Approximate Mean Curvature Vector

• Approximate Mean Curvature Vector is defined as techniques for recovering the extrinsic curvature vector from indirect representations like point clouds, triangulated meshes, and volumetric discretizations.
• It encompasses diverse approaches including discrete cotangent formulas, kernel-regularized varifold methods, and stabilized finite-element strategies that address normalization and sign ambiguities.
• These methods provide practical insights for geometric flow simulations and curvature-driven evolution in both smooth and nonsmooth, anisotropic, or high-dimensional contexts.

In the literature surveyed here, the expression approximate mean curvature vector denotes a family of constructions for recovering, estimating, or evolving the extrinsic mean curvature vector when the underlying geometry is available only through local boundary data, triangulated meshes, point clouds, volumetric discretizations, smoothings of nonsmooth sets, or other indirect representations. The common target is the vector quantity that, in smooth Euclidean geometry, is normal to the surface and is represented by conventions such as $\mathbf{N}B_\alpha^{\ \alpha}$, $H\nu$, $2H\nu$, or $-\Delta_\Sigma \mathbf{x}$, depending on normalization and sign. Taken together, the cited works indicate that no single standardized definition dominates across all settings: some constructions are exact in the smooth limit, some are discrete analogues on meshes, some are kernel-regularized varifold quantities, and some papers use only scalar or indirectly reconstructed vector surrogates (Grinfeld, 2011).

1. Terminological scope and competing meanings

The phrase does not denote one universally fixed object. In the varifold framework, an approximate mean curvature vector is defined by regularizing both the first variation and the mass, then taking a normalized ratio; this makes sense for arbitrary varifolds, including point clouds and volumetric discretizations, and converges to generalized mean curvature under appropriate assumptions (Buet et al., 2016). In triangulated-surface papers, the same phrase often refers instead to a vertex-based discrete vector built from opposite-edge normals, cotangent weights, or stabilized finite-element solves on piecewise linear meshes (Das et al., 2019). In smoothing constructions for convex polytopes, by contrast, the emphasis is overwhelmingly on the scalar mean curvature of smooth approximating hypersurfaces $Y_\varepsilon$, with the corresponding vector interpretation left heuristic rather than formalized as weak convergence of vector-valued measures (2207.13346).

Other works depart even further from a literal vector-valued meaning. In the crystalline-anisotropic setting, the relevant generalized curvature is the scalar divergence of a Cahn–Hoffman field, canonically selected as the minimal $L^2$-divergence; the vector object is the Cahn–Hoffman field itself, not a Euclidean mean curvature vector in the standard sense (Giga et al., 2017). In high-dimensional data analysis, the computed quantity may be only a nonnegative scalar score $$\mathcal K_i=\left|\operatorname{tr}(\widehat{\mathcal S}_i)\right|$$, explicitly not a full signed differential-geometric vector $\vec H$ (Levada, 4 Jun 2026). A persistent source of confusion is therefore terminological: some papers approximate the full vector, some approximate a scalar that determines the vector once a normal is known, and some approximate only a magnitude-like surrogate.

A second ambiguity is convention. For smooth surfaces, the same geometric object appears as $\mathbf{N}B_\alpha^{\ \alpha}$, $H\nu$, $H\nu$0, or $H\nu$1, depending on whether $H\nu$2 is the sum or average of principal curvatures and on the sign chosen for the normal and Laplace–Beltrami operator (Grinfeld, 2011). This suggests that vector-valued formulations are often preferred because they avoid the scalar sign and factor-of-two ambiguities that recur across geometric analysis and discrete differential geometry.

2. Smooth geometric prototypes and exact limiting identities

A particularly compact smooth prototype is the contour identity for a regular patch $H\nu$3 with smooth boundary $H\nu$4: $H\nu$5 Here $H\nu$6 is the unit surface normal, $H\nu$7 is the outward unit normal to the boundary curve within the surface, and $H\nu$8 is the trace of the second fundamental form. Shrinking the patch gives the pointwise formula

$H\nu$9

The identity is exact for smooth patches, and the pointwise value is exact as a limit. Its derivation is a surface-divergence argument based on

$2H\nu$0

with $2H\nu$1 (Grinfeld, 2011).

This formula is important because it expresses an extrinsic curvature vector through boundary data alone. The paper also emphasizes that $2H\nu$2 is more fundamental than separating normal and scalar curvature, precisely because conventions differ. Under the convention $2H\nu$3 with $2H\nu$4, one has $2H\nu$5 up to sign. Under Laplace–Beltrami conventions such as $2H\nu$6, the same contour average recovers the corresponding signed vector (Grinfeld, 2011).

The contour identity also yields a boundary characterization of minimality: $2H\nu$7 for any contour $2H\nu$8 lying in a minimal surface, since $2H\nu$9 there (Grinfeld, 2011). A plausible implication is that many discrete formulas based on summing local in-surface normals can be read as direct descendants of this exact smooth statement rather than as ad hoc estimators.

