WorksMathematics
A Cut Finite Element Method with Boundary Value Correction for Robin, Neumann and Interface Conditions
Abstract
We extend the cut finite element method with boundary value correction of Burman, Hansbo and Larson [Math.\ Comp.\ 87 (2018) 633–657] to Robin and Neumann conditions and to interface problems with imperfect bonding. The formulation retains a piecewise affine boundary for all integration, while using the exact geometry only at boundary quadrature points. The correction uses Taylor expansions of the solution and its gradient along the discrete normal, with the exact normal used to evaluate the transported flux. At an interface, a further correction to the normal flux jump accounts for the difference between these two normals. Under the stated trace and Poincaré bounds, the corrected form is an perturbation of a symmetric form, whose matrix can therefore be used as a preconditioner. For a fixed finite Robin coefficient, we obtain the energy error bound , whereas the bound uniform in the Robin coefficient is , provided the Poincaré constant is uniform. With , the sufficient Taylor orders are therefore and , respectively. Numerical examples for illustrate these rates and the effect of the flux-jump correction in the interface problem.
Introduction
Unfitted finite element methods use meshes that need not align with the boundary or interface of the physical domain, which considerably simplifies mesh generation. The geometric approximation must nevertheless be sufficiently accurate to support the desired convergence rate. If the boundary is replaced by a piecewise affine surface , imposing the boundary condition directly on limits this rate, regardless of the polynomial degree.
There are two main approaches to this problem. A higher order geometric representation can be used, with curved elements or quadrature on the exact domain. Alternatively, the affine approximation can be retained and the boundary values corrected through a Taylor expansion from to . This idea was introduced by Bramble, Dupont and Thomée [6] for polygonal approximations of curved domains. Burman, Hansbo and Larson [9] subsequently developed it for unfitted methods with Nitsche boundary conditions and ghost penalty stabilisation. Their analysis shows that a correction of order gives the optimal energy error for elements of degree when . Thus one Taylor term suffices for and two for [9 Tab. 1]. Without correction, the geometric distance must satisfy for the energy estimate and for the estimate [9 Rem. 3.5]. With one correction term, a piecewise affine boundary with suffices for degrees up to three.
The shifted boundary method of Main and Scovazzi [20, 21] uses the same principle, with a surrogate boundary formed by mesh faces. The computation then takes place on the resulting fitted submesh, and we refer to [1] for the analysis. Related methods on polytopal approximations were developed in [12, 14], while the virtual element methods in [4, 2] also allow staircase boundaries from pixel tessellations. For unfitted meshes, boundary value correction has also been combined with Lagrange multipliers [10], a penalty-free Nitsche method [5], and a mixed formulation [17]. An alternative based on isoparametric mappings of cut elements is given in [19]. These developments mainly concern Dirichlet boundary conditions.
For Neumann conditions, Bertoluzza, Montardini and Prada [3] develop a high order virtual element method on approximate domains. Their formulation allows non-aligned physical and numerical normals, with optimal error estimates under suitable assumptions for arbitrary approximation order. This provides a related treatment of the geometric difficulties that arise when the boundary condition involves the flux.
A Dirichlet condition prescribes the solution trace, so its transport requires only a Taylor expansion of that trace. The normal flux enters through integration by parts on , where the resulting terms use the normal of and are exact on the discrete domain.
For a Robin condition, the boundary data also involve the flux, and we consider
where is the exact normal. Since the discrete normal generally differs from by , transporting the evaluation point while retaining leaves an error in the flux. We remove this error by transporting the vector field and then taking its scalar product with the exact normal at the image point. The resulting operator, denoted by , has an error of , as for the trace correction (Lemma 3.2). The construction uses the polynomial gradient on each cut element together with pointwise evaluations of the exact normal, without requiring curved quadrature or additional unknowns.
We combine this flux correction with a formulation that is uniform in the Robin coefficient , connecting the Neumann problem at to the Dirichlet limit as . Following Juntunen and Stenberg [18], we set and introduce the weights
which give a single formulation for the whole family. At and , it reduces to the Dirichlet method with boundary value correction from [9], while gives a corrected Neumann formulation. The pair is equivalent to the weights in [18 (2.5)–(2.6)], but makes the roles of the trace and flux corrections explicit. The estimates have constants independent of under the stated coercivity and regularity assumptions.
We also consider an elliptic interface problem, where the prescribed flux jump is taken with respect to the exact normal, . Integration by parts on the two discrete subdomains instead gives the jump with respect to . Their difference contains
This term depends on the tangential flux jump, which is not prescribed by the interface data and therefore cannot be removed by transporting those data alone. We retain the corresponding integration-by-parts term, , in the formulation on . Although this term vanishes on the exact interface for a homogeneous normal flux jump, it need not vanish on the surrogate interface. We control it using two different trace estimates (Lemma 5.2). The elementwise gradient requires a discrete trace estimate with a factor , whereas the test function belongs to and satisfies a continuous trace estimate without this factor. Together with , this gives an perturbation.
Finally, we consider imperfect bonding, where a spring layer with stiffness joins the two phases, as in [16, 15]. For this interface counterpart of the Robin condition, the weight gives a single formulation for all bonding stiffnesses. Uniform estimates additionally require the phasewise Poincaré bounds stated below. In §7, we interpret the boundary value correction as an artificial compliance of order , given by the geometric offset weighted by the jump of the inverse diffusivity. This correction adds to the compliance in the physical bonding law.
