heat kernel

the multi-scale smoothing operator in the tri-kernel

$$H_\tau = \exp(-\tau L)$$

where $L$ is the graph Laplacian and $\tau$ is the temperature parameter. one of three components of the tru computation alongside diffusion and springs

the heat equation on graphs

the operator arises from the graph heat equation:

$$\frac{\partial H}{\partial \tau} = -L H, \quad H_0 = I$$

the solution is the matrix exponential $H_\tau = \exp(-\tau L)$. applied to a signal $\phi$ on the cybergraph, $H_\tau \phi$ returns a smoothed version where each particle's value blends with its neighborhood — Gaussian-like decay over graph distance

temperature as scale

$\tau$ controls the spatial reach of smoothing:

regime behavior analogy
$\tau \to 0$ identity — each particle sees only itself crystallization, commitment
small $\tau$ local smoothing — nearby particles blend fine-grained context
large $\tau$ global smoothing — distant particles contribute annealing, exploration
$\tau \to \infty$ uniform distribution — all structure erased maximum entropy

high $\tau$ explores. low $\tau$ commits. the temperature parameter is the thermostat of collective focus — controlling how much the system adapts versus how much it preserves local structure

spectral decomposition

in the eigenbasis of $L$ with eigenvalues $0 = \lambda_1 \leq \lambda_2 \leq \cdots \leq \lambda_n$:

$$H_\tau = \sum_{k=1}^{n} e^{-\tau \lambda_k} \, v_k v_k^\top$$

low-frequency eigenvectors (small $\lambda_k$) persist. high-frequency eigenvectors (large $\lambda_k$) are exponentially suppressed. this is why the operator acts as a multi-scale smoother: it preserves global community structure while erasing local noise

the spectral gap $\lambda_2$ (the Fiedler eigenvalue) governs the contraction rate:

$$\|H_\tau \phi - H_\tau \psi\|_2 \leq e^{-\tau \lambda_2} \|\phi - \psi\|_2$$

properties

positivity-preserving: if $\phi \geq 0$ then $H_\tau \phi \geq 0$. the operator never creates negative values from positive inputs

semigroup: $H_{\tau_1} H_{\tau_2} = H_{\tau_1 + \tau_2}$. smoothing at scale $\tau_1$ then at scale $\tau_2$ equals smoothing at scale $\tau_1 + \tau_2$

locality: Gaussian tail decay means the effect at distance $d$ falls as $\exp(-d^2 / 4\tau)$. for precision $\varepsilon$, only $h = O(\log(1/\varepsilon))$ hops are needed. Chebyshev polynomial approximation computes $H_\tau \phi$ locally with bounded error

contraction: the collective focus theorem proves $H_\tau$ contracts with rate $e^{-\tau \lambda_2} < 1$, contributing to the composite contraction of the full tri-kernel

role in the tri-kernel

the tri-kernel blends three operators:

$$\phi^{(t+1)} = \text{norm}\big[\lambda_d \cdot D(\phi^t) + \lambda_s \cdot S(\phi^t) + \lambda_h \cdot H_\tau(\phi^t)\big]$$

diffusion explores via random walks. springs enforce structural consistency via the screened Laplacian. the heat kernel provides adaptive context — what the graph looks like at scale $\tau$. together the three forces converge to focus: the fixed point that is cyberank

see cyber/heat for the full specification. see tri-kernel for the composite operator. see collective focus theorem for the convergence proof. see Fan Chung for the foundational work on heat kernels and PageRank. see spectral gap for the eigenvalue that governs contraction rate

Local Graph