3. Triangulated surfaces, cotangent formulas, and stabilized finite elements

On triangulated meshes, one widely used approximation is the vertex formula

$-\Delta_\Sigma \mathbf{x}$0

where the sum is over triangles $-\Delta_\Sigma \mathbf{x}$1 incident on vertex $-\Delta_\Sigma \mathbf{x}$2, $-\Delta_\Sigma \mathbf{x}$3 is the length of the edge opposite $-\Delta_\Sigma \mathbf{x}$4, $-\Delta_\Sigma \mathbf{x}$5 is the outward edge-perpendicular unit vector in the triangle plane, and $-\Delta_\Sigma \mathbf{x}$6 is the modified Voronoi area. The same paper shows that this physics-based force-balance derivation is algebraically identical to the cotangent Laplace–Beltrami formula

$-\Delta_\Sigma \mathbf{x}$7

and identifies the computed quantity as $-\Delta_\Sigma \mathbf{x}$8, not $-\Delta_\Sigma \mathbf{x}$9, under its conventions (Das et al., 2019).

A more geometric one-ring formula appears in the contour-based note: if triangles $Y_\varepsilon$0 meet at a vertex $Y_\varepsilon$1, with opposite edge lengths $Y_\varepsilon$2, triangle areas $Y_\varepsilon$3, and in-triangle outward normals $Y_\varepsilon$4 perpendicular to those opposite edges, then

$Y_\varepsilon$5

is proposed as the discrete analogue of $Y_\varepsilon$6. The paper stresses that $Y_\varepsilon$7 should be interpreted as a discrete vector mean curvature, not as a normal estimator, because it vanishes on minimal or nearly minimal surfaces and is then directionally unreliable. It also states that $Y_\varepsilon$8 is the gradient of total area with respect to the vertex position, so $Y_\varepsilon$9 for minimal triangulated surfaces (Grinfeld, 2011).

For piecewise linear triangulated surfaces, several finite-element papers replace local formulas by global variational solves. One stabilized formulation defines $L^2$0 by

$L^2$1

with an edge-jump stabilization

$L^2$2

The key claim is that the unstabilized discrete Laplace–Beltrami approximation generally cannot be expected to converge in $L^2$3 on piecewise linear geometry, while the stabilized method yields

$L^2$4

for smooth closed surfaces (Cenanovic et al., 2017). The related closed-surface and cut-surface formulation

$L^2$5

proves the same first-order $L^2$6 estimate and, in the cut case, shows that face stabilization alone is sufficient (Hansbo et al., 2014).

These mesh-based approaches share a common structure: the approximate mean curvature vector is either a local sum of geometric edge contributions or a weak finite-element representation of the surface Laplacian of the embedding. This suggests that the principal divide inside discrete geometry is not between “local” and “global” methods, but between raw discretizations and structure-preserving ones that enforce area variation or stabilized weak consistency.

4. Kernel-regularized varifold approximations

In the varifold setting, the approximate mean curvature vector is defined for any $L^2$7-varifold $L^2$8 by regularizing the first variation and the mass with radial kernels $L^2$9. With

$$\mathcal K_i=\left|\operatorname{tr}(\widehat{\mathcal S}_i)\right|$$0

the central formula is

$$\mathcal K_i=\left|\operatorname{tr}(\widehat{\mathcal S}_i)\right|$$1

whenever $$\mathcal K_i=\left|\operatorname{tr}(\widehat{\mathcal S}_i)\right|$$2. For rectifiable varifolds with locally bounded first variation,

$$\mathcal K_i=\left|\operatorname{tr}(\widehat{\mathcal S}_i)\right|$$3

and in the smooth $$\mathcal K_i=\left|\operatorname{tr}(\widehat{\mathcal S}_i)\right|$$4 manifold case this limit is the classical mean curvature vector (Buet et al., 2016).

For point-cloud varifolds

$$\mathcal K_i=\left|\operatorname{tr}(\widehat{\mathcal S}_i)\right|$$5

the same construction becomes an explicit neighbor sum: $$\mathcal K_i=\left|\operatorname{tr}(\widehat{\mathcal S}_i)\right|$$6 Because tangential artifacts can be induced by nonuniform sampling, the paper also defines the orthogonal approximate mean curvature by projecting onto the normal component. In the smooth-manifold case, this orthogonal version enjoys an improved error estimate of order

$$\mathcal K_i=\left|\operatorname{tr}(\widehat{\mathcal S}_i)\right|$$7

under the stated approximation assumptions (Buet et al., 2016).

A later generalization enlarges the class of admissible linear operators $$\mathcal K_i=\left|\operatorname{tr}(\widehat{\mathcal S}_i)\right|$$8 in

$$\mathcal K_i=\left|\operatorname{tr}(\widehat{\mathcal S}_i)\right|$$9

Beyond the Buet–Rumpf choices, the paper proves that

$\vec H$0

while

$\vec H$1

and it extends the convergence theory from $\vec H$2 hypersurfaces to $\vec H$3-integral varifolds with unit density, $\vec H$4 for $\vec H$5, and $\vec H$6 (Sagueni, 8 Sep 2025). This operator viewpoint suggests that approximate mean curvature can be understood not only as a kernel formula, but as a linear algebra of admissible ambient projections acting on the same regularized first-variation template.

