Integrals; transport lags;disturbance; indeterminacy

[From Bill Powers (940413.1800 MDT)]

Martin Taylor (940412.1800) --

RE: integrals

The second equation above could equally well be written
do/dt = k*(r - p)

No it couldn't. The integral equation entails the differential
one, but the reverse is not true. It misses precisely the
constant of integration, which is where the effects of past
values of p are found.

The constant of integration has to do with the _initial_ value of
the output of the integral, not with past values of the input. The
output value is a consequence of past values of the integrand, but
being a single number it can't be decomposed into effects of
specific inputs at specific times in the past.

···

----------------------------------------

Let's change the reference signal by a step. What does e do? >It

goes up by a step, instantaneously.
....

The rates are all based on current values, but the levels are >not.

They are based on what has happened in the past.

What does "based on" mean? This is not a mathematical or scientific
term, but a term from informal verbal reasoning. Are you trying to
say that from the present state of the output, it is possible to
deduce what the input was at any particular time in the past? Are
you saying that given the input at a particular time, it is possible
to predict the state of the integral at some future time? If you're
saying either of these things, you are wrong. If you're not, I can't
imagine what you mean by "based on," which implies some particular
relationship between the integral's output and a previous input. An
integration erases all information about specific values of the
input in the past, and preserves only the cumulative sum of effects.
All information about particular occurrances in the past is lost.

If the output function is an integrator, it can take ANY value,
given only the current values of p and r.... Do you really intend

to deny this?

No. It is also true that the output of the integrator can take ANY
value, given only the value of p and r at ANY one time, present or
past.
----------------------------------------
RE: transport lags

What goes into the input now may not affect what comes out of
the output until much later, or it may affect the output almost
immediately.

That would be a very peculiar transport lag, one in which the delay
for any given input signal is unpredictable. I don't think that such
a device is physically realizable. In all real processes involving a
transport lag, the delay between an input and its effects at the
output is one fixed number set by the physical properties of the
device, such as the length of an axon. What goes into the input will
come out of the output exactly tau seconds later, neither more nor
less. In the type of transport lag you seem to be imagining, you
couldn't even predict the order in which input impulses would appear
at the output.

Is this why you have been speaking about "convolutions" (actually,
transport lags -- you haven't described any convolutions) as if they
introduced some sort of randomness into a process? If so, you've
been suffering a misconception.

We are each trying to get the other to see that the variables
that are effective NOW are those that exist NOW at the various
places in the loop. I think we agree on that, but not on its
implications. To me, the implication is that what exists NOW
in the loop can in no way be affected by what exists NOW at any
other place in the loop, and may be most strongly affected by
what existed some time ago at a different place in the loop.

If you have defined the functions correctly, this will take care of
itself. Trying to analyze how one variable depends on past values of
other variables is guaranteed to lead to confusion; the sequential
view is simply not viable. After you have seen how the whole loop
looks in continuous operation, you can go back and look at time or
phase relationships. If you try to start with the time or phase
relationships, you will start chasing cause and effect around and
around the loop, which is a blind alley first explored by the
ancient Greeks or earlier peoples, Zeno among them.
-----------------------------------------------
RE: defining the disturbance

(I'll call it the "wydhawf." The wydhawf is the value of the
output of the disturbance function.)

...

I think we need a word for the output of the disturbance
function, which is in the dimensions of the CEV, just as is the
output of the feedback function.

...

What I want a word for is the value of H(d). It isn't the
fluctuation of the CEV, because that is dC/dt, which is the sum
of the derivatives of the other two, if the effect is linear.
Above, I used "wydhawf," but since there is already a word for
"disturbing variable," why not use the more familiar
"disturbance," as I had (apparently wrongly) thought we
previously agreed?

Because the disturbing variable and the word "disturbance" have been
defined since the beginning of PCT as d, not H(d). Except where I
have accidentally been slipshod,l I have used the same term in the
same way consistently.

But if you really want to use two words for one thing, I'll try >to

remember to use "wydhawf" in future--if I can remember it

tomorrow.

