[Martin Taylor 980303 10:20]
(The date stamp is when I created the first draft of this message. I have
witheld
posting it because I am not 100% sure of its correctness. I post it now so that
people more mathematically adept than I can check its correctness, and so that
those with simulation data can see whether in fact the prediction is accurate,
even if it is mathematically flawed).
I have about 100 messages unread since Saturday, but I thought I'd better
send this
out anyway.
···
-----------------------
This message was stimulated by the following exchange between Bill Powers
and Bruce Abbott.
+[From Bill Powers (980226.0148 MST)]
+
+Bruce Abbott (980225.2005 EST)--
+
+>>You haven't yet replied to my description of the correlations that are
+>>actually observed, which contradict what your common sense is telling you.
+>
+>I thought that this was it:
+>
+>>>In an
+>>>ideal control system, this immediate response means that e would never get
+>>>large enough even to detect, even though d and o varied over a wide range.
+>>>At these levels of variation, _observed_ changes in e would constitute
+>>>nothing but measurement error, and variation in e would appear uncorrelated
+>>>(except by chance) with d and with o.
+
+This isn't the actual explanation of the correlation phenomenon to which I
+refer. It's not a question of measurement error in observing an ideal
+control system. We see this effect in real data, measures of definitely
+non-ideal human control. We can measure the cv and the output more
+accurately than they can be controlled (cv) or varied (output) by the
+organism. The real explanation lies in system noise -- the fact that
+outputs and other variables show small spontaneous variations uncorrelated
+with changes in the disturbance.
I wanted to see if it was possible to say anything about the correlation
between the
perceptual signal and the disturbance signal of a control loop in the
absence of system
noise. It turned out to be possible, unless I've made some serious mistake.
So here's
the message as originally drafted.
-----------------------
The recent reincarnation of the discussion on the relation between the
disturbance signal and the perceptual signal has made possible an
interesting analysis. We can compute, at
least for the simplest control loop with a perfect integrator output
function, the maximum
possible correlation between the two signals as a function of the degree of
control.
If my analysis is correct, this maximum is 1/CR, where CR is the control
ratio--the
ratio between the magnitude of the amplitude of the variation in the perceptual
signal in the absence of control to its amplitude when control is in operation.
The correlation would be lower, if there is noise in the system.
I'll make a large caveat here. To me, this analysis seems too simple, and
I'd be
grateful if anyone can find an error that _increases_ the maximum
correlation possible
between p and d in this idealized maximally simple control loop, or show by
simulation
that the actual correlation is higher than the analysis shows. I submit the
analysis
because I can't by myself, see anything wrong with it, other than
approximations that
work in the direction of increasing the maximum.
----------------------------
We start with the classic diagram of a control system:
perceptual signal (p) ^ | reference signal (r)
> V
>____________-___error signal(e)__
> comparator |
perceptual signal (p) | |
> output function (G)
input function |
> >output signal (o)
CCEV + -----environment function-------|
> >
disturbance signal (d)| |output signal (o)
^ V
And we will make the usual assumptions that are made when using the
expression "p = o + d". That is to say, all the functions in the loop
except the output function are unit transforms.
Since it is short, I will repeat the loop derivation I wrote for Rick.
p = o + d = Ge + d = G(r-p) + d = Gr - Gp + d (1)
which yields d = p + Gp - Gr
That's starting point 1.
In the equation, d = p + Gp - Gr, all the "variables" are actually Laplace
transforms, which can be treated as if they were algebraic variables if the
system is linear (an integrator is a linear system). A Laplace transform is
a way of describing a time function, to put it crudely. The output
function, G, is a pure integrator in the example I
analyze below.
------------------------
Starting point 2 comes from an interchange between Bill Powers and me,
summarized by Bill (980224.1755 MST):
If the output function is a pure integrator, all frequency components are
phase-shifted by >90 degrees, which reduces the correlation to zero. For a
system with a large leakage term, >the correlation would be high unless
there was noise added in the output function or >comparator.
For now, I'll consider only the case with a perfect integrator output
function, but the
argument works (giving a different final result) with any function. One
just has to
know the correlation between the function's input and output. For G a pure
integrator,
x is uncorrelated with Gx if Bill's and my speculation is correct. For
other functions,
the correlation is likely to depend on the frequency spectrum of the input
of the function.
-----------------------
The analysis:
Firstly, it may help you to understand the analysis if you note that
symbols like "p"
and "d" represent waveforms extended over a long (notionally infinite)
time. I use
the same symbol also to represent the Laplace transform of the signal. When
the signals
are treated as vectors and we talk of correlation, the time variation (the
waveform) is
what you should think of. In developing the equations (e.g. "d = p + Gp -
Gr") the
Laplace transforms are used. The two are essentially interchangeable in the
current
context, which is why I haven't worried about using a script font for the
transform
(as if I could, in an e-mail message:-).
If G = 0 (no control), then p = d. Clearly then the correlation between p
and d is 1.0.
To start the next phase of the derivation, consider the case of r = 0
forever. In this
case, Gr (the integral of r) is also zero forever, and we have
d = p + Gp
Now we use the assertion that a variable is uncorrelated with its integral.
The
variable "d" is composed of two orthogonal components, p and Gp (remember,
"G" is
the output function, which is a perfect integrator).
