Heat Kernel Signature in Geometry

• Heat Kernel Signature is a spectral descriptor that measures self-similarity on a manifold via the exact solution of the heat equation using hyperspherical harmonics.
• The method employs an exact series expansion with exponential decay of coefficients, ensuring robust convergence and numerical stability for high-dimensional data.
• HKS is instrumental in machine learning applications such as kernel SVMs, offering improved discrimination and lower VC-dimensions by leveraging hyperspherical geometry.

The heat kernel signature (HKS) quantifies self-similarity of a point under diffusion on a manifold and provides a canonical, spectrally grounded descriptor for geometric and machine learning applications. In the context of hyperspherical geometry, the HKS arises from the exact solution to the heat equation associated with the Laplace–Beltrami operator on the unit sphere $S^{n-1}\subset\mathbb{R}^n$. The unnormalized HKS at a point $p\in S^{n-1}$ is given by the diagonal of the heat kernel, while a normalized version yields constant value $1$ due to self-similarity normalization. Advances in exact series expansions and critical comparison with commonly used heuristic (parametrix) heat kernels clarify the mathematical properties, implementation, and discrimination power of the HKS in high-dimensional settings (Zhao et al., 2017).

1. Spectral Definition of the Heat Kernel on the Hypersphere

Let $S^{n-1}$ denote the unit sphere in $\mathbb{R}^n$ with $n\geq2$, and consider the Laplace–Beltrami operator $-\widehat{L}^2$. The angular momentum eigenfunctions, the degree-$\ell$ hyperspherical harmonics $Y_{\ell,\{m\}}$, satisfy

$$\widehat{L}^2 Y_{\ell,\{m\}}(x) = \ell(\ell+n-2) Y_{\ell,\{m\}}(x),$$

where $\ell\geq0$ and $\{m\}$ denotes the set of magnetic quantum numbers for each $\ell$. The heat kernel $G^{\rm ext}(x,y;t)$, which solves the heat equation with initial condition $\delta(x,y)$, is the fundamental solution:

$$G^{\rm ext}(x,y;t) = \left(e^{-t\widehat{L}^2}\right)(x,y).$$

By spectral expansion, this admits the absolutely and uniformly convergent series

$$G^{\rm ext}(x,y;t) = \sum_{\ell=0}^\infty e^{-\ell(\ell+n-2)\,t} \sum_{\{m\}} Y_{\ell,\{m\}}(x)\;\overline{Y_{\ell,\{m\}}(y)}.$$

Utilizing hyperspherical addition theorems, the kernel can be written in terms of Gegenbauer polynomials $C_{\ell}^{\alpha}(w)$ and the geodesic cosine $x\cdot y = \cos\theta$ as

$$G^{\rm ext}(x,y;t) = \sum_{\ell=0}^\infty e^{-\ell(\ell+n-2)\,t}\;\frac{2\ell+n-2}{n-2}\;\frac{1}{A_{S^{n-1}}}\;C_{\ell}^{\frac n2-1}(x\cdot y),$$

where the surface area $A_{S^{n-1}}=\frac{2\pi^{n/2}}{\Gamma(n/2)}$. Uniform and absolute convergence for all $x\cdot y\in[-1,1]$ and $t>0$ is guaranteed by the Weierstrass M-test and polynomial growth bounds.

2. Heat Kernel Signature Formulation

The (unnormalized) HKS at $p\in S^{n-1}$ is the self-similarity observed under heat evolution:

$$\widetilde H(p;t) = G^{\rm ext}(p,p;t) = \sum_{\ell=0}^\infty e^{-\ell(\ell+n-2)\,t} \frac{2\ell+n-2}{n-2} \frac{1}{A_{S^{n-1}}}\,C_{\ell}^{\frac n2-1}(1).$$

Using $$C_{\ell}^{\alpha}(1) = \frac{\Gamma(\ell+n-2)}{\Gamma(\ell+1)\,\Gamma(n-2)}$$, this becomes

$$\widetilde H(p;t) = \frac{1}{A_{S^{n-1}}} \sum_{\ell=0}^\infty e^{-\ell(\ell+n-2)t} \frac{2\ell+n-2}{n-2} \frac{\Gamma(\ell+n-2)}{\Gamma(\ell+1)\,\Gamma(n-2)}.$$

The normalized heat kernel $K^{\rm ext}$ is defined by

$$K^{\rm ext}(x,y;t) = \frac{G^{\rm ext}(x,y;t)}{G^{\rm ext}(x,x;t)},$$

yielding the self-similarity normalization $H(p;t) = K^{\rm ext}(p,p;t) = 1$ at all $p$ by construction.

3. Comparison: Exact Series and Parametrix Expansion

Heuristic approaches typically use the zeroth-order "parametrix" or Gaussian kernel,

$$K^{\rm prx}_0(\theta;t) = \exp\left(-\frac{\theta^2}{4t}\right),$$

where $\theta = \arccos(x\cdot y)$. Higher-order corrections attempt to better approximate the true heat kernel,

$$K^{\rm prx}(\theta;t) = G(\theta,t)\,\bigl[u_0(\theta) + u_1(\theta)\,t + u_2(\theta)\,t^2 + \cdots\bigr],$$

where $$G(\theta,t) = (4\pi t)^{-\frac{d}{2}}e^{-\theta^2/(4t)}$$ with $d=n-1$. Correction terms, such as

$$u_0(\theta) \propto \left(\frac{\sin\theta}{\theta}\right)^{-\frac{d-1}{2}},$$

and higher $u_k$, diverge as $\theta\to\pi$ and produce unphysical singularities for large $d$. The full parametrix is valid only for $\theta\to0$, $t\downarrow0$, and $(n-2)t \ll O(1)$. In typical machine learning practice, all corrections are dropped and only the RBF factor is used. The exact series, in contrast, has a firm mathematical foundation and uniform convergence for all relevant $(\theta, t)$.

