Heat Kernel Signature¶
A pointwise multiscale shape descriptor obtained by sampling diagonal heat-kernel return values, combining local-to-global intrinsic geometry with isometry invariance.
Core Idea¶
HKS asks how much heat starting at a point returns to that same point after different diffusion times. The diagonal heat kernel, sampled as a vector, captures fine local geometry early and broader shape context later.
Its spectral formula uses Laplace–Beltrami eigenvalues and squared eigenfunctions. Intrinsic isometry invariance and perturbation stability make it useful for matching deformable shapes, but time range, scale normalization, mesh operator, boundary conditions, and spectral truncation govern practical meaning.
Scope of Application¶
- Shape matching. Compares intrinsic point neighborhoods.
- Segmentation. Provides multiscale vertex features.
- Shape retrieval. Builds local descriptors for indexing.
- Geometry processing. Uses diffusion spectra on meshes.
Clarity¶
State domain and metric, Laplacian discretization, boundary condition, eigenpair count, time range and spacing, normalization, mesh quality, and the transformation invariance actually required. Inclusion test: Require an intrinsic diffusion operator, its diagonal heat kernel, a declared time schedule, and a pointwise sampled vector with numerical and scale conventions. Exclusion test: Exclude the full pairwise heat kernel, a global heat trace, generic spectral embeddings, and extrinsic curvature descriptors that merely use eigenvectors. Nearest boundary: The wave kernel signature also uses Laplacian eigenpairs but weights frequencies by oscillatory quantum evolution rather than heat diffusion. Exit condition: Changing from diagonal return values to off-diagonal transport or global trace changes the descriptor; inadequate spectral/time resolution makes only an approximation whose limits must be stated. Common misclassifications: It is not the full two-point heat kernel. It is not a global heat trace. It is not automatically scale invariant. A truncated numerical HKS is not the exact continuum signature. Nearest named distinctions: Heat kernel: Is the full two-point fundamental solution. Heat trace: Aggregates globally over the diagonal. Wave kernel signature: Uses oscillatory rather than diffusive spectral weights. Shape context: Is an extrinsic/statistical descriptor family.
Manages Complexity¶
The descriptor reduces a two-point, continuous-time kernel to a manageable pointwise vector while retaining a controlled sequence of geometric scales.
Abstract Reasoning¶
- Construct a consistent intrinsic Laplacian.
- Compute a sufficiently resolved eigenbasis.
- Choose times supported by mesh and spectrum.
- Evaluate diagonal return values at each point.
- Normalize and test stability and distinctiveness for the task.
Knowledge Transfer¶
HKS transfers between meshes only with compatible metric scale, Laplacian convention, boundary treatment, time schedule, eigenbasis accuracy, and sampling density; raw coordinates or indices are not comparable.
Relationships to Other Abstractions¶
Current abstraction Heat Kernel Signature Domain-specific
Parents (1) — more general patterns this builds on
-
Heat Kernel Signature is a kind of Representation Prime
Heat Kernel Signature is a strict kind of Representation: it represents local shape by a multiscale vector of heat-kernel return values.
Hierarchy path (1) — routes to 1 parentless root
- Heat Kernel Signature → Representation → Abstraction
Neighborhood in Abstraction Space¶
Heat Kernel Signature sits in a moderately populated region (42nd percentile for distinctiveness): it has near-neighbors but no dense thicket of look-alikes.
Family — Physical & Geometric Dynamical Quantities (29 abstractions)
Nearest neighbors
- Fourier–Bros–Iagolnitzer Transform — 0.88
- Computational electromagnetics — 0.87
- P-Laplacian — 0.87
- Solid Modeling — 0.87
- Endothermic Process — 0.87
Computed from structural-signature embeddings · 2026-10-08