They are two words for two different things. H(d) is not d. You have
allowed yourself to be misled by a specific case, the case seen in
our tracking experiments where H is a multiplier of 1. In that case
there is a trivially simple relationship between the value of the
disturbance and its contribution to the state of the controlled
variable. But the general case is the one in which H is a general
function, not a constant of proportionality. H(d) could be, for
example d^2 - 200 + dd/dt, in which case it would be obvious that
the value of H(d) is not d. In the case of a crosswind, d would be
the velocity of the wind, while H(d) would be a complex nonlinear
function of the velocity of the car, the velocity of the wind, and
the angle between them.

If you decide to use "disturbance" for H(d) despite my steadfast
objections, which have never changed, you will be no better off than
before, because as you point out, H(d) is NOT the observed state of
the controlled variable. The actual state is produced by the sum of
H(d) and G(o). So "the disturbance" would still not be a fluctuation
in the controlled variable, and would still not be directly
detectable by the input function.

If you refer to H(d) as "the disturbance," you will leave the
impression that the disturbance is the same as the change in the
controlled variable, which it is not. So you gain nothing with
respect to popular usages.

But I do not accept this usage of "the disturbance" because to do so
would require going back and rewriting everything I have written
about the disturbance for 40 years (which you seem not to have
read). The term "disturbance" has been defined in PCT; it already
has a meaning; it has been used up. You can say "wydhawf" for the
quantity you mean, but H(d) is much simpler and conveys its meaning
directly. H(d) represents the _effect_ of d, the _contribution_ of d
to the state of the controlled variable, as I have put it many times
before. You can't waltz into this game and arbitrarily start
changing the rules without obtaining the consent of the other
players.

You follow this with a discussion of why the wydhawf is an
unnecessary construct, a computational fiction. I don't think
it is unnecessary, any more than is the error signal, which may
not exist as a measurable quantity in a real control system,
such as one in which the output function is based on a
differential input Op-Amp.

I said it was a computational fiction, which it is, but not that it
is unnecessary. I use H(d) (or more often, D(d)) to indicate one of
two converging effects on the controlled variable. In our equations,
you will quite often find H(d) -- Rick employs that notation
consistently. It is necessary to indicate H(d) somehow. But this
does not mean that there is a physically separate H(d)
distinguishable from G(o).

In the case of the massive controlled position variable, the
position is given, simplified, by

cv = int(int((o+d)/m),

with m being the mass. This can be decomposed into

cv = int(int(o/m)) + int(int(d/m)).

Then H(d) is int(int(()/m) where () is the placeholder where the
disturbance goes. But in fact, the two forces add to produce a net
force, and that force produces a single acceleration of a single
object, and the acceleration of that object is integrated twice to
produce the position that the input function senses. The first
representation above reflects, in the grammar of algebra, the actual
physical process. The second creates an impression of two apparent
processes where only one exists. The two representations are
conceptually, but not physically, equivalent. Given only the second
expression, one could make the mistake of thinking that in general a
different mass could be involved in the two terms.

As to the error signal, if the reference signal enters a physically
distinct comparator, an explicit error signal is required. If
instead it enters an output function, no error signal is required
but we would still want to identify the important quantity r - p,
which we could do simply by writing
(r - p). H(d) has the same kind of existence as the error signal,
except that there are hardly any cases, perhaps no cases, where H(d)
could be examined by itself, whereas there are known control systems
where the error signal can be identified.

The wydhawf, as I once before described it, is the effect the
disturbance would have, were it not opposed by the output
effect (which is exactly as unobservable).

I have long defined H(d) in exactly the same way. So what? The
effect that the disturbance would have if it were not opposed by the
output effect is not what we see in an intact control system, nor
what the input function senses.

Even knowing the effect that the disturbance would have in the
absence of an opposing output does not permit you to deduce the
disturbance from observing the now-uncontrolled variable. That is
because H(d) is not d. Neither is it H. You can't deduce either the
form of a function or the values of its arguments by looking only at
the value of the function.

