is satisfied.
The proof is organized around a classical spectral convergence theorem for families of operators, which we state first. Throughout, μ = μ k > 0 \mu= \mu_k > 0 μ = μ k > 0 denotes the spectral gap of the continuous operator A k = Δ η , β k A^k = \Delta_{\eta,\beta}^k A k = Δ η , β k provided by Theorem 2.1 (for k = n k = n k = n through the duality with the 0-form operator with reversed drift, whose hypotheses are invariant under β ↦ − β \beta\mapsto-\beta β ↦ − β ): the spectrum of A k A^k A k is contained in { 0 } ∪ { Re z ≥ μ } \{0\} \cup\{\operatorname{Re} z \ge\mu\} { 0 } ∪ { Re z ≥ μ } and 0 is an algebraically simple eigenvalue. We consider a shift c c c for k ∈ { 0 , n } k \in\{0,n\} k ∈ { 0 , n } to be determined later and write A ~ k : = A k + c I \widetilde{A}^{k}:=A^{k}+cI A k := A k + c I , A ~ h k : = A h k + c I \widetilde{A}_{h}^{k}:=A_{h}^{k}+cI A h k := A h k + c I .
The spectral-preservation result is based on the following classical theorem for collectively compact operator approximations; see [2 ] .
Theorem 6.2 (spectral convergence under collective compactness). Let X X X be a complex Hilbert space, let T ∈ L ( X ) T \in\mathcal{L}(X) T ∈ L ( X ) be compact, and let { T h } 0 < h ≤ h ˉ ⊂ L ( X ) \{T_h\}_{0<h\leq\bar{h}} \subset\mathcal{L}(X) { T h } 0 < h ≤ h ˉ ⊂ L ( X ) satisfy
(H1) pointwise convergence: T h f → T f T_hf \to Tf T h f → T f in X X X for every f ∈ X f \in X f ∈ X ;
(H2) collective compactness: for every sequence h j → 0 h_j \to0 h j → 0 and every bounded sequence ( f j ) ⊂ X (f_j) \subset X ( f j ) ⊂ X , the sequence ( T h j f j ) (T_{h_j}f_j) ( T h j f j ) contains a strongly convergent subsequence.
The spectrum of T h T_h T h converges to that of T T T in the following sense. Let Ω ⊂ C \Omega\subset\mathbb{C} Ω ⊂ C be open and bounded with rectifiable Jordan boundary Γ \Gamma Γ such that Γ ∩ σ ( T ) = ∅ \Gamma\cap\sigma(T)=\varnothing Γ ∩ σ ( T ) = ∅ and 0 ∉ Ω ‾ 0 \notin\overline{\Omega} 0 ∈ / Ω . Then there exists h Γ > 0 h_{\Gamma}>0 h Γ > 0 such that for all h < h Γ h<h_{\Gamma} h < h Γ :
Γ ⊂ ρ ( T h ) \Gamma\subset\rho(T_h) Γ ⊂ ρ ( T h ) and sup z ∈ Γ ∥ ( z I − T h ) − 1 ∥ ≤ C Γ < ∞ \sup_{z\in\Gamma}\|(zI-T_h)^{-1}\| \leq C_{\Gamma}<\infty sup z ∈ Γ ∥ ( z I − T h ) − 1 ∥ ≤ C Γ < ∞ ;
the Riesz spectral projections E : = 1 2 π i ∮ Γ ( z I − T ) − 1 d z E:=\frac{1}{2\pi i}\oint_{\Gamma}(zI-T)^{-1}\,dz E := 2 π i 1 ∮ Γ ( z I − T ) − 1 d z and E h : = 1 2 π i ∮ Γ ( z I − T h ) − 1 d z E_h:=\frac{1}{2\pi i}\oint_{\Gamma}(zI-T_h)^{-1}\,dz E h := 2 π i 1 ∮ Γ ( z I − T h ) − 1 d z satisfy ∥ ( E h − E ) ∣ R ( E ) ∥ → 0 \|(E_h-E)|_{R(E)}\|\to0 ∥ ( E h − E ) ∣ R ( E ) ∥ → 0 and rank E h = rank E < ∞ \operatorname{rank}E_h=\operatorname{rank}E<\infty rank E h = rank E < ∞ ;
consequently, the total algebraic multiplicity of σ ( T h ) ∩ Ω \sigma(T_h)\cap\Omega σ ( T h ) ∩ Ω equals that of σ ( T ) ∩ Ω \sigma(T)\cap\Omega σ ( T ) ∩ Ω , and σ ( T h ) ∩ Ω → σ ( T ) ∩ Ω \sigma(T_h)\cap\Omega\to\sigma(T)\cap\Omega σ ( T h ) ∩ Ω → σ ( T ) ∩ Ω in the Hausdorff sense.
We therefore apply Theorem 6.2 to the shifted resolvent operators
T h = ( A ~ h k ) − 1 , T = ( A ~ k ) − 1 , k ∈ { 0 , n } . (85) T_h=(\widetilde{A}_{h}^{k})^{-1},\qquad T=(\widetilde{A}^{k})^{-1},\qquad k\in\{0,n\}. \tag*{(85)} T h = ( A h k ) − 1 , T = ( A k ) − 1 , k ∈ { 0 , n } . (85)
Once (H1)–(H2) are established for T h T_h T h , the theorem yields spectral convergence of the shifted resolvents. Through the spectral mapping
λ ∈ σ ( A h k ) ⟺ ( λ + c ) − 1 ∈ σ ( T h ) , (86) \lambda\in\sigma(A_{h}^{k})\Longleftrightarrow(\lambda+c)^{-1}\in\sigma(T_h), \tag*{(86)} λ ∈ σ ( A h k ) ⟺ ( λ + c ) − 1 ∈ σ ( T h ) , (86)
this gives the corresponding convergence and preservation of the eigenvalues of A h k A_h^k A h k for sufficiently small h h h .
Firstly, for k = 0 k=0 k = 0 . Let X : = L 2 Λ 0 ( M ) X:=L^2\Lambda^0(M) X := L 2 Λ 0 ( M ) and consider the shifted forms
a c ( u , v ) : = η ( d u , d v ) + ( i β d u , v ) + c ( u , v ) , u , v ∈ H Λ 0 ( M ) . (87) a_c(u,v):=\eta(du,dv)+(i_{\beta}du,v)+c(u,v),\qquad u,v\in H\Lambda^0(M). \tag*{(87)} a c ( u , v ) := η ( d u , d v ) + ( i β d u , v ) + c ( u , v ) , u , v ∈ H Λ 0 ( M ) . (87)
Define T ∈ L ( X ) T\in\mathcal{L}(X) T ∈ L ( X ) by T f : = u Tf:=u T f := u , where u ∈ H Λ 0 ( M ) u\in H\Lambda^0(M) u ∈ H Λ 0 ( M ) solves a c ( u , v ) = ( f , v ) a_c(u,v)=(f,v) a c ( u , v ) = ( f , v ) for all v ∈ H Λ 0 ( M ) v\in H\Lambda^0(M) v ∈ H Λ 0 ( M ) , and let T h f : = u h ∈ V h 0 T_hf:=u_h\in V_h^0 T h f := u h ∈ V h 0 be its Galerkin approximation. Since
∣ a c ( u , u ) ∣ ≥ η ∥ d u ∥ 2 − C β ∥ d u ∥ ∥ u ∥ + c ∥ u ∥ 2 ≥ η 2 ∥ d u ∥ 2 + ( c − C β 2 2 η ) ∥ u ∥ 2 ≥ α ∥ u ∥ H Λ 0 2 |a_c(u,u)|\geq\eta\|du\|^2-C_{\beta}\|du\|\|u\|+c\|u\|^2\geq\frac{\eta}{2}\|du\|^2+\left(c-\frac{C_{\beta}^2}{2\eta}\right)\|u\|^2\geq\alpha\|u\|_{H\Lambda^0}^2 ∣ a c ( u , u ) ∣ ≥ η ∥ d u ∥ 2 − C β ∥ d u ∥∥ u ∥ + c ∥ u ∥ 2 ≥ 2 η ∥ d u ∥ 2 + ( c − 2 η C β 2 ) ∥ u ∥ 2 ≥ α ∥ u ∥ H Λ 0 2
with α : = min { η / 2 , c − C β 2 / ( 2 η ) } > 0 \alpha:=\min\{\eta/2,c-C_{\beta}^2/(2\eta)\}>0 α := min { η /2 , c − C β 2 / ( 2 η )} > 0 if c ≥ C β 2 / η c\geq C_{\beta}^2/\eta c ≥ C β 2 / η , the form a c a_c a c is uniformly coercive on H Λ 0 H\Lambda^0 H Λ 0 and on every subspace V h 0 V_h^0 V h 0 . Then both problems are well posed for every h h h , with
∥ T f ∥ H Λ 0 + ∥ T h f ∥ H Λ 0 ≤ 2 α ∥ f ∥ . (88) \|Tf\|_{H\Lambda^0}+\|T_hf\|_{H\Lambda^0}\leq\frac{2}{\alpha}\|f\|. \tag*{(88)} ∥ T f ∥ H Λ 0 + ∥ T h f ∥ H Λ 0 ≤ α 2 ∥ f ∥. (88)
(H1). By Céa’s lemma, we have ∥ T f − T h f ∥ H Λ 0 ≤ M a α inf v h ∈ V h 0 ∥ T f − v h ∥ H Λ 0 \|Tf-T_hf\|_{H\Lambda^0}\leq\frac{M_a}{\alpha}\inf_{v_h\in V_h^0}\|Tf-v_h\|_{H\Lambda^0} ∥ T f − T h f ∥ H Λ 0 ≤ α M a inf v h ∈ V h 0 ∥ T f − v h ∥ H Λ 0 with the continuity constant M a M_a M a of a c a_c a c . The right-hand side tends to 0 for each fixed f f f by the approximation property of the family { V h 0 } \{V_h^0\} { V h 0 } [5 ] .
(H2) and compactness of T T T . By (88) , for any h j → 0 h_j\to0 h j → 0 and ∥ f j ∥ ≤ 1 \|f_j\|\leq1 ∥ f j ∥ ≤ 1 the sequence T h j f j T_{h_j}f_j T h j f j is bounded in H 1 ( M ) H^1(M) H 1 ( M ) . The compact embedding H 1 ( M ) ↪ ↪ L 2 ( M ) H^1(M)\hookrightarrow\hookrightarrow L^2(M) H 1 ( M ) ↪↪ L 2 ( M ) extracts a strongly convergent subsequence. The compact embedding also shows that T T T is compact.
Then for k = n k=n k = n . Let X : = L 2 Λ n ( M ) X := L^2\Lambda^n(M) X := L 2 Λ n ( M ) . The shifted problem in mixed form reads: given f ∈ X f \in X f ∈ X , find ( σ , u ) ∈ H Λ n − 1 ( M ) × L 2 Λ n ( M ) (\sigma,u) \in H\Lambda^{n-1}(M) \times L^2\Lambda^n(M) ( σ , u ) ∈ H Λ n − 1 ( M ) × L 2 Λ n ( M ) with
( σ , τ ) − ( i β u , τ ) − η ( u , d τ ) = 0 ∀ τ ∈ H Λ n − 1 ( M ) , ( d σ , v ) + c ( u , v ) = ( f , v ) ∀ v ∈ L 2 Λ n ( M ) , (89) \begin{aligned}
(\sigma,\tau) - (i\beta u,\tau) - \eta(u,d\tau) &= 0 &&\forall\tau\in H\Lambda^{n-1}(M),\\
(d\sigma,v) + c(u,v) &= (f,v) &&\forall v \in L^2\Lambda^n(M), \tag*{(89)}
\end{aligned} ( σ , τ ) − ( i β u , τ ) − η ( u , d τ ) ( d σ , v ) + c ( u , v ) = 0 = ( f , v ) ∀ τ ∈ H Λ n − 1 ( M ) , ∀ v ∈ L 2 Λ n ( M ) , (89)
and set T f : = u Tf := u T f := u ; the discrete solution operator T h f : = u h T_hf := u_h T h f := u h is defined by the same equations on V h n − 1 × V h n V_h^{n-1} \times V_h^n V h n − 1 × V h n , with i β i\beta i β replaced by i β h = Q h n − 1 i β i\beta_h = Q_h^{n-1}i\beta i β h = Q h n − 1 i β (the projection is invisible in (89) tested against τ h ∈ V h n − 1 \tau_h \in V_h^{n-1} τ h ∈ V h n − 1 ). Writing b ( ( σ , u ) , ( τ , v ) ) b((\sigma,u),(\tau,v)) b (( σ , u ) , ( τ , v )) for the sum of the two left-hand sides, the choice of test functions
( τ , v ) : = ( σ , η u + γ d σ ) , γ : = η 2 c , (90) (\tau,v) := (\sigma,\eta u+\gamma d\sigma), \qquad\gamma:= \frac{\eta}{2c}, \tag*{(90)} ( τ , v ) := ( σ , η u + γ d σ ) , γ := 2 c η , (90)
gives, after absorbing the advective and coupling terms by Young’s inequality and selecting c ≥ C β 2 / η c \ge C_\beta^2/\eta c ≥ C β 2 / η ,
∣ b ( ( σ , u ) , ( τ , v ) ) ∣ ≥ 1 2 ∥ σ ∥ 2 + γ 2 ∥ d σ ∥ 2 + η c 4 ∥ u ∥ 2 , ∥ ( τ , v ) ∥ H Λ × L 2 ≤ C ∥ ( σ , u ) ∥ H Λ × L 2 , (91) \lvert b((\sigma,u),(\tau,v))\rvert\ge\frac{1}{2}\lVert\sigma\rVert^2 + \frac{\gamma}{2}\lVert d\sigma\rVert^2 + \frac{\eta c}{4}\lVert u\rVert^2,\qquad\lVert(\tau,v)\rVert_{H\Lambda\times L^2} \le C\lVert(\sigma,u)\rVert_{H\Lambda\times L^2}, \tag*{(91)} ∣ b (( σ , u ) , ( τ , v ))∣ ≥ 2 1 ∥ σ ∥ 2 + 2 γ ∥ d σ ∥ 2 + 4 η c ∥ u ∥ 2 , ∥( τ , v ) ∥ H Λ × L 2 ≤ C ∥( σ , u ) ∥ H Λ × L 2 , (91)
i.e. a uniform inf-sup condition. Hence (91) holds on V h n − 1 × V h n V_h^{n-1} \times V_h^n V h n − 1 × V h n , with constants independent of h h h . In particular both problems are well posed for every h h h , with
∥ σ ∥ H Λ + ∥ u ∥ + ∥ σ h ∥ H Λ + ∥ u h ∥ ≤ C s ∥ f ∥ . (92) \lVert\sigma\rVert_{H\Lambda}+\lVert u\rVert+\lVert\sigma_h\rVert_{H\Lambda}+\lVert u_h\rVert\le C_s\lVert f\rVert. \tag*{(92)} ∥ σ ∥ H Λ + ∥ u ∥ + ∥ σ h ∥ H Λ + ∥ u h ∥ ≤ C s ∥ f ∥ . (92)
(H1). By the standard argument of mixed finite elements, by (91) , we have
∥ σ − σ h ∥ H Λ + ∥ u − u h ∥ ≤ C inf ( τ h , v h ) ∈ V h n − 1 × V h n ( ∥ σ − τ h ∥ H Λ + ∥ u − v h ∥ ) . \lVert\sigma-\sigma_h\rVert_{H\Lambda}+\lVert u-u_h\rVert\le C\inf_{(\tau_h,v_h)\in V_h^{n-1}\times V_h^n}\left(\lVert\sigma-\tau_h\rVert_{H\Lambda}+\lVert u-v_h\rVert\right). ∥ σ − σ h ∥ H Λ + ∥ u − u h ∥ ≤ C ( τ h , v h ) ∈ V h n − 1 × V h n inf ( ∥ σ − τ h ∥ H Λ + ∥ u − v h ∥ ) .
with C C C independent of h h h , and the right-hand side tends to 0 for each fixed f f f by density of smooth forms and the approximation property of the finite element complex [5 ] .
H ∗ Λ n H^{*}\Lambda^n H ∗ Λ n regularity. The first equation in (89) gives ( u , d τ ) = ( g , τ ) (u,d\tau)=(g,\tau) ( u , d τ ) = ( g , τ ) for g = ( σ − i β u ) / η ∈ L 2 Λ n − 1 g=(\sigma-i\beta u)/\eta\in L^2\Lambda^{n-1} g = ( σ − i β u ) / η ∈ L 2 Λ n − 1 and ∀ τ ∈ H Λ n − 1 \forall\tau\in H\Lambda^{n-1} ∀ τ ∈ H Λ n − 1 , thus u ∈ H ∗ Λ n u \in H^{*}\Lambda^n u ∈ H ∗ Λ n and δ u = g \delta u=g δ u = g . This also implies that
∥ u ∥ H ∗ Λ n ≲ ∥ u ∥ + ∥ σ ∥ ≲ ∥ f ∥ . (93) \lVert u\rVert_{H^{*}\Lambda^n} \lesssim\lVert u\rVert+\lVert\sigma\rVert\lesssim\lVert f\rVert. \tag*{(93)} ∥ u ∥ H ∗ Λ n ≲ ∥ u ∥ + ∥ σ ∥ ≲ ∥ f ∥ . (93)
Together with the compact embedding H Λ n ∩ H ∗ Λ n ↪ ↪ L 2 Λ n H\Lambda^n \cap H^{*}\Lambda^n \hookrightarrow\hookrightarrow L^2\Lambda^n H Λ n ∩ H ∗ Λ n ↪↪ L 2 Λ n , this shows in particular that T : X → X T:X\to X T : X → X is compact.