We give the main arguments in Lemma 3.2, Lemma 5.2, Theorem 5.2 and Theorem 6.1, while proofs that follow from standard arguments are sketched.
Throughout, means , where is independent of the mesh size, the position of the boundary relative to the mesh, , and the contrast, unless stated otherwise. It may depend on the polynomial degree, the shape regularity, the curvature of , and the stated Poincaré bounds. Uniformity in the physical parameters is therefore conditional on uniform control of these bounds and of the solution norms.
The Robin problem and its two limits
Let , , be a bounded domain with a boundary and outward unit normal . For constant , reaction coefficient , and given , , and , we consider
The case gives the Neumann problem, while finite positive gives the Robin problem. The Dirichlet limit is obtained by setting and letting with fixed. We assume or . For the pure Neumann problem (), the continuous solution is defined modulo constants and the data must satisfy the usual compatibility condition. We restrict both the discrete trial and test spaces to functions with zero mean on , and normalise the exact solution to have the same mean on in the error analysis. This restriction is needed because flux transport need not preserve the discrete compatibility condition for constant test functions. In [18], the condition is written in terms of the compliance . We use so that the Neumann condition corresponds to the finite value and the notation agrees with the bonding stiffness in §6.
We denote the normal flux by
The boundary condition then reads , while multiplication of the equation by a test function and integration by parts give
The method combines this identity with the boundary condition through the following substitution.
Lemma 2.1 (The uniform substitution). Let , let , and let and be given by (2). Then and . For any , , and satisfying , we have
Proof. The first identity follows from . The second is . ◻
The first identity replaces part of the flux in (6) by the trace and the data. The second supplies a residual that may be tested without changing the exact solution. At , the identities become and , while in the Dirichlet limit they become and .
Remark 2.1 (Relation to the method of Juntunen and Stenberg). Applying Lemma 2.1 on the exact domain with and , without geometric correction, gives the symmetric formulation . It agrees with [18 (2.5)–(2.6)] under the correspondence and . Their stability condition [18 Thm. 3.2] requires sufficiently large in our notation. We extend this form by introducing a surrogate boundary, a ghost penalty, and the two transport operators.
The discrete geometry and the two Taylor operators.
The surrogate boundary.
Let be a quasi-uniform, shape-regular simplicial mesh of a fixed polytopal domain containing , with mesh size . We describe the boundary by a level set function , with near , such that
The cut elements are . In two dimensions, we assume that each cut element meets in one connected arc with endpoints on two element edges, and replace this arc by the chord joining its endpoints. In three dimensions we use a piecewise planar surface and assume the geometric estimates stated below. We set
We assume that is a closed Lipschitz surface bounding a polytopal domain , with uniformly bounded Lipschitz character as . The active mesh consists of the elements for which , and on their union we define
The outward unit normal of , denoted by , is constant on each affine piece of . Since all integrals are taken over , , or faces of the active mesh, the method requires quadrature only on polytopal sets.
The three pointwise maps.
Let be a point where is defined. For sufficiently small , the equation has a unique root near zero. We define
Thus transfers points from to along the discrete normal, while gives the exact normal at the image point . The geometric errors are measured by
Lemma 3.1 (Geometric consistency). In two dimensions, let be of class with curvature bounded by . For sufficiently small , the chord construction satisfies
The surface Jacobian of satisfies on each affine piece of . In three dimensions, these estimates are assumptions on the piecewise planar approximation.
Proof. For a chord of length with endpoints on , consider a second-order expansion of the corresponding arc. This gives a distance of order and a normal error of order , from which the first two estimates follow for sufficiently small . Integration of the distance estimate over then gives the bound on the symmetric difference. On each chord, the surface Jacobian differs from one by , which is with a constant depending on the curvature.
Other expansion directions are possible, and [9] allows a field satisfying . With our choice , the offset can be computed by Newton iteration for , starting at . Since is not the closest point map, we evaluate separately from for use in the flux transport (15). All three maps are needed only at quadrature nodes on .
Transporting a trace and transporting a flux
Write for differentiation in the discrete normal direction. For and a function smooth near , define
This is the Taylor polynomial along the segment from to , evaluated at , with . The operator transports the trace, as in [6, 9], and we also apply it componentwise to vector fields.
To transport the flux in the Robin condition, we define
We first expand along and then take its scalar product with the exact normal at the image point. Since is a fixed vector at each quadrature node, this is equivalent to expanding the scalar function . In particular, no derivative of the normal is required. For a discrete function, uses derivatives up to order and uses derivatives up to order , evaluated from the polynomial on the cut element. We therefore take .
Lemma 3.2 (Transport error). Let , where is a neighbourhood of the segment . Then
Proof. The first estimate follows by applying Taylor’s theorem with remainder to between and . For the second, apply the same argument to . Its value at is , since is the exact normal at .
The discrete flux differs from the exact flux at the image point by , which is generally for an affine boundary approximation. A Taylor expansion that retains in the scalar product leaves this directional error. By using in , we retain only the transport error . This construction uses the elementwise polynomial gradient and does not require gradient continuity between elements.
Assumption 3.1 (Extension). The solution of (4) and the datum extend to a fixed neighbourhood of , with there. The extensions satisfy pointwise in . The datum also extends to a neighbourhood of .