5. Approximate curvature vectors in geometric flows

Several papers use approximate mean curvature vectors not merely as static estimators but as velocities in geometric evolution. One general-varifold construction fixes $\vec H$7 and defines

$\vec H$8

The field $\vec H$9 is then inserted into the explicit Euler push-forward map

$\mathbf{N}B_\alpha^{\ \alpha}$0

and iterated to produce a time-discrete approximate mean curvature flow for arbitrary finite-mass varifolds, including point-cloud varifolds. For fixed $\mathbf{N}B_\alpha^{\ \alpha}$1, the time-discrete flows converge as $\mathbf{N}B_\alpha^{\ \alpha}$2 to a unique $\mathbf{N}B_\alpha^{\ \alpha}$3-approximate flow satisfying an exact Brakke-type identity and the mass dissipation formula

$\mathbf{N}B_\alpha^{\ \alpha}$4

(Buet et al., 1 Oct 2025).

A closely related volumetric-varifold paper studies the approximate mean curvature

$\mathbf{N}B_\alpha^{\ \alpha}$5

for cellwise constant-in-orientation volumetric discretizations

$\mathbf{N}B_\alpha^{\ \alpha}$6

Its main result is an integral Brakke approximate equality with consistency defect

$\mathbf{N}B_\alpha^{\ \alpha}$7

showing that the regularized curvature vector can replace the true mean curvature in a discretized flow law while retaining the correct dissipation structure up to explicit errors (Sagueni, 8 Sep 2025).

For surfaces with fixed boundary, the approximation is indirect. The semidiscrete scheme evolves a parametrized surface by

$\mathbf{N}B_\alpha^{\ \alpha}$8

where $\mathbf{N}B_\alpha^{\ \alpha}$9 and $H\nu$0 are separate discrete approximations of scalar mean curvature and unit normal. The continuous problem imposes the boundary conditions

$H\nu$1

together with a conormal derivative identity for $H\nu$2. The paper proves optimal $H\nu$3-error estimates of order $H\nu$4 for $H\nu$5 for $H\nu$6, $H\nu$7, $H\nu$8, and $H\nu$9, so $H\nu$00 inherits the same order as an approximation of the mean curvature vector (Ivaniszyn et al., 25 Apr 2025).

For planar curves, an Onsager-principle discretization uses the polygonal length $H\nu$01 and dissipation $H\nu$02 to derive the nodal ODE

$H\nu$03

Here $H\nu$04 is the discrete first variation of length and $H\nu$05 is the discrete $H\nu$06-gradient-flow velocity. The semidiscrete system preserves the exact discrete dissipation law

$H\nu$07

which identifies $H\nu$08 as the paper’s discrete curvature-vector force (Liu et al., 2024).

6. Singular, polyhedral, anisotropic, and high-dimensional variants

Approximate mean curvature becomes more heterogeneous when the underlying geometry is nonsmooth or the ambient setting is nonclassical. For convex polytopes and mean convex manifolds with corners, smooth hypersurfaces $H\nu$09 are constructed so that scalar mean curvature remains positive and the weighted boundary metric

$H\nu$10

recovers the polyhedral angular metric in the limit. The sharp local statement is that mean curvature on the smoothing graph blows up like $H\nu$11 near codimension-two faces. The paper explicitly does not prove weak convergence of vector-valued mean curvature measures, but it strongly suggests a picture in which curvature concentrates near the $H\nu$12-skeleton of the polytope (2207.13346).

In crystalline anisotropy, the approximation target is scalar from the outset. A facet is regularized not by assigning it a Euclidean mean curvature vector, but by producing a Cahn–Hoffman vector field $H\nu$13 with $H\nu$14, subordinate to the anisotropy through $H\nu$15. The canonical generalized curvature is then

$H\nu$16

the minimal $H\nu$17-divergence among admissible Cahn–Hoffman fields. Every bounded facet can be approximated by one admitting such a field, which makes generalized crystalline mean curvature available by approximation, but this remains a scalar theory rather than a Euclidean vector theory (Giga et al., 2017).

At the opposite end of the spectrum, high-dimensional data-manifold methods may discard direction entirely. One recent estimator computes

$H\nu$18

from a $H\nu$19-nearest-neighbor covariance eigensystem, then reduces the cost from $H\nu$20 to $H\nu$21 by an exact algebraic identity and further to

$H\nu$22

by truncated SVD and a Haar-expectation approximation of null-space contributions. The paper explicitly states that this is not a full signed differential-geometric mean-curvature vector, but a scalar nonnegative local mean-curvature score (Levada, 4 Jun 2026).

A final misconception concerns normals. Because several discrete constructions actually approximate the product “curvature times normal,” not the normal alone, they become directionally unstable near minimality. The triangulated contour analogue makes this explicit: if the mean curvature vector is small, the construction cannot robustly determine a normal direction (Grinfeld, 2011). This suggests that in both classical and data-driven settings, vector-valued approximations are strongest when interpreted as curvature forces or first-variation densities, not as normal estimators detached from curvature magnitude.