Let me put that another way:

YOU CAN'T DEDUCE EITHER THE FORM OF A FUNCTION OR THE VALUES OF ITS
ARGUMENTS BY LOOKING ONLY AT THE VALUE OF THE FUNCTION.
--------------------------------------------
RE: indeterminacy

I said

just as obviously, it DOES NOT contain information about the
latter because the relationship is indeterminate.

You said

It is quite wrong to say that an indeterminate relationship
eliminatesinformation. It may, but more commonly it reduces
rather than eliminates information. I quote from my previous
posting:

This is an example of a function of the
form X = F(S+N).It is wrong to say of such a function that X
_inherently_ is independent of S or of N.

...

The same holds if N is partitioned into N1, N2, N3, ..., Nk.

But this says that you know that there are k N's, and that there are
none in addition that you don't know about. I am talking about the
kind of mathematical indeterminacy that is involved in being asked
to calculate the area of a rectangle, and being given only the
length of one side. Knowledge of that one side does not in the
slightest reduce your uncertainty about the area. That is what I
mean by "indeterminate" as opposed to "uncertain."

As far as a control system is concerned, the "most common" case is
the one where the system has no indication of the form of H or the
value of d, and where even the analyst is incapable of anticipating
all possible causes of disturbances and the paths through which they
will act. Fortunately, the control system does not have to know
either H or d to control successfully, because its actions work
directly on the controlled variable, and are based entirely on
perception of the controlled variable and the setting of the
reference signal. Identical control actions can follow upon an
infinite variety of different H's and d's, so knowledge of H and d
can't be important in achieving control. It is impossible for the
perceptual signal to contain information about which set of H's and
d's happen to be operating at a given moment, because that set can
change without any change in the effects on the controlled variable
and hence on the perception.

A control system can be designed strictly on the basis of the
properties of the available components, without taking into
consideration the kind of disturbances that might arise. The
properties of the system, if utilized to the maximum, will determine
which disturbances can be resisted and which can't. Disturbance
waveforms, spectra, distributions, average values, or statistical
properties are irrelevant to the design, if the design is already
the best that can be obtained from the available materials.
-------------------------------------------------------------
Best,

Bill P.

<Martin Taylor 940414 15:30>

Bill Powers (940413.1800 MDT)

I may be only sporadically on-line for the next 4 or 5 weeks. We have had
a 2-week respite from the sleep study, which enabled me to address some
issues on CSG-L, but that has come to an end. In the week between and
the week after the next 2 sleep runs, I have workshops to attend, and
possibly also the next week as well (5th from now). So there won't be
much time for more than occasional comments here, if I even get time to
read mail. At present, there are still 120 items in my mailbox that I
had hoped to do something about, so I have to apologize to people who might
reasonably have expected a response and didn't get one. They probably never
will.

···

===================

Horrible misunderstandings on simple Maths. I don't know where they
come from, so it's hard to respond effectively.

An
integration erases all information about specific values of the
input in the past, and preserves only the cumulative sum of effects.
All information about particular occurrances in the past is lost.

But if any values in the past had been different from what they actually
were, the integral would have been different now. So their effect is
still there NOW. Nobody, I hope, thinks that any single number can
provide values for more than a single degree of freedom. For many
degrees of freedom, you need many numbers. So what?

The second equation above could equally well be written
do/dt = k*(r - p)

No it couldn't. The integral equation entails the differential
one, but the reverse is not true. It misses precisely the
constant of integration, which is where the effects of past
values of p are found.

The constant of integration has to do with the _initial_ value of
the output of the integral, not with past values of the input. The
output value is a consequence of past values of the integrand, but
being a single number it can't be decomposed into effects of
specific inputs at specific times in the past.

I don't really know how this response is supposed to relate to my comment,
but I'll try to deal with it, anyway.

An indefinite integral has a constant of integration, which is indeed
the initial value of the output, not involving values prior to the beginning
of the integration.

If you have an integral from a to b of some function, it has a defined value,
which is the value of the indefinite integral at b minus the value of the
integral at a. Since each of these has the same constant of integration,
it is cancelled out in the subtraction. The definite integral has a defined
value.

