Tangent-Bundle Penalty: Elasticity & Surface FEM

• Tangent-bundle penalty is a soft geometric constraint that penalizes deviations from a desired bundle structure, ensuring near-injectivity in elasticity and almost tangentiality in surface finite elements.
• In nonlinear elasticity, it approximates the Ciarlet–Nečas condition with a double-integral form to prevent self-interpenetration and to enforce global injectivity in deformations.
• In surface finite elements, a mesh-dependent penalty on the normal components weakly enforces tangentiality while preserving optimal convergence rates in discrete approximations.

Searching arXiv for the cited papers and closely related terminology to ground the article in current arXiv records. In the cited arXiv literature, a tangent-bundle penalty denotes a soft geometric constraint that supplements a variational formulation by penalizing configurations that violate an intended bundle structure. Two distinct constructions appear under this label. In second-gradient nonlinear elasticity, Krömer and Valdman introduce a double-integral penalty that approximates the Ciarlet–Nečas condition and acts as a measure of self-penetration for hyperelastic deformations (Krömer et al., 2018). In tensor surface finite elements, Hardering and Praetorius introduce a mesh-dependent penalty on the normal component of discrete fields so as to enforce almost tangentiality on a discrete surface while retaining optimal tangential convergence behavior (Hardering et al., 2021). The term therefore identifies a geometric penalization paradigm rather than a single universal formula.

1. Scope and conceptual role

The two cited constructions share a common structural idea: admissibility is not imposed by an exact constraint at the discrete or penalized level, but by an additional energy term that vanishes, or becomes negligible, only when the relevant geometric inconsistency is absent.

In nonlinear elasticity, the inconsistency is global non-injectivity. The base model consists of the standard hyperelastic energy

$E^{el}(y)=\int_{\Omega}W(x,\nabla y(x))\,dx$

subject to the hard constraint

$\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$

that is, the Ciarlet–Nečas condition. The tangent-bundle penalty replaces this hard global condition by a soft term built from a double integral over $\Omega\times\Omega$, while a second-gradient regularization supplies additional smoothness and a local invertibility radius (Krömer et al., 2018).

In the surface finite-element setting, the inconsistency is the presence of non-tangential components in an ambient $\mathbb{R}^{d+1}$ representation of a tangential tensor field. The discrete bilinear form is therefore augmented by a penalty on $Q_hu_h$, where $Q_h$ is the orthogonal complement of the discrete tangential projector $P_h=\mathrm{Id}-n_h\otimes n_h$. This weakly pushes the solution toward tangentiality without imposing a pointwise hard constraint on the discrete space (Hardering et al., 2021).

2. Nonlinear-elasticity formulation

The nonlinear-elasticity construction of "Global injectivity in second-gradient Nonlinear Elasticity and its approximation with penalty terms" (Krömer et al., 2018) separates the total energy into three contributions: $$E_{(\varepsilon_1,\varepsilon_2),\sigma}(y) = E^{el}_{\varepsilon_1}(y) + E^{CN}_{\varepsilon_2}(y) + E^{reg}_{\sigma}(y).$$ Here

$$E^{reg}_{\sigma}(y)=\sigma\int_{\Omega}|D^2y(x)|^s\,dx, \qquad \sigma>0,\quad s>d,$$

and $E^{el}_{\varepsilon_1}$ is based on an everywhere-finite, polyconvex approximation $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$0. The approximation is chosen so that, for suitable $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$1 and $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$2,

$\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$3

As $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$4, this recovers the determinant blow-up that enforces $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$5.

The tangent-bundle penalty itself is defined by fixing a strictly increasing $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$6 with $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$7, for instance $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$8 or $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$9, and an exponent $\Omega\times\Omega$0, and setting

$\Omega\times\Omega$1

where $\Omega\times\Omega$2 and $\Omega\times\Omega$3.

