betaFeedback

WorksMathematics

An Incidence-Based Edge-Averaged Method for Convection-Diffusion Problems on Polytopal Meshes

Abstract

We develop an incidence-based edge-averaged method for scalar convection-diffusion problems on convex polytopal meshes in arbitrary dimensions. The proposed method extends the classical edge-averaged finite-element scheme through a graph-theoretic formulation based on vertex-edge incidence matrices. We first establish this formulation on simplices in arbitrary dimensions and then extend it to convex polytopes by introducing complementary edges and generalized barycentric coordinates. A key ingredient is a matrix identity that relates the generalized Whitney edge mass matrix to a diagonal representation of the Poisson stiffness matrix, enabling an exponentially-fitted discretization expressed entirely in terms of nodal values and edge weights. We prove that the resulting stiffness matrix is a nonsingular MM-matrix under suitable positivity and connectivity conditions, ensuring monotonicity and the discrete maximum principle. Numerical experiments on two- and three-dimensional simplicial and nonsimplicial meshes demonstrate first-order convergence in the H1H^1 semi-norm and stable, nonoscillatory solutions in convection-dominated regimes.

Introduction

Convection-diffusion equations arise in many applications involving the combined effects of transport and diffusion. A major numerical difficulty occurs in the convection-dominated regime, where solutions may develop sharp boundary or interior layers. Unless the mesh is sufficiently refined and suitably aligned with these layers, standard Galerkin finite-element discretizations often produce nonphysical oscillations that substantially deteriorate the quality of the numerical solution [17, 18, 25]. This difficulty has motivated the development of stabilized discretizations for convection-dominated problems [5, 21].

Methods to address such issues have been developed mainly in two directions. The first introduces additional stabilization terms into the weak formulation to obtain stability and convergence estimates that remain robust as the diffusion coefficient approaches zero (see [11, 12, 19] for examples). Although these methods can reduce spurious oscillations, they generally do not guarantee a discrete maximum principle (DMP), which ensures the discrete solution is free of nonphysical over- and undershoots. Hence, the second direction seeks to enforce the DMP. Algebraic flux correction methods [8] do so directly, but typically lead to nonlinear algebraic systems. Linear monotone discretizations offer another approach, as monotonicity is a sufficient condition for the discrete maximum principle. At the algebraic level, a stronger but convenient sufficient condition for monotonicity is that the stiffness matrix be a nonsingular MM-matrix, which can be checked directly from the assembled matrix.

Another monotone approach, the edge-averaged finite element (EAFE) method [28], has been developed for convection-diffusion problems on simplicial meshes. Motivated by the Scharfetter-Gummel scheme [24], EAFE constructs an edgewise approximation of the exponentially fitted flux. Related edge-averaging and exponential-fitting ideas have subsequently been used to develop a variety of numerical schemes [6, 10, 14, 22]. The monotonicity of EAFE is guaranteed when the stiffness matrix for the Poisson equation in the same conforming finite-element space is an MM-matrix, a condition that is satisfied, for example, on Delaunay triangulations in two dimensions. An important feature of EAFE is that the final formulation can be expressed in terms of nodal differences and diffusion stiffness weights without explicitly computing the associated edge basis functions.

Several recent works have extended different aspects of EAFE. The simplex-averaged finite element (SAFE) method [27] generalizes the approach from scalar H(grad⁡)H(\operatorname{grad}) problems to vector convection-diffusion problems in H(curl⁡)H(\operatorname{curl}) and H(div⁡)H(\operatorname{div}), using graph coboundary operators on simplicial meshes. A related stable mimetic finite-difference (MFD) method [2] employs primal-dual meshes, incidence-relation matrices, discrete differential operators, and shifted flux operators to obtain scalar and vector schemes in a finite-difference setting. The edge-averaged virtual element (EAVE) method [13], on the other hand, extends EAFE to scalar convection-diffusion problems on polygonal meshes in two dimensions. It includes both a monotone formulation on suitable Voronoi meshes and a general formulation on arbitrary polygonal meshes. However, these developments do not yet provide a graph-based, edge-averaged method for general polytopal meshes in three or more dimensions. Both SAFE and MFD have primarily been developed on simplicial meshes in two and three dimensions, whereas the existing EAVE construction is limited to two-dimensional polygonal meshes. These limitations are particularly relevant for convection-dominated problems, where adaptive or anisotropic refinement and layer-aligned mesh generation may benefit from greater mesh flexibility and the use of non-simplicial elements. This need becomes more pronounced in three-dimensional problems, where constructing high-quality simplicial meshes can be challenging, and in space-time discretizations [7], where treating time as an additional coordinate may require four-dimensional meshes for time-dependent three-dimensional problems [3].

Therefore, in this work, we propose an incidence-based edge-averaged method (IBEAM) for scalar convection-diffusion problems on polytopal meshes in arbitrary dimensions. We first develop an incidence-based formulation of EAFE on simplices in arbitrary dimensions and then extend this framework to general convex polytopes. The key observation underlying our construction is that the one-skeleton of a simplex is a complete graph, whereas that of a general polytope is not. We therefore augment the local edge graph of each polytopal element to the complete graph on its vertex set. The additional connections, which correspond geometrically to element diagonals, are referred to as complementary edges and are used as auxiliary local objects in the construction. On the resulting complete graph, we define an extended incidence matrix that includes both geometrical and complementary edges. We also consider a nodal finite-element space generated by generalized barycentric coordinates, primarily Wachspress coordinates [15, 26] in this work, together with an associated edge space constructed from Whitney one-forms [20] on the extended edge set. The extended incidence matrix represents the gradients of the nodal basis functions in terms of these edge functions, thereby connecting the nodal stiffness matrix to the mass matrix of the extended edge space. We then replace the generally dense edge mass matrix by a diagonal representation that retains the required relation with the nodal stiffness matrix. Combining this representation with an exponentially-fitted flux operator yields the IBEAM discretization without requiring the explicit computation of the extended edge basis functions. Since its derivation is based on incidence relations between compatible local vertex and edge spaces, the algebraic construction is not intrinsically restricted to a particular spatial dimension and can be extended to higher dimensions. As in EAFE, when the stiffness matrix of the associated nodal discretization for the Poisson equation is an MM-matrix, so is the resulting IBEAM stiffness matrix, providing a sufficient condition for monotonicity. Numerical experiments on simplicial and nonsimplicial meshes demonstrate first-order convergence in the H1H^1 semi-norm and stable, nonoscillatory behavior in convection-dominated regimes.

The remainder of this paper is organized as follows. Section 2 introduces the convection-diffusion model problem and reviews the classical EAFE scheme. Section 3 develops an incidence-based formulation on simplices in arbitrary dimensions and extends the framework to general convex polytopes using complementary edges and generalized barycentric coordinates. Section 4 presents the IBEAM discretization and establishes sufficient conditions for its monotonicity. Section 5 investigates the convergence behavior and numerical performance of IBEAM on various simplicial and polytopal meshes, including convection-dominated problems with boundary layers. Finally, Section 6 summarizes our contributions and discusses future research directions.

Preliminaries

This section introduces the convection-diffusion model problem and reviews the classical EAFE scheme [28], providing the necessary background for the graph-theoretic framework developed in the next section.

Model Problem

Let Ω⊂Rd\Omega\subset\mathbb{R}^{d} be a bounded Lipschitz domain. We denote by H1(Ω)H^{1}(\Omega) and H01(Ω)H_{0}^{1}(\Omega) the standard Sobolev spaces, and by L2(Ω)L^{2}(\Omega) the space of square-integrable functions. The norm, semi-norm, and inner product associated with a Hilbert space XX are denoted by ∥⋅∥X\| \cdot\|_{X}, ∣⋅∣X|\cdot|_{X}, and ⟨⋅,⋅⟩X\langle\cdot,\cdot\rangle_{X}, respectively.

We consider the following steady-state convection-diffusion problem:

−∇⋅(α∇u+βu)=fin Ω,(1)-\nabla\cdot(\alpha\nabla u + \boldsymbol{\beta}u) = f \quad\text{in }\Omega, \tag*{(1)}
u=0on ∂Ω,(2)u = 0 \quad\text{on }\partial\Omega, \tag*{(2)}

where f∈L2(Ω)f \in L^{2}(\Omega), α∈C0(Ω‾)\alpha\in C^{0}(\overline{\Omega}) is the diffusion coefficient satisfying

0<αmin⁡≤α(x)≤αmax⁡∀x∈Ω‾,0 < \alpha_{\min} \le\alpha(x) \le\alpha_{\max} \qquad\forall x \in\overline{\Omega},

and β∈[C0(Ω‾)]d\boldsymbol{\beta} \in[C^{0}(\overline{\Omega})]^{d} is the convection field. We are particularly interested in convection-dominated regimes, where the magnitude of the convective transport is large relative to diffusion. Defining the flux by

J(u)=α∇u+βu,J(u) = \alpha\nabla u + \boldsymbol{\beta}u,

the weak formulation is to find u∈H01(Ω)u \in H_{0}^{1}(\Omega) such that

B(u,v)=F(v)∀v∈H01(Ω),B(u,v) = F(v) \qquad\forall v \in H_{0}^{1}(\Omega),

where

B(u,v):=∫ΩJ(u)⋅∇v dx,F(v):=∫Ωfv dx.(3)B(u,v) := \int_{\Omega} J(u) \cdot\nabla v \, \mathrm{d}x, \qquad F(v) := \int_{\Omega} fv \, \mathrm{d}x. \tag*{(3)}

Finite-Element Spaces

We introduce the finite-element spaces and notation underlying the classical EAFE scheme. Although EAFE is formulated in two and three dimensions, we restrict our discussion to triangular meshes for simplicity. In particular, we describe the linear Lagrange and lowest-order Nédélec finite-element spaces and the relationship between their basis functions.