If you use the integral form of the control loop equations, the time
of integration goes from when the control loop started operation until
now. The integral is a definite integral starting from when the loop was
turned on. Or, if you want, from ANY prior moment until now. But it does
not start NOW. If NOW is taken as time 0, then the integration has to have
started at time t0, where t0 is some negative number, not 0. If t0=0,
the output of the integrator is exactly the constant of integration, and
there is no point in mentioning the existence of the integration function.

If you analyze the integration from some arbitrary t0, there is an initial
output value, o(t0), to which the integral between t0 and 0 must be added.
To do this is to state that whatever happened to the control system before
t0 is of no interest. The system might not have existed, and was placed in
the t0 state at t0 by some God-like experimenter. But after that, it did
exist, and if the output at t0 was far from balancing the wydhawf at t0,
there will be an initial error induced by that imbalance. The error will
be corrected by outputs that have some time course, but that integrate
to an output that does properly balance the disturbance after a while.
That time course of output will be different from what would happen if the
t0 state had been set up so that the t0 output properly balanced the
wydhawf. The initial constant of integration, PLUS the integrated
k*(r-p) has to balance the wydhawf if the perceptual signal is properly
under control.

None of this is apparent in the differential form, because the constant
of integration (and the value of the current integral) is not represented.

If the output function is an integrator, it can take ANY value,
given only the current values of p and r.... Do you really intend
to deny this?

No. It is also true that the output of the integrator can take ANY
value, given only the value of p and r at ANY one time, present or
past.

Which seems to agree with the above.

----------------------------------------
RE: transport lags

What goes into the input now may not affect what comes out of
the output until much later, or it may affect the output almost
immediately.

That would be a very peculiar transport lag, one in which the delay
for any given input signal is unpredictable.

Given the context of this quote, reproduced below, I find that a remarkable
non-sequitur.

Even if Fo were defined as
SUM-over-tau{e(t-tau)*f(tau)}, as in the Artificial Cerebellum, only
the _present_ value of e enters that function, and only the
_present_ value of output comes out of it. ...
Everything that has to do with time-dependent effects takes place
INSIDE the functions, not outside them.

Yes, that's been at the heart of what I've been talking about. What goes
into the input now may not affect what comes out of the output until much
later, or it may affect the output almost immediately. It happens inside
the function, and what happens is represented in f(tau), whether it be an
immediate impulse function (not physically realizable), a unit step
(an integrator) or a unit step after a delay (an integrator with transport
lag). What comes out NOW is what is used now at the next stage in the
loop.

I don't know why you extracted that one sentence, and assumed it to refer
to a physically impossible device. I referred to the fact that time-dependent
effects occur inside a function, in total agreement with you.

Parenthetically, is it really necessary to find the most marginally plausible
interpretations of all my agreements with you? It seems to me that you
find it very difficult to accept that I know what you are talking about,
and agree with it in almost all respects. The places where we disagree
would be much more readily dealt with if you could accept the places
where we do agree, instead of twisting my words into ways that are very
hard to relate to the original intent.

Is this why you have been speaking about "convolutions" (actually,
transport lags -- you haven't described any convolutions)

Why do you make that parenthetical observation? I described several
specific convolutions, but only those used in your specific model--a
transport lag followed by an integrator. The transport lag is a delayed
impulse, so not a very interesting convolution, and the integrator is
a unit step function, as you earlier pointed out (as if I had not already
so described it). A convolution of the two is a delayed step function
(convolutions being associative, this is legitimate to do). I not only
described them, I graphed them in a second posting intended to clarify
the issue. I then suggested some perhaps more plausible forms, such as
that the integrator should be leaky (graphed as a step that decays
exponentially to zero after the initial step), a suggestion you said you had
tried, with some minor improvement in the results.

