Spectral graph convolution defines filtering in the frequency domain of graphs. The graph Laplacian L = D − A has eigenvectors U that form a Fourier basis for graph signals. A graph convolution is x * g = U · diag(g_hat) · U^T · x, where g_hat are learnable spectral filter coefficients. ChebNet approximates filters as Chebyshev polynomials for efficiency; GCN (Kipf & Welling 2017) further simplifies to the first-order approximation D^{-1/2} A_hat D^{-1/2}, which is a localized low-pass filter. The visualization shows the original graph signal (left) and its filtered version (right) using the selected filter.