Let Th\mathcal{T}_{h} be a conforming triangulation of Ω\Omega, with vertex set Vh\mathcal{V}_{h} and edge set Eh\mathcal{E}_{h}. For each triangle T∈ThT \in\mathcal{T}_{h}, let {xi}i=13\{x_{i}\}_{i=1}^{3} denote its vertices and {el}l=13\{e_{l}\}_{l=1}^{3} its edges. We orient each edge el=(i,j)e_l=(i,j) from the lower-indexed vertex, xix_i to the higher-indexed vertex, xjx_j and define the corresponding oriented edge vector by

τl=xj−xi,i<j.(4)\tau_l=x_j-x_i,\qquad i<j. \tag*{(4)}

We define the conforming finite-element spaces

Vh={vh∈H01(Ω):vh∣T∈P1(T), ∀T∈Th},V_h=\{v_h\in H_0^1(\Omega):v_h|_T\in P_1(T),\ \forall T\in\mathcal{T}_h\},
Qh={wh∈H(curl⁡;Ω):wh∣T∈N0(T), ∀T∈Th},Q_h=\{w_h\in H(\operatorname{curl};\Omega):w_h|_T\in\mathcal{N}_0(T),\ \forall T\in\mathcal{T}_h\},

where P1(T)P_1(T) denotes the space of linear polynomials on TT, and N0(T)\mathcal{N}_0(T) denotes the local lowest-order Nédélec space, given by

N0(T)=[P0(T)]2+x⊥P0(T),\mathcal{N}_0(T)=[P_0(T)]^2+x^\perp P_0(T),

where x⊥x^\perp is a 90° rotation of xx satisfying x⋅x⊥=0x\cdot x^\perp=0. The degrees of freedom for any w∈N0(T)w\in\mathcal{N}_0(T) are defined by

dof⁡l(w)=1∣el∣∫elw⋅τl ds=(w⋅τl)∣el,\operatorname{dof}_l(w)=\frac{1}{|e_l|}\int_{e_l}w\cdot\tau_l\,\mathrm{d}s=(w\cdot\tau_l)|_{e_l},

where el=(i,j)e_l=(i,j).

Let {λi}i=13\{\lambda_i\}_{i=1}^3 denote the barycentric coordinates on TT, with the local vertices {xi}i=13\{x_i\}_{i=1}^3 of TT numbered in increasing order of their global indices, and let {ϕl}l=13\{\phi_l\}_{l=1}^3 denote the local canonical Nédélec basis functions satisfying dof⁡l(ϕm)=δlm\operatorname{dof}_l(\phi_m)=\delta_{lm}. For each edge el=(i,j)e_l=(i,j), these basis functions are given by

ϕl=λi∇λj−λj∇λi.\phi_l=\lambda_i\nabla\lambda_j-\lambda_j\nabla\lambda_i.

Edge-Averaged Finite-Element Scheme

We now review the classical edge-averaged finite-element (EAFE) scheme [28], focusing on the key steps in its derivation that are relevant to the subsequent generalization.

For each triangle T∈ThT\in\mathcal{T}_h, we define the local bilinear form

BT(uh,vh)=∫TJ(uh)⋅∇vh dx,B_T(u_h,v_h)=\int_T J(u_h)\cdot\nabla v_h\,\mathrm{d}x,

where uh,vh∈Vhu_h,v_h\in V_h. We approximate the flux J(uh)J(u_h) by an exponentially-fitted flux JT(uh)∈N0(T)J_T(u_h)\in\mathcal{N}_0(T), yielding

BTN(uh,vh)=∫TJT(uh)⋅∇vh dx,B_T^N(u_h,v_h)=\int_T J_T(u_h)\cdot\nabla v_h\,\mathrm{d}x,

where

JT(uh)=∑l=13dof⁡l(JT(uh))ϕl=∑l=13(JT(uh)⋅τl)∣elϕl,J_T(u_h)=\sum_{l=1}^3\operatorname{dof}_l(J_T(u_h))\phi_l =\sum_{l=1}^3(J_T(u_h)\cdot\tau_l)|_{e_l}\phi_l,

and JT(uh)∈N0(T)J_T(u_h)\in\mathcal{N}_0(T). Since ∇vh∣T∈N0(T)\nabla v_h|_T\in\mathcal{N}_0(T), we express the local bilinear form as

BTN(uh,vh)=∫TJT(uh)⋅∇vh dx=∑l=13∑m=13(JT(uh)⋅τl)∣el(∇vh⋅τm)∣em∫Tϕl⋅ϕm dx.\begin{aligned} B_T^N(u_h,v_h) &=\int_T J_T(u_h)\cdot\nabla v_h\,\mathrm{d}x\\ &=\sum_{l=1}^3\sum_{m=1}^3(J_T(u_h)\cdot\tau_l)|_{e_l}(\nabla v_h\cdot\tau_m)|_{e_m}\int_T\phi_l\cdot\phi_m\,\mathrm{d}x. \end{aligned}

To express these edge quantities in terms of nodal values, we introduce the edge difference operator

δluh:=uh(xj)−uh(xi)=uj−ui,\delta_l u_h := u_h(x_j)-u_h(x_i)=u_j-u_i,

where el=(i,j)e_l=(i,j) is edge ll connecting vertex ii to vertex jj. Since vhv_h is linear on TT, the gradient term satisfies

(∇vh⋅τm)∣em=δmvh.(\nabla v_h\cdot\boldsymbol{\tau}_m)|_{e_m}=\delta_m v_h.

Next, we consider the flux term. For each edge ele_l, we introduce a locally defined potential function γ\gamma satisfying

∇γ⋅τl=α−1β⋅τlon el.\nabla\gamma\cdot\boldsymbol{\tau}_l=\alpha^{-1}\boldsymbol{\beta}\cdot\boldsymbol{\tau}_l\quad\text{on }e_l.

From the definition of the flux, we have

α−1eγ(J(uh)⋅τl)=∇(eγuh)⋅τl.\alpha^{-1}e^\gamma(J(u_h)\cdot\boldsymbol{\tau}_l)=\nabla(e^\gamma u_h)\cdot\boldsymbol{\tau}_l.

Applying dof⁡l\operatorname{dof}_l to both sides through the weighted flux interpolation and using the fact that the tangential component of JT(uh)∈N0(T)J_T(u_h)\in N_0(T) is constant along ele_l, we obtain

(JT(uh)⋅τl)∣el1∣el∣∫elα−1eγ dx=1∣el∣∫el∇(eγuh)⋅τl dx=δl(eγuh).(J_T(u_h)\cdot\boldsymbol{\tau}_l)|_{e_l}\frac{1}{|e_l|}\int_{e_l}\alpha^{-1}e^\gamma\,\mathrm{d}x =\frac{1}{|e_l|}\int_{e_l}\nabla(e^\gamma u_h)\cdot\boldsymbol{\tau}_l\,\mathrm{d}x =\delta_l(e^\gamma u_h).

Therefore,

(JT(uh)⋅τl)∣el=κlδl(eγuh),(5)(J_T(u_h)\cdot\boldsymbol{\tau}_l)|_{e_l}=\kappa_l\delta_l(e^\gamma u_h), \tag*{(5)}

where

κl=(1∣el∣∫elα−1eγ dx)−1.\kappa_l=\left(\frac{1}{|e_l|}\int_{e_l}\alpha^{-1}e^\gamma\,\mathrm{d}x\right)^{-1}.

