Report

Introduction

• Graph: generic data representation form describing the geometric structures of data domains.

  We denote an undirected, connected, weighted graph as $\mathcal{G}=\{\mathcal{V},\mathcal{E},\mathcal{W}\}$, and the number of vertices is $N=|\mathcal{V}|$.

• Graph Signal: data that reside on the vertices of a graph, usually represented as a vector $ f \in \mathbb{R}^N $.

• Edge Weights: usually represent similarity between the two vertices.

  Common graph construction methods when edge weights are not naturally defined:

  • Gaussian kernel weighting function
  • k-nearest neighbors

The Graph Laplacian

• (Unnormalized) Graph Laplacian: $\mathcal{L}=D-W$

  $\mathcal{L}$ works as a difference operator for any signal $f$ .

$$\begin{array}{c}Lf(i)=\sum _j{w}_{ij}(f(i)-f(j))\end{array}$$

Since $\mathcal{L}$ is a real symmetric matrix for an undirected graph, it has $N$ eigenvalues denoted as $\{{\lambda }_{\iota }{\}}_{\iota =0,1,...,N-1}$. Also, since the graph has at least one connected component, there is at least one eigenvalue that is equal to zero.

• Normalized Graph Laplacian:

$$\begin{array}{c}\hat{\mathcal{L}}=D^{-1/2}\mathcal{L}{D}^{-1/2}=I-D^{-1/2}W{D}^{-1/2}\end{array}$$

the eigenvalues of $\hat{\mathcal{L}}$ belongs to $[0,2]$ , with $\hat{\lambda }=2$ if and only if $\mathcal{G}$ is bipartite.

• Random Walk Matrix: $P=D^{-1}W$

There is not a clear answer as to when to use each of these basis.

Graph Fourier Transform

The classical Fourier transform works in the time domain and the frequency domain. Analogously we define the graph Fourier transform, which works in the vertex domain and the graph spectral domain.

For any signal $f$ in the vertex domain, we have its corresponding signal (kernel) $\hat{f}$ in the graph spectral domain.

In matrix form, we have $U$ ($n\ast n$) as the Fourier basis of the graph Laplacian (eigenvectors as columns), $F$ ($n\ast m$) as a list of $m$ signals on $n$ vertices (each signal is a column).

$$\begin{array}{c}\hat{f}({\lambda }_{\iota })=\sum _{i=1}^{N}f(i){\mathcal{u}}_{\iota }^{\ast }(i)\\ \hat{F}=U^{\prime }F\end{array}$$

and the inverse graph Fourier transform:

$$\begin{array}{c}f(i)=\sum _{\iota =0}^{N-1}\hat{f}({\lambda }_{\iota })u_{\iota }(i)\\ F=U\hat{F}\end{array}$$

The eigenvalues and eigenvectors in graph spectral domain provide a similar notation of frequency: the eigenvectors with larger eigenvalues contains more zero crossings.

Discrete Calculus and Signal Smoothness

The edge derivative of a signal $f$ with respect to edge $e_{ij}$ at vertex $i$ is defined as :

$$\begin{array}{c}\frac{\partial f}{\partial e}{|}_i=\sqrt{W_{i,j}}|f(j)-f(i)|\end{array}$$

The gradient of $f$ at vertex $i$ (${\mathrm{\nabla }}_{i}f$) is therefore the vector of all derivatives with edges starting from $i$ in the graph.

The discrete $p$-Dirichlet form of $f$ is defined as:

$$\begin{array}{c}{S}_p(f)=\frac{1}{p}\sum _{i\in V}\parallel {\mathrm{\nabla }}_{i}f{\parallel }_2^p\end{array}$$

When $p=2$ , it is known as the graph Laplacian quadratic form.

Seminorm of $f$ is defined as $\parallel f{\parallel }_{\mathcal{L}}=\sqrt{S_2(f)}$ .

The smoother the graph, the smaller the $S_2(f)$ .

The discrete regularization framework:

$$\begin{array}{c}{argmin}_f\{\parallel f-y{\parallel }_2^2+\gamma S_p(f)\}\end{array}$$

When $p=2$ ,it is called Tikhonov Regularization and can be used for image denoising.

Generalized Graph Signal Operators

Filtering

• graph spectral domain:

  let $\hat{h}(1\ast n)$ be the transfer function in spectral domain. let $H=diag(\hat{h})$ .

$$\begin{array}{c}{\hat{f}}_{out}({\lambda }_{\iota })={\hat{f}}_{in}({\lambda }_{\iota })\hat{h}({\lambda }_{\iota })\\ {\hat{F}}_{out}=H{\hat{F}}_{in}\end{array}$$

$ f_{out} $ can then be calculated through an inverse graph Fourier transform.

$$\begin{array}{c}{U}^{\prime }\ast F_{out}=H{U}^{\prime }{F}_{in}\\ F_{out}=UH{U}^{\prime }{F}_{in}\end{array}$$

• vertex domain: linear combination within K-hop neighborhoods

$$\begin{array}{c}{f}_{out}(i)=b_{i,i}{f}_{in}(i)+\sum _{j\in \mathcal{N}(i,K)}{b}_{i,j}{f}_{in}(j)\end{array}$$

When the transfer function in graph spectral domain is an order K polynomial, the two forms of filtering can be related.

Convolution

Convolution in the vertex domain (the time domain) is equivalent to multiplication in the graph spectral domain (the frequency domain).

Since in the graph setting there is no $h(t-\tau )$, we use multiplication in the graph spectral domain to generalize the definition of convolution:

$$\begin{array}{c}(f{\ast }_{g}h)(i)=\sum _{\iota =0}^{N-1}\hat{f}({\lambda }_{\iota })\hat{h}({\lambda }_{\iota })u_{\iota }(i)\\ f{\ast }_{g}h=U((U^{\prime }f)\odot (U^{\prime }g)\\ F{\ast }_{g}h=UH{U}^{\prime }F\end{array}$$

so convolution with $h$ is just filter with $h$.

Translation

It's not clear what it means to translate a graph signal, but we can generalize translation as a convolution with a delta centered at $n$:

$$\begin{array}{c}(T_{n}g)(i)=\sqrt{N}(g\ast {\delta }_n)(i)\end{array}$$

Modulation

$$\begin{array}{c}(M_{k}g)(i)=\sqrt{N}{u}_k(i)g(i)\end{array}$$

Dilation

$$\begin{array}{c}(\widehat{{\mathcal{D}}_{s}g)}(\lambda )=\hat{g}(s\lambda )\end{array}$$

Coarsening

this operation can be separated in to two subtasks:

• Down Sampling: identifying a reduced set of vertices
• Reduction: assigning edges and weights

For a bipartite graph, it is natural to down sample it by a factor of 2.

For non-bipartite graphs, conditions are much more complex.

Localized Multiscale Transforms

Wavelet transform can localize signal in both time and frequency simultaneously, which is better than Fourier transform.

• vertex domain
  • random transforms
  • graph wavelets
  • lifting-based transforms
  • tree wavelets
• graph spectral domain
  • diffusion wavelets
  • spectral graph wavelet
  • graph-quadrature mirror filterbanks

Open Issues & Extension

Open Issues

• How the construction of graph affects properties of the localized, multiscale transforms for signals on graphs.
• When and why to use normalized, unnormalized graph Laplacian or other basis.
• What kind of distance to use in the vertex domain.
• Approximate computational techniques for signal processing on graphs.
• How to link structural properties of graph signals and their underlying graphs to properties.

Extensions

• Directed graph.
• Time series of signal data.
• Time series of underlying graphs.