This is an archived course. A more recent version may be available at ocw.mit.edu.

2.7 Eigenvalue Stability of Finite Difference Methods

2.7.3 Circulant Matrices

Measurable Outcome 2.10, Measurable Outcome 2.11

Though the eigenvalues of \(A\) typically require numerical techniques for the general problem, a special case of practical interest occurs when the matrix is ‘periodic'. That is, the column entries shift a column every row. Thus, the matrix has the form,

\[A = \left(\begin{array}{ccccc} a_1 & a_2 & a_3 & \ldots & a_{N } \\ a_ N & a_1 & a_2 & \ldots & a_{N-1} \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ a_2 & a_3 & a_4 & \ldots & a_1 \end{array}\right)\] (2.140)

This type of matrix is known as a circulant matrix. Circulant matrices have eigenvalues given by,

\[\lambda _ n = \sum _{j=1}^{N} a_ j e^{i\, 2\pi (j-1) \frac{n}{N}} \qquad \mbox{for} \quad n = 0, 1, \ldots , N-1 \label{equ:circulant_ eig}\] (2.141)

As we saw in Section 2.7.2, when periodic boundary conditions are assumed, the central space discretization of one-dimensional convection gives purely imaginary eigenvalues, and when scaled by a timestep for which the CFL number is one, the eigenvalues stretch along the axis until \(\pm i\). Since for a convection problem with constant velocity and periodic boundary conditions gives a circulant matrix, we can use Equation (2.141) to determine the eigenvalues analytically. We begin by finding the coefficients, \(a_ j\). For a central space discretization, we find,

\[a_2 = -\frac{u}{2{\scriptstyle \Delta } x}, \quad a_ N = \frac{u}{2{\scriptstyle \Delta } x}, \mbox{ and for all other } j, \quad a_ j = 0.\] (2.142)

Then, substituting these \(a_ j\) into Equation (2.141) gives,

  \(\displaystyle \lambda _ n\) \(\displaystyle =\) \(\displaystyle -\frac{u}{2{\scriptstyle \Delta } x} e^{i\, 2\pi \frac{n}{N}} + \frac{u}{2{\scriptstyle \Delta } x} e^{i\, 2\pi (N-1) \frac{n}{N}},\)   (2.143)
    \(\displaystyle =\) \(\displaystyle -\frac{u}{2{\scriptstyle \Delta } x} e^{i\, 2\pi \frac{n}{N}} + \frac{u}{2{\scriptstyle \Delta } x} e^{i\, 2\pi n} e^{-i\, 2\pi \frac{n}{N}}.\)   (2.144)

Since \(e^{i\, 2\pi n} = 1\) (because \(n\) is an integer), then,

  \(\displaystyle \lambda _ n\) \(\displaystyle =\) \(\displaystyle -\frac{u}{2{\scriptstyle \Delta } x} e^{i\, 2\pi \frac{n}{N}} + \frac{u}{2{\scriptstyle \Delta } x} e^{-i\, 2\pi \frac{n}{N}},\)   (2.145)
    \(\displaystyle =\) \(\displaystyle -\frac{u}{2{\scriptstyle \Delta } x}\left(e^{i\, 2\pi \frac{n}{N}} - e^{-i\, 2\pi \frac{n}{N}}\right),\)   (2.146)
    \(\displaystyle =\) \(\displaystyle -i\frac{u}{{\scriptstyle \Delta } x}\sin \left(2\pi \frac{n}{N}\right).\)   (2.147)

Multiplying by the timestep,

\[\lambda _ n {\Delta t}= -i\frac{u{\Delta t}}{{\scriptstyle \Delta } x}\sin \left(2\pi \frac{n}{N}\right).\] (2.148)

As observed in Section 2.7.2, the eigenvalues are purely imaginary and will extend to \(\pm i\) when \(\mathrm{CFL} = |u|{\Delta t}/{\scriptstyle \Delta } x= 1\).