This assumption allows us to integrate by parts on without a volume consistency term. Under the weaker extension assumption in [9], the residual outside is retained and estimated on the strip [9 (3.14), (3.49)]. We use Assumption 3.1, which holds exactly for all test cases in §8, to focus on the transport of the boundary condition.
The method
The form
Let be a Nitsche parameter and set
We use a ghost penalty on the interior faces of the active mesh adjacent to cut elements,
The discrete problem is to find such that for all , where
If , both trial and test functions are restricted to the subspace . The datum is evaluated at , while and approximate the corresponding solution trace and flux from the polynomial on the cut element.
The term comes from integration by parts on and therefore uses the normal of that domain. It is present for every , while the terms involving and impose the transported boundary condition through Lemma 2.1. The ghost penalty provides uniform control of the active mesh norm with respect to the boundary cuts.
The two limits and the Taylor order.
Let and keep fixed as . Then , , , and . The method tends to Nitsche’s method with the boundary value correction of [9], acting on the trial function only. The flux transport disappears because the Dirichlet condition prescribes only the trace.
At , we have , , and . Equations (19)–(20) become
The two terms with coefficient form a Barbosa–Hughes stabilisation whose residual vanishes at the exact solution on the exact geometry, as in [18 (2.14)]. On , the residual instead has the Taylor transport error from Lemma 3.2. Dropping these two terms gives a natural Neumann formulation with a flux correction: the flux from integration by parts is replaced by before the datum is imposed.
Remark 4.1 (Where the Robin coefficient acts). The trace correction is multiplied by , while the flux correction is multiplied by or by . Thus the Dirichlet limit uses only the trace correction and the Neumann case only the flux correction, with both entering for finite positive .
The uniform consistency bound in Corollary 5.1 is , which becomes under . It is therefore bounded by the approximation error when
This is the rule in [9 Tab. 1] for the Dirichlet problem, and also suffices for the expected estimate (Remark 5.2 and [9 Tab. 2]). The estimate can be sharpened when is fixed and finite, as shown below.
Analysis.
Norms and trace inequalities.
We use the energy norm
The boundary weight is at the Dirichlet limit and vanishes at the Neumann limit. For fixed , as , so the norm approaches the natural Robin energy norm. When , we restrict the discrete space to functions with zero mean on .
The ghost penalty controls the gradient on the full active elements. Combined with the polynomial inverse inequality, this gives
The constants are independent of the cut position, and analogous estimates hold on active-mesh faces. These are standard consequences of the ghost penalty, for which we refer to [7, 9].
For the trace of the function itself, we also use
The constant is independent of because the domains have a uniform Lipschitz character. For a resolved chordal approximation of a curve, neighbouring segments meet at an angle , which allows local boundary charts to be chosen with uniform bounds. In three dimensions we assume the same uniform domain property.
We make the Poincaré bound explicit:
We assume that is independent of and make its role in the analysis explicit. If , we may take . For , a Poincaré inequality with a boundary term also gives a uniform bound for sufficiently small . For , the zero-mean restriction gives the usual Poincaré bound. Since the constants below may depend on , uniformity in requires a uniform bound on . In particular, such a bound does not hold on the full space as when .
Combining (5.3) and (5.4) gives
This estimate has no negative power of and will be used in both the stability and consistency arguments.
Lemma 5.1 (Taylor operators on discrete functions). Let and assume is bounded. Then
Both left-hand sides vanish when .
Proof. On each full active element, we apply the polynomial inverse inequality and use the ghost penalty to control the gradient. The th trace and flux terms are then bounded, respectively, by
and
Summing over proves the result.
The normal mismatch
Integration by parts uses the normal , whereas the boundary condition uses . For , the resulting flux difference is
The term occurs in the stability estimate.
Lemma 5.2 (Bound for the normal mismatch). Under (5.4),
The constant depends on and but is independent of and of the cut position.
Proof. By the Cauchy–Schwarz inequality,
We apply the discrete trace estimate (23) to the gradient and the continuous estimate (26) to the function itself. Together they give
The result follows from .
Remark 5.1 (The Neumann case). The volume terms and (5.4) control the norm without a boundary penalty. Hence (5.5) and the argument above also apply when . If , both trial and test functions must satisfy the zero-mean restriction specified above. The corrected form need not annihilate constant test functions, so a quotient-space statement alone is not enough.
Stability
Theorem 5.1 (Stability). Let be sufficiently large and let . Under (25), there are and such that
The constants may depend on , the mesh regularity, the geometry and , but are independent of the cut position. They are also independent of whenever is uniformly bounded.
Proof. For , we have and . Using , we can therefore write the form as
where
Young’s inequality gives
Since , the first estimate in (23) absorbs the flux terms when . The remaining positive terms control .
By Lemma 5.2, the first term in is bounded by . The bound gives for the second term, so both can be absorbed for small .
For , Lemma 5.1 bounds the trace-correction terms and the flux-flux term by . Applying (5.5) to the remaining flux-correction term gives the bound . Since , these terms can also be absorbed.
Consistency
Define the transported boundary residual by
Proof. The extension assumption allows integration by parts on , giving
Since the extension is smooth across mesh faces, . Subtracting the load from (19) therefore leaves
Using gives (33), while the boundary condition at gives
The residual bound follows by applying Lemma 3.2 to these two remainders.