Substituting this expression and the corresponding edge difference for vhv_h into the local bilinear form yields

BTN(uh,vh)=∑l=13∑m=13κlδl(eγuh)δmvh∫Tϕl⋅ϕm dx.(6)B_T^N(u_h,v_h)=\sum_{l=1}^{3}\sum_{m=1}^{3}\kappa_l\delta_l(e^\gamma u_h)\delta_m v_h\int_T\boldsymbol{\phi}_l\cdot\boldsymbol{\phi}_m\,\mathrm{d}x. \tag*{(6)}

To derive the EAFE scheme, we employ a mass-lumping technique for the Nédélec mass matrix that preserves the following identity for the Poisson bilinear form in the linear Lagrange finite-element space [16]:

∫T∇uh⋅∇vh dx=∑l=13∑m=13δluhδmvh∫Tϕl⋅ϕm dx=∑l=13ωlTδluhδlvh,(7)\int_T\nabla u_h\cdot\nabla v_h\,\mathrm{d}x =\sum_{l=1}^{3}\sum_{m=1}^{3}\delta_lu_h\delta_mv_h\int_T\boldsymbol{\phi}_l\cdot\boldsymbol{\phi}_m\,\mathrm{d}x =\sum_{l=1}^{3}\omega_l^T\delta_lu_h\delta_lv_h, \tag*{(7)}

where

ωlT=−∫T∇λi⋅∇λj dx,(8)\omega_l^T=-\int_T\nabla\lambda_i\cdot\nabla\lambda_j\,\mathrm{d}x, \tag*{(8)}

for the edge el=(i,j)e_l=(i,j). Motivated by this identity, we approximate the double-sum formulation in (2.5) by replacing the Nédélec mass matrix with the edge weights ωlT\omega_l^T, leading to the local EAFE bilinear form

BTh(uh,vh)=∑l=13ωlTκlδl(eγuh)δlvh.(9)B_T^h(u_h,v_h)=\sum_{l=1}^{3}\omega_l^T\kappa_l\delta_l(e^\gamma u_h)\delta_lv_h. \tag*{(9)}

We note that when JT(uh)J_T(u_h) is constant on TT, it can be expressed as JT(uh)=∇ψJ_T(u_h)=\nabla\psi for some linear function ψ\psi. In this case, the Poisson identity implies

BTN(uh,vh)=BTh(uh,vh).(10)B_T^N(u_h,v_h)=B_T^h(u_h,v_h). \tag*{(10)}

For a nonconstant JT(uh)J_T(u_h), BThB_T^h instead provides an approximation to BTNB_T^N. The EAFE scheme has been shown to achieve first-order convergence with respect to the mesh size hh, under suitable regularity assumptions and for sufficiently small hh [28]. The convergence analysis relies on a discrete inf-sup condition and a consistency estimate, the latter of which exploits the identity (10) for constant interpolated fluxes.

The EAFE scheme also guarantees monotonicity under the Delaunay mesh condition, preventing spurious oscillations in convection-dominated regimes. Moreover, its discrete bilinear form involves only edge differences and the corresponding weights, allowing for a simple implementation using the Bernoulli function [13]. Although originally formulated on simplicial meshes in two and three dimensions, the EAFE scheme can be naturally interpreted from a graph-theoretic perspective. This perspective not only provides a better understanding of the underlying algebraic structure of EAFE but also enables its extension to simplices in arbitrary dimensions. Furthermore, it provides a systematic approach to deriving generalized EAFE schemes for convection-diffusion problems on non-simplicial polytopal meshes.

Graph-Theoretic Framework

This section introduces a graph-theoretic representation of the EAFE scheme and extends this representation to polytopal elements through their embedding into a complete graph. This representation extends the edge-based structure underlying EAFE by introducing complementary edges, which allow the diagonal representation of the Poisson stiffness matrix (2.6) to be generalized from simplices to arbitrary convex polytopes.

Graph-Based EAFE in Arbitrary Dimensions

We begin by considering a dd-dimensional simplex T∈ThT \in\mathcal{T}_h with vertices {xi}i=1nv⊂Rd\{x_i\}_{i=1}^{n_v} \subset\mathbb{R}^d and edges {el}l=1ne\{e_l\}_{l=1}^{n_e}, where nv=d+1n_v=d+1 and ne=d(d+1)/2n_e=d(d+1)/2. Each edge is oriented from the lower-indexed vertex to the higher-indexed vertex, consistent with the definition of τl\boldsymbol{\tau}_l in (2.3).

Let {λi}i=1nv\{\lambda_i\}_{i=1}^{n_v} denote the barycentric coordinates on TT, satisfying

λi(xj)=δij,∑i=1nvλi=1.\lambda_i(x_j)=\delta_{ij}, \qquad\sum_{i=1}^{n_v}\lambda_i=1.

For each oriented edge el=(i,j)e_l=(i,j), which connects vertex ii to vertex jj, we define the corresponding canonical Whitney edge basis function by

ϕl=λi∇λj−λj∇λi,i<j.\phi_l=\lambda_i\nabla\lambda_j-\lambda_j\nabla\lambda_i,\qquad i<j.

We define the vertex-edge incidence matrix G∈Rne×nvG\in\mathbb{R}^{n_e\times n_v} by

Glk={1,if k=j,−1,if k=i,0,otherwise,G_{lk}= \begin{cases} 1, & \text{if } k=j,\\ -1, & \text{if } k=i,\\ 0, & \text{otherwise}, \end{cases}

for each oriented edge el=(i,j)e_l=(i,j). The incidence matrix represents the discrete gradient operator, mapping nodal values to their differences along oriented edges. In particular,

δluh=(Guh)l,\delta_l u_h=(G u_h)_l,

where uh\mathbf{u}_h denotes the vector of nodal values of uhu_h. Moreover, the partition-of-unity property of the barycentric coordinates implies

∑j=1nv∇λj=0.\sum_{j=1}^{n_v} \nabla\lambda_j = \mathbf{0}.

Together with the definition of the Whitney edge functions, this yields (see, e.g., [4])

∇λi=∑l=1neGliϕl.(11)\nabla\lambda_i = \sum_{l=1}^{n_e} G_{li}\boldsymbol{\phi}_l. \tag*{(11)}

This identity expresses the gradients of the nodal basis functions in terms of the Whitney edge functions through the incidence matrix. Using this relation, the Poisson bilinear form can be expressed as (see [1])

∫T∇uh⋅∇vh dx=∑l=1ne∑m=1neδluhδmvh∫Tϕl⋅ϕm dx=vhTGTMGuh,(12)\int_T \nabla u_h \cdot\nabla v_h \, \mathrm{d}x = \sum_{l=1}^{n_e}\sum_{m=1}^{n_e} \delta_l u_h \delta_m v_h \int_T \boldsymbol{\phi}_l \cdot\boldsymbol{\phi}_m \, \mathrm{d}x = \mathbf{v}_h^{\mathsf{T}}G^{\mathsf{T}}M G\mathbf{u}_h, \tag*{(12)}

where M∈Rne×neM \in\mathbb{R}^{n_e \times n_e} is the edge mass matrix defined by

Mlm=∫Tϕl⋅ϕm dx.M_{lm} = \int_T \boldsymbol{\phi}_l \cdot\boldsymbol{\phi}_m \, \mathrm{d}x.

On the other hand, the Poisson bilinear form admits the following representation in terms of edge differences [28]:

∫T∇uh⋅∇vh dx=∑l=1neωlTδluhδlvh=vhTGTWGuh,(13)\int_T \nabla u_h \cdot\nabla v_h \, \mathrm{d}x = \sum_{l=1}^{n_e} \omega_l^T \delta_l u_h \delta_l v_h = \mathbf{v}_h^{\mathsf{T}}G^{\mathsf{T}}W G\mathbf{u}_h, \tag*{(13)}

where

W=diag⁡(ω1T,…,ωneT),W = \operatorname{diag}\left(\omega_1^T,\ldots,\omega_{n_e}^T\right),

with the edge weights ωlT\omega_l^T defined as in (2.7), now on a dd-dimensional simplex. Since both representations (3.2) and (3.3) hold for arbitrary nodal vectors uh\mathbf{u}_h and vh\mathbf{v}_h, we obtain the matrix identity,

GTMG=GTWG,(14)G^{\mathsf{T}}MG = G^{\mathsf{T}}WG, \tag*{(14)}

which we use later. A corresponding identity for diagonal mass lumping on tetrahedral meshes can also be found in [9].

We now express the EAFE bilinear form (2.8) in terms of the graph incidence matrix. Recall that the edge differences of vhv_h satisfy δlvh=(Gvh)l\delta_l v_h = (Gv_h)_l. Moreover, following the same argument used to derive (2.4),

(JT(uh)⋅τl)∣el=κlδl(eγluh).\left.(J_T(u_h)\cdot\boldsymbol{\tau}_l)\right|_{e_l} = \kappa_l \delta_l(e^{\gamma_l u_h}).

To obtain an explicit representation of the flux term, we approximate the diffusion and convection coefficients by constants αl\alpha_l and βl\beta_l, respectively, along each edge el=(i,j)e_l = (i,j). The corresponding edgewise potential γl\gamma_l is defined by

γl(x)=αl−1βl⋅(x−xi),\gamma_l(x) = \alpha_l^{-1}\boldsymbol{\beta}_l \cdot(x-x_i),

so that ∇γl=αl−1βl\nabla\gamma_l = \alpha_l^{-1}\beta_l and γl(xi)=0\gamma_l(x_i)=0. Introducing the Bernoulli function

B(s)={ses−1,s≠0,1,s=0,\mathbf{B}(s)= \begin{cases} \frac{s}{e^s-1}, & s\ne0,\\ 1, & s=0, \end{cases}

the edge flux can be expressed as

κlδl(eγluh)=αl[B(−pl)uh(xj)−B(pl)uh(xi)],(15)\kappa_l\delta_l(e^{\gamma_l}u_h)=\alpha_l\left[\mathbf{B}(-p_l)u_h(x_j)-\mathbf{B}(p_l)u_h(x_i)\right], \tag*{(15)}

where pl=αl−1βl⋅τlp_l=\alpha_l^{-1}\beta_l\cdot\tau_l. Motivated by this representation, we define the Bernoulli-weighted incidence matrix Gb∈Rne×nvG_b\in\mathbb{R}^{n_e\times n_v} by

(Gb)lk={B(−pl),if k=j,−B(pl),if k=i,0,otherwise,(G_b)_{lk}= \begin{cases} \mathbf{B}(-p_l), & \text{if } k=j,\\ -\mathbf{B}(p_l), & \text{if } k=i,\\ 0, & \text{otherwise}, \end{cases}

for each oriented edge el=(i,j)e_l=(i,j). This modification incorporates the exponential fitting of the EAFE scheme into the discrete gradient operator.

