Graph Shift Operator

Graph Shift Operator
Author

SEOYEON CHOI

Published

July 2, 2023

Definition(Djuric and Richard 2018): Given a normal shift operator \({\bf S}\), we say that a graph signal \({\bf y}\) is weakly stationary with respect to \({\bf S}\) if, for all \(a\), \(b\), and \(c \leq b\), the following equality holds:

Djuric, Petar, and Cédric Richard. 2018. Cooperative and Graph Signal Processing: Principles and Applications. Academic Press.

\[\mathbb{E} \bigg[ \big({\bf S}^a{\bf y}\big)\Big(\big({\bf S}^H)^b {\bf y}\Big)^H \bigg]=\mathbb{E}\bigg[\big({\bf S}^{a+c}{\bf y}\big)\Big(\big({\bf S}^H\big)^{b-c}{\bf y} \Big)^H \bigg].\]

Note

Using \({\bf S}\) as the periodic shift operator \({\bf S}=\begin{cases} 1 & j-j = 1 \\ 0 & o.w.\end{cases}.\), the definition is equivalent to the traditional stationarity definition in time series analysis.

Conjugate

\(E(y)=0\)을 가정하고, \(y = (y_1,y_2)\)의 벡터를 가정했을 때,

\(Cov(y) = \begin{pmatrix} cov(y_1,y_2) & cov(y_1,y_2) \\ cov(y_2,y_1) & cov(y_2,y_2) \end{pmatrix}\)

  • \(cov(y_1,y_1) = V(y_1) = E(y_1 - \mu_1)^2 = E(y_1)^2 (\therefore \mu = 0)\)
  • \(cov(y_2,y_1) = E(y_2 - \mu_2)E(y_1 - \mu_1) = E(y_2-y_1)\)
  • \(cov(y_2,y_2) = V(y_2) = E(y_2 - \mu_2)^2 = E(y_2)^2 (\therefore \mu = 0)\)
  • \(cov(y_1,y_2) = E(y_1 - \mu_1)E(y_2 - \mu_2) = E(y_1-y_2)\)

\(= E \begin{pmatrix} y_1^2 & y_1y_2 \\ y_2 y_1 & y_2^2 \end{pmatrix} = E(y y^\top)\)

\(y = (y_q y_2)^\top\)

\(y^\top = (y_1,y_2)\)

\(y y^\top = \begin{bmatrix} y1 \\ y_2 \end{bmatrix} \begin{bmatrix} y2 & y_2 \end{bmatrix} = \begin{bmatrix} y_1^2 & y_1y_2 \\ y_2y_1 & y_2^2 \end{bmatrix}\)

\(cov(y) = E(y t^\top) = E(y y^H)\) -> 확률변수가 복소수일 경우 가정 가능하다

\(cov(y\text{의 } a \text{만큼 평행이동}) = cov(y\text{의 } b \text{만큼 평행이동})\)

\(cov((S^ay)(S^b y)^\top) = cov((S^c y)(S^d y)^\top)\)

  • 결국, normal GSO가 주어질 때, \(y\)는 약정상성을 S에 대해 가지고 있다는 말이 된다.
  • 정상성 조건: 평균, 분산이 일정할때, 자기 공분산이 시차 t에만 의존할 때