(H2). We now prove collective compactness of the discrete solution operators. Let h j → 0 h_j \to0 h j → 0 , let { f j } ⊂ X \{f_j\}\subset X { f j } ⊂ X be bounded, and let
( σ j , u j ) : = ( σ h j , T h j f j ) ∈ V h j n − 1 × V h j n (\sigma_j,u_j):=(\sigma_{h_j},T_{h_j}f_j)\in V_{h_j}^{n-1}\times V_{h_j}^n ( σ j , u j ) := ( σ h j , T h j f j ) ∈ V h j n − 1 × V h j n
denote the corresponding discrete solutions. By (92) ,
sup j ( ∥ σ j ∥ H Λ + ∥ u j ∥ ) < ∞ . (94) \sup_j\left(\lVert\sigma_j\rVert_{H\Lambda}+\lVert u_j\rVert\right)<\infty. \tag*{(94)} j sup ( ∥ σ j ∥ H Λ + ∥ u j ∥ ) < ∞. (94)
By the definition of the discrete coderivative, the first discrete equation is equivalent to
σ j = i β , h j n u j + η δ h j n u j . \sigma_j=i_{\beta,h_j}^{n}u_j+\eta\delta_{h_j}^{n}u_j. σ j = i β , h j n u j + η δ h j n u j .
It follows from the uniform boundedness of ι β , h j n \iota^{n}_{\beta,h_j} ι β , h j n that
∥ δ h j n u j ∥ ≤ ∥ σ j ∥ + ∥ ι β , h j n u j ∥ ≤ ∥ σ j ∥ + C β ∥ u j ∥ , (95) \|\delta^{n}_{h_j}u_j\| \le\|\sigma_j\|+\|\iota^{n}_{\beta,h_j}u_j\|
\le\|\sigma_j\|+C_\beta\|u_j\|, \tag*{(95)} ∥ δ h j n u j ∥ ≤ ∥ σ j ∥ + ∥ ι β , h j n u j ∥ ≤ ∥ σ j ∥ + C β ∥ u j ∥ , (95)
and therefore
sup j ( ∥ u j ∥ + ∥ δ h j n u j ∥ ) < ∞ . (96) \sup_j\left(\|u_j\|+\|\delta^{n}_{h_j}u_j\|\right)<\infty. \tag*{(96)} j sup ( ∥ u j ∥ + ∥ δ h j n u j ∥ ) < ∞. (96)
For the standard FEEC families considered here on shape-regular mesh families, the uniformly L 2 L^2 L 2 -bounded commuting projections and the compact embedding
H ∗ Λ n ( M ) ≃ H 1 Λ 0 ( M ) ↪ ↪ L 2 Λ 0 ( M ) ≃ L 2 Λ n ( M ) H^*\Lambda^n(M)\simeq H^1\Lambda^0(M)\hookrightarrow\hookrightarrow L^2\Lambda^0(M)\simeq L^2\Lambda^n(M) H ∗ Λ n ( M ) ≃ H 1 Λ 0 ( M ) ↪↪ L 2 Λ 0 ( M ) ≃ L 2 Λ n ( M )
yield the top-degree discrete Rellich compactness property in the δ h n \delta^n_h δ h n -graph norm; see [3 , 14 ] . Since discrete compactness results are often formulated in terms of d d d , we briefly verify the corresponding δ h n \delta^n_h δ h n -version needed here.
Let π h ℓ : L 2 Λ ℓ ( M ) ⟶ V h ℓ \pi^\ell_h:L^2\Lambda^\ell(M)\longrightarrow V^\ell_h π h ℓ : L 2 Λ ℓ ( M ) ⟶ V h ℓ for ℓ = n − 1 , n \ell=n-1,n ℓ = n − 1 , n be uniformly L 2 L^2 L 2 -bounded commuting projections satisfying π h ℓ ∣ V h ℓ = I \pi^\ell_h|_{V^\ell_h}=I π h ℓ ∣ V h ℓ = I and d n − 1 π h n − 1 = π h n d n − 1 d^{n-1}\pi^{n-1}_h=\pi^n_hd^{n-1} d n − 1 π h n − 1 = π h n d n − 1 . We define their L 2 L^2 L 2 -adjoints
R h ℓ : V h ℓ ⟶ L 2 Λ ℓ ( M ) , R h ℓ : = ( π h ℓ ) ∗ , (97) R^\ell_h:V^\ell_h\longrightarrow L^2\Lambda^\ell(M),\qquad R^\ell_h:=(\pi^\ell_h)^*, \tag*{(97)} R h ℓ : V h ℓ ⟶ L 2 Λ ℓ ( M ) , R h ℓ := ( π h ℓ ) ∗ , (97)
by
( w , R h ℓ v h ) = ( π h ℓ w , v h ) , ∀ w ∈ L 2 Λ ℓ ( M ) , ∀ v h ∈ V h ℓ . (98) (w,R^\ell_hv_h)=(\pi^\ell_hw,v_h),\qquad\forall w\in L^2\Lambda^\ell(M),\quad\forall v_h\in V^\ell_h. \tag*{(98)} ( w , R h ℓ v h ) = ( π h ℓ w , v h ) , ∀ w ∈ L 2 Λ ℓ ( M ) , ∀ v h ∈ V h ℓ . (98)
By the Riesz representation theorem, R h ℓ R^\ell_h R h ℓ is well defined, and the uniform L 2 L^2 L 2 boundedness of π h ℓ \pi^\ell_h π h ℓ gives
∥ R h ℓ v h ∥ ≤ ∥ π h ℓ ∥ L ( L 2 , L 2 ) ∥ v h ∥ ≤ C π ∥ v h ∥ . (99) \|R^\ell_hv_h\|\leq\|\pi^\ell_h\|_{L(L^2,L^2)}\|v_h\|\leq C_\pi\|v_h\|. \tag*{(99)} ∥ R h ℓ v h ∥ ≤ ∥ π h ℓ ∥ L ( L 2 , L 2 ) ∥ v h ∥ ≤ C π ∥ v h ∥. (99)
Moreover, for every τ ∈ H Λ n − 1 \tau\in H\Lambda^{n-1} τ ∈ H Λ n − 1 and v h ∈ V h n v_h\in V^n_h v h ∈ V h n ,
( d n − 1 τ , R h n v h ) = ( π h n d n − 1 τ , v h ) = ( d n − 1 π h n − 1 τ , v h ) = ( π h n − 1 τ , δ h n v h ) = ( τ , R h n − 1 δ h n v h ) . (100) \begin{aligned}
(d^{n-1}\tau,R^n_hv_h)&=(\pi^n_hd^{n-1}\tau,v_h)=(d^{n-1}\pi^{n-1}_h\tau,v_h)\\
&=(\pi^{n-1}_h\tau,\delta^n_hv_h)=(\tau,R^{n-1}_h\delta^n_hv_h). \tag*{(100)}
\end{aligned} ( d n − 1 τ , R h n v h ) = ( π h n d n − 1 τ , v h ) = ( d n − 1 π h n − 1 τ , v h ) = ( π h n − 1 τ , δ h n v h ) = ( τ , R h n − 1 δ h n v h ) . (100)
Since δ n = ( d n − 1 ) ∗ \delta^n=(d^{n-1})^* δ n = ( d n − 1 ) ∗ and the right-hand side is bounded with respect to the L 2 L^2 L 2 -norm of τ \tau τ : ∣ ( τ , R h n − 1 δ h n u h ) ∣ ≤ C π ∥ δ h n u h ∥ ∥ τ ∥ |(\tau,R^{n-1}_h\delta^n_hu_h)|\leq C_\pi\|\delta^n_hu_h\|\|\tau\| ∣ ( τ , R h n − 1 δ h n u h ) ∣ ≤ C π ∥ δ h n u h ∥∥ τ ∥ , this proves
R h n v h ∈ H ∗ Λ n ( M ) , δ n R h n v h = R h n − 1 δ h n v h , (101) R^n_hv_h\in H^*\Lambda^n(M),\qquad\delta^nR^n_hv_h=R^{n-1}_h\delta^n_hv_h, \tag*{(101)} R h n v h ∈ H ∗ Λ n ( M ) , δ n R h n v h = R h n − 1 δ h n v h , (101)
and hence
∥ R h n v h ∥ H ∗ Λ n 2 = ∥ R h n v h ∥ 2 + ∥ δ n R h n v h ∥ 2 = ∥ R h n v h ∥ 2 + ∥ R h n − 1 δ h n v h ∥ 2 ≤ C π 2 ( ∥ v h ∥ 2 + ∥ δ h n v h ∥ 2 ) . \begin{aligned}
\|R^n_hv_h\|^2_{H^*\Lambda^n}
&=\|R^n_hv_h\|^2+\|\delta^nR^n_hv_h\|^2\\
&=\|R^n_hv_h\|^2+\|R^{n-1}_h\delta^n_hv_h\|^2\\
&\leq C_\pi^2\left(\|v_h\|^2+\|\delta^n_hv_h\|^2\right).
\end{aligned} ∥ R h n v h ∥ H ∗ Λ n 2 = ∥ R h n v h ∥ 2 + ∥ δ n R h n v h ∥ 2 = ∥ R h n v h ∥ 2 + ∥ R h n − 1 δ h n v h ∥ 2 ≤ C π 2 ( ∥ v h ∥ 2 + ∥ δ h n v h ∥ 2 ) .
Taking v h = u j v_h=u_j v h = u j and using (96) , we have that { R h j n u j } \{R^n_{h_j}u_j\} { R h j n u j } is bounded in H ∗ Λ n ( M ) H^*\Lambda^n(M) H ∗ Λ n ( M ) . The compact embedding above therefore gives, after passing to a subsequence,
R h j n u j ⟶ u strongly in L 2 Λ n ( M ) as h j → 0 (102) R^n_{h_j}u_j\longrightarrow u\qquad\text{strongly in }L^2\Lambda^n(M)\text{ as }h_j\to0 \tag*{(102)} R h j n u j ⟶ u strongly in L 2 Λ n ( M ) as h j → 0 (102)
for some u ∈ L 2 Λ n ( M ) u \in L^2\Lambda^n(M) u ∈ L 2 Λ n ( M ) .
Let Q h n Q_h^n Q h n be the L 2 L^2 L 2 -orthogonal projection onto V h n V_h^n V h n . Because π h n \pi_h^n π h n is a projection onto V h n V_h^n V h n , one has
Q h n R h n = I on V h n . (103) Q_h^n R_h^n = I \quad\text{on } V_h^n. \tag*{(103)} Q h n R h n = I on V h n . (103)
Indeed, for v h , w h ∈ V h n v_h,w_h \in V_h^n v h , w h ∈ V h n ,
( w h , Q h n R h n v h ) = ( π h n w h , v h ) = ( w h , v h ) . (w_h,Q_h^nR_h^nv_h)=(\pi_h^nw_h,v_h)=(w_h,v_h). ( w h , Q h n R h n v h ) = ( π h n w h , v h ) = ( w h , v h ) .
Furthermore, the standard L 2 L^2 L 2 -approximation property gives Q h n w → w Q_h^nw \to w Q h n w → w strongly for every w ∈ L 2 Λ n ( M ) w \in L^2\Lambda^n(M) w ∈ L 2 Λ n ( M ) as h → 0 h \to0 h → 0 . Consequently, using u j = Q h j n R h j n u j u_j=Q_{h_j}^nR_{h_j}^nu_j u j = Q h j n R h j n u j ,
∥ u j − u ∥ ≤ ∥ Q h j n ( R h j n u j − u ) ∥ + ∥ ( Q h j n − I ) u ∥ ≤ ∥ R h j n u j − u ∥ + ∥ ( Q h j n − I ) u ∥ ⟶ 0 , as h j → 0. (104) \begin{aligned}
\|u_j-u\| &\leq\|Q_{h_j}^n(R_{h_j}^nu_j-u)\|+\|(Q_{h_j}^n-I)u\| \\
&\leq\|R_{h_j}^nu_j-u\|+\|(Q_{h_j}^n-I)u\|\longrightarrow0,\quad\text{as }h_j\to0. \tag*{(104)}
\end{aligned} ∥ u j − u ∥ ≤ ∥ Q h j n ( R h j n u j − u ) ∥ + ∥ ( Q h j n − I ) u ∥ ≤ ∥ R h j n u j − u ∥ + ∥ ( Q h j n − I ) u ∥ ⟶ 0 , as h j → 0. (104)
Thus { u j } = { T h j f j } \{u_j\}=\{T_{h_j}f_j\} { u j } = { T h j f j } contains a strongly convergent subsequence in X X X , which proves (H2). Thus { T h } \{T_h\} { T h } is collectively compact.
Therefore, we obtain the spectral convergence for any eigenvalue of T T T as h → 0 h\to0 h → 0 . However, it does not exclude the possibility that, for a fixed h h h , discrete eigenvalues far from the origin lie in the left half-plane and hence violate Assumption 5.3, even though such eigenvalues disappear in the limit h → 0 h\to0 h → 0 . To rule out this possibility for sufficiently small h h h , the next lemma shows that all discrete eigenvalues that could violate Assumption 5.3 lie in a fixed compact subset of C \mathbb{C} C , uniformly in h h h . Once such a compact confinement is available, one may choose a fixed bounded region containing only the zero eigenvalue of the continuous operator. The preservation of algebraic multiplicity in Theorem 6.2 then excludes any additional discrete eigenvalues in this region for all sufficiently small h h h , thereby proving Assumption 5.3. To provide a unified proof for this lemma, we set D h u h : = d 0 u h D_hu_h:=d^0u_h D h u h := d 0 u h for k = 0 k=0 k = 0 and D h u h : = δ h u h D_hu_h:=\delta_hu_h D h u h := δ h u h for k = n k=n k = n , the discrete codifferential is defined by ( δ h u h , τ h ) = ( u h , d n − 1 τ h ) (\delta_hu_h,\tau_h)=(u_h,d^{n-1}\tau_h) ( δ h u h , τ h ) = ( u h , d n − 1 τ h ) for all τ h ∈ V h n − 1 \tau_h\in V_h^{n-1} τ h ∈ V h n − 1 .
Lemma 6.3 (distribution of eigenvalues). Define R d : = C β + C β 2 + 2 η μ 2 η R_d:=\frac{C_\beta+\sqrt{C_\beta^2+2\eta\mu}}{2\eta} R d := 2 η C β + C β 2 + 2 η μ . Let k ∈ { 0 , n } k\in\{0,n\} k ∈ { 0 , n } and let ( λ ~ h , u h ) (\widetilde{\lambda}_h,u_h) ( λ h , u h ) be an eigenpair of A ~ h k \widetilde{A}_h^k A h k with Re λ ~ h ≤ μ 2 + c \operatorname{Re}\widetilde{\lambda}_h\leq\frac{\mu}{2}+c Re λ h ≤ 2 μ + c for fixed c ≥ η R d 2 + C β R d + 1 c\geq\eta R_d^2+C_\beta R_d+1 c ≥ η R d 2 + C β R d + 1 . Then
∥ D h u h ∥ ≤ R d ∥ u h ∥ , ∣ λ ~ h − c ∣ ≤ η R d 2 + C β R d . (105) \|D_hu_h\|\leq R_d\|u_h\|,\qquad|\widetilde{\lambda}_h-c|\leq\eta R_d^2+C_\beta R_d. \tag*{(105)} ∥ D h u h ∥ ≤ R d ∥ u h ∥ , ∣ λ h − c ∣ ≤ η R d 2 + C β R d . (105) Consequently
σ ( A ~ h k ) ⊂ K ∪ { Re z > μ 2 + c } , K : = { z ∈ C : ∣ z − c ∣ ≤ η R d 2 + C β R d , Re z ≤ μ 2 + c } , (106) \sigma(\widetilde{A}_h^k)\subset K\cup\left\{\operatorname{Re}z>\frac{\mu}{2}+c\right\},\qquad K:=\left\{z\in\mathbb{C}:|z-c|\leq\eta R_d^2+C_\beta R_d,\ \operatorname{Re}z\leq\frac{\mu}{2}+c\right\}, \tag*{(106)} σ ( A h k ) ⊂ K ∪ { Re z > 2 μ + c } , K := { z ∈ C : ∣ z − c ∣ ≤ η R d 2 + C β R d , Re z ≤ 2 μ + c } , (106) and 0 ∉ K 0\notin K 0 ∈ / K .
