Source-linked AI summary
The Vector Heat Method
Nicholas Sharp, Yousuf Soliman, Keenan Crane
TL;DR
Computing a most-parallel vector field over a curved domain is useful for geometric algorithms, but explicit geodesic paths are costly to construct. The paper uses short-time connection-Laplacian heat flow to compute transport along shortest geodesics, and presents an efficient method with applications including whole-surface logarithmic maps and surface means. Its numerical behavior and limitations depend on solver scalability and mesh quality.
Problem
Extending a vector from a source to its most-parallel values elsewhere is useful for geometric computing, yet practical algorithms for accurate Karcher means on general surfaces were lacking.
Source: §1 · p. 1, §8.3 · p. 14Method
The method relates short-time vector heat flow to parallel transport along minimal geodesics and reduces computation to sparse linear systems.
Source: §1 · p. 1Results
The paper presents a first method for efficiently and reliably computing fields formed by parallel transport along shortest geodesics.
Source: §10 · p. 17Takeaways & Limitations
The method supports whole-surface logarithmic maps, Karcher means, geometric medians, velocity extrapolation, and intrinsic landmarks.
Source: §1 · p. 1Solver scalability and numerical accuracy constrain use; poor mesh quality can produce indefinite matrices and spurious flipped vectors, while too few elements can limit accuracy.
Source: §9 · p. 16
Abstract
from arXiv · showhide
This paper describes a method for efficiently computing parallel transport of tangent vectors on curved surfaces, or more generally, any vector-valued data on a curved manifold. More precisely, it extends a vector field defined over any region to the rest of the domain via parallel transport along shortest geodesics. This basic operation enables fast, robust algorithms for extrapolating level set velocities, inverting the exponential map, computing geometric medians and Karcher/Fr\'{e}chet means of arbitrary distributions, constructing centroidal Voronoi diagrams, and finding consistently ordered landmarks. Rather than evaluate parallel transport by explicitly tracing geodesics, we show that it can be computed via a short-time heat flow involving the connection Laplacian. As a result, transport can be achieved by solving three prefactored linear systems, each akin to a standard Poisson problem. To implement the method we need only a discrete connection Laplacian, which we describe for a variety of geometric data structures (point clouds, polygon meshes, etc). We also study the numerical behavior of our method, showing empirically that it converges under refinement, and augment the construction of intrinsic Delaunay triangulations (iDT) so that they can be used in the context of tangent vector field processing.
1 INTRODUCTION
The paper asks how to extend a vector at one point to a most-parallel field elsewhere on a curved domain. It uses short-time vector heat flow to compute transport along shortest geodesics, then applies the method to geometric algorithms and varied vector data.
1 INTRODUCTION
- Short-time vector heat flow relates diffusion to parallel transport along minimal geodesics, avoiding explicit path construction.The same principle applies to point sources and vector fields over curves or other domain subsets.Source: §1 · p. 1
- The approach reduces transport to sparse linear systems that can use mature high-performance solvers.Discrete Laplacians support use on polygon meshes and point clouds, and the strategy generalizes to vector data such as symmetric direction fields and differential forms.Source: §1 · p. 1
- The paper introduces a fast source-set transport method, an augmented intrinsic Delaunay algorithm, and methods for whole-surface logarithmic maps, Karcher means, and geometric medians.The logarithmic map, Karcher means, and geometric medians are described as first methods of their kind for the stated general-surface settings.Source: §1 · p. 1
- The paper also describes discretizations for multiple geometric structures and applies the method to velocity extrapolation, geodesic centroidal Voronoi tessellations, and ordered intrinsic landmarks.Source: §1 · p. 1
2 RELATED WORK
Prior approaches include discrete transport constructions and explicit geodesic tracing, while the vector heat method instead computes transport using linear diffusion equations. The paper distinguishes this task from scalar distance computation and connects it to connection-Laplacian methods and global geometric applications.