Its interpretation is local in pairwise distance but global in effect. If two reference points $\Omega\times\Omega$4 are a distance $\Omega\times\Omega$5 apart, then $\Omega\times\Omega$6 records their reference-space separation. If under deformation the points become closer than $\Omega\times\Omega$7, then $\Omega\times\Omega$8 is too small, and the positive part produces a penalty. The prefactor $\Omega\times\Omega$9 is designed so that the total contribution blows up unless the deformation stays globally injective down to $\mathbb{R}^{d+1}$0 scales. The second-gradient term with $\mathbb{R}^{d+1}$1 forces $\mathbb{R}^{d+1}$2, $\mathbb{R}^{d+1}$3, and yields a uniform local invertibility radius $\mathbb{R}^{d+1}$4 (Krömer et al., 2018).

3. Convergence and finite-$\mathbb{R}^{d+1}$5 injectivity

For fixed $\mathbb{R}^{d+1}$6, the penalized functionals in the elasticity setting $\mathbb{R}^{d+1}$7-converge, in the weak $\mathbb{R}^{d+1}$8-topology, to the original second-gradient hyperelastic energy with exact Ciarlet–Nečas admissibility (Krömer et al., 2018). The limit functional is

$\mathbb{R}^{d+1}$9

The convergence statement has the standard two parts: if $Q_hu_h$0 in $Q_hu_h$1, then

$Q_hu_h$2

and for every admissible $Q_hu_h$3 there exists a recovery sequence $Q_hu_h$4 strongly in $Q_hu_h$5 such that

$Q_hu_h$6

A distinctive feature of this penalty is that it is not merely asymptotic. If one chooses $Q_hu_h$7, then a finite bound

$Q_hu_h$8

already forces true global injectivity, provided $Q_hu_h$9 is sufficiently small in terms of $Q_h$0 and the a priori local-invertibility radius $Q_h$1. The paper states the lower bound

$Q_h$2

where $Q_h$3 is the multiplicity function. Consequently,

$Q_h$4

so that any low-energy state with bounded penalty is already self-contact-free for sufficiently small $Q_h$5. This is the point at which the soft constraint becomes, for low-energy competitors, an exact exclusion of interpenetration.

4. Surface finite elements and weak tangentiality enforcement

In "Tangential Errors of Tensor Surface Finite Elements" (Hardering et al., 2021), the tangent-bundle penalization idea is used in a different geometric setting. The continuous problem is: find $Q_h$6 such that

$Q_h$7

with tangential projection $Q_h$8.

The discrete unknown $Q_h$9 lives in a surface-finite element space $P_h=\mathrm{Id}-n_h\otimes n_h$0 on a discrete surface $P_h=\mathrm{Id}-n_h\otimes n_h$1 of geometric order $P_h=\mathrm{Id}-n_h\otimes n_h$2, and the bilinear form is

$P_h=\mathrm{Id}-n_h\otimes n_h$3

where

$P_h=\mathrm{Id}-n_h\otimes n_h$4

and

$P_h=\mathrm{Id}-n_h\otimes n_h$5

Here $P_h=\mathrm{Id}-n_h\otimes n_h$6 is the element-wise tangential projector, $P_h=\mathrm{Id}-n_h\otimes n_h$7 its orthogonal complement, $P_h=\mathrm{Id}-n_h\otimes n_h$8 a penalty pre-factor, $P_h=\mathrm{Id}-n_h\otimes n_h$9 the scaling exponent, and $$E_{(\varepsilon_1,\varepsilon_2),\sigma}(y) = E^{el}_{\varepsilon_1}(y) + E^{CN}_{\varepsilon_2}(y) + E^{reg}_{\sigma}(y).$$0 the mesh-size.