Proof. For k = 0 k=0 k = 0 ,
λ ~ h ∥ u h ∥ 2 = ( A ~ h 0 u h , u h ) = η ∥ d u h ∥ 2 + ( i β d u h , u h ) + c ∥ u h ∥ 2 . (107) \widetilde{\lambda}_h\|u_h\|^2=(\widetilde{A}_h^0u_h,u_h)=\eta\|du_h\|^2+(i\beta du_h,u_h)+c\|u_h\|^2. \tag*{(107)} λ h ∥ u h ∥ 2 = ( A h 0 u h , u h ) = η ∥ d u h ∥ 2 + ( i β d u h , u h ) + c ∥ u h ∥ 2 . (107)
For k = n k=n k = n , the mixed formulation gives σ h = Q h n − 1 ( i β u h ) + η δ h u h \sigma_h=Q_h^{n-1}(i\beta u_h)+\eta\delta_hu_h σ h = Q h n − 1 ( i β u h ) + η δ h u h and A ~ h n u h = d σ h + c u h \widetilde{A}_h^nu_h=d\sigma_h+cu_h A h n u h = d σ h + c u h , whence, using the definition of δ h \delta_h δ h with τ h = σ h \tau_h=\sigma_h τ h = σ h , we have
λ ~ h ∥ u h ∥ 2 = ( A ~ h n u h , u h ) = ( σ h , δ h u h ) + c ∥ u h ∥ 2 = η ∥ δ h u h ∥ 2 + ( i β u h , δ h u h ) + c ∥ u h ∥ 2 . (108) \widetilde{\lambda}_h\|u_h\|^2=(\widetilde{A}_h^nu_h,u_h)=(\sigma_h,\delta_hu_h)+c\|u_h\|^2=\eta\|\delta_hu_h\|^2+(i\beta u_h,\delta_hu_h)+c\|u_h\|^2. \tag*{(108)} λ h ∥ u h ∥ 2 = ( A h n u h , u h ) = ( σ h , δ h u h ) + c ∥ u h ∥ 2 = η ∥ δ h u h ∥ 2 + ( i β u h , δ h u h ) + c ∥ u h ∥ 2 . (108)
In both cases, taking the real part gives
Re λ ~ h ∥ u h ∥ 2 ≥ η ∥ D h u h ∥ 2 − C β ∥ D h u h ∥ ∥ u h ∥ + c ∥ u h ∥ 2 . \operatorname{Re}\widetilde{\lambda}_h \lVert u_h\rVert^2 \ge\eta\lVert D_hu_h\rVert^2-C_\beta\lVert D_hu_h\rVert\lVert u_h\rVert+c\lVert u_h\rVert^2. Re λ h ∥ u h ∥ 2 ≥ η ∥ D h u h ∥ 2 − C β ∥ D h u h ∥ ∥ u h ∥ + c ∥ u h ∥ 2 .
If Re λ ~ h ≤ μ 2 + c \operatorname{Re}\widetilde{\lambda}_h \le\frac{\mu}{2}+c Re λ h ≤ 2 μ + c , the quadratic inequality η t 2 − C β t − μ 2 ≤ 0 \eta t^2-C_\beta t-\frac{\mu}{2}\le0 η t 2 − C β t − 2 μ ≤ 0 for t : = ∥ D h u h ∥ / ∥ u h ∥ t:=\lVert D_hu_h\rVert/\lVert u_h\rVert t := ∥ D h u h ∥ / ∥ u h ∥ yields t ≤ R d t\le R_d t ≤ R d . Then both (107) and (108) imply
∣ λ ~ h − c ∣ ≤ η ∥ D h u h ∥ 2 + C β ∥ D h u h ∥ ∥ u h ∥ ∥ u h ∥ 2 ≤ η R d 2 + C β R d . \lvert\widetilde{\lambda}_h-c\rvert\le\frac{\eta\lVert D_hu_h\rVert^2+C_\beta\lVert D_hu_h\rVert\lVert u_h\rVert}{\lVert u_h\rVert^2}\le\eta R_d^2+C_\beta R_d. ∣ λ h − c ∣ ≤ ∥ u h ∥ 2 η ∥ D h u h ∥ 2 + C β ∥ D h u h ∥ ∥ u h ∥ ≤ η R d 2 + C β R d .
Finally, if 0 ∈ K 0\in K 0 ∈ K , then c = ∣ 0 − c ∣ ≤ η R d 2 + C β R d < c c=\lvert0-c\rvert\le\eta R_d^2+C_\beta R_d<c c = ∣ 0 − c ∣ ≤ η R d 2 + C β R d < c by the choice of c c c , which is a contradiction.
Finally, we present the proof of Lemma 6.1 .
Proof of Lemma 6.1 . Take c ≥ c ∗ : = max { η R d 2 + C β R d + 1 , C β 2 η } c\ge c_*:=\max\left\{\eta R_d^2+C_\beta R_d+1,\frac{C_\beta^2}{\eta}\right\} c ≥ c ∗ := max { η R d 2 + C β R d + 1 , η C β 2 } . Fix k ∈ { 0 , n } k\in\{0,n\} k ∈ { 0 , n } and let T T T , T h T_h T h be as above, so that T = ( A ~ k ) − 1 T=(\widetilde{A}^k)^{-1} T = ( A k ) − 1 and T h = ( A ~ h k ) − 1 Q h k T_h=(\widetilde{A}_h^k)^{-1}Q_h^k T h = ( A h k ) − 1 Q h k on X = L 2 Λ k X=L^2\Lambda^k X = L 2 Λ k . The inverses exist for every h h h by (88) and (91) . The nonzero eigenvalues of the two operators are reciprocal to each other and have the same multiplicities.
By Theorem 2.1 , σ ( A ~ k ) ⊂ { c } ∪ { Re z ≥ μ + c } \sigma(\widetilde{A}^k)\subset\{c\}\cup\{\operatorname{Re}z\ge\mu+c\} σ ( A k ) ⊂ { c } ∪ { Re z ≥ μ + c } with c c c algebraically simple. The compact set K K K of (106) therefore satisfies σ ( A ~ k ) ∩ K = { c } \sigma(\widetilde{A}^k)\cap K=\{c\} σ ( A k ) ∩ K = { c } and ∂ K ∩ σ ( A ~ k ) = ∅ \partial K\cap\sigma(\widetilde{A}^k)=\varnothing ∂ K ∩ σ ( A k ) = ∅ . Moreover, c c c lies in the interior of K K K and 0 ∉ K 0\notin K 0 ∈ / K . The inversion ι ( z ) : = 1 / z \iota(z):=1/z ι ( z ) := 1/ z is a homeomorphism of a neighborhood of K K K . Set Ω : = ι ( int K ) \Omega:=\iota(\operatorname{int}K) Ω := ι ( int K ) and Γ : = ι ( ∂ K ) \Gamma:=\iota(\partial K) Γ := ι ( ∂ K ) , then Ω \Omega Ω is open and bounded with rectifiable Jordan boundary Γ \Gamma Γ , 0 ∉ Ω 0\notin\Omega 0 ∈ / Ω (as ∞ ∉ K \infty\notin K ∞ ∈ / K ), and
σ ( T ) ∩ Ω = { 1 / c } with algebraic multiplicity 1 , Γ ∩ σ ( T ) = ∅ . \sigma(T)\cap\Omega=\{1/c\}\quad\text{with algebraic multiplicity }1,\qquad\Gamma\cap\sigma(T)=\varnothing. σ ( T ) ∩ Ω = { 1/ c } with algebraic multiplicity 1 , Γ ∩ σ ( T ) = ∅ .
The hypotheses of Theorem 6.2 were verified above for both degrees. Hence there is h Γ > 0 h_\Gamma>0 h Γ > 0 such that for h < h Γ h<h_\Gamma h < h Γ the total algebraic multiplicity of σ ( T h ) ∩ Ω \sigma(T_h)\cap\Omega σ ( T h ) ∩ Ω equals 1.
As shown in the beginning, the value c c c is an eigenvalue of A ~ h k \widetilde{A}_h^k A h k for every h h h . Since 1 / c ∈ Ω 1/c\in\Omega 1/ c ∈ Ω , it is the only eigenvalue in Ω \Omega Ω and has multiplicity of 1, then we have
σ ( A ~ h k ) ∩ K = { c } algebraically simple . \sigma(\widetilde{A}_h^k)\cap K=\{c\}\quad\text{algebraically simple}. σ ( A h k ) ∩ K = { c } algebraically simple .
Let h 0 : = h Γ h_0:=h_\Gamma h 0 := h Γ (minimized over the two degrees). For h < h 0 h<h_0 h < h 0 and any eigenvalue λ ~ h ≠ c \widetilde{\lambda}_h\ne c λ h = c of A ~ h k \widetilde{A}_h^k A h k , we have λ ~ h ∉ K \widetilde{\lambda}_h\notin K λ h ∈ / K . Then Lemma 6.3 forces Re λ ~ h > μ 2 + c \operatorname{Re}\widetilde{\lambda}_h>\frac{\mu}{2}+c Re λ h > 2 μ + c . Finally undoing the shift, we have that 0 is an algebraically simple eigenvalue of A h k A_h^k A h k , and every other eigenvalue satisfies Re λ h > μ / 2 > 0 \operatorname{Re}\lambda_h>\mu/2>0 Re λ h > μ /2 > 0 . This is Assumption 5.3 , with the quantitative gap μ / 2 \mu/2 μ /2 .
Choice 2: M-matrix discretization for P1-k for arbitrary h
An alternative way to guarantee Assumption 5.3 for arbitrary mesh size is to construct, for the degrees k ∈ { 0 , n } k\in\{0,n\} k ∈ { 0 , n } , discretizations based on the lowest-order Whitney forms whose system matrices are singular irreducible M M M -matrices: by the Perron–Frobenius theorem, the zero eigenvalue of such a matrix is simple, its left and right eigenvectors are strictly positive, and all remaining eigenvalues have strictly positive real parts, which is exactly the required spectral distribution. The construction is restricted to the Euclidean setting because it relies on Delaunay or angle conditions. Throughout this subsection, M M M is a polyhedral domain with periodic boundary conditions (for instance a flat torus), so that every facet is interior and is endowed with the Euclidean metric. T h \mathcal{T}_h T h is an affine simplicial triangulation.
In this singular case the discrete maximum principle takes the form A h ω h ≤ 0 ⟹ A h ω h = 0 A_h\omega_h\le0\implies A_h\omega_h=0 A h ω h ≤ 0 ⟹ A h ω h = 0 . If essential boundary conditions render the matrix nonsingular, it degenerates to the usual monotonicity A h ω h ≤ 0 ⟹ ω h ≤ 0 A_h\omega_h \le0 \Longrightarrow\omega_h \le0 A h ω h ≤ 0 ⟹ ω h ≤ 0 . Throughout this subsection, β ∈ W 1 , ∞ ( M ) \beta\in W^{1,\infty}(M) β ∈ W 1 , ∞ ( M ) . The construction combines the edge-averaged finite element (EAFE) method [52 ] and its simplex extension (SAFE) [51 ] with an exponential fitting of the flux.
Exponential-fitting discretization
Grouping the diffusive and convective terms, the mixed scheme (54) can be written in flux form,
( σ h , τ h ) − ( u h , ( η d + j β ) τ h ) = 0 , ∀ τ h ∈ V h k − 1 , (\sigma_h,\tau_h)-(u_h,(\eta d+j_\beta)\tau_h)=0,\qquad\forall\tau_h\in V_h^{k-1}, ( σ h , τ h ) − ( u h , ( η d + j β ) τ h ) = 0 , ∀ τ h ∈ V h k − 1 ,
( ∂ t u h , v h ) + ( d σ h , v h ) + ( d u h , ( η d + j β ) v h ) = 0 , ∀ v h ∈ V h k , (\partial_tu_h,v_h)+(d\sigma_h,v_h)+(du_h,(\eta d+j_\beta)v_h)=0,\qquad\forall v_h\in V_h^k, ( ∂ t u h , v h ) + ( d σ h , v h ) + ( d u h , ( η d + j β ) v h ) = 0 , ∀ v h ∈ V h k ,
where J η , β : = η d + j β J_{\eta,\beta}:=\eta d+j_\beta J η , β := η d + j β acts as a ( k + 1 ) (k+1) ( k + 1 ) -form flux operator. Our goal is to construct a discrete flux operator J h k : V h k → V h k + 1 J_h^k:V_h^k\to V_h^{k+1} J h k : V h k → V h k + 1 leading to an M M M -matrix discretization. The discrete contraction is then recovered as the formal adjoint of the flux remainder,
( ι β , h k + 1 u h , v h ) = ( u h , ( J h k − η d k ) v h ) , u h ∈ V h k + 1 , v h ∈ V h k , (109) (\iota_{\beta,h}^{k+1}u_h,v_h)=(u_h,(J_h^k-\eta d^k)v_h),\qquad u_h\in V_h^{k+1},\quad v_h\in V_h^k, \tag*{(109)} ( ι β , h k + 1 u h , v h ) = ( u h , ( J h k − η d k ) v h ) , u h ∈ V h k + 1 , v h ∈ V h k , (109)
so that both contributions of diffusion and advection are encoded in the single operator J h k J_h^k J h k .
When β \beta β is a gradient field, the Witten deformation of § 3 motivates an exponential fitting of the flux: if β = ( d ψ ) ♯ \beta=(d\psi)^\sharp β = ( d ψ ) ♯ , then
η e − ψ / η d ( e ψ / η u ) = η d u + d ψ ∧ u = η d u + j β u , (110) \eta e^{-\psi/\eta}d(e^{\psi/\eta}u)=\eta du+d\psi\wedge u=\eta du+j_\beta u, \tag*{(110)} η e − ψ / η d ( e ψ / η u ) = η d u + d ψ ∧ u = η d u + j β u , (110)
so the flux is a conjugated exterior derivative. In general, β \beta β is not a gradient field. For k = 0 k=0 k = 0 , we use an edgewise potential as in [52 ] ; for k > 0 k>0 k > 0 , we use the elementwise construction of [51 ] .
Denote by Σ h k \Sigma_h^k Σ h k and Σ h k \Sigma_h^k Σ h k the sets of oriented k k k -dimensional subsimplices of T ∈ T h T\in\mathcal{T}_h T ∈ T h and T h \mathcal{T}_h T h , respectively, and let ℓ S \ell_S ℓ S be the canonical degree of freedom associated with S S S . We write φ S \varphi_S φ S for the corresponding Whitney basis form.
For k = 0 k=0 k = 0 , we use the edgewise construction of EAFE [52 ] . For each oriented edge E = a i a j → E=\overrightarrow{a_i a_j} E = a i a j , define ψ E \psi_E ψ E , up to a constant, by
∂ ψ E ∂ τ E = β ⋅ τ E on E , (111) \frac{\partial\psi_E}{\partial\tau_E}=\beta\cdot\tau_E\quad\text{on }E, \tag*{(111)} ∂ τ E ∂ ψ E = β ⋅ τ E on E , (111)
where τ E \tau_E τ E is the unit tangent vector of E E E . Set
E E : = ( ∫ E e ψ E / η ) − 1 , δ E ( w ) : = w ( a j ) − w ( a i ) . \mathcal{E}_E:=\left(\int_E e^{\psi_E/\eta}\right)^{-1},\qquad\delta_E(w):=w(a_j)-w(a_i). E E := ( ∫ E e ψ E / η ) − 1 , δ E ( w ) := w ( a j ) − w ( a i ) .
The fitted flux J ~ h 0 v h ∈ V h 1 \widetilde{J}_h^0v_h\in V_h^1 J h 0 v h ∈ V h 1 is defined by
ℓ E ( J ~ h 0 v h ) : = η E E δ E ( e ψ E / η v h ) , E ∈ Σ h 1 . (112) \ell_E(\widetilde{J}_h^0v_h):=\eta\mathcal{E}_E\delta_E(e^{\psi_E/\eta}v_h),\qquad E\in\Sigma_h^1. \tag*{(112)} ℓ E ( J h 0 v h ) := η E E δ E ( e ψ E / η v h ) , E ∈ Σ h 1 . (112)
These degrees of freedom are independent of the additive constant in ψ E \psi_E ψ E and are single-valued on shared edges.
For k > 0 k>0 k > 0 , an analogous construction of the fitted flux on S ∈ Σ h k + 1 S\in\Sigma_h^{k+1} S ∈ Σ h k + 1 would require solving a PDE for a local potential, whose compatibility conditions need not hold for a general vector field β \beta β . We use the elemental construction of SAFE [51 ] . On each element T T T , set
β T : = ∫ T β d x , ψ T ( x ) : = β T ⋅ x . (113) \beta_T:=\int_T\beta\,dx,\qquad\psi_T(x):=\beta_T\cdot x. \tag*{(113)} β T := ∫ T β d x , ψ T ( x ) := β T ⋅ x . (113)
Let Π T k \Pi_T^k Π T k be the canonical Whitney interpolation operator and define
W T k v : = Π T k ( e ψ T / η v ) , E T k : = ( W T k ) − 1 . (114) \mathcal{W}_T^k v := \Pi_T^k(e^{\psi_T/\eta}v), \qquad\mathcal{E}_T^k := (\mathcal{W}_T^k)^{-1}. \tag*{(114)} W T k v := Π T k ( e ψ T / η v ) , E T k := ( W T k ) − 1 . (114)
The inverse is taken on P 1 − Λ k ( T ) P_1^-\Lambda^k(T) P 1 − Λ k ( T ) . Both operators are diagonal in the Whitney basis; explicitly,
ℓ S ( E T k w h ) = E S T ℓ S ( w h ) , E S T : = ( ∫ S e ψ T / η ) − 1 , S ∈ Σ T k . \ell_S(\mathcal{E}_T^k w_h)=\mathcal{E}_S^T\ell_S(w_h), \qquad\mathcal{E}_S^T:=\left(\int_S e^{\psi_T/\eta}\right)^{-1}, \qquad S\in\Sigma_T^k. ℓ S ( E T k w h ) = E S T ℓ S ( w h ) , E S T := ( ∫ S e ψ T / η ) − 1 , S ∈ Σ T k .
The elemental fitted flux is
J ~ T k : = η E T k + 1 d k W T k , 1 ≤ k ≤ n − 1. (115) \widetilde{J}_T^k:=\eta\mathcal{E}_T^{k+1}d^k\mathcal{W}_T^k,\qquad1\leq k\leq n-1. \tag*{(115)} J T k := η E T k + 1 d k W T k , 1 ≤ k ≤ n − 1. (115)
We define J ~ h k \widetilde{J}_h^k J h k elementwise by
( J ~ h k v h ) ∣ T : = J ~ T k ( v h ∣ T ) . (\widetilde{J}_h^k v_h)|_T:=\widetilde{J}_T^k(v_h|_T). ( J h k v h ) ∣ T := J T k ( v h ∣ T ) .
