\[\dot{x} = f(x),\]
which says that the rate of change of \(x\) is just some arbitrary function \(f(\cdot)\) of \(x\). In block diagram format, the system looks like this:
Specifically, let's let our nonlinear function take the form
System block diagram. \(1/s\) is an integrator. The derivative of \(x\) is its input, and \(x\) is its output. |
\[f(x) = e^{-x}-1.\]
Do we see this stable point?
System state trajectory over time for initial states scattered between -1.5 and 1.5 . |
What if we add some noise?
\[\dot{x} = f(x) + u \]
\[u \text{ is } \mathcal{N}(0, \sigma)\]
That is, we've added Gaussian white noise of variance \(\sigma^{2}\) to our system.
Our system block diagram now looks like this:
Block diagram of system with noise added to input of integrator. |
Do we still see the system go to same stable point?
System simulation with noise added to the input of the integrator. Noise is Gaussian white noise with variance indicated by \(\sigma\) above each subplot. |
It looks like the system goes to the same stable point at low levels of noise, but system get's pretty swamped with noise at higher variances, which makes it hard to tell if the system settles around the predicted stable point. Running the simulation many times and averaging across trials helps us see the long term behavior more clearly:
Each line is the average of 25 trials. |
That's better. It looks like the system settles, on average, at the predicted state for low noise levels, but at higher levels of noise, the system seems to be averaging at a higher state value. If we do the same averaging over trials but with much larger noise, we see this effect more clearly:
The system no longer averages out anywhere close to the predicted stable point without noise. The system doesn't even look like it reaches its average stable value by the end of the simulation for very high levels of noise. Why does the stable point shift?
Does the average stable point change because we are now inputting a low pass filtered version of the noise into the nonlinearity? I'm not sure of an intuitive way of understanding this yet.
Now let's tweak the system a little:
Does the average stable point change because we are now inputting a low pass filtered version of the noise into the nonlinearity? I'm not sure of an intuitive way of understanding this yet.
Now let's tweak the system a little:
\[\dot{x} = f(x+u)\]
Now, noise is passed through the nonlinearity before integration.Show system diagram.
Block diagram of system with noise added to the input to the nonlinearity. |
Running the simulation for this system yields:
System simulation with noise added to the input to the nonlinearity. Each line is a single trial, not averaged. |
Think about it this way. If you are the system, and you are at a state x, what do you see? Because of the noise you see the everything around you. So if we take the noise distribution and average each point of the phase portrait with our noise (i.e. convolve our system function \(f(\cdot)\) with the pdf (probability density function) of our noise), we get a better sense of the fixed points of our system:
Phase portrait of the the system after convolving the system function, \(f\), with the noise distribution pdf. |
\[f(x+u)=f(x)+f'(x)u+\frac{1}{2}f''(x)u^{2}+\frac{1}{3!}f^{(3)}(x)u^3+...\]
This is neat! We can analytically see that noise changes the phase portrait.
Taking the expected value,
\[E[f(x+u)]=f(x)+f'(x)E[u]+\frac{1}{2}f''(x)E[u^{2}]+\frac{1}{3!}f^{(3)}(x)E[u^3]+...\]
Since we're assuming \(u\) is symmetric about 0, the expected value of the odd powers of \(u\) are also 0, so
\[E[f(x+u)]=f(x)+\frac{1}{2}f''(x)E[u^{2}]+\frac{1}{4!}f^{(4)}(x)E[u^4]+...\].
Before moving forward, we can see here why the phase portrait changes with noise. When noise entering the nonlinearity, \(\dot{x}\) is no longer just \(f(\cdot)\), but \(f(\cdot)\) summed with its derivatives weighted by the higher order moments of \(u\). In our example case,
\[f(x)=e^{-x}-1\]
\[f'(x)=-e^{-x}\]
\[f''(x)=e^{-x}\]
\[f''(x)=e^{-x}\]
\[f^{(3)}(x)=-e^{-x}\]
\[f^{(4)}(x)=e^{-x}\]
\[\vdots\]
Notice a pattern? All of the odd ordered derivatives are the same, and all of the even ordered derivatives are the same! So for our example,
\[E[f(x+u)]=e^{-x}-1+\frac{1}{2}e^{-x}E[u^{2}]+\frac{1}{4!}e^{-x}E[u^4]+...\]
\[E[f(x+u)]=e^{-x}-1+e^{-x}\left(\frac{\sigma^{2}}{2}+\frac{3\sigma^{4}}{4!}+...\right)\]
What's different between the two noise cases?
Both noise cases change the dynamics of the system, but the system is more sensitive to noise input to the nonlinearity.
Keep in mind that in reality we can really see both kinds of noise:.
Block diagram of system with noise added to both |
I WANT TO SING ALONG
ReplyDelete