4. Numerical Implementation and Series Truncation

The coefficients of the exact kernel decay exponentially:

$$a_\ell(t) = e^{-\ell(\ell+n-2)t}\,\frac{2\ell+n-2}{n-2}\,\frac{|C_{\ell}^{\frac n2-1}(x\cdot y)|}{A_{S^{n-1}}},$$

with $$a_{\ell+1}/a_\ell \sim \exp[-(2\ell+n-1)t] O(\ell^0)\to0$$ as $\ell\to\infty$. For very small $t$ the series converges slowly; for very large $t$ the kernel saturates to $\approx1/A_{S^{n-1}}$. A balance between convergence speed and kernel discrimination is achieved for $t\sim(\log n)/n$. Series truncation at $\ell_{\rm max}$ is recommended when

$$e^{-\ell_{\max}(\ell_{\max}+n-2)t}\,\frac{2\ell_{\max}+n-2}{n-2} \frac{\Gamma(\ell_{\max}+n-2)}{\Gamma(\ell_{\max}+1)\,\Gamma(n-2)} < \varepsilon$$

for desired precision $\varepsilon$ (e.g. $10^{-6}$), typically requiring $\ell_{\max}=O(\sqrt{1/t})$.

Parameter	Expression	Description
Series Term ($a_\ell$)	$$e^{-\ell(\ell+n-2)t} \frac{2\ell+n-2}{n-2} \frac{|C_{\ell}^{n/2-1}(x\cdot y)|}{A_{S^{n-1}}}$$	Decaying mode contribution
Optimal $t$ choice	$t\simeq(\log n)/n$	Balances convergence/discrimination
Truncation threshold	$$e^{-\ell_{\max}(\ell_{\max}+n-2)t}\dots < \varepsilon$$	Determines $\ell_{\max}$

5. Spectral and Geometric Insights

Completeness and addition theorems for hyperspherical harmonics underlie the theoretical framework:

$$\sum_{\ell=0}^{\infty}\sum_{m} Y_{\ell,m}(x) Y_{\ell,m}^*(y) = \delta(x,y),$$

and

$$\sum_{m} Y_{\ell,m}(x) Y_{\ell,m}^*(y) = \frac{2\ell+n-2}{n-2}\frac{1}{A_{S^{n-1}}}\,C_{\ell}^{\frac n2-1}(x\cdot y).$$

Uniform absolute convergence of the heat kernel series follows from Gegenbauer polynomial bounds, $$|C_{\ell}^{\alpha}(w)|\leq M_\ell\sim O(\ell^{2\alpha-1})$$, and the Weierstrass M-test.

A notable property is the reduction in VC-dimension when mapping $\mathbb{R}^n_+\to S^{n-1}$, eliminating the radial degree of freedom and yielding a lower VC-bound $\mu_{VC}$. This effect enhances generalization, particularly in small-sample, high-dimensional regimes. Furthermore, eigenvalues $\lambda_\ell = \ell(\ell+n-2)$ induce rapid decay for higher spectral modes, leading to stable numerical computation for moderate $t$.

6. Summary of Implementation and Application Guidance

• The exact unnormalized HKS at $p\in S^{n-1}$ is given by

$$H(p;t) = \frac{1}{A_{S^{n-1}}}\sum_{\ell=0}^{\ell_{\max}} e^{-\ell(\ell+n-2)t}\;\frac{2\ell+n-2}{n-2}\;\frac{\Gamma(\ell+n-2)}{\Gamma(\ell+1)\,\Gamma(n-2)}$$

with truncation at $\ell_{\max}$ for the desired precision.

• Select $t\approx(\log n)/n$ to balance spatial localization and mixing.
• For cross-similarity ($x \neq y$), use

$$K(x,y;t) = \frac{\sum_{\ell=0}^{\ell_{\max}} e^{-\ell(\ell+n-2)t}\,\frac{2\ell+n-2}{n-2}\,C_{\ell}^{n/2-1}(x\cdot y)} {\sum_{\ell'=0}^{\ell_{\max}} e^{-\ell'(\ell'+n-2)t}\frac{2\ell'+n-2}{n-2}C_{\ell'}^{n/2-1}(1)}$$

• Avoid parametrix corrections in high dimensions except for asymptotic ($\theta\to0$, $t\to0$) analyses; these terms lead to singularities and unreliable approximations elsewhere.
• The provided LaTeX expressions are directly suited for computational and publication use.

The HKS, derived from the exact hyperspherical heat kernel, provides a theoretically principled, numerically stable, and discriminative similarity measure, especially advantageous in kernel SVM applications where data are mapped to $S^{n-1}$, such as text mining and high-dimensional statistical learning (Zhao et al., 2017).