2 RELATED WORK
- Discrete transport precedents include Schild’s ladder, Regge’s simplicial calculus, and discrete connections; this method instead places vectors at mesh vertices.The cited work also includes spectral transport of one field along another, rather than transport along shortest geodesics.Source: §2 · p. 2
- Explicit geodesic approaches require expensive path computation or tracing, while the vector heat method builds Laplace matrices and solves sparse systems.The paper estimates path-based computation at O(n^1.5) versus sparse diffusion solving with complexity approaching O(n).Source: §2 · p. 2
- 15x slower: tracing geodesics to a source takes longer than applying the vector heat method, even without counting distance computation.The comparison reports this timing for the illustrated case; the related-work discussion also describes path tracing as more than an order of magnitude slower in practice.Source: Figure 2 · p. 2, §2 · p. 2
- The scalar heat method computes geodesic distance, whereas the vector heat method computes parallel transport along shortest geodesics.The vector version also needs no discrete divergence or gradient operators, which makes it easier to apply to point clouds and potentially general graphs.Source: §2 · p. 2
- Connection Laplacians account for parallel transport during diffusion on curved domains, and related work applies vector diffusion to data analysis and learning.The paper builds primarily on a discrete exterior-calculus connection Laplacian for triangle meshes.Source: §2 · p. 2
- Global logarithmic maps support Karcher means and geometric medians on arbitrary surfaces, where prior approaches lacked practical accurate methods.The paper also contrasts its fixed matrix-factorization traversal with Dijkstra-like methods that use source-dependent branches and memory access patterns.Source: §2 · p. 2
3 PRELIMINARIES
The preliminaries establish how heat diffusion and parallel transport extend from scalar fields to vectors on curved domains. The connection Laplacian makes vector diffusion geometrically meaningful, and its short-time behavior encodes transport along shortest geodesics.
3.4 Connection Laplacian
- On curved domains, componentwise scalar diffusion does not account for the lack of a global coordinate system; vector diffusion instead uses the connection Laplacian.The connection Laplacian is a negative semidefinite, self-adjoint, elliptic operator on vector fields.Source: §3 · p. 3, §3.4 · pp. 3–4
3.1 Notation
- The notation defines geodesic distance as shortest-path length and the cut locus as points without a unique closest point in a subset.It also uses X|p for the vector at p and δx for the Dirac delta centered at x.Source: §3.1 · p. 3
3.2 Heat Diffusion
- The scalar heat equation diffuses an initial heat distribution over time; for a point source, its heat kernel is the fundamental solution, a Gaussian in Euclidean space.Convolving initial data with the heat kernel yields the solution at time t.Source: §3.2 · p. 3
3.3 Parallel Transport and Connections
- Parallel transport depends on the chosen path, and its infinitesimal variation defines the covariant derivative associated with the Levi-Civita connection.The covariant derivative describes how a vector field changes when traveling in a specified tangent direction.Source: §3.3 · p. 3
3.4 Connection Laplacian
- As time approaches zero, the vector heat kernel behaves like parallel transport along shortest geodesics, with magnitude decay matching the scalar heat kernel.This short-time behavior is the relationship the method uses to compute parallel transport through diffusion.Source: §3.4 · p. 4
4 SMOOTH FORMULATION
The smooth formulation uses short-time vector diffusion to approximate parallel transport from a source set, while scalar diffusion supplies the transported magnitudes. Its interpolation construction extends to varied source sets, and separating magnitude from direction avoids cancellation errors in the direct quotient.