Gp * *d
> /
> /
> /
> /
> /
> /
> /
>/_______* p
The squares of the lengths of the vectors represents the mean-square
amplitude variation
in the signal values. When you think of the variations on this diagram that
occur as
the output gain changes, or as the reference signal is allowed to vary,
remember that
it is the "d" vector that stays constant, while the others may change their
magnitude
and direction--not the other way around.
The correlation between any two vectors is the cosine of the angle between
them. That
is why Gp and p are drawn at right angles. Their correlation is zero. If Gp
is large
compared to p, d has a correlation of nearly 1.0 with Gp and nearly zero
with p. Since
we are dealing only with the case in which the reference signal is fixed at
zero, the
output signal is Ge, which is -Gp. So the disturbance signal is correlated
almost
-1.0 with the output signal--as we know to be the case for good control.
The control ratio (CR) is the ratio between the fluctuations that would
occur in p in
the absence of control and the fluctuations in p when the perception is
controlled. In
other words, CR = d/p when the transform between qi and p is the unit
transform.
The cosine of an angle in a right-angled triangle is the length of the
opposite side
over the hypotenuse. That is to say Corr(d:p) = p/d.
From this, the correlation between d and p is 1/CR.
-------------------
Variation in the reference signal
Why did I say this was a maximum correlation rather than the precise
correlation?
In part this is because there is always noise, as Bill pointed out. But
more because
of the role of the reference signal. Above, we considered the case of r
permanently
zero. Now we let r vary, independently of d, of course.
The control system is linear. What this means is that the superposition
theorem holds--
the contributions of different components can be added in the time domain,
in the
frequency domain, and in the Laplace domain. In particular, the relation d
= p + Gp
can be used as a starting point, onto which can be added any effects due to the
variation of r. Most importantly, p and Gp will change when r varies.
Gp (the integral of p in the simple system we are analyzing) can, because
of superposition,
be divided into two parts, which I will label G_d.p and G_r.p. G_d.p is
just what
we had before, when r was fixed permanently at zero. G_r.p is the variation
in Gp that
is extra, due to the variation in r.
Now we can look at the full expressions that was shown at the head of this
message:
d = p + Gp - Gr
and rewrite it
d = p + G_d.p + (G_r.p - Gr)
The first part of this is exactly what we had before, when r was
permanently fixed
at zero. The part in brackets is the contribution of variation in r.
What does the part in brackets contribute to the correlation between d and
p? Since the
reference signal varies independently of the variation in the disturbance
signal, and
any contribution of (G_r.p - Gr) is due to the reference signal, that
contribution is
orthogonal to d (and to G_d.p). It cannot increase the correlation between
d and p,
except by accident over the (very) short term. Furthermore, the better the
control,
the more nearly does p match r, and therefore the more nearly does G_r.p
match Gr.
The two tend to cancel one another, so if they have an effect, it tends to
become
small when control is good.
What this means is that even when the reference signal is allowed to vary
freely, the
maximum correlation that should be observed between the perceptual signal
and the
disturbance signal is 1/CR.
-------------------------
Extension to more realistic control loops.
The above applies mathematically only to a control system that is linear,
has no loop
transport delay, has a perfect integrator as its output/feedback function,
and has no
other time-binding functions in the loop (i.e. all the other functions are
simple
summations or multiplications by constant factors). Most control systems
are not
like that. What then?
Some cases can be examined heuristically. For example, if the output
function G is a
leaky integrator rather than a perfect integrator, the angle between p and
Gp depends
on the frequency spectrum of p. If low frequencies dominate, then Gp
correlates well
with p, but if high frequencies dominate, G acts like a good integrator. So the
correlation between d and p can be greater than 1/CR if the output function
is a
leaky integrator.
If there is some loop transport delay, it has to be incorporated into the
initial
expression for d. It could all be included in the output function, G. With loop
delay, the correlation between the input to G and its output will vary between
positive and negative correlation as a function of the frequency of p. This
will
show up in a correlation between Gp and p that varies with the spectrum of
p. In the
diagram, Gp leans left and right as the frequency of d varies. If the
magnitude of
G is large enough, this can lead at some frequencies to p being larger than
d--there
is no control, and the cosine takes on an imaginary (or at least complex)
value; the
loop is oscillating, not controlling.
Other cases can be examined similarly, provided the system is composed of
linear
components that allow the use of Laplace transforms. And for some
non-linear systems
one can make reasonable heuristic approximations by appealing to
small-amplitude
linearity.
--------------------------
All of the above depends on the correctness of two things: (1) the
statement that the
output of an integrator is, in general, uncorrelated with its input, and
(2) whether
I have done the analysis correctly. As I said up front, the result seems
too simple
not to be well known, and from prior discussion it clearly is not well
known. I'm
sure Bill has masses of simulation data that could be used to test the result
claimed, even though simulation data necessarily include transport loop delay
(which I think tends on balance to lower the correlation, but in a way that
depends on the spectrum of p). Perhaps Bill's data could show that the
correlation
control loops between the disturbance signal and the perceptual signal in such
"perfect integrator" is characteristically greater than 1/CR. If it is, it
would
show that there is a flaw in the analysis somewhere, even though I can't see it
right now, and I've asked a couple of control engineers, who also seem to
think it
is OK.
Martin