Suppose that JT(uh)J_T(u_h) is constant on TT. Then, there exists a linear function ψT\psi_T such that JT(uh)=∇ψTJ_T(u_h)=\nabla\psi_T. Taking the edge degrees of freedom and using (15), we obtain

GψT=DαGbuh,G\psi_T=D_\alpha G_bu_h,

where ψT\psi_T denotes the vector of nodal values of ψT\psi_T and

Dα=diag⁡(α1,…,αne).D_\alpha=\operatorname{diag}(\alpha_1,\ldots,\alpha_{n_e}).

Consequently, using the matrix identity (14), we have

BTN(uh,vh):=vhTGTMDαGbuh=vhTGTMGψT=vhTGTWGψT=vhTGTWDαGbuh.B_T^N(u_h,v_h):=v_h^{\mathsf T}G^{\mathsf T}MD_\alpha G_bu_h =v_h^{\mathsf T}G^{\mathsf T}MG\psi_T =v_h^{\mathsf T}G^{\mathsf T}WG\psi_T =v_h^{\mathsf T}G^{\mathsf T}WD_\alpha G_bu_h.

We, therefore, obtain the local EAFE bilinear form,

BTh(uh,vh)=vhTGTWDαGbuh.(16)B_T^h(u_h,v_h)=v_h^{\mathsf T}G^{\mathsf T}WD_\alpha G_bu_h. \tag*{(16)}

In the case where the flux JT(uh)J_T(u_h) is not constant, BThB_T^h only approximates BTNB_T^N.

Remark 3.1. The graph-based formulation (3.6) provides a simple implementation of the EAFE scheme. Starting from the edge weights WW associated with the Poisson stiffness matrix, the convection-diffusion bilinear form can be assembled by approximating the coefficients as constants along each edge and incorporating the Bernoulli weights into the incidence matrix. In particular, the matrix representation GTWDαGbG^{\mathsf T}WD_\alpha G_b makes explicit how the EAFE scheme approximates the flux J(uh)J(u_h) through an exponentially-fitted, edge-based discrete gradient. Moreover, since this formulation is constructed from barycentric coordinates and incidence relations on dd-dimensional simplices, it extends the classical EAFE scheme beyond two and three dimensions while preserving its underlying algebraic structure.

The extension to convex polytopal elements in arbitrary dimensions requires an additional consideration: unlike simplices, their vertex-edge graphs are not necessarily complete. Nevertheless, the incidence-based formulation (3.6) can be extended using generalized barycentric coordinates and complementary edges. In the following subsection, we introduce these complementary edges to construct a complete graph and extend the formulation to general convex polytopes.

Extension to Convex Polytopal Meshes

Let Kh\mathcal{K}_h be a conforming mesh of Ω\Omega consisting of convex polytopal elements. For each K∈KhK \in\mathcal{K}_h, let {xi}i=1nv⊂Rd\{x_i\}_{i=1}^{n_v} \subset\mathbb{R}^d denote its vertices and {el}l=1ne\{e_l\}_{l=1}^{n_e} its edges, where nvn_v and nen_e are the numbers of vertices and edges of KK, respectively. We orient each edge from the lower-indexed vertex to the higher-indexed vertex, as in (4). Let {λi}i=1nv\{\lambda_i\}_{i=1}^{n_v} denote the generalized barycentric coordinates on KK, satisfying

λi(xj)=δij,∑i=1nvλi=1,∑i=1nvλixi=x.\lambda_i(x_j) = \delta_{ij}, \qquad\sum_{i=1}^{n_v} \lambda_i = 1, \qquad\sum_{i=1}^{n_v} \lambda_i x_i = x.

We define the corresponding conforming finite-element space by

Vh={vh∈H01(Ω):vh∣K∈W(K), ∀K∈Kh},V_h = \{v_h \in H_0^1(\Omega) : v_h|_K \in W(K),\ \forall K \in\mathcal{K}_h\},

where

W(K)=span⁡{λi}i=1nv.W(K) = \operatorname{span}\{\lambda_i\}_{i=1}^{n_v}.

To extend the matrix identity (14) to polytopal elements, let K∗K^* denote the complete graph formed by the vertices of KK, with edge set EK∗\mathcal{E}_K^* containing

ne∗=nv(nv−1)2n_e^* = \frac{n_v(n_v-1)}{2}

edges. Since the original edge set EK\mathcal{E}_K is a subset of EK∗\mathcal{E}_K^*, we define the set of complementary edges by

EKc=EK∗∖EK.\mathcal{E}_K^c = \mathcal{E}_K^* \setminus\mathcal{E}_K.

These complementary edges are auxiliary connections introduced for the local algebraic construction and do not modify the underlying polytopal mesh. We define the vertex-edge incidence matrix G∈Rne∗×nvG \in\mathbb{R}^{n_e^* \times n_v} on the complete graph K∗K^* by

Glk={1,if k=j,−1,if k=i,0,otherwise,G_{lk} = \begin{cases} 1, & \text{if } k=j,\\ -1, & \text{if } k=i,\\ 0, & \text{otherwise}, \end{cases}

for each oriented edge el=(i,j)∈EK∗e_l = (i,j) \in\mathcal{E}_K^*.

Associated with each edge of the complete graph, including complementary edges, we define the generalized Whitney edge function

ϕl=λi∇λj−λj∇λi,i<j.(17)\phi_l = \lambda_i \nabla\lambda_j - \lambda_j \nabla\lambda_i,\qquad i<j. \tag*{(17)}

Unlike the simplex case, these generalized Whitney edge functions do not necessarily form a canonical basis with respect to the edge degrees of freedom. However, since the generalized barycentric coordinates satisfy the partition-of-unity property, the same argument as in the simplex case in (11) gives

∇λi=∑l=1ne∗Gliϕl.(18)\nabla\lambda_i = \sum_{l=1}^{n_e^*} G_{li}\phi_l. \tag*{(18)}

Thus, introducing complementary edges preserves the incidence-based representation of the nodal basis gradients. This allows us to extend the matrix identity (14) to convex polytopal elements. We first illustrate this construction with an example before establishing the general result.

Example. Consider a quadrilateral element KK with four vertices and four edges, as illustrated in Figure 1. Its vertex-edge graph is not complete, since nv=4n_v=4 and ne=4n_e=4, whereas the complete graph contains ne∗=4(4−1)/2=6n_e^*=4(4-1)/2=6 edges. We therefore introduce two complementary edges, e5e_5 and e6e_6, connecting the two pairs of nonadjacent vertices.

Figure 1. Original quadrilateral KK (left) and its complete graph K∗K^* (right). Dashed lines represent complementary edges.

Following the edge orientation convention, we assign:

Original: e1=(1,2)e_1=(1,2), e2=(1,3)e_2=(1,3), e3=(2,4)e_3=(2,4), e4=(3,4)e_4=(3,4),

Complementary: e5=(1,4)e_5=(1,4), e6=(2,3)e_6=(2,3).

The corresponding incidence matrix of the complete graph is

G=[−1100−10100−10100−11⋯⋯⋯⋯−10010−110].G=\begin{bmatrix} -1 & 1 & 0 & 0\\ -1 & 0 & 1 & 0\\ 0 & -1 & 0 & 1\\ 0 & 0 & -1 & 1\\ \cdots& \cdots& \cdots& \cdots\\ -1 & 0 & 0 & 1\\ 0 & -1 & 1 & 0 \end{bmatrix}.

The first four rows correspond to the original edges of KK, while the last two rows correspond to the complementary edges.

For the unit square K=[0,1]2K=[0,1]^2, the generalized barycentric coordinates are given by the bilinear Lagrange basis functions

λ1(x,y)=(1−x)(1−y),λ2(x,y)=x(1−y),\lambda_1(x,y)=(1-x)(1-y),\qquad\lambda_2(x,y)=x(1-y),
λ3(x,y)=(1−x)y,λ4(x,y)=xy,\lambda_3(x,y)=(1-x)y,\qquad\lambda_4(x,y)=xy,

satisfying the partition-of-unity property. The corresponding generalized Whitney edge functions ϕl\phi_l are constructed as in (3.7). In particular, the two complementary edges give rise to the additional edge functions

ϕ5=λ1∇λ4−λ4∇λ1,ϕ6=λ2∇λ3−λ3∇λ2.\phi_5=\lambda_1\nabla\lambda_4-\lambda_4\nabla\lambda_1,\qquad\phi_6=\lambda_2\nabla\lambda_3-\lambda_3\nabla\lambda_2.

