8 Feedback loops

Feedback loops are usually encountered in an atomic physics lab when trying to stabi-
lize the frequency and intensity of lasers. As control engineering is an old profession
with its share of jargon and idiosyncracies, here is a short introduction to control
theory that is aimed at the problems most relevant to AMO folks. To make it easy to
follow up in the standard textbooks (e.g. [AM12]), I have retained the standard symbols.
Consider the block diagram of a laser system in Fig. 2. This can be recast in the form

Laser current current light Frequency voltage

Laser hardware
controller discriminator

voltage Feedback -
controller
+

reference
setpoint

Figure 2: Block diagram of a laser system.

of the standard control engineering feedback loop in Fig. 3. The output y of the laser
system P is affected by environmental factors, represented collectively by the noise
term ne . The standard laser control problem is to stabilize the lasers output y against
fluctuations in the environmental variable ne .9 To do this, the output of the laser is
measured and compared to the reference setpoint r, and the error signal, e = r y,
is fed into the feedback controller C. In practice, the error signal that enters the
feedback controller is never noiseless there is always some piece of it that is due to
the measurement systems noise nm . Based on the error signal fed into it, the feedback
controller computes a correction u which is fed into the laser system P .

nm measurement ne environmental
noise noise

r + e u y
reference error control output
- controller process

Figure 3: Standard form of a feedback control loop.

9 The output y can refer to the intensity or to the frequency of the laser the analysis is general enough

to describe either frequency or intensity stabilization.

21
 As read off the control loops block diagram, the output of the laser is

y = P [C(r y + nm ) + ne ] (70)

where P (u) and C(e) are the laser and feedback controllers response functions re-
spectively. In particular, P (u) is never really known explicitly.10 However, thanks to
the amazing properties of feedback, the output can be stabilized even without knowing
all the gory details of the laser systems response P (u).11
The main simplification results from the axiom that everything is linear to first-order.
We can linearize the process and control response functions P, C to understand their
behaviour for small perturbations around the operating point. For example, the output
of the controller for a small change in the error signal is

u0 + u = C(e0 + e)
(71)
 
C
u e C e,
e e0

where the controllers action on fluctuations of the error signal is effectively just a
multiplicative gain C rather than a complicated non-linear function, and it can be
modeled as a simple amplifier block.12 This is why the use of linear loop analysis is
justified in most practical situations.

0 +
- process
controller

Figure 4: Linearized version of the standard feedback control loop.

So let us examine the feedback loop for small fluctuations, assuming that the
reference setpoint is fixed. The linearized equation for the feedback loop is y =
P[C(y + nm ) + ne ] which leads to

P PC
y = ne + nm . (72)
1 + PC 1 + PC

First, we observe that the loop gain L P C should never approach -1. If that
happens, the loop becomes unstable because the noise sources get severely amplified.
10 If P (u) is explicitly known, then a controller that is better tailored to the system than a general (e.g.

PID) controller can be used to push it efficiently to the reference setpoint. This is never really worthwhile
or practical though.
11 This should not be too surprising if you recall that you can often build perfectly good amplifier circuits

without really knowing an op-amps open loop gain explicitly.

12 The gain C = C(f ) can still be an arbitrary function of fourier frequency i.e. the linearized response of

the system can be different for fluctuations at different frequencies. Only its dependence on e has been
linearized.

22
Figure 5: Nyquist plots showing two examples of unstable loops. For robust loop stability the gain margin
G and the phase margin should both be positive.

D
te
rm
do
s m
e in
at at
es
in
dom
m
er
It

 
1
Figure 6: Noise filtering response of a PID loop, with C(f ) = K 1 + j2f Ti
+ j2f Td . The plot
shows the noise response for a loop with K=1, Ti = 1 ms, Td = 5 ms.

Recall that the transfer functions P, C are frequency-dependent complex numbers, as

they can affect both the magnitude and phase of their inputs. Ensuring that L(f ) 6= 1
at all frequencies is the content of the Nyquist stability criterion.13 To check this,
we plot the path of L(f ) in the complex plane as f is varied. From such a plot, we can
also readily read off the gain and phase margin of the control loop (i.e. how well the
13 The practical criterion for robust loop stability is that the gain margin G and the phase margin are

both positive, to ensure that stability is not compromised by time delays, attenuation or transients when the
systems gain is turned on.

23
loop gain keeps away from -1 on the complex plane).
Next we see why a good feedback loop kills noise. In the limit that L(f )  1,
regardless of the systems response P, the output fluctuation is

ne (f )
y(f ) + nm (f ) . (73)
C(f )

The environmental noise ne (f ) is suppressed by a factor equal to the controllers gain

C(f ). In other words, 1/C acts as a filter on the noise. Figure 6 shows
 the noise filtering
1
response of a PID loop, where the controllers gain is C(f ) = K 1 + j2f Ti + j2f Td .
Note that the integral term ensures that the noise suppression is nominally infinite
as f 0.14 The eventual limit to the fluctuations in the lasers output is the un-
suppressed measurement noise nm (f ) that enters the error signal.

14 An offset is the same as noise at DC (f =0).

24