By the end of this lesson, you should be able to:
- Construct delay-coordinate vectors from a scalar time series.
- Explain how delay τ changes redundancy in a reconstruction.
- Interpret embedding dimension m as the number of delayed coordinates used to unfold the dynamics.
From one signal to a state space
In experiments we often observe only one scalar measurement, even though the underlying system has several interacting state variables. Delay-coordinate reconstruction builds a new state vector from delayed copies of that one signal, for example \(\mathbf{y}_n=[x_n,x_{n-\tau},x_{n-2\tau},\ldots,x_{n-(m-1)\tau}]\).
Takens embedding idea
Takens' theorem shows that, under generic conditions and with a sufficiently large embedding dimension, delay coordinates can preserve the qualitative geometry of the original attractor up to a smooth change of coordinates. In practice, the delay \(\tau\), embedding dimension \(m\), sampling, noise, and record length all matter.
Choosing the delay and dimension
If \(\tau\) is too small, successive coordinates are highly redundant and the reconstruction collapses near the diagonal. If it is too large, dynamically related points may become poorly coordinated. The dimension \(m\) must also be large enough to unfold self-overlaps caused by projection. Practical methods often use autocorrelation or mutual information for delay selection and false-nearest-neighbor ideas for choosing \(m\).
What should stay with you
- Delay embedding can recover state-space structure from one measured variable.
- The delay controls how different successive coordinates are.
- A sufficient embedding dimension helps unfold projection overlaps, but finite data and noise limit practical reconstruction quality.
What happens when the delay τ is chosen much too small?
With a very small delay, x(t) and x(t−τ) are nearly the same, so the reconstructed coordinates contain little new information.