Corollary 5.1 (Weighted consistency). Let denote the norm of the extension of and set . Then
The constant is independent of , , and the position of the cut, with the dependence on the Poincaré estimate displayed through . For each fixed finite , the right-hand side is . At the Dirichlet limit it is .
Proof. Write
This gives the decomposition , with the transport estimates and . We estimate the two residuals separately, starting with the trace residual. The boundary term in the energy norm gives
This inequality also holds at , when both sides vanish. The discrete flux estimate gives . Since , , and , these estimates bound the part containing by
For the flux residual, the continuous trace estimate and (25) give . Combining this with yields
Multiplication by proves (5.14). Finally,
This proves the estimate for fixed finite , with constants that may depend on and
The energy error.
Theorem 5.3 (Energy error estimate). Under the hypotheses of Theorem 5.1 and Assumption 3.1, let and . Then
Constants are independent of , , and the position of the cut when the Poincaré constant is uniformly bounded. For fixed this implies
Thus gives the optimal energy order for fixed finite , while remains sufficient uniformly up to the Dirichlet limit.
Proof. Let be an interpolant on the active mesh and set and . Standard interpolation and trace estimates on the full elements give
They also give , with the extension used in all these norms. For a zero-mean space, we adjust the interpolant by a constant without changing the approximation order.
We split as in (40) and first consider . The volume and ghost-penalty terms have the usual interpolation bounds, while the boundary terms are controlled by
The first estimate is interpreted as zero at . Together with the interpolation bounds, these estimates give
In particular, no inverse power of is needed.
To estimate , write and . Since and , the same trace estimates imply
For , the first expression vanishes and the term in the second expression is absent. We apply the continuous trace estimate to in the first term of (41) and use the boundary penalty and the discrete flux estimate in the other three terms. This gives
Its constant depends on the stated geometric and Poincaré constants. Consequently, .
Coercivity and the discrete equation now give
Applying Corollary 5.1 and then the triangle inequality proves (5.15). The last assertions follow from and for finite .
The error.
Remark 5.2 (Expected estimate). The expected estimate is
For the Dirichlet problem, the corresponding estimate is proved in [9 Thm. 3.2] under the assumptions made there. We do not prove (38) for finite in this paper. Since the trial-side corrections make the form non-symmetric, a duality proof requires a separate adjoint-consistency estimate. It must also control the geometric remainders using the regularity available for the dual solution, usually . The remainder bound above does not provide this directly.
These points remain open after the improvement of the energy estimate. For example, and give order two in energy at fixed finite , but the geometric term in (38) would still be of order two. An additional Taylor term is therefore needed to obtain the expected third order in , in agreement with the numerical results.
A symmetric preconditioner.
Setting and replacing by in (4.3) gives the symmetric form
This is the uncorrected Juntunen–Stenberg form on , with the ghost penalty added. Define
Then
Proposition 5.1 (A small perturbation of a symmetric form). Let and . Under the mesh assumptions, the symmetric form satisfies
Under (5.4),
The constants are independent of the cut position, although may depend on .
Proof. Coercivity follows from the argument preceding in (31), and continuity follows from (23) and the bounds and . By (27) and (28),
Substitution in (41) bounds the two trace-correction terms by and the flux-flux term by . The first term is bounded by using (26). The geometric estimates now give (43).
Corollary 5.2 (Preconditioning by the uncorrected form). Let . In the inner product induced by ,
Thus preconditioned Richardson iteration converges with contraction factor at most , and the operator condition number in this norm is at most .
Proof. For a discrete function ,
This proves the operator bound. When the preconditioned operator acts on a unit vector, the resulting norm lies between and , which gives the condition-number estimate. □
The symmetric matrix uses the same quadrature points as the corrected method but does not require or . The correction can then be applied through matrix–vector products with , as illustrated by the numerical results in § 8.3. An analogous splitting is available for the interface form, subject to the phasewise Poincaré bounds used in § 6.
The interface problem with imperfect bonding.
The problem.
Let be a closed interface in a fixed domain , dividing into an enclosed subdomain and a surrounding subdomain . We take the normal to point from into and let . For a function defined on both sides, set . Given weights , we also define and the complementary average . With this notation, we consider the problem
with a Dirichlet condition on the fitted outer boundary . The third equation describes imperfect bonding through a spring layer with stiffness . For homogeneous data, the flux is proportional to the jump in . The limit gives perfect bonding, with a prescribed jump obtained by scaling the data as in the boundary Robin problem. At , the two fluxes are prescribed independently and vanish when . This is the scalar counterpart of the spring-layer law of Zhong and Meguid [22], which A. Hansbo and P. Hansbo [16] analysed and discretised in elasticity. We return to this relation in § 7.
The bonding law has the form of the Robin condition, with replacing the boundary value and replacing the boundary flux. We can therefore apply Lemma 2.1 directly, although the flux-jump condition requires an additional correction.
Three transports instead of two.