4 SMOOTH FORMULATION
- Directly quotienting vector and scalar diffusion can incur cancellation errors, so the method obtains direction from vector diffusion and magnitude from scalar diffusion.The authors report that this separation provides numerical robustness even when direction has small errors.Source: §4 · p. 4, §4.2 · p. 5
- The smooth formulation applies beyond two-dimensional surfaces and tangent bundles to vector bundles over Riemannian manifolds of any dimension.Source: §4 · p. 4
4.1 Scalar Interpolation
- As Gaussian width tends to zero, weighted averaging gives the value at the closest source; on a manifold, heat-kernel weights extend this interpolation to points, curves, and regions.At points on the cut locus, the interpolated value approaches an average of the closest-source values.Source: §4.1 · p. 5
- The interpolation construction accommodates source sets made of points, curves, and regions on a surface.The corresponding initial conditions are Dirac-type measures concentrated on the source set.Source: §4.1 · p. 5
4.2 Vector Heat Method
- The algorithm diffuses the vector field and its magnitude with heat flows, diffuses the source-set indicator, then combines the results to form the extended field.The vector flow supplies direction; the scalar flows support magnitude interpolation and normalization.Source: §4.1 · p. 5, §4.2 · p. 5
- For small t, vector diffusion gives directions close to those of the closest source vectors, and rescaling by interpolated source magnitudes yields the parallel-transport approximation.The stated approximation applies at points outside the cut locus of the source set.Source: §4.2 · p. 5
- At cut-locus points, the output approaches an average of all closest vectors rather than a unique closest-source transport.The paper treats the cut locus as a natural feature of globally accurate approximation, not as a target to approximate itself.Source: §4.2 · p. 5
5 DISCRETE FORMULATION
The discrete formulation encodes tangent vectors in local coordinates and uses scalar and connection Laplacians to compute transport on triangle meshes. Intrinsic Delaunay flips improve robustness while preserving the geometry and tangent-coordinate mappings.
5.1 Discrete Surface
- The method is first formulated independently of a particular discretization, then specialized to oriented triangle meshes with positive edge lengths satisfying each face’s triangle inequality.The authors note that the method can in principle apply to nonorientable domains, but assume orientation for simpler exposition and implementation.Source: §5 · p. 6, §5.1 · p. 6
5.2 Intrinsic Tangent Spaces
- Tangent vectors at each vertex use local polar coordinates, encoded as complex numbers whose magnitude and angle specify the vector.A canonical reference edge sets the zero direction, and outgoing edge directions are represented relative to it.Source: §5.2 · p. 6
- Discrete parallel transport along an edge is represented by a rotation between the endpoint coordinate systems, encoded as a unit complex number for the connection Laplacian.The rotation keeps a vector parallel as it moves between vertices.Source: §5.2 · p. 6
5.3 Discrete Laplace Operators
- The triangle-mesh connection Laplacian is a complex matrix formed by multiplying scalar Laplacian off-diagonal entries by edge rotations; it is Hermitian.The scalar operator uses cotangent weights and a lumped mass matrix; the heat equations are discretized with one-step backward Euler.Source: §5.3 · pp. 6–7
- Transport is computed from source data using short-time heat solves, then normalized pointwise by the scalar heat solution.For point sources, the source vectors and scalar data are represented using vertex impulses.Source: §5.3 · p. 7
5.4 Tangent Intrinsic Delaunay
- Intrinsic Delaunay preprocessing flips edges with negative cotangent weights while preserving the piecewise Euclidean surface and vertex set.After each flip, the algorithm updates tangent directions so vectors can be mapped back by copying their polar coordinates.Source: §5.4 · p. 7
- Intrinsic Delaunay triangulation makes the connection Laplacian semidefinite and prevents vectors from being flipped; it improves reliability without changing input geometry.The condition is not strictly necessary for high-quality results, but helps avoid problems on pathological inputs.Source: §5.4 · p. 7
- The augmented flips can correct improperly oriented transported fields on non-Delaunay meshes, improving the resulting log map without changing input geometry.Fig. 10 contrasts an erroneous result with the solution obtained by tracking tangent spaces during flips.Source: §5.4 · p. 7, Figure 10 · p. 8
6 GENERALIZATIONS
The vector heat method extends to different vector bundles, connections, and geometric data structures when their scalar and connection Laplacians can be discretized. The paper illustrates constructions for 1-forms, symmetric direction fields, and several mesh and non-mesh representations.
