Source

Adapted from my own handwritten and hand-typed notes from Year 1 to Year 3 of my EEE degree

A discrete time dynamical system updates a state vector each step by . Diagonalisation shows where the state ends up as .

Idea

Computing directly is hard. With :

As :

  • : , that mode dies out
  • : , that mode stays, giving the steady state
  • : that mode blows up

So the long run state is , where keeps only the entries.

Method

Steps:

  1. Find the eigenpairs of and form ,
  2. Replace each by its limit (0 or 1)
  3. Compute , or simply scale the eigenvector to match a conserved total

Worked Example: Early Bird Scheme

Each month 30% of "early bird" MRT commuters go back to peak hour, and 20% of peak hour commuters switch to early bird. Total 100,000 commuters, starting with 20,000 early birds.

Eigenpairs:

  • : gives , so
  • : , which decays away

The total is conserved (each column sums to 1), so the steady state is the eigenvector scaled to 100,000:

In the long run 40,000 commuters use the early bird scheme, whatever the starting split.

Note

A matrix whose columns are nonnegative and sum to 1 (a Markov or transition matrix) always has , and the steady state is its eigenvector. Building and is covered in Diagonalisation.