We construct as in § 3, so that it divides into two polytopal subdomains and . Let denote the cut elements. We use the Hansbo–Hansbo space, with separate unknowns on the two sides,
where each factor is defined on the elements meeting its subdomain. Thus each cut element carries two independent sets of degrees of freedom. We choose the harmonic weights and mean coefficient
following Dryja [13] and Burman and Zunino [11, (2.22)], with a related discussion in [8 §2]. With , we have . Thus the weighted flux average is times the arithmetic average of the normal gradients [11 §2.2]. Although other weights with can be used in the analysis [11 Rem. 2.13], the present choice controls the dependence on the coefficient contrast. This property is used in Lemma 6.1. Integration by parts on the polygonal subdomains gives
where we used the identity . In addition to and , we need a transport of the flux jump,
which is the analogue of (15) for the flux jump. We expand componentwise along and contract the result with the fixed vector , so that only is differentiated. Applying Lemma 3.2 on each side then shows that approximates with an error of order .
The method
As in the boundary problem, we introduce the weights
which satisfy and . We can therefore apply Lemma 2.1 with , , and . The discrete problem is to find such that for all , where
Here transports the weighted average flux, while the ghost penalty acts separately on the two active meshes. On each side, it is applied to faces belonging to cut elements.
The second line of (6.7) corrects the flux jump, while the remaining terms follow from the Robin method in § 4. The value of enters only through and . At , we have and , so that the two terms vanish. With the corresponding scaling of the data, this gives the Hansbo–Hansbo method [15] for , with the jump corrected by and the flux jump by . At , the values , and remove the terms, and the two flux conditions prescribe the one-sided fluxes. For , both transports contribute with weights and . The experiments in §8.2 cover nine orders of magnitude in .
Remark 6.1 (The flux-jump term is the misalignment correction). On the exact interface, the term vanishes when the flux is continuous. Its counterpart on must nevertheless be retained, as shown by the exact identity on ,
Here approximates with an error of order , while the second term contains the tangential flux jump and a normal mismatch of size . The tangential flux jump is generally non-zero, since on it satisfies . This term therefore depends on the solution and cannot be replaced by data. A fixed subdivision of each chord reduces its size but does not improve its order in . We instead retain the computable flux jump in the method. After integration by parts, the correction by leaves only the residual , which is of order at the exact solution.
Consistency and error estimate
In this section we use the interface energy norm
Here is the sum of the two ghost penalties. At , we fix the mean on the enclosed phase by imposing zero mean there on both the trial and test spaces. For the error estimate, we use the same normalization for the exact solution.
We define the two transported residuals
Theorem 6.1 (Consistency). Let and let solve (45), with each satisfying Assumption 3.1 on its side. Then, for all ,
and .
Proof. As in the proof of Theorem 5.2, we use (6.4) to cancel the volume terms and the two boundary terms. The flux-jump correction and its datum then leave . Applying Lemma 2.1 to the remaining terms, with and , gives the first two terms of (6.12). The residual bound follows from Lemma 3.2 on each side.
Theorem 6.2 (Energy error estimate). Assume the phasewise Poincaré bounds of Lemma 6.1 and Assumption 3.1 on each side, and let . For sufficiently large and , the discrete problem has a unique solution when . When , impose the Neumann compatibility condition and fix the mean in the enclosed phase, restricting both trial and test spaces as specified above. Then
Here is the fixed extension neighbourhood of , and carries the same coefficient weights. The constants and may depend on the phasewise Poincaré bounds but are independent of the position of relative to the mesh. They are also independent of and the contrast , provided that the Poincaré bounds are uniform in these parameters.
Remark 6.2 (What contrast independence means here). Since the norms on both sides of (57) contain coefficient weights, contrast independence refers to the constants in these norms, subject to the stated Poincaré bounds. For perfect bonding, Burman, Guzmán, Sánchez and Sarkis [8, Thm. 4.3] prove a stronger flux estimate whose constant is independent of the coefficients. Their proof uses a discrete extension from one phase to the whole domain. We have not examined whether boundary value correction preserves that estimate.
Sketch of proof. We follow the coercivity proof of Theorem 5.1, using the weights (6.3) to control the coefficient factors [15, 11]. The last two lines of (6.7) are treated as in the boundary problem, with in place of . The second line gives the additional misalignment term , which is bounded in Lemma 6.1 below. Consistency follows from Theorem 6.1, and we bound the first two terms of (6.12) as in Corollary 5.1. For the flux-jump residual, we retain the coefficient weights before applying the Cauchy–Schwarz inequality. A Taylor remainder from phase paired with the trace from phase has coefficient . Extracting the solution weight leaves the factor
The calculation is the same as in Lemma 6.1 below, so the continuous trace estimate gives an bound for this residual in the weighted norms. We treat the average-flux residual coefficient by coefficient as well, and use the boundary penalty for the trace residual. Together, these estimates imply the stated uniform bound. The perfect-bonding case follows by taking the scaled-data limit, while the interpolation argument is the phasewise version of Theorem 5.3. □
Lemma 6.1 (The interface misalignment remainder). Let the weights be those of (6.3) and assume, as in (5.4), that for . Then
where depends on , but, for a fixed bound on , is independent of , and the contrast .
Proof. As in Lemma 5.2, we apply the discrete trace inequality to the gradients and the continuous trace inequality (5.3) to the functions. This gives
where and . The product contains the four coefficients . The two diagonal ones are and , both at most 1. The two off-diagonal ones are both equal to
by the arithmetic–geometric mean inequality. Thus all four coefficients are bounded by 1 independently of the contrast, and multiplication by gives the result. The harmonic weights [11] are essential here, since arithmetic weights would introduce the off-diagonal factors and its reciprocal. □
Remark 6.3 (The debonded limit). At , the enclosed phase solves a pure Neumann problem, whose solution is determined only up to a constant. We therefore fix its mean before applying Poincaré’s inequality, as in the related normalisation used in [8 Thm. 4.3]. For , the bonding term controls this constant, but that control weakens as tends to zero. Thus Theorem 6.2 does not assert uniformity in this limit unless the phasewise Poincaré constants remain bounded.