Thus, the example incidence matrix implies

∇λi=∑l=16Gliϕl,i=1,2,3,4,\nabla\lambda_i=\sum_{l=1}^{6}G_{li}\phi_l,\qquad i=1,2,3,4,

so the complementary edges provide the additional terms needed to recover the incidence-based representation of the nodal basis gradients.

This construction extends beyond rectangular elements. For general convex polytopes, generalized barycentric coordinates, such as Wachspress coordinates [15, 26], can be used to construct the corresponding nodal and generalized Whitney edge functions. The same incidence-based gradient identity then follows from the partition-of-unity property on the complete graph.

Matrix Identity on Convex Polytopes

We now establish the polytopal counterpart of the matrix identity (14) using the complementary edges introduced in Subsection 3.2.

Lemma 3.2. Let G∈Rne∗×nvG \in\mathbb{R}^{n_e^* \times n_v} be the incidence matrix of the complete graph K∗K^*, and let M∈Rne∗×ne∗M \in\mathbb{R}^{n_e^* \times n_e^*} be the generalized Whitney edge mass matrix defined by

Mlm=∫Kϕl⋅ϕm dx.M_{lm} = \int_K \boldsymbol{\phi}_l \cdot\boldsymbol{\phi}_m \,\mathrm{d}x.

Then,

GTMG=GTWG,(19)G^{\mathsf{T}} M G = G^{\mathsf{T}} W G, \tag*{(19)}

where W∈Rne∗×ne∗W \in\mathbb{R}^{n_e^* \times n_e^*} is the diagonal matrix with entries

Wll=−∫K∇λi⋅∇λj dx,el=(i,j).W_{ll} = -\int_K \nabla\lambda_i \cdot\nabla\lambda_j \,\mathrm{d}x, \qquad e_l = (i,j).

Proof. We establish the identity by comparing the entries of GTMGG^{\mathsf{T}} M G and GTWGG^{\mathsf{T}} W G. For i≠ji \ne j, the complete graph K∗K^* contains exactly one edge connecting vertices ii and jj. Since GliGlj=−1G_{li}G_{lj}=-1 for this edge and vanishes otherwise, we have

(GTWG)ij=∑l=1ne∗GliWllGlj=−Wll=∫K∇λi⋅∇λj dx.\left(G^{\mathsf{T}} W G\right)_{ij} = \sum_{l=1}^{n_e^*} G_{li}W_{ll}G_{lj} = -W_{ll} = \int_K \nabla\lambda_i \cdot\nabla\lambda_j \,\mathrm{d}x.

For i=ji=j, using the partition-of-unity property of the generalized barycentric coordinates, we obtain

(GTWG)ii=∑l=1ne∗Gli2Wll=−∑k=1k≠inv∫K∇λi⋅∇λk dx=∫K∣∇λi∣2 dx.\left(G^{\mathsf{T}} W G\right)_{ii} = \sum_{l=1}^{n_e^*} G_{li}^{2}W_{ll} = -\sum_{\substack{k=1 \\ k \ne i}}^{n_v} \int_K \nabla\lambda_i \cdot\nabla\lambda_k \,\mathrm{d}x = \int_K \lvert\nabla\lambda_i\rvert^{2}\,\mathrm{d}x.

Thus, for all i,ji,j,

(GTWG)ij=∫K∇λi⋅∇λj dx.\left(G^{\mathsf{T}} W G\right)_{ij} = \int_K \nabla\lambda_i \cdot\nabla\lambda_j \,\mathrm{d}x.

On the other hand, using the incidence-based gradient representation (18), we have

(GTMG)ij=∑l=1ne∗∑m=1ne∗GliMlmGmj=∫K(∑l=1ne∗Gliϕl)⋅(∑m=1ne∗Gmjϕm) dx=∫K∇λi⋅∇λj dx.\begin{aligned} \left(G^{\mathsf{T}} M G\right)_{ij} = \sum_{l=1}^{n_e^*}\sum_{m=1}^{n_e^*}G_{li}M_{lm}G_{mj} \\ &= \int_K \left(\sum_{l=1}^{n_e^*}G_{li}\boldsymbol{\phi}_l\right)\cdot\left(\sum_{m=1}^{n_e^*}G_{mj}\boldsymbol{\phi}_m\right)\,\mathrm{d}x \\ &= \int_K \nabla\lambda_i \cdot\nabla\lambda_j \,\mathrm{d}x. \end{aligned}

Therefore,

G⊤MG=G⊤WG.G^{\top}MG = G^{\top}WG.

The identity extends the diagonal representation of the Poisson stiffness matrix from simplices to convex polytopal elements, generalizing (14). It provides the algebraic foundation for the edge-averaged discretization developed in the following section.

Recall from the example in Figure 1 where we introduce two new edges with corresponding basis functions, ϕ5\phi_5 and ϕ6\phi_6. This allows us to establish the necessary relationship between the edge basis functions and the barycentric coordinates in order to ensure the matrix identity in Lemma 3.2. However, in the implementation, since we replace the mass matrix with the weight matrix, there is no reason to ever construct the edge basis explicitly, and we can instead rely entirely on the barycentric coordinates in the original graph.

Incidence-Based Edge-Averaged Method

We next extend the EAFE scheme to convex polytopal meshes using the incidence-based framework developed in Section 3. By combining the Bernoulli-weighted incidence matrix with the diagonal representation of the Poisson stiffness matrix, we construct the incidence-based edge-averaged method (IBEAM) for convex polytopes in arbitrary dimensions. We then establish its monotonicity under suitable conditions on the edge weights.

Following the EAFE construction in Subsection 2.3, we approximate the flux J(uh)J(u_h) locally by JK(uh)J_K(u_h) and consider the corresponding bilinear form

BKN(uh,vh)=∫KJK(uh)⋅∇vh dx.B_K^N(u_h,v_h)=\int_K J_K(u_h)\cdot\nabla v_h\,\mathrm{d}x.

For each oriented edge el=(i,j)e_l=(i,j) of the complete graph K∗K^*, including complementary edges, we approximate the tangential flux by a constant value along ele_l. Using the exponential-fitting construction in (5), we impose the weighted edge degrees of freedom

dof⁡l(α−1eγJK(uh))=dof⁡l(α−1eγJ(uh))=δl(eγuh).\operatorname{dof}_l\left(\alpha^{-1}e^\gamma J_K(u_h)\right)=\operatorname{dof}_l\left(\alpha^{-1}e^\gamma J(u_h)\right)=\delta_l(e^\gamma u_h).

Since JK(uh)⋅τlJ_K(u_h)\cdot\tau_l is constant along ele_l, it follows that

dof⁡l(JK(uh))=(JK(uh)⋅τl)∣el=κlδl(eγuh).\operatorname{dof}_l(J_K(u_h))=\left.(J_K(u_h)\cdot\tau_l)\right|_{e_l}=\kappa_l\delta_l(e^\gamma u_h).

Using the edgewise constant-coefficient approximations αl\alpha_l and βl\beta_l, together with (15), we obtain

dof⁡l(JK(uh))=αl[B(−pl)uh(xj)−B(pl)uh(xi)]=(DαGbuh)l,(20)\operatorname{dof}_l(J_K(u_h))=\alpha_l[\mathbf{B}(-p_l)u_h(x_j)-\mathbf{B}(p_l)u_h(x_i)]=(D_\alpha G_bu_h)_l, \tag*{(20)}

where pl=αl−1βl⋅τlp_l=\alpha_l^{-1}\beta_l\cdot\tau_l. Thus, the exponentially-fitted edge fluxes are represented by the vector DαGbuhD_\alpha G_bu_h.

We now establish an identity for the local bilinear form under the assumption of an elementwise constant approximated flux.

Lemma 4.1. Suppose that JK(uh)J_K(u_h) is constant on KK and that its edge degrees of freedom are given by the exponentially-fitted fluxes in (20). Then,

BKN(uh,vh)=vh⊤G⊤WDαGbuh.B_K^N(u_h,v_h)=v_h^{\top}G^{\top}WD_\alpha G_bu_h.

Proof. Since JK(uh)J_K(u_h) is constant on KK, there exists an affine function ψK∈W(K)\psi_K\in\mathcal{W}(K) such that

JK(uh)=∇ψK.J_K(u_h)=\nabla\psi_K.

For each oriented edge el=(i,j)e_l=(i,j), applying the edge degree of freedom gives

dof⁡l(JK(uh))=1∣el∣∫el∇ψK⋅τl ds=ψK(xj)−ψK(xi)=(GψK)l.\operatorname{dof}_l(J_K(u_h))=\frac{1}{|e_l|}\int_{e_l}\nabla\psi_K\cdot\tau_l\,\mathrm{d}s=\psi_K(x_j)-\psi_K(x_i)=(G\psi_K)_l.

where ψK\psi_K denotes the vector of nodal values of ψK\psi_K. Combining this relation with (4.1), we obtain

GψK=DαGbuh.G\psi_K=D_\alpha G_bu_h.