Since ψ T \psi_T ψ T depends on the element, the flux degrees of freedom on shared subsimplices need not agree. For k = 0 k=0 k = 0 , J ~ T 0 \widetilde{J}_T^0 J T 0 denotes the restriction of the edgewise construction (6.32) to T T T .
The following result quantifies the exponential fitting by showing that the degrees of freedom of the fitted flux are obtained from the exponentially weighted continuous flux.
Lemma 6.4 (DOF representation). For a smooth 0-form v v v and E ∈ Σ h 1 E\in\Sigma_h^1 E ∈ Σ h 1 ,
ℓ E ( J ~ h 0 v ) = E E ∫ E e ψ E / η tr E ( η d v + β ♭ v ) . (116) \ell_E(\widetilde{J}_h^0v)=\mathcal{E}_E\int_E e^{\psi_E/\eta}\operatorname{tr}_E(\eta\,dv+\beta^\flat v). \tag*{(116)} ℓ E ( J h 0 v ) = E E ∫ E e ψ E / η tr E ( η d v + β ♭ v ) . (116) For a smooth k k k -form v v v on T T T , 1 ≤ k ≤ n − 1 1\leq k\leq n-1 1 ≤ k ≤ n − 1 , and S ∈ Σ T k + 1 S\in\Sigma_T^{k+1} S ∈ Σ T k + 1 ,
ℓ S ( J ~ T k v ) = E S T ∫ S tr S [ e ψ T / η ( η d k v + β T ♭ ∧ v ) ] . (117) \ell_S(\widetilde{J}_T^k v)=\mathcal{E}_S^T\int_S\operatorname{tr}_S\left[e^{\psi_T/\eta}\left(\eta\,d^k v+\beta_T^\flat\wedge v\right)\right]. \tag*{(117)} ℓ S ( J T k v ) = E S T ∫ S tr S [ e ψ T / η ( η d k v + β T ♭ ∧ v ) ] . (117) For the elemental SAFE fluxes, J ~ T k + 1 J ~ T k = 0 \widetilde{J}_T^{k+1}\widetilde{J}_T^k=0 J T k + 1 J T k = 0 for 1 ≤ k ≤ n − 2 1\leq k\leq n-2 1 ≤ k ≤ n − 2 .
Proof. By (6.31),
η ∂ ∂ τ E ( e ψ E / η v ) = e ψ E / η ( η ∂ v ∂ τ E + ( β ⋅ τ E ) v ) . \eta\frac{\partial}{\partial\tau_E}\left(e^{\psi_E/\eta}v\right)=e^{\psi_E/\eta}\left(\eta\frac{\partial v}{\partial\tau_E}+(\beta\cdot\tau_E)v\right). η ∂ τ E ∂ ( e ψ E / η v ) = e ψ E / η ( η ∂ τ E ∂ v + ( β ⋅ τ E ) v ) .
Integrating along E E E and using (6.32) gives (6.36).
For k > 0 k>0 k > 0 , the definition of Π T k \Pi_T^k Π T k and Stokes’ theorem give ℓ S ( d Π T k w ) = ∫ ∂ S tr Π T k w = ∫ ∂ S tr w = ℓ S ( d w ) \ell_S(d\Pi_T^k w)=\int_{\partial S}\operatorname{tr}\Pi_T^k w=\int_{\partial S}\operatorname{tr}w=\ell_S(dw) ℓ S ( d Π T k w ) = ∫ ∂ S tr Π T k w = ∫ ∂ S tr w = ℓ S ( d w ) for smooth w w w . Taking w = e ψ T / η v w=e^{\psi_T/\eta}v w = e ψ T / η v and using η d ( e ψ T / η v ) = e ψ T / η ( η d v + d ψ T ∧ v ) \eta\,d(e^{\psi_T/\eta}v)=e^{\psi_T/\eta}(\eta\,dv+d\psi_T\wedge v) η d ( e ψ T / η v ) = e ψ T / η ( η d v + d ψ T ∧ v ) with d ψ T = β T ♭ d\psi_T=\beta_T^\flat d ψ T = β T ♭ yields (6.37). Finally, for 1 ≤ k ≤ n − 2 1\leq k\leq n-2 1 ≤ k ≤ n − 2 ,
J ~ T k + 1 J ~ T k = η 2 E T k + 2 d ( W T k + 1 E T k + 1 ) d W T k = η 2 E T k + 2 d d W T k = 0 , \widetilde{J}_T^{k+1}\widetilde{J}_T^k=\eta^2\mathcal{E}_T^{k+2}d(\mathcal{W}_T^{k+1}\mathcal{E}_T^{k+1})d\mathcal{W}_T^k=\eta^2\mathcal{E}_T^{k+2}dd\mathcal{W}_T^k=0, J T k + 1 J T k = η 2 E T k + 2 d ( W T k + 1 E T k + 1 ) d W T k = η 2 E T k + 2 dd W T k = 0 ,
since d W T k v ∈ P 1 − Λ k + 1 ( T ) d\mathcal{W}_T^k v\in P_1^-\Lambda^{k+1}(T) d W T k v ∈ P 1 − Λ k + 1 ( T ) and W T k + 1 E T k + 1 = I d \mathcal{W}_T^{k+1}\mathcal{E}_T^{k+1}=\mathrm{Id} W T k + 1 E T k + 1 = Id . □ \square □
For k = 0 k=0 k = 0 , we exploit the graph-Laplacian representation of the P 1 P_1 P 1 stiffness matrix. The direct L 2 L^2 L 2 pairing of d 0 u h d^0u_h d 0 u h with the fitted flux J ~ h 0 v h \widetilde{J}_h^0v_h J h 0 v h does not give the desired M M M -matrix structure. We therefore insert a local projection onto constant 1-forms. More precisely, for u h , v h ∈ P 1 Λ 0 ( T ) u_h,v_h \in\mathcal{P}_1\Lambda^0(T) u h , v h ∈ P 1 Λ 0 ( T ) ,
( d 0 u h , d 0 v h ) T = ∑ E ∈ Σ T 1 ω E T δ E ( u h ) δ E ( v h ) , δ E ( u h ) = ℓ E ( d 0 u h ) = u h ( a j ) − u h ( a i ) , (118) (\mathrm{d}^0u_h,\mathrm{d}^0v_h)_T = \sum_{E\in\Sigma_T^1}\omega_E^T\delta_E(u_h)\,\delta_E(v_h), \qquad\delta_E(u_h)=\ell_E(\mathrm{d}^0u_h)=u_h(a_j)-u_h(a_i), \tag*{(118)} ( d 0 u h , d 0 v h ) T = E ∈ Σ T 1 ∑ ω E T δ E ( u h ) δ E ( v h ) , δ E ( u h ) = ℓ E ( d 0 u h ) = u h ( a j ) − u h ( a i ) , (118)
where ω E T : = − ( d 0 φ i , d 0 φ j ) T \omega_E^T:=-(\mathrm{d}^0\varphi_i,\mathrm{d}^0\varphi_j)_T ω E T := − ( d 0 φ i , d 0 φ j ) T for the oriented edge E = a i a j ‾ E=\overline{a_i a_j} E = a i a j and Σ T k \Sigma_T^k Σ T k denotes the k k k -dimensional subsimplices of T T T . Accordingly, define Π ‾ T 1 : P 1 Λ 1 ( T ) → P 0 Λ 1 ( T ) \overline{\Pi}_T^1:\mathcal{P}_1\Lambda^1(T)\to\mathcal{P}_0\Lambda^1(T) Π T 1 : P 1 Λ 1 ( T ) → P 0 Λ 1 ( T ) as the unique element satisfying
( Π ‾ T 1 u h , q ) T = ∑ E ∈ Σ T 1 ω E T ℓ E ( u h ) ℓ E ( q ) , ∀ q ∈ P 0 Λ 1 ( T ) , (119) (\overline{\Pi}_T^1u_h,q)_T=\sum_{E\in\Sigma_T^1}\omega_E^T\ell_E(u_h)\,\ell_E(q), \qquad\forall q\in\mathcal{P}_0\Lambda^1(T), \tag*{(119)} ( Π T 1 u h , q ) T = E ∈ Σ T 1 ∑ ω E T ℓ E ( u h ) ℓ E ( q ) , ∀ q ∈ P 0 Λ 1 ( T ) , (119)
and Π ‾ h 1 ∣ T : = Π ‾ T 1 \overline{\Pi}_h^1|_T:=\overline{\Pi}_T^1 Π h 1 ∣ T := Π T 1 . Then
( d 0 u h , Π ‾ h 1 J ~ h 0 v h ) = ∑ T ∈ T h ∑ E ∈ Σ T 1 ω E T δ E ( u h ) ℓ E ( J ~ h 0 v h ) , (120) (\mathrm{d}^0u_h,\overline{\Pi}_h^1\widetilde{J}_h^0v_h)=\sum_{T\in\mathcal{T}_h}\sum_{E\in\Sigma_T^1}\omega_E^T\delta_E(u_h)\,\ell_E(\widetilde{J}_h^0v_h), \tag*{(120)} ( d 0 u h , Π h 1 J h 0 v h ) = T ∈ T h ∑ E ∈ Σ T 1 ∑ ω E T δ E ( u h ) ℓ E ( J h 0 v h ) , (120)
where, for the oriented edge E = a i a j ‾ E=\overline{a_i a_j} E = a i a j , the flux degree of freedom expands as
ℓ E ( J ~ h 0 v h ) = η E δ E ( e ψ E / η v h ) = η E ( e ψ E ( a j ) η v h ( a j ) − e ψ E ( a i ) η v h ( a i ) ) . (121) \ell_E(\widetilde{J}_h^0v_h)=\eta_E\delta_E(e^{\psi_E/\eta}v_h)=\eta_E\left(\frac{e^{\psi_E(a_j)}}{\eta}v_h(a_j)-\frac{e^{\psi_E(a_i)}}{\eta}v_h(a_i)\right). \tag*{(121)} ℓ E ( J h 0 v h ) = η E δ E ( e ψ E / η v h ) = η E ( η e ψ E ( a j ) v h ( a j ) − η e ψ E ( a i ) v h ( a i ) ) . (121)
Furthermore, to transfer the M M M -matrix structure of the spatial discretization to the semidiscrete evolution operator, we replace the mass matrix associated with the temporal term by a positive diagonal one. This avoids the mixing of degrees of freedom caused by the inverse of a mass matrix and ensures that multiplication by the inverse lumped mass matrix preserves the M M M -matrix structure. We therefore use the nodal mass-lumped pairing:
( ∂ t u h , v h ) ≈ ( ∂ t u h , v h ) h : = ∑ z ∈ Σ h 0 m z ∂ t u h ( z ) v h ( z ) , m z = ∫ M φ z > 0 , (122) (\partial_tu_h,v_h)\approx(\partial_tu_h,v_h)_h:=\sum_{z\in\Sigma_h^0}m_z\,\partial_tu_h(z)\,v_h(z), \qquad m_z=\int_M\varphi_z>0, \tag*{(122)} ( ∂ t u h , v h ) ≈ ( ∂ t u h , v h ) h := z ∈ Σ h 0 ∑ m z ∂ t u h ( z ) v h ( z ) , m z = ∫ M φ z > 0 , (122)
and denote by M ~ h 0 : V h 0 → V h 0 \widetilde{M}_h^0:V_h^0\to V_h^0 M h 0 : V h 0 → V h 0 the induced self-adjoint operator,
( u h , v h ) h = ( M ~ h 0 u h , v h ) = ( u h , M ~ h 0 v h ) . (u_h,v_h)_h=(\widetilde{M}_h^0u_h,v_h)=(u_h,\widetilde{M}_h^0v_h). ( u h , v h ) h = ( M h 0 u h , v h ) = ( u h , M h 0 v h ) .
The lumping contributes only a positive diagonal and will allow us to identify the eigenvalues of A h 0 A_h^0 A h 0 below.
For the volume-form case k = n k=n k = n , the auxiliary variable σ h ∈ V h n − 1 \sigma_h\in V_h^{n-1} σ h ∈ V h n − 1 is eliminated through the inverse of its facet mass matrix. A consistent mass matrix would introduce nonlocal couplings. We therefore replace it by a positive diagonal facet mass matrix, so that the elimination remains local and preserves the off-diagonal sign pattern required for the M M M -matrix property. Following [8 , 13 , 44 ] , we define
( σ h , τ h ) ≈ ( σ h , τ h ) h : = ∑ F ∈ Σ h n − 1 M F ℓ F ( σ h ) ℓ F ( τ h ) , M F : = 1 ∣ F ∣ ∑ T ⊃ F d T , F , (\sigma_h,\tau_h)\approx(\sigma_h,\tau_h)_h:=\sum_{F\in\Sigma_h^{n-1}}\mathsf{M}_F\,\ell_F(\sigma_h)\,\ell_F(\tau_h), \qquad\mathsf{M}_F:=\frac{1}{|F|}\sum_{T\supset F}d_{T,F}, ( σ h , τ h ) ≈ ( σ h , τ h ) h := F ∈ Σ h n − 1 ∑ M F ℓ F ( σ h ) ℓ F ( τ h ) , M F := ∣ F ∣ 1 T ⊃ F ∑ d T , F ,
where d T , F d_{T,F} d T , F is the signed distance from the circumcenter of T T T to the hyperplane of F F F (positive when the circumcenter lies on the same side as T T T ). This quadrature diagonalizes the mass matrix with an O ( h ) O(h) O ( h ) consistency error. Its positivity is ensured under a strict Delaunay triangulation [8 , 13 , 44 ] . Indeed, M F M_F M F is the diagonal entry of the circumcentric discrete Hodge star, and the resulting bilinear form ( ⋅ , ⋅ ) h (\cdot,\cdot)_h ( ⋅ , ⋅ ) h is the associated Discrete Exterior Calculus (DEC) inner product on primal ( n − 1 ) (n-1) ( n − 1 ) -cochains [29 , 30 ] . Denote by M ~ h n − 1 : V h n − 1 → V h n − 1 \widetilde{M}_h^{n-1}:V_h^{n-1}\to V_h^{n-1} M h n − 1 : V h n − 1 → V h n − 1 the induced self-adjoint operator, ( σ h , τ h ) h = ( M ~ h n − 1 σ h , τ h ) = ( σ h , M ~ h n − 1 τ h ) (\sigma_h,\tau_h)_h=(\widetilde{M}_h^{n-1}\sigma_h,\tau_h)=(\sigma_h,\widetilde{M}_h^{n-1}\tau_h) ( σ h , τ h ) h = ( M h n − 1 σ h , τ h ) = ( σ h , M h n − 1 τ h ) . Introducing the change of variables ζ h = M ~ h n − 1 τ h \zeta_h=\widetilde{M}_h^{n-1}\tau_h ζ h = M h n − 1 τ h , the first equation in (55) can be rewritten as
( σ h , τ h ) h − ( u h , J ~ h n − 1 τ h ) = 0 , ⟺ ( σ h , ζ h ) − ( u h , J ~ h n − 1 ( M ~ h n − 1 ) − 1 ζ h ) = 0. (123) (\sigma_h,\tau_h)_h-(u_h,\widetilde{J}_h^{n-1}\tau_h)=0,\qquad\Longleftrightarrow\qquad(\sigma_h,\zeta_h)-(u_h,\widetilde{J}_h^{n-1}(\widetilde{M}_h^{n-1})^{-1}\zeta_h)=0. \tag*{(123)} ( σ h , τ h ) h − ( u h , J h n − 1 τ h ) = 0 , ⟺ ( σ h , ζ h ) − ( u h , J h n − 1 ( M h n − 1 ) − 1 ζ h ) = 0. (123)
No lumping is needed for the temporal term, since the mass matrix of P 1 − Λ n \mathcal{P}_1^-\Lambda^n P 1 − Λ n is already diagonal and positive. Note that the flux operators for 0 0 0 -forms and ( n − 1 ) (n-1) ( n − 1 ) -forms require additional modification, whereas for the other forms an L 2 L^2 L 2 -projection is used to obtain a conforming flux. The unified flux operator J h k : V h k → V h k + 1 J_h^k:V_h^k\to V_h^{k+1} J h k : V h k → V h k + 1 is thus
J h k : = { Q h 1 Π ‾ h 1 J ~ h 0 ( M ~ h 0 ) − 1 , k = 0 , Q h k + 1 J ~ h k , 1 ≤ k ≤ n − 2 , J ~ h n − 1 ( M ~ h n − 1 ) − 1 , k = n − 1. (124) J_h^k:=
\begin{cases}
Q_h^1\overline{\Pi}_h^1\widetilde{J}_h^0(\widetilde{M}_h^0)^{-1}, & k=0,\\
Q_h^{k+1}\widetilde{J}_h^k, & 1\le k\le n-2,\\
\widetilde{J}_h^{n-1}(\widetilde{M}_h^{n-1})^{-1}, & k=n-1.
\end{cases} \tag*{(124)} J h k := ⎩ ⎨ ⎧ Q h 1 Π h 1 J h 0 ( M h 0 ) − 1 , Q h k + 1 J h k , J h n − 1 ( M h n − 1 ) − 1 , k = 0 , 1 ≤ k ≤ n − 2 , k = n − 1. (124)
where Q h k + 1 Q_h^{k+1} Q h k + 1 is the L 2 L^2 L 2 -orthogonal projection onto V h k + 1 V_h^{k+1} V h k + 1 , and the discrete contraction is given by (109) . First-order convergence results for the corresponding EAFE and SAFE model problems are established under the assumptions of [51 , 52 ] .