Imperfect bonding as the interface Robin condition
A. Hansbo and P. Hansbo [16] considered elasticity problems with strong and weak discontinuities, coupling the two phases through a cohesive law that includes the limit of perfect bonding. The third equation in (45) is the scalar analogue of this law. Our formulation (6.7) uses the same doubled unknowns and weighted averages, with Taylor transports and a flux-jump correction to account for the surrogate interface.
Boundary value correction also has a direct interpretation as an interface compliance. To see this, take and assume continuous flux, so that . Then, on ,
where . Thus the transported condition gives, to this order,
This has the algebraic form of an imperfect bonding law with compliance , although the geometric compliance can have either sign. The geometric offset thus acts as an artificial compliance, weighted by the jump of the inverse diffusivity. Since , this contribution has order and, at this order, adds to the physical compliance .
Two features distinguish this correction from a physical compliance. Because may change sign with , we treat it as a perturbation in the stability analysis. Moreover, when , so the first-order correction vanishes under the assumed flux continuity. At a boundary, the weights and give a corresponding interpretation. The Dirichlet limit corrects the value and the Neumann limit corrects the flux, while both corrections contribute for a Robin condition.
Numerical examples
We use structured triangulations of the unit square and continuous elements, with one chord per cut element for the surrogate boundary. The ghost penalty (18) uses , and the Nitsche parameter is set to . The first geometry is a circle of radius 0.3137, centred at (0.5017, 0.4983). The second is a four-petal flower with level set , where and , about the same centre. Neither boundary is aligned with the mesh, as shown in Figure 1.

Figure 1. The two geometries at : the background mesh, the active elements (shaded), the surrogate boundary (thick) and the exact boundary (thin).
For the boundary value problems, we take with and , compute and from this solution, and evaluate boundary data at . Cut cells are integrated using a centroid fan with signed triangle areas. With sufficiently accurate triangle quadrature, this gives exact integration of polynomials on a simple polygon. On each element, we obtain the nodal basis and the derivatives required by the Taylor operators by inverting the Vandermonde matrix of a scaled monomial basis.
We first apply a patch test on a curved boundary. For a polynomial solution of degree and , the operators , and are exact, so the residuals (32) and (55) vanish. This algebraic test uses , whereas the error analysis assumes . The relative error is below for , for and for . These bounds hold for in both the boundary and interface tests, with the larger error for consistent with the conditioning of the monomial Vandermonde matrix. The patch test thus checks the geometry maps, Taylor operators, Nitsche terms, ghost penalty and quadrature together.
We also measure the geometric errors for . The ratio stays below 0.26 on the circle and 1.02 on the flower, while the corresponding bounds for are 1.18 and 5.78. These measurements are consistent with the orders in Lemma 3.1.
The Robin family: convergence and the Taylor order.
Table 1 reports least-squares convergence rates in the seminorm and in over the four meshes , for , and the three representative values (Neumann), (Robin) and (Dirichlet).
| 1 | 0 | 0.98 | 1.79 | 0.98 | 1.92 | 0.99 | 2.01 |
| 1 | 1 | 0.98 | 1.79 | 0.98 | 1.92 | 0.99 | 2.02 |
| 2 | 0 | 2.15 | 1.77 | 2.20 | 1.82 | 1.54 | 1.87 |
| 2 | 1 | 2.22 | 2.96 | 2.22 | 3.16 | 2.22 | 3.29 |
| 2 | 2 | 2.22 | 2.96 | 2.22 | 3.16 | 2.22 | 3.29 |
| 3 | 0 | 1.81 | 1.78 | 1.85 | 1.82 | 1.33 | 1.90 |
| 3 | 1 | 3.40 | 4.07 | 3.40 | 4.24 | 3.41 | 4.06 |
| 3 | 2 | 3.40 | 4.00 | 3.40 | 4.19 | 3.41 | 4.36 |
Table 1. Least-squares convergence rates on the circle over . Optimal is in and in .
The Dirichlet results agree with the rule (4.5). For and , the predicted energy rate is , compared with measured rates of 1.54 for and 1.33 for . Adding one Taylor term raises these rates to 2.22 and 3.41.
At fixed finite , the improved estimate in Theorem 5.3 predicts an energy rate of . For and , the predicted order 2 agrees with the measured rates of 2.15 at and 2.20 at . The optimal energy convergence in these cases therefore follows from the analysis. For and , the predicted rate is again 2, while the measured rates are 1.81 and 1.85 on the tested meshes. The distinction between finite and the Dirichlet limit comes from the complementary weights in Remark 4.1, which confine the half-order loss in the uniform estimate to the -weighted part of the residual.
Figure 2. Energy-norm convergence on the circle for . Dashed curves use and solid curves use . The reference slopes are and .
The improved energy estimate does not imply optimal convergence. For and , the measured rates range from 1.77 to 1.87, increasing with to between 2.96 and 3.29. This agrees with the need for a correction discussed in Remark 5.2. At , where only the flux transport is active, increasing from 0 to 1 raises the rate from 1.77 to 2.96.