Consequently, using the incidence-based gradient representation and Lemma 3.2, we have

BKN(uh,vh)=∫K∇ψK⋅∇vh dx=vhTGTMGψK=vhTGTWGψK=vhTGTWDαGbuh,\begin{aligned} B_K^N(u_h,v_h)&=\int_K\nabla\psi_K\cdot\nabla v_h\,\mathrm{d}x\\ &=v_h^{\mathsf{T}}G^{\mathsf{T}}MG\psi_K\\ &=v_h^{\mathsf{T}}G^{\mathsf{T}}WG\psi_K\\ &=v_h^{\mathsf{T}}G^{\mathsf{T}}WD_\alpha G_bu_h, \end{aligned}

which completes the proof. □

Again, for a general approximated flux JK(uh)J_K(u_h), motivated by the constant-flux case, we define the local IBEAM bilinear form by

BKh(uh,vh):=vhTGTWDαGbuh.(21)B_K^h(u_h,v_h):=v_h^{\mathsf{T}}G^{\mathsf{T}}WD_\alpha G_bu_h. \tag*{(21)}

The proposed IBEAM scheme is then to find uh∈Vhu_h\in V_h such that

∑K∈KhBKh(uh,vh)=F(vh),∀vh∈Vh.\sum_{K\in\mathcal{K}_h}B_K^h(u_h,v_h)=F(v_h),\qquad\forall v_h\in V_h.

The corresponding local stiffness matrix is

AK=GTWDαGb,A_K=G^{\mathsf{T}}WD_\alpha G_b,

and the global stiffness matrix AA is assembled from these local contributions in the standard finite-element manner.

A Sufficient Condition for Monotonicity Property

Finally, we establish a sufficient condition for the monotonicity of the proposed IBEAM scheme. Specifically, we show that the resulting stiffness matrix is a nonsingular MM-matrix under suitable positivity and connectivity conditions.

Lemma 4.2. Suppose that ωlK>0\omega_l^K>0 for every edge ele_l of the complete graph K∗K^* associated with each K∈KhK\in\mathcal{K}_h, and that αl>0\alpha_l>0. Then, the global stiffness matrix AA, restricted to the interior degrees of freedom, is a nonsingular MM-matrix.

Proof. The global stiffness matrix is assembled from local contributions of the form GTWDαGbG^{\mathsf{T}}WD_\alpha G_b. For an oriented edge el=(i,j)e_l=(i,j), the corresponding contribution, restricted to the two vertices ii and jj, is

AK(l)=ωlKαl[−11][−B(pl)B(−pl)]=ωlKαl[B(pl)−B(−pl)−B(pl)B(−pl)].A_K^{(l)}=\omega_l^K\alpha_l \begin{bmatrix} -1\\ 1 \end{bmatrix} \begin{bmatrix} -\mathbf{B}(p_l)&\mathbf{B}(-p_l) \end{bmatrix} =\omega_l^K\alpha_l \begin{bmatrix} \mathbf{B}(p_l)&-\mathbf{B}(-p_l)\\ -\mathbf{B}(p_l)&\mathbf{B}(-p_l) \end{bmatrix}.

Since B(s)>0\mathbf{B}(s)>0 for all s∈Rs\in\mathbb{R}, ωlK>0\omega_l^K>0, and αl>0\alpha_l>0 by assumption, the diagonal entries of each edge contribution are positive, while its off-diagonal entries are negative.

Therefore, the assembled global stiffness matrix has nonpositive off-diagonal entries and is a ZZ-matrix.

Summing the columns of each local edge contribution gives

[11][B(pl)−B(−pl)−B(pl)B(−pl)]=[00].\begin{bmatrix} 1 & 1 \end{bmatrix} \begin{bmatrix} \mathbf{B}(\mathbf{p}_l) & -\mathbf{B}(-\mathbf{p}_l) \\ -\mathbf{B}(\mathbf{p}_l) & \mathbf{B}(-\mathbf{p}_l) \end{bmatrix} = \begin{bmatrix} 0 & 0 \end{bmatrix}.

Thus, the fully assembled stiffness matrix A‾\overline{A}, before imposing Dirichlet boundary conditions, satisfies

1TA‾=0T.\mathbf{1}^{\mathsf{T}}\overline{A}=\mathbf{0}^{\mathsf{T}}.

Since A‾\overline{A} is a ZZ-matrix with positive diagonal entries, it is weakly column-diagonally dominant. After imposing homogeneous Dirichlet boundary conditions, the reduced matrix AA remains a weakly column-diagonally dominant ZZ-matrix, with strict dominance in columns corresponding to interior vertices adjacent to the boundary. Since every connected component of the interior vertex graph is connected to the Dirichlet boundary, AA is a nonsingular MM-matrix by the irreducible diagonal dominance criterion applied to each component. □

Remark 4.3. The nonsingular MM-matrix property implies that A−1≥0A^{-1} \ge0 entrywise, ensuring the existence and uniqueness of the discrete IBEAM solution and the discrete maximum principle. However, this monotonicity result does not by itself guarantee convergence of the proposed scheme. On simplicial meshes, IBEAM extends the classical EAFE scheme from two and three dimensions to arbitrary dimensions. Since the underlying finite-element structure remains unchanged, the first-order convergence analysis of EAFE can be extended to arbitrary-dimensional simplices under analogous regularity and mesh assumptions. We therefore omit the proof. For general convex polytopal meshes, however, the same argument cannot be applied directly, since the generalized barycentric basis functions are not necessarily piecewise linear and may have nonzero second derivatives. Nevertheless, the numerical experiments in the following section demonstrate first-order convergence in the H1H^1 semi-norm on various polytopal meshes.

Numerical Experiments

To investigate the efficacy of the IBEAM method, we perform two types of numerical experiments in two and three dimensions. The first is a convection-dominated problem with a boundary layer and no closed-form solution, which we use to assess qualitatively whether the scheme avoids the spurious oscillations that afflict the standard Galerkin method. The second set of tests is a convergence study with a manufactured solution, in which we measure the L2L^2 norm and H1H^1 semi-norm errors under mesh refinement. In the latter set of test problems, we simulate on square, quadrilateral, and polygonal meshes in two dimensions and on tetrahedral and hexahedral meshes in three dimensions. We again note that polygonal and hexahedral meshes have not been previously treated by EAFE-type methods [27] (nor hexahedral for EAVE-type schemes [13]), and therefore we highlight the performance of IBEAM on these meshes. We include tetrahedral meshes as a validation that pre-existing results would remain consistent.

All computations were performed in MATLAB using Wachspress coordinates [15, 26] for the barycentric coordinates and a five-point Gaussian quadrature rule. A direct solver was used to solve the resulting linear systems. Simulations were run on the CPU of a System76 Pangolin laptop with an AMD Ryzen 7 6800U processor (8 cores, 16 threads, up to 4.7 GHz) and 32 GB of LPDDR5 memory.

Two-Dimensional Results

For the two-dimensional results, we consider three mesh types: square, quadrilateral, and general polygonal. We choose the square mesh for ease of construction while still demonstrating the method’s effectiveness for non-simplicial grids. Quadrilaterals are chosen to see how the method is affected when presented with an unstructured mesh, and convex polygons reinforce the schemes performance on more general meshes. We present a visual representation of each of these grids in Figure 2.

Two-dimensional square, quadrilateral, and polygonal meshes

Figure 2. (a) Square mesh. (b) Quadrilateral mesh. (c) Polygonal mesh. Two-dimensional meshes used in numerical experiments.

Boundary Layer Example

Our first test is a convection-dominated convection-diffusion problem with α=10−3\alpha= 10^{-3} and 10−610^{-6}, β=[1 2]T\boldsymbol{\beta} = [1\ 2]^{\mathsf{T}}, and f=1f = 1. Problems of this type arise in many applications, including heat and mass transport, pollutant dispersion, and semiconductor device modeling, where diffusion is weak relative to convection and the solution develops thin boundary layers. The relative strength of convection and diffusion is measured by the mesh Péclet number,

Peh=∣β∣h2α,\mathrm{Pe}_{h} = \frac{|\boldsymbol{\beta}|h}{2\alpha},

which compares the amount of convective transport across one element to the amount of diffusive transport over the same distance. When Peh≤1\mathrm{Pe}_{h} \le1, diffusion dominates on the mesh scale, and a standard Galerkin method performs well. When Peh>1\mathrm{Pe}_{h} > 1, convection dominates at the mesh scale, the stiffness matrix may lose its MM-matrix property, and Galerkin solutions develop spurious oscillations [21]. Equivalently, Peh>1\mathrm{Pe}_{h} > 1 means the mesh is too coarse to resolve the boundary layer, whose width is of order α/∣β∣\alpha/|\boldsymbol{\beta}| [23]. Large Péclet numbers suggest a demanding test of the monotonicity properties of the IBEAM system established in Subsection 4.1.