6 GENERALIZATIONS
- The method generalizes to other vector-valued data and data structures, provided the relevant Laplacians can be formulated.A vector bundle is a manifold with a copy of the same vector space at each point; the connection determines what nearby vectors mean by parallel.Source: §6 · p. 8, §6.1 · p. 8
6.1.1 Differential 1-Forms.
- For differential 1-forms, the connection Laplacian is built by adding a curvature term to a discretized Hodge Laplacian.A diagonal DEC mass-matrix approach does not capture sufficient directional information; an approximate Galerkin mass matrix with larger support appears to yield the correct solution.Source: §6.1.1 · p. 8
6.1 Other Vector Bundles
- For symmetric direction fields, raising edge-rotation coefficients to the symmetry degree enables the usual algorithm, with final vectors recovered by taking complex roots.Examples include line fields and cross fields.Source: §6.1.2 · p. 9
6.1.3 Different Connection.
- Changing the connection changes what parallel means; setting all edge rotations to one yields closest-point interpolation of complex functions.An augmented connection can instead produce a local field-aligned parameterization centered at a source point.Source: §6.1.3 · p. 9
6.2 Other Discretizations
- Any representation with scalar and connection Laplacians can support the method; the paper discusses polygon meshes, point clouds, and voxelized surfaces.For embedded surfaces, nearby tangent spaces are related by the smallest rotation between their tangent planes.Source: §6.2 · p. 9
6.2.1 Polygon Meshes.
- For general, possibly non-planar polygon meshes, the method augments the discrete Laplacian with transport coefficients for every pair of vertices in each polygon.The coefficients use chosen vertex normals and the smallest rotation between the corresponding tangent planes.Source: §6.2.1 · p. 9
6.2.2 Point Clouds.
- For point clouds, the method uses a positive semidefinite point-cloud Laplacian, estimated tangent planes, transport coefficients, and Voronoi-area masses.Tangent planes are already needed to build the scalar Laplacian.Source: §6.2.2 · p. 9
6.2.3 Voxelizations.
- For voxelized surfaces, values are associated with boundary faces, with normals and areas estimated from voxelization geometry.The construction follows the point-cloud case; Fig. 11 shows an example.Source: §6.2.3 · p. 9, Figure 11 · p. 10
7 DISCUSSION AND VALIDATION
The method offers a performance–quality trade-off, with efficient repeated-source transport and accuracy that improves under refinement. Its discrete operator preserves several basic transport properties, though some properties are not exact for multiple sources or near the cut locus.
7.1 Basic Properties
- The method preserves linearity, magnitude, rotation covariance, and round-trip symmetry for a single point source.The Hermitian connection Laplacian underlies the symmetry result.Source: §7.1 · p. 10
- For multiple sources or curves, properties may not hold exactly because diffusion can cancel vectors pointing in different directions.Further properties, including curvature–monodromy equivalence around closed loops, may also fail to hold exactly.Source: §7.1 · p. 10
7.2 Implementation and Performance
- On a 100k-triangle mesh, transport takes about 30 ms after roughly 1 second of preprocessing, while exact polyhedral transport takes about 40x longer.The authors report that even tracing geodesics alone takes about 10–20x longer than the full vector heat method.Source: §7.2 · p. 11
7.3 Convergence and Accuracy
- The vector heat method converges roughly as O(h), versus O(h^2) for exact polyhedral schemes; their fields differ mainly near the cut locus, where neither is guaranteed accurate.Accuracy depends on tessellation resolution and quality; the heat method has better amortized performance, while window-based schemes can be more accurate.Source: §7.3 · p. 11
8 APPLICATIONS
The applications use parallel transport and globally accurate logarithmic maps to extend geometric algorithms to curved surfaces. The method supports velocity extrapolation, log-map construction, surface means, geodesic centroidal Voronoi tessellations, and intrinsic landmarks.