We also vary over ten orders of magnitude at , with the results shown in Figure 3, left. For , the error grows monotonically by a factor 2.69 between the Neumann and Dirichlet limits. The transition occurs near , which is of order on this mesh. For , the variation is at most 0.1 per cent over the tested range, so the correction removes most of the observed dependence on in this example.
Figure 3. Left: the error at as varies from the Neumann to the Dirichlet limit. The curves are nearly flat for . Right: the interface problem with (solid) and without (dashed) the flux-jump term of (6.7), for three contrasts.
On the non-convex flower, where changes sign along , the energy rates for and are 2.24, 2.24 and 2.23 for . For , the Dirichlet rate falls to 1.22. The corrected method therefore retains its convergence rate despite the changing sign of .
The interface problem and the flux-jump correction
We use the same circular interface with and , and choose the exact solutions and . Both the solution jump and the flux jump are non-zero, as is the tangential flux jump in Remark 6.1.
For perfect bonding, the rate is 1.00 for . For , it increases from 1.40 with to 2.17 with , while for the corresponding rates are 1.36 and 3.32. These results agree with the rule (21), and the case gives similar results on these meshes. At and , we also vary over nine orders of magnitude. With , the error varies by 0.3 per cent, indicating weak dependence on the bonding stiffness in the tested range.
To isolate the flux-jump correction, we run the method with and without the second line of (51). In the latter case, the flux-jump datum is imposed directly on . Both runs use , and perfect bonding, with the results given in Table 2.
| contrast | with | without | ||
| rate | rate | rate | rate | |
| 1 | 2.16 | 3.24 | 1.82 | 1.88 |
| 10 | 2.17 | 3.28 | 1.40 | 1.88 |
| 100 | 2.17 | 3.27 | 1.31 | 1.88 |
Table 2. With and without the flux-jump term of (6.7), , , perfect bonding.
Omitting the flux-jump correction reduces the rate at each contrast. At contrast 1, the rate falls from 2.16 to 1.82, while at contrasts 10 and 100 it falls from 2.17 to 1.40 and 1.31, respectively. The tangential flux jump is , which need not vanish at unit contrast. In this example, it remains non-zero because the jump varies along the interface. Equal coefficients and a constant jump would make the term vanish. The corrected rates are almost unchanged over the three contrasts, supporting the weighted estimate in Theorem 6.2 for these cases. This correction adds two terms on the cut elements.
The symmetric preconditioner.
Corollary 5.2 gives a contraction factor of order when preconditions the corrected form, a prediction we test with in Table 3. The observed factor is obtained from the iteration on the homogeneous problem, starting from a random vector. We also report iteration counts for Richardson iteration and GMRES, both using the same factorisation of and a stopping tolerance of for the relative residual.
| case | observed | iterations (max) Richardson | GMRES | ||
| , | 0.0206 | 0.0041 | 5 | 4 | |
| , | 0.0226 | 0.0069 | 5 | 5 | |
| , | 0.0181 | 0.0037 | 5 | 4 | |
| , | 0.0222 | 0.0073 | 5 | 4 | |
| , | 0.0186 | 0.0041 | 4 | 4 | |
| , | 0.0238 | 0.0044 | 4 | 4 |
Table 3. The uncorrected symmetric form as a preconditioner, : observed contraction factor on the coarsest and finest meshes, the largest iteration count over to a relative residual , and the largest discrepancy with the direct solution of the non-symmetric system.
The contraction factor never exceeds 0.028, and its smallest value on the finest mesh is 0.0037. Over the four meshes, the least-squares rates range from 0.58 to 0.88 across the six cases, consistent with the bound of Corollary 5.2. At the Dirichlet limit, , so the term responsible for the half-order bound in Proposition 5.1 vanishes from (5.20), leaving an perturbation.
No more than five iterations are needed in these tests, and the counts decrease under refinement. The last column compares the iterative and direct solutions of the non-symmetric system, with small discrepancies consistent with the stopping tolerance and finite precision arithmetic. The method therefore uses one symmetric factorisation and applies the geometric correction at each iteration.
Conclusions
We have presented a cut finite element method with boundary value correction for Robin and Neumann conditions, including an extension to interfaces with imperfect bonding. By transporting the solution and its gradient to the exact boundary and evaluating the flux with the exact normal, the method retains a piecewise affine geometry for all integration.
The energy analysis gives a geometric error of uniformly through the Dirichlet limit, but the continuous trace inequality sharpens this to for fixed finite . The distinction comes from the trace residual, which carries the Dirichlet weight, and leads to the sufficient conditions and when . In particular, and give second-order energy convergence for fixed finite , explaining the optimal energy rates in Table 1 under the stated assumptions. The analysis does not establish the larger slopes measured on these meshes, and the corresponding estimate still requires the adjoint-consistency argument discussed in Remark 5.2.
For interface problems, the additional flux-jump correction accounts for the different normals used by integration by parts and by the prescribed jump condition. The numerical examples show a substantial loss of accuracy when this term is omitted, whereas the corrected method gives energy rates close to two for quadratic elements at all three tested contrasts. Uniform stability remains subject to the Poincaré bounds stated in the analysis, with a normalization for any uncoupled constant mode at zero bonding stiffness.
The perturbation estimate also suggests a simple preconditioner based on the symmetric form assembled on the affine geometry. The preconditioned solvers require at most five iterations in the numerical examples, with the geometric correction applied without a separate factorisation.