For the following experiments, we compute a solution using IBEAM on a square mesh and the standard Galerkin method on a uniform triangular mesh, both with a mesh size of h=2−5h = 2^{-5}. For the present parameters, Peh=5⋅2−5/(2⋅10−3)≈35\mathrm{Pe}_{h} = \sqrt{5}\cdot2^{-5}/(2\cdot10^{-3}) \approx35 for α=10−3\alpha= 10^{-3} and 35,000 for α=10−6\alpha= 10^{-6}, so the problem is well inside the convection-dominated regime.

Figure 3b and Figure 3c show that the IBEAM solution is free of spurious oscillations, consistent with the known behavior of EAFE-type schemes [28] and with the MM-matrix and discrete maximum principle properties in Subsection 4.1. By contrast, the standard Galerkin method oscillates near the boundary layer [21], even with the larger value of α\alpha, rendering the solution nonphysical.

Convergence Study

Next, we perform a convergence study using the manufactured solution u(x,y)=sin⁡(πx)sin⁡(πy)u(x,y) = \sin(\pi x)\sin(\pi y), with convection vector β=[0 1]T\boldsymbol{\beta} = [0\ 1]^{\mathsf{T}} and right-hand side chosen so that uu satisfies the convection-diffusion equation. The diffusion coefficient takes values α∈{1,0.1,0.01,0.001}\alpha\in\{1, 0.1, 0.01, 0.001\}, and the mesh size varies from h=2−2h = 2^{-2} to h=2−8h = 2^{-8}. For each mesh family discussed above, we report the L2L^2 norm and H1H^1 semi-norm error at each refinement level. We compare the observed H1H^1 semi-norm rates with those known for EAFE: an O(h)O(h) estimate is proven in [28].

FEM solution surface

(a) (a) FEM solution, α=10−3\alpha= 10^{-3}.

IBEAM solution surface for alpha equal to 10 to the minus 3

(b) (b) IBEAM solution, α=10−3\alpha= 10^{-3}.

IBEAM solution surface for alpha equal to 10 to the minus 6

(c) (c) IBEAM solution, α=10−6\alpha= 10^{-6}.

Figure 3. Comparison of FEM and IBEAM solutions for α=10−3\alpha= 10^{-3}, β=[12]⊤\beta= \begin{bmatrix}1\\2\end{bmatrix}^{\top}, f=1f = 1, and h=1/32h = 1/32, where the IBEAM solution refers to the solution utilizing local stiffness matrix (21). Unlike the FEM solution, which exhibits nonphysical oscillations, the IBEAM solution avoids oscillatory behavior, even for significantly smaller α=10−6\alpha= 10^{-6}.

On square meshes (Figure 4), in the diffusion-dominated regime (α=1\alpha= 1), the computed rates agree with the established rate [28] of O(h)O(h) in the H1H^1 semi-norm from the coarsest mesh onward, and we observe O(h2)O(h^2) convergence in the L2L^2 norm. As α\alpha decreases, the order is reduced on coarse meshes, and additional refinement is needed before the optimal rate is reached. We call this range of mesh sizes the pre-asymptotic regime.

IBEAM error convergence rates on square meshes

Figure 4. IBEAM error convergence rates on square meshes for α∈{1,0.1,0.01,0.001}\alpha\in\{1, 0.1, 0.01, 0.001\}, plotted against mesh size hh, for the H1H^1 semi-norm (left) and L2L^2 norm (right) error.

On quadrilateral meshes (Figure 5), the pre-asymptotic regime of convergence appears at α=10−1\alpha= 10^{-1}, earlier than on square meshes. As the mesh becomes less structured, we expect the condition number of the element stiffness matrix to increase and the conditions required for monotonicity, as shown in Subsection 4.1, to be violated. For general convex polygonal meshes (Figure 6), we approximate the mesh size hh by Nel−1/2N_{\mathrm{el}}^{-1/2}. Here, we see that the pre-asymptotic regime is now noticeable across all diffusion values, requiring further refinement to correct it. The difference from a structured mesh is most pronounced: some elements do not satisfy the conditions established for monotonicity in Subsection 4.1. In particular, large variations in edge lengths and interior angles cause stability issues. This manifests as a degradation in quality for the L2L^2 norm convergence order, but the O(h)O(h) convergence in the H1H^1 semi-norm is partially preserved. In particular, for α=10−3\alpha= 10^{-3}, the error does not yet appear to converge at O(h)O(h), but we expect the asymptotic regime to emerge with further mesh refinement.

IBEAM error convergence rates on quadrilateral meshes

Figure 5. IBEAM error convergence rates on quadrilateral meshes for α∈{1,0.1,0.01,0.001}\alpha\in\{1, 0.1, 0.01, 0.001\}, plotted against mesh size hh, for the H1H^1 semi-norm (left) and L2L^2 norm (right) error.

IBEAM error convergence rates on convex polygonal meshes

Figure 6. IBEAM error convergence rates on convex polygonal meshes for α∈{1,0.1,0.01,0.001}\alpha\in\{1, 0.1, 0.01, 0.001\}, plotted against mesh size hh, for the H1H^1 semi-norm (left) and L2L^2 norm (right) error.

Three-Dimensional Results

For the three-dimensional results, we consider two mesh types: tetrahedral and hexahedral. The tetrahedral mesh allows for direct comparison to a variety of schemes that function on simplicial meshes and serves as a baseline result. The hexahedral mesh is chosen for its simplicity of construction and to demonstrate the scheme’s performance on non-simplicial meshes.

Boundary Layer Example

We return to the boundary layer model problem, but in three dimensions. For the following experiments, we compute a solution using IBEAM on a hexahedral mesh and the standard Galerkin method on a uniform tetrahedral mesh. The parameter regime is kept consistent: α=10−3\alpha= 10^{-3} and α=10−6\alpha= 10^{-6}, f=1f = 1, and β=[1 2 0]T\boldsymbol{\beta} = [1\ 2\ 0]^{\mathsf{T}}, but we reduce the mesh resolution by a factor of two, h=2−4h = 2^{-4}. The mesh Péclet numbers associated with these parameter regimes are Peh=5⋅2−4/(2⋅10−3)≈70\mathrm{Pe}_h = \sqrt{5}\cdot2^{-4}/(2\cdot10^{-3}) \approx70 for α=10−3\alpha= 10^{-3} and 70,000 for 10−610^{-6}, so we expect a boundary layer to form which the standard Galerkin finite-element method would fail to resolve.

Figures 7b and 7c show a contour plot of the level curve at z=0.5z = 0.5, where we observe a smooth solution unaffected by spurious oscillations; however, spurious oscillations appear in Figure 7a, even for a larger value of α\alpha. Again, this validates the monotonicity properties established in Subsection 4.1.

FEM solution at z = 0.5, alpha = 10^-3

(a) FEM solution at z=0.5z = 0.5, α=10−3\alpha= 10^{-3}.

IBEAM solution at z = 0.5, alpha = 10^-3

(b) IBEAM solution at z=0.5z = 0.5, α=10−3\alpha= 10^{-3}.

IBEAM solution at z = 0.5, alpha = 10^-6

(c) IBEAM solution at z=0.5z = 0.5, α=10−6\alpha= 10^{-6}.

Figure 7. Comparison of FEM and IBEAM solutions at the level curve z=0.5z = 0.5 for α=10−3\alpha= 10^{-3}, β=[1 2 0]T\boldsymbol{\beta} = [1\ 2\ 0]^{\mathsf{T}}, f=1f = 1, and h=2−4h = 2^{-4}, where IBEAM solution refers to the solution utilizing local stiffness matrix (4.2). Unlike the FEM solution, which exhibits nonphysical oscillations, the IBEAM solution avoids oscillatory behavior, even for significantly smaller α=10−6\alpha= 10^{-6}.

Convergence Study

Next, we perform a convergence study using the manufactured solution u(x,y)=sin⁡(πx)sin⁡(πy)sin⁡(πz)u(x,y)=\sin(\pi x)\sin(\pi y)\sin(\pi z), with convection vector β=[0 0 1]T\boldsymbol{\beta}=[0\ 0\ 1]^{\mathsf{T}}, and the right-hand side chosen so that uu satisfies the convection-diffusion equation. We let the diffusion parameter take values α∈{1,0.1,0.01,0.001}\alpha\in\{1, 0.1, 0.01, 0.001\} and consider meshes with mesh size h=2−2h=2^{-2} to h=2−6h=2^{-6}. For each mesh family discussed above, we again report the L2L^2 norm and H1H^1 semi-norm error at each refinement level. We compare the H1H^1 semi-norm error results against O(h)O(h) as established in [28].

On tetrahedral meshes (Figure 8), the computed rates agree with or exceed the established O(h)O(h) rate in the H1H^1 semi-norm [28] starting from the coarsest mesh for all but the smallest α\alpha. We see that in the case of strongest convection dominance, the computed error rates in both the H1H^1 semi-norm and the L2L^2 norm fall short of the O(h)O(h) and O(h2)O(h^2) bounds, respectively. We expect that further refinement, while maintaining monotonicity, ensures better convergence. Hexahedral meshes (Figure 9) demonstrate performance similar to that observed on tetrahedral meshes. These results show that IBEAM retains its predicted convergence behavior on non-simplicial three-dimensional meshes, provided sufficient refinement, which has not been addressed by EAFE or EAVE-type methods [13, 27, 28].