8 APPLICATIONS
- Fast parallel transport provides a shared foundation for applications implemented mainly by setting up and solving linear systems.Prefactorization gives low amortized cost, and the approach generalizes to point clouds and polygon meshes.Source: §8 · p. 12
8.1 Velocity Extrapolation
- The method extrapolates tangential velocity globally, while the closest-point property keeps signed distance functions nearly preserved over long integration times.With finite time t, interface data may not be exactly preserved, though it will generally be very close.Source: §8.1 · p. 12
8.2 Logarithmic Map
- Global log maps remain accurate across the surface up to the cut locus, unlike local approximations that can fail over longer distances.Carefully projected initial conditions produce a smooth radial field; prior approximations can have large global errors, especially in the angular coordinate.Source: §8.2 · p. 13, Figure 20 · p. 14
8.3 Karcher Means and Geometric Medians
- The vector-heat log map enables Karcher means and geometric medians on general surfaces, with examples converging in fewer than 20 iterations.The Karcher mean is significantly influenced by outliers, while the geometric median is not; previous local log-map approximations can fail to converge or converge only after line search.Source: §8.3 · p. 15, Figure 21 · p. 15
8.4 Geodesic Centroidal Voronoi Diagrams
- Using each cell’s Karcher mean yields a true geodesic centroidal Voronoi tessellation, including diagrams with large, multiply connected cells.Euclidean or diffusion-based centroids can produce different results, particularly for large curved cells or when the Karcher mean lies outside the cell.Source: §8.4 · pp. 15–16, Figure 24 · p. 15
8.5 Ordered Intrinsic Landmarks
- Geometric medians reduce landmark variability from values near the cut locus, which are effectively treated as outliers.The method progressively adds landmarks by furthest-point sampling, beginning from geometric medians of the surface.Source: §8.5 · p. 16
9 LIMITATIONS
The method’s practical challenges are solver scalability and numerical accuracy, and the applications leave questions about boundary handling and landmark robustness open.
9 LIMITATIONS
- Solver scalability and numerical accuracy are the main challenges; direct solvers work well up to millions of elements, but poor mesh quality can cause indefinite matrices and flipped vectors.Intrinsic Delaunay triangulations help, while low-order elements limit convergence to linear order.Source: §9 · p. 16
- Meshes with too few elements or poorly distributed vertices can still yield low-quality results, even with intrinsic Delaunay triangulation.More sophisticated spline or subdivision bases might improve accuracy.Source: §9 · p. 16
- Open application questions include preserving data along level sets, defining the log map for domains with boundary, and improving landmark robustness.Source: §9 · p. 17
10 CONCLUSION
The paper presents an efficient, reliable method for computing parallel transport along shortest geodesics, an operation with broad potential in geometry processing. It identifies open directions while anticipating that implementation through Laplace-like systems may facilitate adoption.
10 CONCLUSION
- The method efficiently and reliably computes vector fields arising from parallel transport along shortest geodesics.The paper notes that these fields enable many fundamental geometry-processing algorithms but have received little prior attention.Source: §10 · p. 17
- Open directions include boundary conditions, accuracy, practical efficiency, transport along non-geodesic curves, and transport along a given vector field.Source: §10 · p. 17
- The authors expect that building and solving Laplace-like systems may facilitate adoption in real applications.Source: §10 · p. 17
A DISTANCE GRADIENT DISCRETIZATION
The radial field for the logarithmic map is constructed by transporting unit vectors initialized on a small circle around the source. Finite-element initial conditions improve the resulting field, and integrating it recovers the radial coordinate.
A DISTANCE GRADIENT DISCRETIZATION
- Finite-element initial conditions, rather than simply sampling vectors at neighboring vertices, yield reliable and accurate radial fields.The construction projects unit normals on a small circle around the source onto piecewise linear basis functions.Source: A DISTANCE GRADIENT DISCRETIZATION · p. 18
- Because the circle-supported initial measure touches only source-adjacent triangles, its integrals need only be evaluated for immediate neighbors.Source: A DISTANCE GRADIENT DISCRETIZATION · p. 18
- After vector diffusion, normalizing the transported vectors gives the radial field; a Poisson solve integrates it to obtain the distance coordinate.The initial conditions do not depend on the circle radius, and the radial vector at the source is set to zero.Source: A DISTANCE GRADIENT DISCRETIZATION · pp. 18–19
- Across tested meshes, t = h^2 is a reasonable parameter choice, and the vector heat field closely matches the exact polyhedral reference where Zhang et al.’s local approximation deviates away from the source.Source: Figure 26 · p. 19, Figure 27 · p. 19