Funding. EB was supported in part by EPSRC grants EP/X042650/1 and EP/V050400/1. PH was supported by the Swedish Research Council under grant 2022-03908. MGL was supported in part by the Swedish Research Council under grants 2021-04925 and 2025-05562, the Knut and Alice Wallenberg Foundation under grant KAW 2025.0277, and the Swedish Research Programme Essence.
References
- [1]N. Atallah, C. Canuto and G. Scovazzi, The second-generation shifted boundary method and its numerical analysis, Comput. Methods Appl. Mech. Engrg. 372 (2020), 113341.DOI
- [2]S. Bertoluzza, M. Montardini, M. Pennacchio and D. Prada, The virtual element method on polygonal pixel-based tessellations, J. Comput. Phys. 518 (2024), Paper No. 113334.arxiv.org/abs/2206.03449
- [3]S. Bertoluzza, M. Montardini and D. Prada, The high order virtual element method on approximate domains: Neumann boundary conditions, arXiv:2608.07100 [math.NA] (2026).arxiv.org/abs/2608.07100
- [4]S. Bertoluzza, M. Pennacchio and D. Prada, Weakly imposed Dirichlet boundary conditions for 2D and 3D virtual elements, Comput. Methods Appl. Mech. Engrg. 400 (2022), Paper No. 115454.DOI
- [5]J. Boiveau, E. Burman, S. Claus and M. G. Larson, Fictitious domain method with boundary value correction using penalty-free Nitsche method, J. Numer. Math. 26 (2018), no. 2, 64–68.arxiv.org/abs/1610.04482
- [6]J. H. Bramble, T. Dupont and V. Thomée, Projection methods for Dirichlet’s problem in approximating polygonal domains with boundary-value corrections, Math. Comp. 26 (1972), 869–879.DOI
- [7]E. Burman, Ghost penalty, C. R. Math. Acad. Sci. Paris 348 (2010), no. 21–22, 1217–1220.DOI
- [8]E. Burman, J. Guzmán, M. A. Sánchez and M. Sarkis, Robust flux error estimation of an unfitted Nitsche method for high-contrast interface problems, IMA J. Numer. Anal. 38 (2018), no. 2, 646–668.arxiv.org/abs/1602.00603
- [9]E. Burman, P. Hansbo and M. G. Larson, A cut finite element method with boundary value correction, Math. Comp. 87 (2018), no. 310, 633–657.
- [10]E. Burman, P. Hansbo and M. G. Larson, Dirichlet boundary value correction using Lagrange multipliers, BIT Numer. Math. 60 (2020), no. 1, 235–260.arxiv.org/abs/1903.07104
- [11]E. Burman and P. Zunino, A domain decomposition method based on weighted interior penalties for advection-diffusion-reaction problems, SIAM J. Numer. Anal. 44 (2006), no. 4, 1612–1638.DOI
- [12]J. Cheung, M. Perego, P. Bochev and M. Gunzburger, Optimally accurate higher-order finite element methods for polytopial approximations of domains with smooth boundaries, Math. Comp. 88 (2019), no. 319, 2187–2219.arxiv.org/abs/1710.05628
- [13]M. Dryja, On discontinuous Galerkin methods for elliptic problems with discontinuous coefficients, Comput. Methods Appl. Math. 3 (2003), no. 1, 76–85.DOI
- [14]T. Dupont, J. Guzmán and L. R. Scott, Obtaining higher-order Galerkin accuracy when the boundary is polygonally approximated, J. Numer. Math. 33 (2025), no. 1, 55–86.arxiv.org/abs/2001.03082
- [15]A. Hansbo and P. Hansbo, An unfitted finite element method, based on Nitsche’s method, for elliptic interface problems, Comput. Methods Appl. Mech. Engrg. 191 (2002), no. 47–48, 5537–5552.DOI
- [16]A. Hansbo and P. Hansbo, A finite element method for the simulation of strong and weak discontinuities in solid mechanics, Comput. Methods Appl. Mech. Engrg. 193 (2004), no. 33–35, 3523–3540.DOI
- [17]Y. Hou, Y. Liu and T. Zhao, A high-order mixed finite element method for second order elliptic equations on curved domain with boundary value corrections, Int. J. Numer. Anal. Model. 23 (2026), no. 3, 442–459.DOI
- [18]M. Juntunen and R. Stenberg, Nitsche’s method for general boundary conditions, Math. Comp. 78 (2009), no. 267, 1353–1374.
- [19]C. Lehrenfeld, High order unfitted finite element methods on level set domains using isoparametric mappings, Comput. Methods Appl. Mech. Engrg. 300 (2016), 716–733.DOI
- [20]A. Main and G. Scovazzi, The shifted boundary method for embedded domain computations. Part I: Poisson and Stokes problems, J. Comput. Phys. 372 (2018), 972–995.DOI
- [21]A. Main and G. Scovazzi, The shifted boundary method for embedded domain computations. Part II: Linear advection-diffusion and incompressible Navier-Stokes equations, J. Comput. Phys. 372 (2018), 996–1026.DOI
- [22]Z. Zhong and S. A. Meguid, On the elastic field of a spherical inhomogeneity with an imperfectly bonded interface, J. Elasticity 55 (1999), no. 2, 101–140.