The motivation is coercivity. If one merely replaces $$E_{(\varepsilon_1,\varepsilon_2),\sigma}(y) = E^{el}_{\varepsilon_1}(y) + E^{CN}_{\varepsilon_2}(y) + E^{reg}_{\sigma}(y).$$1 by $$E_{(\varepsilon_1,\varepsilon_2),\sigma}(y) = E^{el}_{\varepsilon_1}(y) + E^{CN}_{\varepsilon_2}(y) + E^{reg}_{\sigma}(y).$$2 in the discrete form, then any normal field $$E_{(\varepsilon_1,\varepsilon_2),\sigma}(y) = E^{el}_{\varepsilon_1}(y) + E^{CN}_{\varepsilon_2}(y) + E^{reg}_{\sigma}(y).$$3 lies in the kernel of $$E_{(\varepsilon_1,\varepsilon_2),\sigma}(y) = E^{el}_{\varepsilon_1}(y) + E^{CN}_{\varepsilon_2}(y) + E^{reg}_{\sigma}(y).$$4, so uniqueness is lost. Penalizing the normal component restores stability in the ambient $$E_{(\varepsilon_1,\varepsilon_2),\sigma}(y) = E^{el}_{\varepsilon_1}(y) + E^{CN}_{\varepsilon_2}(y) + E^{reg}_{\sigma}(y).$$5 representation. The penalty acts only weakly: it pushes $$E_{(\varepsilon_1,\varepsilon_2),\sigma}(y) = E^{el}_{\varepsilon_1}(y) + E^{CN}_{\varepsilon_2}(y) + E^{reg}_{\sigma}(y).$$6 to zero as $$E_{(\varepsilon_1,\varepsilon_2),\sigma}(y) = E^{el}_{\varepsilon_1}(y) + E^{CN}_{\varepsilon_2}(y) + E^{reg}_{\sigma}(y).$$7, but it does not enforce exact tangentiality at the discrete level. The analysis interprets $$E_{(\varepsilon_1,\varepsilon_2),\sigma}(y) = E^{el}_{\varepsilon_1}(y) + E^{CN}_{\varepsilon_2}(y) + E^{reg}_{\sigma}(y).$$8 in the penalty energy norm, and this scaling is central to the subsequent error estimates.

5. Geometric assumptions and a priori error structure

The surface-finite element analysis assumes that $$E_{(\varepsilon_1,\varepsilon_2),\sigma}(y) = E^{el}_{\varepsilon_1}(y) + E^{CN}_{\varepsilon_2}(y) + E^{reg}_{\sigma}(y).$$9 is a smooth, closed, orientable hypersurface in $$E^{reg}_{\sigma}(y)=\sigma\int_{\Omega}|D^2y(x)|^s\,dx, \qquad \sigma>0,\quad s>d,$$0, with $$E^{reg}_{\sigma}(y)=\sigma\int_{\Omega}|D^2y(x)|^s\,dx, \qquad \sigma>0,\quad s>d,$$1, and that a triangulated reference mesh $$E^{reg}_{\sigma}(y)=\sigma\int_{\Omega}|D^2y(x)|^s\,dx, \qquad \sigma>0,\quad s>d,$$2 is lifted by a degree-$$E^{reg}_{\sigma}(y)=\sigma\int_{\Omega}|D^2y(x)|^s\,dx, \qquad \sigma>0,\quad s>d,$$3 interpolation of the closest point map to obtain $$E^{reg}_{\sigma}(y)=\sigma\int_{\Omega}|D^2y(x)|^s\,dx, \qquad \sigma>0,\quad s>d,$$4 of geometric order $$E^{reg}_{\sigma}(y)=\sigma\int_{\Omega}|D^2y(x)|^s\,dx, \qquad \sigma>0,\quad s>d,$$5 (Hardering et al., 2021). The resulting geometric approximation satisfies

$$E^{reg}_{\sigma}(y)=\sigma\int_{\Omega}|D^2y(x)|^s\,dx, \qquad \sigma>0,\quad s>d,$$6

The normal entering $$E^{reg}_{\sigma}(y)=\sigma\int_{\Omega}|D^2y(x)|^s\,dx, \qquad \sigma>0,\quad s>d,$$7 may be chosen more accurately than $$E^{reg}_{\sigma}(y)=\sigma\int_{\Omega}|D^2y(x)|^s\,dx, \qquad \sigma>0,\quad s>d,$$8, by requiring

$$E^{reg}_{\sigma}(y)=\sigma\int_{\Omega}|D^2y(x)|^s\,dx, \qquad \sigma>0,\quad s>d,$$9