IBEAM error convergence rates on tetrahedral meshes

Figure 8. IBEAM error convergence rates on tetrahedral meshes for α∈{1,0.1,0.01,0.001}\alpha\in\{1, 0.1, 0.01, 0.001\}, plotted against mesh size hh, for the H1H^1 semi-norm (left) and L2L^2 norm (right) error.

IBEAM error convergence rates on hexahedral meshes

Figure 9. IBEAM error convergence rates on hexahedral meshes for α∈{1,0.1,0.01,0.001}\alpha\in\{1, 0.1, 0.01, 0.001\}, plotted against mesh size hh, for the H1H^1 semi-norm (left) and L2L^2 norm (right) error.

Conclusion

We developed an incidence-based edge-averaged method for scalar convection-diffusion problems on convex polytopal meshes in arbitrary dimensions. The proposed method extends the classical EAFE scheme through a graph-theoretic formulation based on vertex-edge incidence matrices. By introducing complementary edges and generalized barycentric coordinates, we established a matrix identity that relates the generalized Whitney edge mass matrix to a diagonal representation of the Poisson stiffness matrix, enabling an exponentially-fitted discretization on general convex polytopes. We also established sufficient conditions under which the resulting stiffness matrix is a nonsingular MM-matrix, ensuring monotonicity and the discrete maximum principle. Numerical experiments on two- and three-dimensional meshes demonstrate first-order convergence in the H1H^1 semi-norm and stable, nonoscillatory behavior in convection-dominated regimes. Approximately second-order convergence in the L2L^2 norm is also observed in most cases.

One promising direction for future research is to apply the proposed method to time-dependent convection-diffusion problems through a space-time formulation [3]. By treating time as an additional coordinate, three-dimensional time-dependent problems can be formulated in four-dimensional space-time domains, providing an opportunity to investigate the numerical performance of IBEAM on four-dimensional polytopal meshes. Furthermore, the incidence-based framework suggests potential extensions to vector convection-diffusion problems in H(curl⁡)H(\operatorname{curl}) and H(div⁡)H(\operatorname{div}), where edge-face and face-element incidence relations may be exploited to construct structure-preserving discretizations.

Acknowledgments

The authors used AI to assist in generating portions of the code, in debugging, and scripting figure generation. All AI-assisted code was reviewed, tested, and validated by the authors, and the numerical results were verified independently. The authors assume full responsibility for all content. The authors also thank Ludmil Zikatanov for his valuable comments, suggestions, and insights toward the manuscript.

References

  1. [1]J. H. Adler, C. C. Cavanaugh, X. Hu, and L. T. Zikatanov, A finite-element framework for a mimetic finite-difference discretization of Maxwell’s equations, SIAM Journal on Scientific Computing, 43 (2021), pp. A2638–A2659.arxiv.org/abs/2012.03148
  2. [2]J. H. Adler, C. R. Cavanaugh, X. Hu, A. Huang, and N. Trask, A stable mimetic finite-difference method for convection-dominated diffusion equations, SIAM Journal on Scientific Computing, 45 (2023), pp. A2973–A3000.arxiv.org/abs/2208.04169
  3. [3]J. H. Adler, X. Hu, and S. Lee, A unified spatiotemporal formulation with physics-preserving structure for time-dependent convection-diffusion problems, arXiv preprint arXiv:2512.24650, (2025).arxiv.org/abs/2512.24650
  4. [4]D. N. Arnold, R. S. Falk, and R. Winther, Finite element exterior calculus, homological techniques, and applications, Acta Numerica, 15 (2006), pp. 1–155.DOI
  5. [5]M. Augustin, A. Caiazzo, A. Fiebach, J. Fuhrmann, V. John, A. Linke, and R. Umla, An assessment of discretizations for convection-dominated convection-diffusion equations, Computer Methods in Applied Mechanics and Engineering, 200 (2011), pp. 3395–3409.DOI
  6. [6]R. E. Bank, W. M. Coughran, Jr., and L. C. Cowsar, The finite volume Scharfetter-Gummel method for steady convection diffusion equations, Computing and Visualization in Science, 1 (1998), pp. 123–136.DOI
  7. [7]R. E. Bank, P. S. Vassilevski, and L. T. Zikatanov, Arbitrary dimension convection-diffusion schemes for space-time discretizations, Journal of Computational and Applied Mathematics, 310 (2017), pp. 19–31.
  8. [8]G. R. Barrenechea, V. John, P. Knobloch, and R. Rankin, A unified analysis of algebraic flux correction schemes for convection-diffusion equations, SeMA Journal, 75 (2018), pp. 655–685.DOI
  9. [9]A. Bossavit and L. Kettunen, Yee-like schemes on a tetrahedral mesh, with diagonal lumping, International Journal of Numerical Modeling: Electronic Networks, Devices and Fields, 12 (1999), pp. 129–142.doi.org/10.1002/(sici)1099-1204(199901/04)12:1/2<129::aid-jnm327>3.0.co;2-g
  10. [10]F. Brezzi, L. D. Marini, and P. Pietra, Two-dimensional exponential fitting and applications to drift-diffusion models, SIAM Journal on Numerical Analysis, 26 (1988), pp. 1342–1355.DOI
  11. [11]A. N. Brooks and T. J. R. Hughes, Streamline upwind/Petrov-Galerkin formulations for convection dominated flows with particular emphasis on the incompressible Navier-Stokes equations, Computer Methods in Applied Mechanics and Engineering, 32 (1982), pp. 199–259.DOI
  12. [12]E. Burman and P. Hansbo, Edge stabilization for Galerkin approximations of convection-diffusion-reaction problems, Computer Methods in Applied Mechanics and Engineering, 193 (2004), pp. 1437–1453.DOI
  13. [13]S. Cao, L. Chen, and S. Lee, Edge-averaged virtual element methods for convection-diffusion and convection-dominated problems, Journal of Scientific Computing, 104 (2025).DOI
  14. [14]W. Dörfler, Uniform error estimates for an exponentially fitted finite element method for singularly perturbed elliptic equations, SIAM Journal on Numerical Analysis, 36 (1999), pp. 1709–1738.DOI
  15. [15]M. Floater, A. Gillette, and N. Sukumar, Gradient bounds for Wachspress coordinates on polytopes, SIAM Journal on Numerical Analysis, 52 (2014), pp. 515–532.
  16. [16]Y. Haugazeau and P. Lacoste, Condensation de la matrice de masse pour les éléments finis mixtes de H(rot), Comptes rendus de l’Académie des sciences. Série I, Mathématique, 316 (1993), pp. 509–512.
  17. [17]V. John and P. Knobloch, On spurious oscillations at layers diminishing (SOLD) methods for convection-diffusion equations: Part I – a review, Computer Methods in Applied Mechanics and Engineering, 196 (2007), pp. 2197–2215.DOI
  18. [18]F. Kikuchi, Discrete maximum principle and artificial viscosity in finite element approximations to convective diffusion equations, ISAS report/Institute of Space and Aeronautical Science, University of Tokyo, 42 (1977), pp. 153–166.
  19. [19]P. Knobloch, A generalization of the local projection stabilization for convection-diffusion-reaction equations, SIAM Journal on Numerical Analysis, 48 (2010), pp. 659–680.DOI
  20. [20]J. Lohi and L. Kettunen, Whitney forms and their extensions, Journal of Computational and Applied Mathematics, 393 (2021), p. 113520.DOI
  21. [21]K. W. Morton, Numerical Solution of Convection-Diffusion Problems, Chapman & Hall, 1996.
  22. [22]T. D. Nguyen, C. Bessem, and F. Rogier, High-order Scharfetter-Gummel-based schemes and applications to gas discharge modeling, Journal of Computational Physics, 461 (2022), p. 111196.DOI
  23. [23]H.-G. Roos, M. Stynes, and L. Tobiska, Robust Numerical Methods for Singularly Perturbed Differential Equations: Convection-Diffusion-Reaction and Flow Problems, Springer, 2nd ed., 2008.
  24. [24]D. L. Schrafter and H. K. Gummel, Large-signal analysis of a silicon Read diode oscillator, IEEE Transactions on Electron Devices, 16 (1969), pp. 64–77.
  25. [25]M. Stynes, Steady-state convection-diffusion problems, Acta Numerica, 14 (2005), pp. 445–508.DOI
  26. [26]E. L. Wachspress, Barycentric coordinates for polytopes, Computers & Mathematics with Applications, 61 (2011), pp. 3319–3321.DOI
  27. [27]S. Wu and J. Xu, Simplex-averaged finite element methods for H(grad), H(curl), and H(div) convection-diffusion problems, SIAM Journal on Numerical Analysis, 58 (2020), pp. 884–906.
  28. [28]J. Xu and L. T. Zikatanov, A monotone finite element scheme for convection-diffusion equations, Mathematics of Computation, 68 (1999), pp. 1429–1446.DOI

Paper details

vlnc:W_avt3zq46
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 6

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

AI

View statement
Looking for related papers…
Contents