We first establish a bound for the discrete contraction operator which is uniform in both h h h and η \eta η .
Lemma 6.5 (uniform elemental bound). There exists C > 0 C>0 C > 0 , depending only on shape regularity and n n n , k k k , such that
∥ ( J ~ T k − η d k ) v h ∥ L 2 ( T ) ≤ C ∣ β T ∣ ∥ v h ∥ L 2 ( T ) , v h ∈ P 1 − Λ k ( T ) , 1 ≤ k ≤ n − 1 , h , η > 0. \left\|(\widetilde{J}_T^k-\eta d^k)v_h\right\|_{L^2(T)}\le C|\beta_T|\|v_h\|_{L^2(T)},\qquad v_h\in\mathcal{P}_1^-\Lambda^k(T),\quad1\le k\le n-1,\quad h,\eta>0. ( J T k − η d k ) v h L 2 ( T ) ≤ C ∣ β T ∣∥ v h ∥ L 2 ( T ) , v h ∈ P 1 − Λ k ( T ) , 1 ≤ k ≤ n − 1 , h , η > 0. For k = 0 k=0 k = 0 , the same estimate holds with ∣ β T ∣ |\beta_T| ∣ β T ∣ replaced by ∥ β ∥ L ∞ ( T ) \|\beta\|_{L^\infty(T)} ∥ β ∥ L ∞ ( T ) . Consequently, for 1 ≤ k ≤ n − 2 1\le k\le n-2 1 ≤ k ≤ n − 2 , the operators (109) satisfy Assumption 5.1 5.1 5.1 with C β = C ∥ β ∥ L ∞ C_\beta=C\|\beta\|_{L^\infty} C β = C ∥ β ∥ L ∞ , uniformly in h h h and η \eta η .
Proof. For k > 0 k>0 k > 0 , the weight g S : = E S T e ψ T / η ∣ S g_S:=\left.\mathcal{E}_S^T e^{\psi^T/\eta}\right|_S g S := E S T e ψ T / η S is positive with unit mean, ∫ S g S = 1 \int_S g_S=1 ∫ S g S = 1 . By (117) ,
ℓ S ( ( J ~ T k − η d k ) v h ) = η ( ∫ S ( g S − 1 ) tr S ( d k v h ) + ∫ S g S tr S ( β T b ∧ v h ) ) . \ell_S((\widetilde{J}_T^k-\eta d^k)v_h)=\eta\left(\int_S(g_S-1)\operatorname{tr}_S(d^k v_h)+\int_S g_S\operatorname{tr}_S(\beta_T^b\wedge v_h)\right). ℓ S (( J T k − η d k ) v h ) = η ( ∫ S ( g S − 1 ) tr S ( d k v h ) + ∫ S g S tr S ( β T b ∧ v h ) ) .
For lowest-order elements, d k v h ∈ P 0 Λ k + 1 ( T ) d^k v_h\in\mathcal{P}_0\Lambda^{k+1}(T) d k v h ∈ P 0 Λ k + 1 ( T ) , so the first term vanishes. The second term satisfies
∣ ∫ S g S tr S ( β T b ∧ v h ) ∣ ≤ C ∣ β T ∣ ∥ v h ∥ L ∞ ( T ) ∣ S ∣ . \left|\int_S g_S\operatorname{tr}_S(\beta_T^b\wedge v_h)\right|\le C|\beta_T|\|v_h\|_{L^\infty(T)}|S|. ∫ S g S tr S ( β T b ∧ v h ) ≤ C ∣ β T ∣∥ v h ∥ L ∞ ( T ) ∣ S ∣.
Expanding ( J ~ T k − η d k ) v h (\widetilde{J}_T^k-\eta d^k)v_h ( J T k − η d k ) v h in the Whitney basis { φ S } S ∈ Σ T k + 1 \{\varphi_S\}_{S\in\Sigma_T^{k+1}} { φ S } S ∈ Σ T k + 1 and combining ∣ S ∣ ≤ C h T k + 1 |S|\le Ch_T^{k+1} ∣ S ∣ ≤ C h T k + 1 , ∥ φ S ∥ L 2 ( T ) ≤ C h T n / 2 − k − 1 \|\varphi_S\|_{L^2(T)}\le Ch_T^{n/2-k-1} ∥ φ S ∥ L 2 ( T ) ≤ C h T n /2 − k − 1 , and ∥ v h ∥ L ∞ ( T ) ≤ C h T − n / 2 ∥ v h ∥ L 2 ( T ) \|v_h\|_{L^\infty(T)}\le Ch_T^{-n/2}\|v_h\|_{L^2(T)} ∥ v h ∥ L ∞ ( T ) ≤ C h T − n /2 ∥ v h ∥ L 2 ( T ) proves the first assertion.
For k = 0 k=0 k = 0 , use g E : = E E T e ψ E / η ∣ E g_E:=\left.\mathcal{E}_E^T e^{\psi^E/\eta}\right|_E g E := E E T e ψ E / η E and (116) . The tangential derivative of v h v_h v h is constant on E E E , so the diffusive term again vanishes after subtraction. Moreover,
∣ ∫ E g E tr E ( β T b v h ) ∣ ≤ ∥ β ∥ L ∞ ( T ) ∥ v h ∥ L ∞ ( T ) ∣ E ∣ . \left|\int_E g_E\operatorname{tr}_E(\beta_T^b v_h)\right|\le\|\beta\|_{L^\infty(T)}\|v_h\|_{L^\infty(T)}|E|. ∫ E g E tr E ( β T b v h ) ≤ ∥ β ∥ L ∞ ( T ) ∥ v h ∥ L ∞ ( T ) ∣ E ∣.
The same scaling argument gives the estimate. The last assertion follows from (109) , the L 2 L^2 L 2 contractivity of Q h k + 1 Q_h^{k+1} Q h k + 1 , and ∣ β T ∣ ≤ ∥ β ∥ L ∞ ( T ) |\beta_T| \le\|\beta\|_{L^\infty(T)} ∣ β T ∣ ≤ ∥ β ∥ L ∞ ( T ) .
Remark 6.6 (regularity of β \beta β ). In this section, we assume that β ∈ W 1 , ∞ \beta\in W^{1,\infty} β ∈ W 1 , ∞ . For k > 0 k>0 k > 0 , the construction only requires a piecewise constant approximation of β \beta β . Element averages remain defined for β ∈ L ∞ \beta\in L^\infty β ∈ L ∞ and converge in L p L^p L p , 1 ≤ p < ∞ 1 \le p < \infty 1 ≤ p < ∞ , as h → 0 h \to0 h → 0 , although no convergence rate need be available. For k = 0 k=0 k = 0 , the edgewise ODE additionally requires a well-defined tangential restriction of β \beta β to each edge. The assumption β ∈ W 1 , ∞ \beta\in W^{1,\infty} β ∈ W 1 , ∞ suffices for both constructions and is consistent with the convergence analyses in [51 , 52 ] .
The M M M -matrix property for spectral preservation
We now show that the exponential-fitting discretization (124) satisfies the spectral preservation property (Assumption 5.3) for arbitrary h h h and η \eta η . The strategy is algebraic: under suitable mesh conditions, the system matrices at the degrees k ∈ { 0 , n } k \in\{0,n\} k ∈ { 0 , n } are singular irreducible M M M -matrices, so the Perron–Frobenius theorem delivers exactly the required spectral distribution. The mesh conditions are as follows.
Assumption 6.7 (mesh condition). The triangulation is conforming and connected. For every edge E = [ a i , a j ] E=[a_i,a_j] E = [ a i , a j ] , the assembled stiffness weight satisfies
ω E : = − ∑ T ⊃ E ( d 0 φ i , d 0 φ j ) T ≥ 0. \omega_E := -\sum_{T\supset E} (d^0\varphi_i,d^0\varphi_j)_T \ge0. ω E := − T ⊃ E ∑ ( d 0 φ i , d 0 φ j ) T ≥ 0. For every facet degree of freedom F F F , the lumped mass weight satisfies
M F : = 1 ∣ F ∣ ∑ T ⊃ F d T , F > 0. \mathsf{M}_F := \frac{1}{|F|}\sum_{T\supset F} d_{T,F} > 0. M F := ∣ F ∣ 1 T ⊃ F ∑ d T , F > 0.
Remark 6.8 (geometric meaning of Assumption 6.7). The relevant conditions are imposed on the assembled weights:
ω E ≥ 0 for every edge E , M F > 0 for every facet F . \omega_E \ge0 \quad\text{for every edge } E,\qquad\mathsf{M}_F>0 \quad\text{for every facet } F. ω E ≥ 0 for every edge E , M F > 0 for every facet F .
Two dimensions. Let E = [ a a , a b ] E=[a_a,a_b] E = [ a a , a b ] be an interior edge shared by T 1 , T 2 T_1,T_2 T 1 , T 2 , and let α i \alpha_i α i be the angle of T i T_i T i opposite E E E . Then
ω E T i = − ( d 0 φ a , d 0 φ b ) T i = 1 2 cot α i , ω E = 1 2 ( cot α 1 + cot α 2 ) = sin ( α 1 + α 2 ) 2 sin α 1 sin α 2 . \omega_E^{T_i}=-(d^0\varphi_a,d^0\varphi_b)_{T_i}=\frac{1}{2}\cot\alpha_i,\qquad\omega_E=\frac{1}{2}(\cot\alpha_1+\cot\alpha_2)=\frac{\sin(\alpha_1+\alpha_2)}{2\sin\alpha_1\sin\alpha_2}. ω E T i = − ( d 0 φ a , d 0 φ b ) T i = 2 1 cot α i , ω E = 2 1 ( cot α 1 + cot α 2 ) = 2 sin α 1 sin α 2 sin ( α 1 + α 2 ) .
Thus ω E ≥ 0 \omega_E\ge0 ω E ≥ 0 is equivalent to the Delaunay condition α 1 + α 2 ≤ π \alpha_1+\alpha_2\le\pi α 1 + α 2 ≤ π , whereas ω E > 0 \omega_E>0 ω E > 0 is equivalent to strict Delaunay. Moreover,
d T i , E = ∣ E ∣ 2 cot α i , M E = d T 1 , E + d T 2 , E ∣ E ∣ = ω E . d_{T_i,E}=\frac{|E|}{2}\cot\alpha_i,\qquad\mathsf{M}_E=\frac{d_{T_1,E}+d_{T_2,E}}{|E|}=\omega_E. d T i , E = 2 ∣ E ∣ cot α i , M E = ∣ E ∣ d T 1 , E + d T 2 , E = ω E .
Consequently, on a periodic triangular mesh, strict Delaunay ensures both requirements.
Three dimensions. For an edge E = [ a a , a b ] E=[a_a,a_b] E = [ a a , a b ] of a tetrahedron T T T , let E T ′ E'_T E T ′ be the opposite edge and let θ E T ′ T \theta_{E'_T}^T θ E T ′ T be the interior dihedral angle along E T ′ E'_T E T ′ . The cotangent formula is
ω E = 1 6 ∑ T ⊃ E ∣ E T ′ ∣ cot θ E T ′ T . \omega_E=\frac{1}{6}\sum_{T\supset E}|E'_T|\cot\theta_{E'_T}^T. ω E = 6 1 T ⊃ E ∑ ∣ E T ′ ∣ cot θ E T ′ T .
Thus the sign requirement is an assembled dihedral-angle condition [52 ] . Elementwise nonobtuse dihedral angles are sufficient for ω E ≥ 0 \omega_E \ge0 ω E ≥ 0 , but this is stronger than the assembled condition.
For an interior facet F = T + ∩ T − F = T^+ \cap T^- F = T + ∩ T − , the precise positivity condition is
M F = d T + , F + d T − , F ∣ F ∣ > 0. M_F = \frac{d_{T^+,F} + d_{T^-,F}}{\lvert F\rvert} > 0. M F = ∣ F ∣ d T + , F + d T − , F > 0.
A strict Delaunay pair T + , T − T^+, T^- T + , T − satisfies this condition in any dimension, by positivity of its signed circumcentric dual length [30 Theorem 3] .
The spectral conclusion is drawn from the following standard consequence of the Perron–Frobenius theorem [21 , 40 ] .
Lemma 6.9 (singular irreducible M M M -matrices). Let A ∈ R N × N A \in\mathbb{R}^{N \times N} A ∈ R N × N , N ≥ 2 N \ge2 N ≥ 2 , have nonpositive off-diagonal entries and zero row sums (or zero column sums). Assume further that A A A has a symmetric sparsity pattern and is irreducible; equivalently, the graph on the vertex set { 1 , … , N } \{1,\ldots,N\} { 1 , … , N } whose edges are the pairs { i , j } \{i,j\} { i , j } with A i j < 0 A_{ij} < 0 A ij < 0 and A j i < 0 A_{ji} < 0 A j i < 0 is connected. Then 0 0 0 is an algebraically simple eigenvalue of A A A and every other eigenvalue satisfies Re λ > 0 \operatorname{Re}\lambda> 0 Re λ > 0 . The hypotheses, in the zero-row-sum case, are preserved under left multiplication by a positive diagonal matrix.
Proof. Without loss of generality, assume A 1 = 0 A\mathbf{1} = 0 A 1 = 0 . Set s : = max i A i i s := \max_i A_{ii} s := max i A ii and B : = s I − A B := sI - A B := s I − A , then B ≥ 0 B \ge0 B ≥ 0 entrywise, B B B is irreducible (its off-diagonal pattern is that of A A A ), and B 1 = s 1 B\mathbf{1} = s\mathbf{1} B 1 = s 1 . By the Perron–Frobenius theorem, an irreducible nonnegative matrix with a strictly positive eigenvector has that eigenvalue equal to its spectral radius, and the latter is algebraically simple. Hence ρ ( B ) = s \rho(B) = s ρ ( B ) = s is algebraically simple and 0 = s − ρ ( B ) 0 = s - \rho(B) 0 = s − ρ ( B ) is an algebraically simple eigenvalue of A A A . For ξ ∈ σ ( B ) ∖ { s } \xi\in\sigma(B) \setminus\{s\} ξ ∈ σ ( B ) ∖ { s } we have ∣ ξ ∣ ≤ s \lvert\xi\rvert\le s ∣ ξ ∣ ≤ s and ξ ≠ s \xi\ne s ξ = s , which forces Re ξ < s \operatorname{Re}\xi< s Re ξ < s . Thus every nonzero λ = s − ξ ∈ σ ( A ) \lambda= s - \xi\in\sigma(A) λ = s − ξ ∈ σ ( A ) has Re λ > 0 \operatorname{Re}\lambda> 0 Re λ > 0 . Meanwhile, left multiplication by a positive diagonal matrix does not alter the required pattern and therefore preserves the same property. □
We next verify that the proposed exponential-fitting discretization satisfies the required spectral preservation assumption.
Lemma 6.10. Under Assumption 6.7, the discretization (54) with i β , h \mathbf{i}_{\beta,h} i β , h defined by (109) and (124) satisfies Assumption 5.3 for every h > 0 h > 0 h > 0 and every η > 0 \eta> 0 η > 0 .
Proof. Case k = 0 k = 0 k = 0 . An eigenpair ( λ h , u h ) (\lambda_h,u_h) ( λ h , u h ) of A h 0 A_h^0 A h 0 satisfies
λ h ( u h , v h ) = ( A h 0 u h , v h ) = ( d h 0 u h , J h 0 v h ) = ( d h 0 u h , Π ˉ h 1 J ~ h 0 ( M ~ h 0 ) − 1 v h ) , ∀ v h ∈ V h 0 . (125) \lambda_h(u_h,v_h) = (A_h^0u_h,v_h) = (d_h^0u_h,J_h^0v_h) = (d_h^0u_h,\bar{\Pi}_h^1\tilde{J}_h^0(\tilde{M}_h^0)^{-1}v_h), \qquad\forall v_h \in V_h^0. \tag*{(125)} λ h ( u h , v h ) = ( A h 0 u h , v h ) = ( d h 0 u h , J h 0 v h ) = ( d h 0 u h , Π ˉ h 1 J ~ h 0 ( M ~ h 0 ) − 1 v h ) , ∀ v h ∈ V h 0 . (125)
Substituting v h ↦ M ~ h 0 v h v_h \mapsto\tilde{M}_h^0v_h v h ↦ M ~ h 0 v h and using (120) –(121) yields the generalized eigenvalue problem
λ h ( u h , v h ) h = ( d h 0 u h , Π ˉ h 1 J ~ h 0 v h ) = η ∑ E ∈ Σ h 1 ω E E E δ E ( u h ) δ E ( e ψ E / η v h ) , (126) \lambda_h(u_h,v_h)_h = (d_h^0u_h,\bar{\Pi}_h^1\tilde{J}_h^0v_h) = \eta\sum_{E\in\Sigma_h^1}\omega_E\mathcal{E}_E\delta_E(u_h)\delta_E(e^{\psi_E/\eta}v_h), \tag*{(126)} λ h ( u h , v h ) h = ( d h 0 u h , Π ˉ h 1 J ~ h 0 v h ) = η E ∈ Σ h 1 ∑ ω E E E δ E ( u h ) δ E ( e ψ E / η v h ) , (126)
where the elemental weights were summed to ω E = ∑ T ⊃ E ω E T \omega_E = \sum_{T\supset E}\omega_E^T ω E = ∑ T ⊃ E ω E T because ψ E \psi_E ψ E and E E \mathcal{E}_E E E are defined edgewise. Let u ∈ R N 0 \mathbf{u} \in\mathbb{R}^{N^0} u ∈ R N 0 denote the coefficient vector of u h u_h u h in the nodal basis. In matrix form, this is A 0 u = λ h M L u A^0\mathbf{u} = \lambda_h M_L\mathbf{u} A 0 u = λ h M L u with M L = diag ( m z ) M_L = \operatorname{diag}(m_z) M L = diag ( m z ) . In particular, for the oriented edge E = a i a j ‾ E = \overline{a_i a_j} E = a i a j , the off-diagonal entry is
A i j 0 = η ω E E E δ E ( φ j ) δ E ( e ψ E / η φ i ) = − η ω E E E e ψ E ( a i ) / η ≤ 0 , (127) A_{ij}^0 = \eta\omega_E\mathcal{E}_E\delta_E(\varphi_j)\delta_E(e^{\psi_E/\eta}\varphi_i) = -\eta\omega_E\mathcal{E}_E e^{\psi_E(a_i)/\eta} \le0, \tag*{(127)} A ij 0 = η ω E E E δ E ( φ j ) δ E ( e ψ E / η φ i ) = − η ω E E E e ψ E ( a i ) / η ≤ 0 , (127)
and likewise A j i 0 = − n ω E E E e ψ E ( a j ) / η ≤ 0 A_{ji}^{0}=-n\omega_E E_E e^{\psi^E(a_j)/\eta}\leq0 A j i 0 = − n ω E E E e ψ E ( a j ) / η ≤ 0 , while the entry vanishes for non-adjacent vertices. The off-diagonal entries are nonpositive, and the sparsity pattern is symmetric, since sign ( A i j ) = sign ( A j i ) = sign ( − ω E ) \operatorname{sign}(A_{ij})=\operatorname{sign}(A_{ji})=\operatorname{sign}(-\omega_E) sign ( A ij ) = sign ( A j i ) = sign ( − ω E ) and ω E ≥ 0 \omega_E\geq0 ω E ≥ 0 . Moreover, since d 0 1 = 0 d^0 1=0 d 0 1 = 0 , the row sums vanish: ∑ j A i j 0 = ( d 0 1 , Π h 0 φ i ) = 0 \sum_j A_{ij}^{0}=(d^0 1,\Pi_h^0\varphi_i)=0 ∑ j A ij 0 = ( d 0 1 , Π h 0 φ i ) = 0 .