Is it necessary to ignore the many kilobytes that have been devoted to
discussing various kinds of convolution kernels in order to say that
I deal only with transport lags of unpredictable length? It's very
hard to carry on a coherent technical discussion when so much gets
lost so quickly. It begins to make me concerned about the effort I
devote to this that could go to more productive work. If it weren't
for the importance I think PCT has, I wouldn't bother, and would just
go my own way (which might suit you very well). But I think I understand
enough to make it worthwhile trying to understand more. However, this
trying becomes very trying, sometimes, when so much is lost, not in
translation, but in so little time.

---------------

But if you really want to use two words for one thing, I'll try >to
remember to use "wydhawf" in future--if I can remember it
tomorrow.

They are two words for two different things. H(d) is not d. You have
allowed yourself to be misled by a specific case, the case seen in
our tracking experiments where H is a multiplier of 1.

As you are insisting on using them, "disturbance" and "disturbing
variable" are both "d". That's one thing. Two words.

I am not misled at all. I know very well that the disturbing variable
is not the effect on the CEV. And that it is not even in the same
dimension, whether the multiplier is 1 or not. That 1 is quite
adventitious, and signals ONLY a choice of units in the dimensions
concerned.

H(d) is NOT the observed state of
the controlled variable. The actual state is produced by the sum of
H(d) and G(o). So "the disturbance" would still not be a fluctuation
in the controlled variable, and would still not be directly
detectable by the input function.

Yes, yes...I know, I know....(Sigh).

If you refer to H(d) as "the disturbance," you will leave the
impression that the disturbance is the same as the change in the
controlled variable, which it is not. So you gain nothing with
respect to popular usages.

Now THAT, I can see as an issue. Leaving false impressions is a bad
idea, and for that reason I now agree that it is a good thing not to
use the term "disturbance" for H(d) or D(d).

But I do not accept this usage of "the disturbance" because to do so
would require going back and rewriting everything I have written
about the disturbance for 40 years (which you seem not to have
read).

Well, I've read BCP and everything in LCS and LCS-2. I don't suppose,
though, that I have read everything you have written about the disturbance
in the last 40 years. I'd be happy to read anything else you might send
me or give me an accessible reference to.

However, as I said before, I had been under the impression that you had
some time ago agreed that "disturbing variable" and "disturbance" were
better used for two different things: "disturbing variable" as "d" and
"disturbance" as H(d). It has been under that impression that I have
been consistently separating their usages in that manner. You say I
misinterpreted what you meant. Fine, I misinterpreted, and have changed
my usage. As you say, it's your bat and ball.

The term "disturbance" has been defined in PCT; it already
has a meaning; it has been used up. You can say "wydhawf" for the
quantity you mean, but H(d) is much simpler and conveys its meaning
directly.

Yep. I'd like a less unwieldy term, not a mathematical expression whose
interpretation depends on a knowledge of the notation of a set of
equations assumed to be known by the reader. Something that sounds like
English, and has the right connotations.

H(d) represents the _effect_ of d, the _contribution_ of d
to the state of the controlled variable, as I have put it many times
before.

So how about we use "disturbance effect" when using English, and H(d)
when we use equations?

You can't waltz into this game and arbitrarily start
changing the rules without obtaining the consent of the other
players.

Didn't think I changed any rules, and thought that the wording I was
using did indeed have the consent of the other main players. Or at
least of one of them. So I was wrong. Sorry.

----------------------

It is necessary to indicate H(d) somehow. But this
does not mean that there is a physically separate H(d)
distinguishable from G(o).

There's often no possibility of measuring directly H(d) and G(o)
(or shall we say, please, "the disturbance effect" and "the output effect"
on the CEV). Sometimes (and still not always) there is the possibility
of inferring the values of the two effects. However, in principle this
cannot be done by the control system that is doing the controlling, since
the two effects are combined into one number. Two degrees of freedom cannot
be described separately by one number. It may sometimes be done by other
independent measurement systems, even ones that observe only the perceptual
signal, if they know other things about the structure of the system.