The discrete tensor-valued space is

$E^{el}_{\varepsilon_1}$0

The error analysis separates tangential and normal parts. For the energy norm, Theorem 4.16 uses

$E^{el}_{\varepsilon_1}$1

and yields

$E^{el}_{\varepsilon_1}$2

For the tangential $E^{el}_{\varepsilon_1}$3-error, Theorem 4.20 introduces

$E^{el}_{\varepsilon_1}$4

and proves

$E^{el}_{\varepsilon_1}$5

The $E^{el}_{\varepsilon_1}$6-analysis likewise splits normal and tangential parts. In particular, for $E^{el}_{\varepsilon_1}$7, the normal component satisfies an estimate of order $E^{el}_{\varepsilon_1}$8, and the tangential component is controlled with $E^{el}_{\varepsilon_1}$9. In the isogeometric case $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$00, the choice $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$01 gives

$\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$02

hence the final rates are the optimal $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$03 in $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$04 and $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$05 in $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$06.

6. Parameter regimes, numerical behavior, and interpretive issues

The two tangent-bundle penalties differ in purpose, scaling, and admissibility mechanism, but both are designed so that the penalty parameter controls a geometric defect rather than merely adding generic regularization.

In the elasticity setting, the exponent $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$07 governs how strongly near-self-contact is amplified. The paper emphasizes that $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$08 is the regime in which bounded penalized energy implies exact injectivity for sufficiently small $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$09 (Krömer et al., 2018). The same work compares this construction with earlier soft-constraint approaches. Miehe–Roubíček used the penalty $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$10, whose numerical implementation is described as awkward and less localized. Ball–Reisner used auxiliary fields for 1D beams, whereas in higher dimensions their method is harder to localize. By contrast, the double-integral form is said to be conveniently discretized because only small auras of width $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$11 around near-contacts actually contribute, it can be built into standard finite-element codes, and it can guarantee exact injectivity on the discrete level if $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$12. The stated practical consequences are that one can dial $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$13, no constraint-enforcing nonlinear solver is needed, and numerical tests in 2D pincers and beams show that as $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$14 the interpenetration gap closes and global injectivity is recovered in the limit.

In the surface-finite element setting, the central tuning parameter is $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$15. If $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$16, the damping is mass-like; if $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$17, it is $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$18-like. The analysis identifies $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$19 as the unique choice giving symmetric convergence rates in the two geometric penalty contributions, and therefore optimal tangential rates even when $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$20 (Hardering et al., 2021). Numerically, the paper reports choosing $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$21–$\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$22 with $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$23 in a genuine isogeometric setup, while $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$24 may be used if a higher-order normal $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$25 is available and the normal part is also of interest. The numerical tests cover vector fields on an ellipsoid and tensor fields on ellipses and the sphere. In the isogeometric case $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$26, $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$27 yields clean slopes of order $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$28 in $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$29 and $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$30 in $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$31, while deviations from $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$32 reduce the tangential $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$33-error slope to $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$34.

A recurrent misconception is to treat tangent-bundle penalties as exact constraint imposition under different notation. The cited works do not support that reading. In elasticity, the term is explicitly a soft approximation to the Ciarlet–Nečas condition, although for $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$35 and sufficiently small $\int_{\Omega}\det\nabla y\,dx=|y(\Omega)|,$36, low-energy states become exactly self-contact-free. In surface finite elements, the penalty ensures almost tangentiality and eliminates kernel modes, but the design goal is not exact discrete tangentiality; it is the preservation of optimal convergence in tangential quantities. This suggests that the decisive feature of tangent-bundle penalization is not hardness of enforcement, but geometric selectivity: only the bundle-incompatible component of a configuration is targeted.