Let G + G_+ G + be the undirected graph whose vertices are the mesh vertices and whose edges are those with ω E > 0 \omega_E>0 ω E > 0 . We first show that G + G_+ G + is connected. By the definition of ω E \omega_E ω E , every v h ∈ V h 0 v_h\in V_h^0 v h ∈ V h 0 satisfies
∥ d 0 v h ∥ L 2 ( M ) 2 = ∑ E = [ a i , a j ] ∈ Σ h 1 ω E ∣ v h ( a i ) − v h ( a j ) ∣ 2 . \lVert d^0v_h\rVert_{L^2(M)}^2=\sum_{E=[a_i,a_j]\in\Sigma_h^1}\omega_E\lvert v_h(a_i)-v_h(a_j)\rvert^2. ∥ d 0 v h ∥ L 2 ( M ) 2 = E = [ a i , a j ] ∈ Σ h 1 ∑ ω E ∣ v h ( a i ) − v h ( a j ) ∣ 2 .
If G + G_+ G + were disconnected, choose v h v_h v h with nodal value 1 on one connected component and 0 on all other components. Every positive-weight edge then joins vertices with equal nodal values, while zero-weight edges contribute nothing. Hence d 0 v h = 0 d^0v_h=0 d 0 v h = 0 . Since M M M is connected and v h v_h v h is continuous, v h v_h v h must be constant, contradicting its construction. Thus G + G_+ G + is connected.
By (127) , an edge E = [ a i , a j ] E=[a_i,a_j] E = [ a i , a j ] belongs to G + G_+ G + if and only if A i j 0 < 0 A_{ij}^{0}<0 A ij 0 < 0 and A j i 0 < 0 A_{ji}^{0}<0 A j i 0 < 0 , i.e., ω E > 0 \omega_E>0 ω E > 0 . Therefore the directed adjacency graph of A 0 A^0 A 0 is obtained from G + G_+ G + by replacing each edge with two oppositely directed edges. It is strongly connected, and hence A 0 A^0 A 0 is irreducible.
Left multiplication by the positive diagonal matrix M L − 1 M_L^{-1} M L − 1 preserves the sign pattern, the zero row sums, and the irreducibility. Hence M L − 1 A L 0 M_L^{-1}A_L^0 M L − 1 A L 0 is a singular irreducible M M M -matrix, and by Lemma 6.9 , its zero eigenvalue is algebraically simple while all other eigenvalues have positive real parts. This proves Assumption 5.3 for k = 0 k=0 k = 0 .
Case k = n k=n k = n . The stationary problem reads
( σ h , τ h ) h − ( u h , J ~ h n − 1 τ h ) = 0 , ∀ τ h ∈ V h n − 1 , (\sigma_h,\tau_h)_h-(u_h,\widetilde{J}_h^{n-1}\tau_h)=0,\qquad\forall\tau_h\in V_h^{n-1}, ( σ h , τ h ) h − ( u h , J h n − 1 τ h ) = 0 , ∀ τ h ∈ V h n − 1 ,
( d n − 1 σ h , v h ) = 0 , ∀ v h ∈ V h n . (d^{n-1}\sigma_h,v_h)=0,\qquad\forall v_h\in V_h^n. ( d n − 1 σ h , v h ) = 0 , ∀ v h ∈ V h n .
For each cell T T T , let φ T \varphi_T φ T be its Whitney basis function, normalized by ∫ T φ T = 1 \int_T\varphi_T=1 ∫ T φ T = 1 , and define the corresponding auxiliary flux σ T ∈ V h n − 1 \sigma_T\in V_h^{n-1} σ T ∈ V h n − 1 by
( σ T , τ h ) h = ( φ T , J ~ h n − 1 τ h ) , ∀ τ h ∈ V h n − 1 . (128) (\sigma_T,\tau_h)_h=(\varphi_T,\widetilde{J}_h^{n-1}\tau_h),\qquad\forall\tau_h\in V_h^{n-1}. \tag*{(128)} ( σ T , τ h ) h = ( φ T , J h n − 1 τ h ) , ∀ τ h ∈ V h n − 1 . (128)
Then A h n φ T = d n − 1 σ T A_h^n\varphi_T=d^{n-1}\sigma_T A h n φ T = d n − 1 σ T .
Writing T h = { T i } \mathcal{T}_h=\{T_i\} T h = { T i } , define
A i j n : = ( d n − 1 σ T j , φ T i ) , M n : = diag ( 1 / ∣ T i ∣ ) . A_{ij}^n:=(d^{n-1}\sigma_{T_j},\varphi_{T_i}),\qquad M^n:=\operatorname{diag}(1/\lvert T_i\rvert). A ij n := ( d n − 1 σ T j , φ T i ) , M n := diag ( 1/ ∣ T i ∣) .
The eigenvalue problem for A h n A_h^n A h n is therefore A n u = λ h M n u A^n u=\lambda_h M^n u A n u = λ h M n u . Hence its eigenvalues are those of A ^ : = ( M n ) − 1 A n \widehat{A}:=(M^n)^{-1}A^n A := ( M n ) − 1 A n . Since φ T i ∣ T i = vol T i / ∣ T i ∣ \varphi_{T_i}|_{T_i}=\operatorname{vol}T_i/\lvert T_i\rvert φ T i ∣ T i = vol T i / ∣ T i ∣ ,
A ^ i j = ∣ T i ∣ ( d n − 1 σ T j , φ T i ) = ∫ T i d n − 1 σ T j . (129) \widehat{A}_{ij}=\lvert T_i\rvert(d^{n-1}\sigma_{T_j},\varphi_{T_i})=\int_{T_i}d^{n-1}\sigma_{T_j}. \tag*{(129)} A ij = ∣ T i ∣ ( d n − 1 σ T j , φ T i ) = ∫ T i d n − 1 σ T j . (129)
We determine the signs of these entries from the facet fluxes. Since the facet mass is diagonal and φ T \varphi_T φ T is supported on T T T , (128) implies that the facet degrees of freedom of σ T \sigma_T σ T vanish outside ∂ T \partial T ∂ T . For a facet F ⊂ ∂ T F\subset\partial T F ⊂ ∂ T , orient F F F outward from T T T and let M F = ( φ F , φ F ) h > 0 M_F=(\varphi_F,\varphi_F)_h>0 M F = ( φ F , φ F ) h > 0 . Testing (128) with φ F \varphi_F φ F , applying Lemma 6.4 with $k=n-1 and S = T S=T S = T , and using d ψ T = β T b \mathrm{d}\psi_T=\beta_T^b d ψ T = β T b and Stokes’ theorem, we obtain
M F ∫ F σ T = ( φ T , J ~ h n − 1 φ F ) = E T ∣ T ∣ ∫ T e ψ T / η ( η d φ F + β T b ∧ φ F ) = η E T ∣ T ∣ ∫ ∂ T e ψ T / η tr φ F = η E T ∣ T ∣ ∫ F e ψ T / η > 0. M_F\int_F \sigma_T=(\varphi_T,\widetilde{J}_h^{n-1}\varphi_F)
=\frac{\mathcal{E}_T}{|T|}\int_T e^{\psi_T/\eta}(\eta\,\mathrm{d}\varphi_F+\beta_T^b\wedge\varphi_F)
=\frac{\eta\mathcal{E}_T}{|T|}\int_{\partial T}e^{\psi_T/\eta}\operatorname{tr}\varphi_F
=\frac{\eta\mathcal{E}_T}{|T|}\int_F e^{\psi_T/\eta}>0. M F ∫ F σ T = ( φ T , J h n − 1 φ F ) = ∣ T ∣ E T ∫ T e ψ T / η ( η d φ F + β T b ∧ φ F ) = ∣ T ∣ η E T ∫ ∂ T e ψ T / η tr φ F = ∣ T ∣ η E T ∫ F e ψ T / η > 0.
Here we used that tr F φ F = vol F / ∣ F ∣ \operatorname{tr}_F\varphi_F=\operatorname{vol}_F/|F| tr F φ F = vol F /∣ F ∣ in the chosen orientation, while its trace vanishes on the other facets of T T T .
Now suppose that T i ≠ T j T_i\ne T_j T i = T j share a facet F F F , oriented outward from T j T_j T j . The induced orientation from T i T_i T i is opposite, so (129) gives
A ^ i j = ∫ ∂ T i tr σ T j = − ∫ F σ T j = − η E T j M F ∣ T j ∣ ∫ F e ψ T j / η < 0. \widehat{A}_{ij}=\int_{\partial T_i}\operatorname{tr}\sigma_{T_j}=-\int_F\sigma_{T_j}=-\frac{\eta\mathcal{E}_{T_j}}{M_F|T_j|}\int_F e^{\psi_{T_j}/\eta}<0. A ij = ∫ ∂ T i tr σ T j = − ∫ F σ T j = − M F ∣ T j ∣ η E T j ∫ F e ψ T j / η < 0.
If T i T_i T i and T j T_j T j do not share a facet, then A ^ i j = 0 \widehat{A}_{ij}=0 A ij = 0 . Thus A ^ \widehat{A} A has nonpositive off-diagonal entries and a symmetric sparsity pattern.
Moreover, since M M M is periodic, Stokes’ theorem yields, for every column j j j ,
∑ i A ^ i j = ∑ i ∫ T i d n − 1 σ T j = ∫ M d n − 1 σ T j = 0. \sum_i\widehat{A}_{ij}=\sum_i\int_{T_i}\mathrm{d}^{n-1}\sigma_{T_j}=\int_M\mathrm{d}^{n-1}\sigma_{T_j}=0. i ∑ A ij = i ∑ ∫ T i d n − 1 σ T j = ∫ M d n − 1 σ T j = 0.
Since M M M is connected and the triangulation is conforming, the cell adjacency graph is connected. The strictly negative entries in both directions between neighboring cells therefore imply that A ^ \widehat{A} A is irreducible.
By Lemma 6.9 , zero is algebraically simple and all remaining eigenvalues have positive real parts. This proves Assumption 5.3 for k = n k=n k = n .
Remark 6.11. The main advantage of the exponential-fitting discretization is that it establishes the spectral property (Assumption 5.3 ) for every η > 0 \eta>0 η > 0 and every h > 0 h>0 h > 0 . The uniform elemental bound in Lemma 6.5 does not directly establish Assumption 5.1 for the cases k = 0 , n − 1 k=0,n-1 k = 0 , n − 1 of (124) due to mass-lumping operation. For these cases, we do not explicitly verify the L 2 L^2 L 2 -boundedness uniformly in h h h and η \eta η , since this condition is mainly needed for Theorem 5.2 . Nevertheless, for each fixed h h h and η \eta η , the resulting operator is a bounded linear operator on a finite-dimensional space. Hence this possible lack of uniform boundedness does not affect the finite element antidynamo result in Theorem 5.10 .
Remark 6.12. If an elementwise constant approximation is also used for k = 0 k=0 k = 0 to construct the flux operator [51 ] , the exponential-fitting factors may differ between elements sharing an edge. In that case, the double sum over elements and edges cannot generally be reduced to (126) using only the assembled weights ω E = ∑ T ⊃ E ω E T \omega_E=\sum_{T\supset E}\omega_E^T ω E = ∑ T ⊃ E ω E T . Instead, nonpositive off-diagonal entries require the corresponding element-dependent weighted sums of ω E T \omega_E^T ω E T to be nonnegative. In this case, a sufficient condition is
ω E T ≥ 0 for every element T and edge E ⊂ T . \omega_E^T\ge0\qquad\text{for every element }T\text{ and edge }E\subset T. ω E T ≥ 0 for every element T and edge E ⊂ T .
This is stronger than the assembled condition ω E ≥ 0 \omega_E\ge0 ω E ≥ 0 used here, although it is not necessary: negative elemental contributions may still be compensated by positive ones after weighted assembly.
Numerical experiments
We present simple numerical experiments to illustrate the importance of an appropriate spatial discretization. Although a complete analysis of the fully discrete problem requires additional work, backward Euler is compatible with the 0- and n n n -form problems whenever the spatial discretization satisfies the spectral preservation requirement. Motivated by this, we use backward Euler in the following two-dimensional 1-form problems to explore the influence of the spatial discretization. We compute
u h m + 1 = ( I + Δ t A h 1 ) − 1 u h m , t m = m Δ t , (130) u_h^{m+1} = (I + \Delta t A_h^1)^{-1}u_h^m,\qquad t_m = m\Delta t, \tag*{(130)} u h m + 1 = ( I + Δ t A h 1 ) − 1 u h m , t m = m Δ t , (130)
where Δ t \Delta t Δ t denotes the time-step size.
We consider the flat torus M = [ 0 , 2 π ) 2 M = [0,2\pi)^2 M = [ 0 , 2 π ) 2 with periodic boundary conditions. We use the lowest-order Whitney finite elements on acute simplicial triangulations satisfying the mesh conditions, and denote the mesh resolution by h = 2 π / N h = 2\pi/N h = 2 π / N . All methods on a given mesh start from the interpolant of
ω 0 = ( 0.6 + 0.25 cos x + 0.15 sin ( x + y ) ) d x + ( 0.6 + 0.2 sin y + 0.1 cos ( x − y ) ) d y . (131) \omega_0 = (0.6 + 0.25\cos x + 0.15\sin(x+y))\,dx
+ (0.6 + 0.2\sin y + 0.1\cos(x-y))\,dy. \tag*{(131)} ω 0 = ( 0.6 + 0.25 cos x + 0.15 sin ( x + y )) d x + ( 0.6 + 0.2 sin y + 0.1 cos ( x − y )) d y . (131)
We plot the normalized L 2 L^2 L 2 norm
R 0 ( t m ) = ∥ u h m ∥ ∥ u h 0 ∥ . (132) R_0(t_m) = \frac{\lVert u_h^m\rVert}{\lVert u_h^0\rVert}. \tag*{(132)} R 0 ( t m ) = ∥ u h 0 ∥ ∥ u h m ∥ . (132)
Weighted Galerkin evolution
We first compare the weighted-inner-product Galerkin discretization with the standard discretization, which uses the L 2 L^2 L 2 -projected contraction. We take η = 0.5 \eta= 0.5 η = 0.5 , N ∈ { 8 , 16 , 32 } N \in\{8,16,32\} N ∈ { 8 , 16 , 32 } , and Δ t = 0.1 \Delta t = 0.1 Δ t = 0.1 on 0 ≤ t ≤ 100 0 \leq t \leq100 0 ≤ t ≤ 100 . The advection field is a potential flow
β = ( 2 sin 2 x , 2 sin 2 y ) = ∇ Φ , Φ = − cos 2 x − cos 2 y , θ = Φ / η , ρ = e − θ . \beta= (2\sin2x,2\sin2y) = \nabla\Phi,\qquad
\Phi= -\cos2x-\cos2y,\qquad
\theta= \Phi/\eta,\qquad
\rho= e^{-\theta}. β = ( 2 sin 2 x , 2 sin 2 y ) = ∇Φ , Φ = − cos 2 x − cos 2 y , θ = Φ/ η , ρ = e − θ .