One of the recurrent problems with our whole interaction (including Rick
in this) is the notion that somehow I believe that somewhere in the
perceptual signal there is a separate representation of the disturbance
effect as distinct from the CEV itself. That's not the case, and never
was. But we come back to "false impressions" again. Something in the
way I express things leads you to that impression, and a few postings
ago I hazarded a guess that it was the initial (trivial) demonstration
that the disturbance effect could be reconstructed from the perceptual
signal plus knowledge of various static parameters relating to the output
and feedback function. Maybe it wasn't that. But if not, I don't know
what it is.

In the case of the massive controlled position variable, the
position is given, simplified, by

cv = int(int((o+d)/m),

with m being the mass. This can be decomposed into

cv = int(int(o/m)) + int(int(d/m)).

Then H(d) is int(int(()/m) where () is the placeholder where the
disturbance goes.
...
The first
representation above reflects, in the grammar of algebra, the actual
physical process. The second creates an impression of two apparent
processes where only one exists. The two representations are
conceptually, but not physically, equivalent. Given only the second
expression, one could make the mistake of thinking that in general a
different mass could be involved in the two terms.

Another nice example of the Whorfian problem--the notation possibly
misleading one's understanding.
-------------

Even knowing the effect that the disturbance would have in the
absence of an opposing output does not permit you to deduce the
disturbance from observing the now-uncontrolled variable. That is
because H(d) is not d. Neither is it H. You can't deduce either the
form of a function or the values of its arguments by looking only at
the value of the function.

Let me put that another way:

YOU CAN'T DEDUCE EITHER THE FORM OF A FUNCTION OR THE VALUES OF ITS
ARGUMENTS BY LOOKING ONLY AT THE VALUE OF THE FUNCTION.

I don't know who wanted to, in the case of H and d, or who proposed that
one might be able to do what the capitalized sentence says can't be done.

Do you know any such person? If not, why waste effort on the argument?

-------------

This is an example of a function of the
form X = F(S+N).It is wrong to say of such a function that X
_inherently_ is independent of S or of N.

...

The same holds if N is partitioned into N1, N2, N3, ..., Nk.

But this says that you know that there are k N's, and that there are
none in addition that you don't know about.

It doesn't really. What I meant was that if you are to talk about the
relation between X and S, you don't need to know anything about how many
N's there are, or what they each contribute. In that sense, the
indeterminacy is of the kind you next describe.

I am talking about the
kind of mathematical indeterminacy that is involved in being asked
to calculate the area of a rectangle, and being given only the
length of one side. Knowledge of that one side does not in the
slightest reduce your uncertainty about the area. That is what I
mean by "indeterminate" as opposed to "uncertain."

Of course, mathematically, it doesn't constrain the area to know one
side of a rectangle. But under certain circumstances, such as knowing
that this object has been CALLED a rectangle rather than a square or
a thick line--meaning that the question was posed in ordinary conversational
context--it makes a difference to the probable areas if one is told that
one side is a micron, a metre, or a kilometre. So although the area is
technically indeterminate, nevertheless the uncertainty about its area
may be reduced by a knowledge of the length of one side.

As far as a control system is concerned, the "most common" case is
the one where the system has no indication of the form of H or the
value of d,

Right. That's normal, and I hope that in future nobody, including you
or Rick, will make the mistake of thinking that I have any interest in
either when we talk about information. (Although, to be complete, here:
if H is known to be monotonic and fixed, the amount of information about
H(d) in any other variable X is the same as the amount of information
in X about d itself. That fact just happens to be of no interest in the
discussion).

Fortunately, the control system does not have to know
either H or d to control successfully, because its actions work
directly on the controlled variable, and are based entirely on
perception of the controlled variable and the setting of the
reference signal. Identical control actions can follow upon an
infinite variety of different H's and d's, so knowledge of H and d
can't be important in achieving control.

Right, but different control actions follow on different H(d)'s.

A control system can be designed strictly on the basis of the
properties of the available components, without taking into
consideration the kind of disturbances that might arise.

Sure, and a bridge can be designed strictly on the basis of the
properties of the available components, without taking into
consideration the kind of loads that might arise.