The weighted method is the Galerkin discretization (49) with the inner product weighted by ρ \rho ρ . In addition to ∥ ⋅ ∥ \lVert\cdot\rVert ∥ ⋅ ∥ , we measure both solutions using
∥ v ∥ θ 2 = ∫ M ρ v ∧ ⋆ v , R θ ( t m ) = ∥ u h m ∥ θ ∥ u h 0 ∥ θ . (133) \lVert v\rVert_\theta^2 = \int_M \rho v\wedge\star v,\qquad
R_\theta(t_m) = \frac{\lVert u_h^m\rVert_\theta}{\lVert u_h^0\rVert_\theta}. \tag*{(133)} ∥ v ∥ θ 2 = ∫ M ρ v ∧ ⋆ v , R θ ( t m ) = ∥ u h 0 ∥ θ ∥ u h m ∥ θ . (133)
Figure 7.1 shows the evolution of the two norms. On the coarsest mesh, the standard discretization exhibits growth in the weighted norm after an initial decrease. This effect is reduced under mesh refinement, and the two discretizations agree more closely on the finest mesh. The mesh dependence is consistent with the need for sufficiently small h h h in Lemma 6.1 .
Figure 7.1. Weighted Galerkin and standard evolution for the potential field with η = 0.5 \eta= 0.5 η = 0.5 and Δ t = 0.1 \Delta t = 0.1 Δ t = 0.1 . Columns correspond to N = 8 , 16 , 32 N = 8, 16, 32 N = 8 , 16 , 32 . The top row shows the normalized L 2 L^2 L 2 norm R 0 R_0 R 0 , and the bottom row shows the normalized weighted norm R θ R_\theta R θ . The main panels cover 0 ≤ t ≤ 100 0 \le t \le100 0 ≤ t ≤ 100 ; the insets show 0 ≤ t ≤ 5 0 \le t \le5 0 ≤ t ≤ 5 .
In contrast, the weighted discretization gives a non-increasing weighted norm on all three meshes. This agrees with the weighted dissipativity in Theorem 3.16 , which is also preserved by backward Euler. The L 2 L^2 L 2 norm need not decrease monotonically and may exhibit transient growth, while the discrete solution still converges exponentially to a stationary state.
Comparison between exponential-fitting and standard discretizations
We next compare the exponential-fitting discretization with the standard method, using the non-potential field
β = ( 2 sin 2 x + 0.5 sin y , 2 sin 2 y − 0.5 sin x ) . \beta= (2\sin2x + 0.5\sin y,\; 2\sin2y - 0.5\sin x). β = ( 2 sin 2 x + 0.5 sin y , 2 sin 2 y − 0.5 sin x ) .
We consider η ∈ { 0.5 , 0.1 } \eta\in\{0.5, 0.1\} η ∈ { 0.5 , 0.1 } on the meshes N ∈ { 8 , 16 , 32 } N \in\{8, 16, 32\} N ∈ { 8 , 16 , 32 } , with Δ t = 0.1 \Delta t = 0.1 Δ t = 0.1 and 0 ≤ t ≤ 100 0 \leq t \leq100 0 ≤ t ≤ 100 .
Figure 7.2 shows a separation between the two trajectories on the coarsest mesh, where the standard discretization exhibits substantial growth that becomes more pronounced as the diffusion coefficient decreases. This large growth is not observed on the two finer meshes over the displayed interval. For both diffusion coefficients, refinement brings the two trajectories closer. In particular, the exponential-fitting discretization produces lower L 2 L^2 L 2 norms than the standard discretization on all three meshes. This behavior may reflect additional numerical diffusion associated with exponential fitting.
Figure 7.2. Exponential-fitting and standard evolution of the normalized L 2 L^2 L 2 norm R 0 R_0 R 0 with Δ t = 0.1 \Delta t = 0.1 Δ t = 0.1 . Rows correspond to η = 0.5 \eta= 0.5 η = 0.5 and η = 0.1 \eta= 0.1 η = 0.1 ; columns correspond to N = 8 , 16 , 32 N = 8, 16, 32 N = 8 , 16 , 32 . The main panels cover 0 ≤ t ≤ 100 0 \le t \le100 0 ≤ t ≤ 100 , and the insets show 0 ≤ t ≤ 5 0 \le t \le5 0 ≤ t ≤ 5 .
It is worth emphasizing that exponential fitting was originally developed and has been widely used to stabilize advection–diffusion problems. Here our motivation is its ability to preserve the correct spectral distribution for the 0- and n n n -form operators without a mesh size restriction, under some mesh conditions. This spectral preservation ensures that the discretization reproduces the correct qualitative long-time asymptotic behavior predicted by the continuous theory, even on coarse meshes and with a small diffusion coefficient.
Influence of mass lumping
Finally, we examine whether exponential fitting alone is sufficient when the mass lumpings are not used. Here nodal lumping refers to the mass on 0-forms, whereas facet lumping refers to the auxiliary flux mass in the 2-form discretization.
We use two fixed smooth periodic velocity fields, each at η ∈ { 0.005 , 0.001 } \eta\in\{0.005, 0.001\} η ∈ { 0.005 , 0.001 } . The first is the trigonometric polynomial
β I ( x , y ) = c + ∑ k ∈ K [ a k cos ( k 1 x + k 2 y ) + b k sin ( k 1 x + k 2 y ) ] , (134) \boldsymbol{\beta}_{I}(x,y) = c + \sum_{k \in\mathcal{K}} \left[a_k \cos(k_1x+k_2y) + b_k \sin(k_1x+k_2y)\right], \tag*{(134)} β I ( x , y ) = c + k ∈ K ∑ [ a k cos ( k 1 x + k 2 y ) + b k sin ( k 1 x + k 2 y ) ] , (134)
where
K = { ( 0 , j ) : 1 ≤ j ≤ 8 } ∪ { ( i , j ) : 1 ≤ i ≤ 4 , − 8 ≤ j ≤ 8 } . \mathcal{K} = \{(0,j): 1 \leq j \leq8\} \cup\{(i,j): 1 \leq i \leq4,\ -8 \leq j \leq8\}. K = {( 0 , j ) : 1 ≤ j ≤ 8 } ∪ {( i , j ) : 1 ≤ i ≤ 4 , − 8 ≤ j ≤ 8 } .
The second field has the simpler form
β I I ( x , y ) = 3 R ( 0.78 ) ( sin ( 5 x + 0.85 ) cos ( 5 y + 2.95 ) cos ( 5 x + 0.85 ) sin ( 5 y + 2.95 ) ) , R ( α ) = ( cos α − sin α sin α cos α ) . (135) \boldsymbol{\beta}_{II}(x,y) = 3\mathcal{R}(0.78)
\begin{pmatrix}
\sin(5x+0.85)\cos(5y+2.95)\\
\cos(5x+0.85)\sin(5y+2.95)
\end{pmatrix},
\qquad
\mathcal{R}(\alpha) =
\begin{pmatrix}
\cos\alpha& -\sin\alpha\\
\sin\alpha& \cos\alpha
\end{pmatrix}. \tag*{(135)} β I I ( x , y ) = 3 R ( 0.78 ) ( sin ( 5 x + 0.85 ) cos ( 5 y + 2.95 ) cos ( 5 x + 0.85 ) sin ( 5 y + 2.95 ) ) , R ( α ) = ( cos α sin α − sin α cos α ) . (135)
For this field, we compare the results obtained with and without mass lumping.
Figure 7.3 and Figure 7.4 show that either partial removal of lumping can lead to sustained spurious growth, depending on the velocity field. For β I \beta_I β I , removing nodal lumping produces large oscillatory growth on the coarsest mesh at both diffusion coefficients. For β I I \beta_{II} β I I , removing facet lumping produces substantial growth on the coarsest mesh within a much shorter time interval. Removing both lumpings also produces growth in these coarse-mesh tests. In contrast, the fully lumped trajectories remain of order one throughout the displayed intervals. The early-time insets show that an initial increase in the physical norm can occur while this transient effect is distinct from the sustained growth of the unlumped variants.
Figure 7.3. Effect of nodal lumping for velocity β I \beta_I β I . Rows correspond to η = 0.005 \eta= 0.005 η = 0.005 and η = 0.001 \eta= 0.001 η = 0.001 ; columns correspond to N = 8 , 16 , 32 N = 8, 16, 32 N = 8 , 16 , 32 . The main panels cover 0 ≤ t ≤ 3000 0 \le t \le3000 0 ≤ t ≤ 3000 with Δ t = 0.1 \Delta t = 0.1 Δ t = 0.1 ; the insets show 0 ≤ t ≤ 100 0 \le t \le100 0 ≤ t ≤ 100 .
Figure 7.4. Effect of facet lumping for velocity β I I \beta_{\mathrm{II}} β II . Rows correspond to η = 0.005 \eta= 0.005 η = 0.005 and η = 0.001 \eta= 0.001 η = 0.001 ; columns correspond to N = 14 , 28 , 42 N = 14, 28, 42 N = 14 , 28 , 42 . The main panels show R 0 R_0 R 0 on 0 ≤ t ≤ 100 0 \le t \le100 0 ≤ t ≤ 100 with Δ t = 0.025 \Delta t = 0.025 Δ t = 0.025 ; the insets show 0 ≤ t ≤ 10 0 \le t \le10 0 ≤ t ≤ 10 .
We compute the spectrum of the discrete operator, which supports this distinction. On the coarsest mesh for β I \beta_I β I , the nodal-unlumped 0-form operator has a conjugate pair with real part approximately − 3.10 × 10 − 3 -3.10 \times10^{-3} − 3.10 × 1 0 − 3 at both diffusion coefficients. For β I I \beta_{II} β I I , the facet-unlumped 2-form operator has a real eigenvalue approximately − 0.1400 -0.1400 − 0.1400 at η = 0.005 \eta= 0.005 η = 0.005 and − 0.1408 -0.1408 − 0.1408 at η = 0.001 \eta= 0.001 η = 0.001 . The corresponding lumped operators have no eigenvalues with negative real part.
These comparisons highlight the importance of mass lumping in the exponential-fitting discretization: replacing the lumped mass matrices by consistent ones need not preserve the same qualitative behavior. Although mesh refinement may prevent such growth, its effect need not be monotone in h h h . For example, for β I \beta_I β I at η = 0.001 \eta= 0.001 η = 0.001 , the variant without any lumping shows growth at N = 32 N = 32 N = 32 , while none of the three variants shows strong growth at N = 16 N = 16 N = 16 over the displayed time interval.
Stationary kernels and dynamo action
The lower bound in Theorem 4.3 allows the dimension of the stationary space to exceed the Betti number. We investigate this possibility numerically and examine its relation to dynamo action. For the induction problem (1) , magnetic fields are represented by closed 1-forms in two dimensions and closed 2-forms in three dimensions. For a mode u h ( t ) = e s h t b h u_h(t) = e^{s_h t}b_h u h ( t ) = e s h t b h , the discrete evolution gives
A h k b h = − s h b h , s h ∈ C . A_h^k b_h = -s_h b_h,\qquad s_h \in\mathbb{C}. A h k b h = − s h b h , s h ∈ C .
Thus, Re s h > 0 \operatorname{Re}s_h > 0 Re s h > 0 indicates exponential growth, Re s h < 0 \operatorname{Re}s_h < 0 Re s h < 0 indicates decay, and s h = 0 s_h = 0 s h = 0 corresponds to a stationary mode. In all examples, we compare
dim ker A h k , dim S h k = dim ( ker A h k ∩ Z h k ) , dim ( ker A h k ∩ B h k ) . \dim\ker A_h^k,\qquad\dim S_h^k = \dim(\ker A_h^k \cap Z_h^k),\qquad\dim(\ker A_h^k \cap B_h^k). dim ker A h k , dim S h k = dim ( ker A h k ∩ Z h k ) , dim ( ker A h k ∩ B h k ) .
Numerical zeros are counted by singular values in the L 2 L^2 L 2 metric, using the common threshold 10 − 12 ∥ A h k ∥ L 2 → L 2 10^{-12}\lVert A_h^k\rVert_{L^2\to L^2} 1 0 − 12 ∥ A h k ∥ L 2 → L 2 .
Three-dimensional comparison
We consider the Roberts flow [42 ] β = ( cos y , sin x , cos x + sin y ) \beta= (\cos y,\sin x,\cos x + \sin y) β = ( cos y , sin x , cos x + sin y ) on T 3 = [ 0 , 2 π ) 3 \mathbb{T}^3 = [0,2\pi)^3 T 3 = [ 0 , 2 π ) 3 , with b 2 = 3 b_2 = 3 b 2 = 3 , using the standard lowest-order FEEC discretization with h = 2 π / N h = 2\pi/N h = 2 π / N . For each mesh, we seek positive values of the diffusion parameter η \eta η for which the stationary problem
A h 2 ( η ) b h = 0 , 0 ≠ b h ∈ B h 2 A_h^2(\eta)b_h = 0,\qquad0 \ne b_h \in\mathcal{B}_h^2 A h 2 ( η ) b h = 0 , 0 = b h ∈ B h 2
admits a solution. We call these values stationary parameter roots and select the largest one (denoted as η h ∗ \eta_h^* η h ∗ ). We then track the corresponding eigenspace at η = 0.99 η h ∗ \eta= 0.99\eta_h^* η = 0.99 η h ∗ and 1.01 η h ∗ 1.01\eta_h^* 1.01 η h ∗ .
As shown in Table 7.1 , the additional stationary modes are exact and satisfy the magnetic closedness constraint. At η = 0.99 η h ∗ \eta= 0.99\eta_h^* η = 0.99 η h ∗ , these modes have positive growth rates, demonstrating discrete dynamo action even though the stationary dimension has returned to b 2 b_2 b 2 . At η = 1.01 η h ∗ \eta= 1.01\eta_h^* η = 1.01 η h ∗ , the same modes decay. The stationary crossing therefore marks a change from decay to growth along this branch as R m = η − 1 R_m = \eta^{-1} R m = η − 1 increases. These observations connect strictness of the topological lower bound with a transition between growth and decay of the selected magnetic modes. A similar mechanism is illustrated by Arnold’s explicit example in [6 ] .
Table 7.1. Three-dimensional Roberts flow numerical tests.
Two-dimensional comparison
As a planar counterpart, we take β = ( cos y , sin x ) \boldsymbol{\beta}=(\cos y,\sin x) β = ( cos y , sin x ) on T 2 = [ 0 , 2 π ) 2 \mathbb{T}^{2}=[0,2\pi)^{2} T 2 = [ 0 , 2 π ) 2 , with b 1 = 2 b_{1}=2 b 1 = 2 . We compare the standard and exponential-fitting methods on acute periodic meshes, with h = 2 π / N h=2\pi/N h = 2 π / N . For the discretization on the meshes N = 8 , 16 , 32 N=8,16,32 N = 8 , 16 , 32 , no positive diffusivity η \eta η was found for which a nonzero stationary mode exists in B h 1 \mathcal{B}_{h}^{1} B h 1 .
The standard discretization nevertheless admits additional stationary modes. Table 7.2 reports one representative critical point selected independently on each of the meshes N = 16 N=16 N = 16 and 32 32 32 . The reported s h s_{h} s h is the growth rate of the additional mode that becomes stationary at η h ∗ \eta_{h}^{*} η h ∗ . This mode grows at η = 0.99 η h ∗ \eta=0.99\eta_{h}^{*} η = 0.99 η h ∗ and decays at η = 1.01 η h ∗ \eta=1.01\eta_{h}^{*} η = 1.01 η h ∗ . The closed stationary space retains dimension b 1 = 2 b_{1}=2 b 1 = 2 at all three parameter values.
Table 7.2. Two-dimensional comparison.
For the exponential-fitting method, the dimensions remain ( 2 , 2 , 0 ) (2,2,0) ( 2 , 2 , 0 ) throughout the parameter search. The table reports the largest real part among the nonzero full-space growth rates, so a negative value means that every nonstationary eigenmode decays. The reported values are negative on both meshes, consistent with the discrete antidynamo result, which guarantees this behavior for every η > 0 \eta>0 η > 0 .
For the standard method, spectral preservation is guaranteed for sufficiently small h h h when η \eta η and β \boldsymbol{\beta} β are fixed. The admissible set of mesh size may depend on these parameters, so the result does not provide a uniform guarantee as they vary. In particular, even on a fixed fine mesh, spurious spectrum and artificial growth may occur as the diffusivity is reduced. Under the stated mesh assumptions, the exponential-fitting discretization satisfies the spectral preservation assumption for every h > 0 h>0 h > 0 and η > 0 \eta>0 η > 0 , making it more robust in advection-dominated regimes.
Conclusion
In this paper, we developed a finite element framework for the long-time analysis of Lie advection–diffusion equations of differential forms. The main objective was to go beyond finite-time convergence as h → 0 h\to0 h → 0 , and to understand whether a fixed finite element discretization preserves the structural asymptotic behavior of the continuous system. Under a spectral preservation assumption for 0 0 0 - and n n n -forms, we established long-time asymptotic results for arbitrary positive diffusion coefficients of advection–diffusion equations. In particular, on two-dimensional compact manifolds, the discrete 1 1 1 -form evolution converges to a stationary closed form for every η > 0 \eta>0 η > 0 . We also verified the structural assumptions in representative cases. For sufficiently fine meshes, the spectral preservation follows from spectral approximation theory. For lowest-order Whitney forms, we adopt an exponential-fitting discretization motivated by the Witten deformation and the edge-averaged finite element method. Under suitable mesh conditions, the resulting operators possess the structure of singular irreducible M M M -matrices for arbitrary mesh sizes.
Several questions remain open. The first concerns boundary conditions. As noted by Arnold and Khesin [7 ] , “the preceding theory has not been settled for manifolds with boundary, though it certainly deserves to be.” Although Subsection 5.4 extends our discrete results to absolute and relative boundary conditions through the underlying de Rham complex structure, this extension is primarily structural. A complete analysis connecting these discrete results with the well-posedness, spectral properties, and long-time asymptotic behavior of the corresponding continuous boundary-value problems remains to be developed. In particular, extending the theory to more general boundary conditions, especially for degrees 0 < k < n 0 < k < n 0 < k < n , is an important direction for future research. A second direction is to identify broader classes of spatial discretizations that satisfy the structural assumptions introduced in this paper. In particular, semi-Lagrangian methods represent a promising candidate: since the Lie derivative is generated by pullback along the flow, a semi-Lagrangian FEEC method based on discrete pullbacks could naturally preserve the underlying cochain structure. These questions point toward a broader global numerical analysis of Lie advection–diffusion and its applications in topological hydrodynamics and magnetohydrodynamics.
Declaration of AI Use. During the preparation of this work, the authors used ChatGPT 6 Pro to set up the numerical experiments for determining some critical velocity fields. In addition, ChatGPT 5.6 Sol and Claude Fable 5 were used to refine the exposition, check mathematical arguments and proofs for potential errors or typographical issues, and suggest improvements to their presentation. The authors take full responsibility for the mathematical content, numerical results, and final manuscript.
Auxiliary proofs
Proof of the scalar spectral theorem
A.1. Proof of the scalar spectral theorem. Throughout, M M M is a smooth, compact, connected Riemannian manifold, either closed or with smooth boundary, and the periodic case is identified with M = T n M = \mathbb{T}^n M = T n . We prove Theorem 2.1 for β ∈ C 0 , 1 ( M ; T M ) \beta\in C^{0,1}(M;TM) β ∈ C 0 , 1 ( M ; T M ) in the closed, periodic, and Neumann cases, where the classical strong maximum principle and the Hopf lemma apply directly. The main simplification is that the principal eigenfunction is explicit, ϵ 0 = 1 \epsilon_0 = 1 ϵ 0 = 1 with λ 0 = 0 \lambda_0 = 0 λ 0 = 0 , so no Krein–Rutman argument is needed. Under Dirichlet conditions, constants are not admissible and the principal eigenpair is instead constructed using Krein–Rutman theory [35 ] , see also [10 ] . For β ∈ L ∞ \beta\in L^\infty β ∈ L ∞ the conclusions also persist. The pointwise arguments are replaced by De Giorgi–Nash regularity, the weak Harnack inequality, and the strong maximum principle for H 1 H^1 H 1 sub- and supersolutions [22 Thms. 8.18–8.19] . In the presence of a boundary, where the classical Hopf argument is not directly available, one may instead apply the Krein–Rutman theory on L 2 ( M ) L^2(M) L 2 ( M ) to compact irreducible positive resolvents and semigroups [17 ] [45 Ch. V, §5] . For brevity, we write L : = Δ η , β 0 = − η Δ g + β ⋅ ∇ L := \Delta^0_{\eta,\beta} = -\eta\Delta_g + \beta\cdot\nabla L := Δ η , β 0 = − η Δ g + β ⋅ ∇ , where Δ g \Delta_g Δ g denotes the Laplace–Beltrami operator.
Proof of Theorem 2.1 in the closed and Neumann cases. On V = H 1 ( M ) V = H^1(M) V = H 1 ( M ) , define
a ( u , v ) : = η ( ∇ u , ∇ v ) + ( β ⋅ ∇ u , v ) . a(u,v) := \eta(\nabla u,\nabla v) + (\beta\cdot\nabla u,v). a ( u , v ) := η ( ∇ u , ∇ v ) + ( β ⋅ ∇ u , v ) .
Writing B ∞ : = ∥ β ∥ L ∞ ( M ) B_{\infty}:=\|\beta\|_{L^{\infty}(M)} B ∞ := ∥ β ∥ L ∞ ( M ) , for every ε > 0 \varepsilon>0 ε > 0 ,
∣ ( β ⋅ ∇ u , u ) ∣ ≤ ε η ∥ ∇ u ∥ 2 + B ∞ 2 4 ε η ∥ u ∥ 2 . |(\beta\cdot\nabla u,u)|\leq\varepsilon\eta\|\nabla u\|^{2}+\frac{B_{\infty}^{2}}{4\varepsilon\eta}\|u\|^{2}. ∣ ( β ⋅ ∇ u , u ) ∣ ≤ ε η ∥∇ u ∥ 2 + 4 ε η B ∞ 2 ∥ u ∥ 2 .
Hence a a a is a closed sectorial form, and Kato’s first representation theorem [34 Ch. VI, Thm. 2.1] yields the sectorial operator L L L . Standard elliptic regularity gives
D ( L ) = { H 2 ( M ) , ∂ M = ∅ , { u ∈ H 2 ( M ) : ∂ ν u = 0 } , ∂ M ≠ ∅ . D(L)=\left\{
\begin{array}{ll}
H^{2}(M), & \partial M=\varnothing,\\
\{u\in H^{2}(M):\partial_{\nu}u=0\}, & \partial M\neq\varnothing.
\end{array}
\right. D ( L ) = { H 2 ( M ) , { u ∈ H 2 ( M ) : ∂ ν u = 0 } , ∂ M = ∅ , ∂ M = ∅ .
Since H 1 ( M ) ↪ ↪ L 2 ( M ) H^{1}(M)\hookrightarrow\hookrightarrow L^{2}(M) H 1 ( M ) ↪↪ L 2 ( M ) , L L L has compact resolvent. Hence its spectrum consists of isolated eigenvalues of finite algebraic multiplicity.
Let L u = λ u Lu=\lambda u Lu = λ u , ∥ u ∥ L 2 = 1 \|u\|_{L^{2}}=1 ∥ u ∥ L 2 = 1 , and X : = ∥ ∇ u ∥ L 2 X:=\|\nabla u\|_{L^{2}} X := ∥∇ u ∥ L 2 . Then
Re λ ≥ η X 2 − B ∞ X ≥ η 2 X 2 − B ∞ 2 2 η , ∣ Im λ ∣ ≤ B ∞ X . \operatorname{Re}\lambda\geq\eta X^{2}-B_{\infty}X\geq\frac{\eta}{2}X^{2}-\frac{B_{\infty}^{2}}{2\eta},\qquad|\operatorname{Im}\lambda|\leq B_{\infty}X. Re λ ≥ η X 2 − B ∞ X ≥ 2 η X 2 − 2 η B ∞ 2 , ∣ Im λ ∣ ≤ B ∞ X .
Thus
∣ Im λ ∣ 2 ≤ c 1 ( Re λ + c 2 ) , c 1 : = 2 B ∞ 2 η , c 2 : = B ∞ 2 2 η . (136) |\operatorname{Im}\lambda|^{2}\leq c_{1}(\operatorname{Re}\lambda+c_{2}),\qquad c_{1}:=\frac{2B_{\infty}^{2}}{\eta},\qquad c_{2}:=\frac{B_{\infty}^{2}}{2\eta}. \tag*{(136)} ∣ Im λ ∣ 2 ≤ c 1 ( Re λ + c 2 ) , c 1 := η 2 B ∞ 2 , c 2 := 2 η B ∞ 2 . (136)
Since L 1 = 0 L1=0 L 1 = 0 , 0 ∈ σ ( L ) 0\in\sigma(L) 0 ∈ σ ( L ) . Let
L u = λ u , u ≠ 0 , Lu=\lambda u,\qquad u\neq0, Lu = λ u , u = 0 ,
and set w : = ∣ u ∣ 2 w:=|u|^{2} w := ∣ u ∣ 2 . Since β ∈ C 0 , 1 ⊂ C 0 , α \beta\in C^{0,1}\subset C^{0,\alpha} β ∈ C 0 , 1 ⊂ C 0 , α , elliptic regularity upgrades the eigenfunction to a classical solution, so that w ∈ C 2 ( M ‾ ) w\in C^{2}(\overline{M}) w ∈ C 2 ( M ) and the maximum principle applies pointwise.
− η Δ g w + β ⋅ ∇ w = 2 Re λ w − 2 η ∣ ∇ u ∣ 2 . -\eta\Delta_{g}w+\beta\cdot\nabla w=2\operatorname{Re}\lambda\,w-2\eta|\nabla u|^{2}. − η Δ g w + β ⋅ ∇ w = 2 Re λ w − 2 η ∣∇ u ∣ 2 .
In the Neumann case,
∂ ν w = 0 on ∂ M . \partial_{\nu}w=0\qquad\text{on }\partial M. ∂ ν w = 0 on ∂ M .
If Re λ ≤ 0 \operatorname{Re}\lambda\leq0 Re λ ≤ 0 , then
η Δ g w − β ⋅ ∇ w ≥ 0. \eta\Delta_{g}w-\beta\cdot\nabla w\geq0. η Δ g w − β ⋅ ∇ w ≥ 0.
The strong maximum principle, together with the Hopf lemma in the Neumann case, implies that w w w is constant. Substituting back gives
Re λ = 0 , ∇ u = 0 , \operatorname{Re}\lambda=0,\qquad\nabla u=0, Re λ = 0 , ∇ u = 0 ,
so u u u is constant and therefore λ = 0 \lambda=0 λ = 0 . Hence
N ( L ) = span { 1 } , Re λ > 0 ∀ λ ∈ σ ( L ) ∖ { 0 } . (137) \mathcal{N}(L)=\operatorname{span}\{1\},\qquad\operatorname{Re}\lambda>0\quad\forall\lambda\in\sigma(L)\setminus\{0\}. \tag*{(137)} N ( L ) = span { 1 } , Re λ > 0 ∀ λ ∈ σ ( L ) ∖ { 0 } . (137)
Suppose that 0 0 0 is not algebraically simple. Since N ( L ) = span { 1 } \mathcal{N}(L)=\operatorname{span}\{1\} N ( L ) = span { 1 } , after rescaling there is v ∈ D ( L ) v\in D(L) v ∈ D ( L ) such that
Because the coefficients are real, q : = Re v q:=\operatorname{Re}v q := Re v satisfies
Proof of lm:expA
with ∂ ν q = 0 \partial_{\nu}q=0 ∂ ν q = 0 in the Neumann case. At an interior minimum of q q q ,
1 = L q = − η Δ g q ≤ 0 , 1=Lq=-\eta\Delta_gq\le0, 1 = L q = − η Δ g q ≤ 0 ,
which is a contradiction. In the Neumann case, if the minimum occurs on the boundary, the Hopf lemma applied to − q -q − q gives ∂ ν q < 0 \partial_{\nu}q<0 ∂ ν q < 0 , again contradicting the boundary condition. Thus 0 is algebraically simple. Consequently,
λ 0 = 0 , e 0 = 1 \lambda_0=0,\qquad e_0=1 λ 0 = 0 , e 0 = 1
is the real, algebraically simple principal eigenpair.
Set
K : = σ ( L ) ∩ { z ∈ C : 0 ≤ Re z ≤ 1 } . K:=\sigma(L)\cap\{z\in\mathbb{C}:0\le\operatorname{Re}z\le1\}. K := σ ( L ) ∩ { z ∈ C : 0 ≤ Re z ≤ 1 } .
By (A.1), K K K is bounded, and compact resolvent implies that K K K is finite. Therefore
μ : = min { 1 , min λ ∈ K ∖ { 0 } Re λ } > 0. \mu:=\min\left\{1,\min_{\lambda\in K\setminus\{0\}}\operatorname{Re}\lambda\right\}>0. μ := min { 1 , λ ∈ K ∖ { 0 } min Re λ } > 0.
Hence
Re λ ≥ μ , λ ∈ σ ( L ) ∖ { 0 } . \operatorname{Re}\lambda\ge\mu,\qquad\lambda\in\sigma(L)\setminus\{0\}. Re λ ≥ μ , λ ∈ σ ( L ) ∖ { 0 } .
Combining this with (A.1),
∣ Im λ ∣ Re λ ≤ c 1 ( r + c 2 ) r , r : = Re λ ≥ μ . \frac{|\operatorname{Im}\lambda|}{\operatorname{Re}\lambda}\le\frac{\sqrt{c_1(r+c_2)}}{r},\qquad r:=\operatorname{Re}\lambda\ge\mu. Re λ ∣ Im λ ∣ ≤ r c 1 ( r + c 2 ) , r := Re λ ≥ μ .
Since the right-hand side is bounded on [ μ , ∞ ) [\mu,\infty) [ μ , ∞ ) , there exists ϑ < π / 2 \vartheta<\pi/2 ϑ < π /2 such that
∣ arg λ ∣ ≤ ϑ ∀ λ ∈ σ ( L ) ∖ { 0 } . |\arg\lambda|\le\vartheta\qquad\forall\lambda\in\sigma(L)\setminus\{0\}. ∣ arg λ ∣ ≤ ϑ ∀ λ ∈ σ ( L ) ∖ { 0 } .
Thus
σ ( L ) ⊂ { 0 } ∪ { z ∈ C ∖ { 0 } : ∣ arg z ∣ ≤ ϑ } , \sigma(L)\subset\{0\}\cup\{z\in\mathbb{C}\setminus\{0\}:|\arg z|\le\vartheta\}, σ ( L ) ⊂ { 0 } ∪ { z ∈ C ∖ { 0 } : ∣ arg z ∣ ≤ ϑ } ,
which proves the claimed spectral gap and sectorial distribution.
Proof of Lemma 4.5
Proof of Lemma 4.5. Since X X X is finite-dimensional and 0 has algebraic multiplicity 1, we have dim N ( A ) = 1 \dim\mathcal{N}(A)=1 dim N ( A ) = 1 and N ( A ) ∩ R ( A ) = { 0 } \mathcal{N}(A)\cap\mathcal{R}(A)=\{0\} N ( A ) ∩ R ( A ) = { 0 } . Hence P P P is a rank-one projection. Set Y : = R ( A ) Y:=\mathcal{R}(A) Y := R ( A ) . By the rank–nullity theorem,
X = N ( A ) ⊕ Y , σ ( A ∣ Y ) = σ ( A ) ∖ { 0 } . X=\mathcal{N}(A)\oplus Y,\qquad\sigma(A|_Y)=\sigma(A)\setminus\{0\}. X = N ( A ) ⊕ Y , σ ( A ∣ Y ) = σ ( A ) ∖ { 0 } .
Now let x ∈ X x\in X x ∈ X . Write uniquely
x = x 0 + y , x 0 ∈ N ( A ) , y ∈ Y . x=x_0+y,\qquad x_0\in\mathcal{N}(A),\ y\in Y. x = x 0 + y , x 0 ∈ N ( A ) , y ∈ Y .
Since A x 0 = 0 Ax_0=0 A x 0 = 0 , we have
e − t A x 0 = x 0 , t ≥ 0. e^{-tA}x_0=x_0,\qquad t\ge0. e − t A x 0 = x 0 , t ≥ 0.
Also, since Y Y Y is A A A -invariant,
e − t A x = x 0 + e − t A ∣ Y y . e^{-tA}x=x_0+e^{-tA|_Y}y. e − t A x = x 0 + e − t A ∣ Y y .
Set
Then
σ ( B ) = σ ( A ) ∖ { 0 } , \sigma(B)=\sigma(A)\setminus\{0\}, σ ( B ) = σ ( A ) ∖ { 0 } ,
so every eigenvalue μ ∈ σ ( B ) \mu\in\sigma(B) μ ∈ σ ( B ) satisfies Re ( μ ) > 0 \operatorname{Re}(\mu)>0 Re ( μ ) > 0 . Since Y Y Y is finite-dimensional, the standard estimate for matrix exponentials implies that for every
0 < γ < min { Re ( μ ) : μ ∈ σ ( B ) } , 0<\gamma<\min\{\operatorname{Re}(\mu):\mu\in\sigma(B)\}, 0 < γ < min { Re ( μ ) : μ ∈ σ ( B )} ,
there exists C > 0 C>0 C > 0 such that
∥ e − t B ∥ L ( Y ) ≤ C e − γ t , t ≥ 0. \lVert e^{-tB}\rVert_{\mathcal{L}(Y)}\leq Ce^{-\gamma t},\qquad t\geq0. ∥ e − tB ∥ L ( Y ) ≤ C e − γ t , t ≥ 0.
Therefore,
e − t A x = x 0 + e − t B y ⟶ x 0 as t → ∞ . e^{-tA}x=x_0+e^{-tB}y\longrightarrow x_0\qquad\text{as }t\to\infty. e − t A x = x 0 + e − tB y ⟶ x 0 as t → ∞.
Since P x = x 0 Px=x_0 P x = x 0 , this proves that
e − t A x ⟶ P x for every x ∈ X . e^{-tA}x\longrightarrow Px\qquad\text{for every }x\in X. e − t A x ⟶ P x for every x ∈ X .
Finally, for x = x 0 + y x=x_0+y x = x 0 + y we have
( e − t A − P ) x = e − t B y . (e^{-tA}-P)x=e^{-tB}y. ( e − t A − P ) x = e − tB y .
Thus e − t A − P e^{-tA}-P e − t A − P vanishes on N ( A ) \mathcal{N}(A) N ( A ) and coincides with e − t B e^{-tB} e − tB on Y Y Y . Since X = N ( A ) ⊕ Y X=\mathcal{N}(A)\oplus Y X = N ( A ) ⊕ Y and all norms are equivalent on the finite-dimensional space X X X , it follows that there exists C > 0 C>0 C > 0 such that
∥ e − t A − P ∥ L ( X ) ≤ C e − γ t , t ≥ 0. \lVert e^{-tA}-P\rVert_{\mathcal{L}(X)}\leq Ce^{-\gamma t},\qquad t\geq0. ∥ e − t A − P ∥ L ( X ) ≤ C e − γ t , t ≥ 0.
This completes the proof.