But it might turn out to be an expensive four-lane bridge for an
occasional foot traveller, or an expensive disaster with big trucks at
the bottom of a river. It's unlikely to be just strong enough to carry
the largest traffic loads encountered, under the maximum wind stress
that occurs during its projected lifetime.

The
properties of the system, if utilized to the maximum, will determine
which disturbances can be resisted and which can't.

After the fact.

Disturbance
waveforms, spectra, distributions, average values, or statistical
properties are irrelevant to the design, if the design is already
the best that can be obtained from the available materials.

A very expensive design, that corrects against forces of megatons that
have rise times of nanoseconds. Much cheaper to look at what kinds of
disturbance waveforms, spectra, distributions, and statistical properties
are likely to occur, and build to deal with them as cheaply as possible
with available materials. That's what evolution must have done, or we
wouldn't be here. Something that dealt as effectively as us with the
same kinds of disturbance, but more cheaply, would probably be here
instead. But nothing that failed to deal with the kinds of disturbance
we have survived would be here, so we are good enough, most of the time.
Some of us are killed in accidents. Evolution doesn't require control
to work against all the disturbances we encounter.

Evolution isn't "survival of the fittest." It is reorganization, carried
on over very long periods of time. What works, stays. What works better
and costs less is likely to stay longer. Working means no more and no
less than passing on the genes that determine the design. Cheap and
effective is in. Faulty and expensive might stick around a while, but
it's more likely not to. Same with reorganization within the individual.
It's a matter of probabilities. Better design has a higher probability
of survival, but nothing is guaranteed.

Same with ideas.

Martin

[From Oded Maler (940415]

I'd love to enter the meta-discussion about the perception our
CEO Marken is controlling inside the PCT enterprise, alas,
I'm deadlining for an important conference, with a paper not yet written.
However the following line caught my eye, and I apologize in advance if my
comment is out of context:

* [From Bill Powers (940413.1800 MDT)]

···

*
* Martin Taylor (940412.1800) --
*
* RE: integrals
*
* >> The second equation above could equally well be written
* >>do/dt = k*(r - p)
*
* >No it couldn't. The integral equation entails the differential
* >one, but the reverse is not true. It misses precisely the
* >constant of integration, which is where the effects of past
* >values of p are found.
*
* The constant of integration has to do with the _initial_ value of
* the output of the integral, not with past values of the input. The
* output value is a consequence of past values of the integrand, but
* being a single number it can't be decomposed into effects of
* specific inputs at specific times in the past.
* ----------------------------------------
* >Let's change the reference signal by a step. What does e do? >It
* goes up by a step, instantaneously.
* ....
* >The rates are all based on current values, but the levels are >not.
* They are based on what has happened in the past.
*
* What does "based on" mean? This is not a mathematical or scientific
* term, but a term from informal verbal reasoning. Are you trying to
* say that from the present state of the output, it is possible to
* deduce what the input was at any particular time in the past? Are
* you saying that given the input at a particular time, it is possible
* to predict the state of the integral at some future time? If you're
* saying either of these things, you are wrong. If you're not, I can't
* imagine what you mean by "based on," which implies some particular
* relationship between the integral's output and a previous input. An
* integration erases all information about specific values of the
* input in the past, and preserves only the cumulative sum of effects.
* All information about particular occurrances in the past is lost.
  ^^^^^^^^^^^^^^^^^^

This is not my understanding of the information contained in an
integrated value. Suppose there is an unknown signal during an
interval [0,t]. Without any observation you may assign equal a-priori
probability to all possible signals on that interval. If however you
know that the integral at t is of some value V, the class of possible
signals decreases dramatically, and contains only those signals whose
integral from 0 to t is V. Hence you cannot say that it loses all
the information, because by knowing V you know more about the signal
than you knew without it. [This is in the framework of passive observation,
maybe in the control framework there is another aspect that I don't see].

Hope this doesn't disturb,

--Oded

--

Oded Maler, VERIMAG, Miniparc ZIRST, 38330 Montbonnot, France
Phone: 76909635 Fax: 76413620 e-mail: Oded.Maler@imag.fr