Related
Would this be a valid implementation of a cross entropy loss that takes the ordinal structure of the GT y into consideration? y_hat is the prediction from a neural network.
ce_loss = F.cross_entropy(y_hat, y, reduction="none")
distance_weight = torch.abs(y_hat.argmax(1) - y) + 1
ordinal_ce_loss = torch.mean(distance_weight * ce_loss)
I'll attempt to answer this question by first fully defining the task, since the question is a bit sparse on details.
I have a set of ordinal classes (e.g. first, second, third, fourth,
etc.) and I would like to predict the class of each data example from
among this set. I would like to define an entropy-based loss-function
for this problem. I would like this loss function to weight the loss
between a predicted class torch.argmax(y_hat) and the true class y
according to the ordinal distance between the two classes. Does the
given loss expression accomplish this?
Short answer: sure, it is "valid". You've roughly implemented L1-norm ordinal class weighting. I'd question whether this is truly the correct weighting strategy for this problem.
For instance, consider that for a true label n, the bin n response is weighted by 1, but the bin n+1 and n-1 responses are weighted by 2. This means that a lot more emphasis will be placed on NOT predicting false positives than on correctly predicting true positives, which may imbue your model with some strange bias.
It also means that examples on the edge will result in a larger total sum of weights, meaning that you'll be weighting examples where the true label is say "first" or "last" more highly than the intermediate classes. (Say you have 5 classes: 1,2,3,4,5. A true label of 1 will require distance_weight of [1,2,3,4,5], the sum of which is 15. A true label of 3 will require distance_weight of [3,2,1,2,3], the sum of which is 11.
In general, classification problems and entropy-based losses are underpinned by the assumption that no set of classes or categories is any more or less related than any other set of classes. In essence, the input data is embedded into an orthogonal feature space where each class represents one vector in the basis. This is quite plainly a bad assumption in your case, meaning that this embedding space is probably not particularly elegant: thus, you have to correct for it with sort of a hack-y weight fix. And in general, this assumption of class non-correlation is probably not true in a great many classification problems (consider e.g. the classic ImageNet classification problem, wherein the class pairs [bus,car], and [bus,zebra] are treated as equally dissimilar. But this is probably a digression into the inherent lack of usefulness of strict ontological structuring of information which is outside the scope of this answer...)
Long Answer: I'd highly suggest moving into a space where the ordinal value you care about is instead expressed in a continuous space. (In the first, second, third example, you might for instance output a continuous value over the range [1,max_place]. This allows you to benefit from loss functions that already capture well the notion that predictions closer in an ordered space are better than predictions farther away in an ordered space (e.g. MSE, Smooth-L1, etc.)
Let's consider one more time the case of the [first,second,third,etc.] ordinal class example, and say that we are trying to predict the places of a set of runners in a race. Consider two races, one in which the first place runner wins by 30% relative to the second place runner, and the second in which the first place runner wins by only 1%. This nuance is entirely discarded by the ordinal discrete classification. In essence, the selection of an ordinal set of classes truncates the amount of information conveyed in the prediction, which means not only that the final prediction is less useful, but also that the loss function encodes this strange truncation and binarization, which is then reflected (perhaps harmfully) in the learned model. This problem could likely be much more elegantly solved by regressing the finishing position, or perhaps instead by regressing the finishing time, of each athlete, and then performing the final ordinal classification into places OUTSIDE of the network training.
In conclusion, you might expect a well-trained ordinal classifier to produce essentially a normal distribution of responses across the class bins, with the distribution peak on the true value: a binned discretization of a space that almost certainly could, and likely should, be treated as a continuous space.
Pattern matching (as found in e.g. Prolog, the ML family languages and various expert system shells) normally operates by matching a query against data element by element in strict order.
In domains like automated theorem proving, however, there is a requirement to take into account that some operators are associative and commutative. Suppose we have data
A or B or C
and query
C or $X
Going by surface syntax this doesn't match, but logically it should match with $X bound to A or B because or is associative and commutative.
Is there any existing system, in any language, that does this sort of thing?
Associative-Commutative pattern matching has been around since 1981 and earlier, and is still a hot topic today.
There are lots of systems that implement this idea and make it useful; it means you can avoid write complicated pattern matches when associtivity or commutativity could be used to make the pattern match. Yes, it can be expensive; better the pattern matcher do this automatically, than you do it badly by hand.
You can see an example in a rewrite system for algebra and simple calculus implemented using our program transformation system. In this example, the symbolic language to be processed is defined by grammar rules, and those rules that have A-C properties are marked. Rewrites on trees produced by parsing the symbolic language are automatically extended to match.
The maude term rewriter implements associative and commutative pattern matching.
http://maude.cs.uiuc.edu/
I've never encountered such a thing, and I just had a more detailed look.
There is a sound computational reason for not implementing this by default - one has to essentially generate all combinations of the input before pattern matching, or you have to generate the full cross-product worth of match clauses.
I suspect that the usual way to implement this would be to simply write both patterns (in the binary case), i.e., have patterns for both C or $X and $X or C.
Depending on the underlying organisation of data (it's usually tuples), this pattern matching would involve rearranging the order of tuple elements, which would be weird (particularly in a strongly typed environment!). If it's lists instead, then you're on even shakier ground.
Incidentally, I suspect that the operation you fundamentally want is disjoint union patterns on sets, e.g.:
foo (Or ({C} disjointUnion {X})) = ...
The only programming environment I've seen that deals with sets in any detail would be Isabelle/HOL, and I'm still not sure that you can construct pattern matches over them.
EDIT: It looks like Isabelle's function functionality (rather than fun) will let you define complex non-constructor patterns, except then you have to prove that they are used consistently, and you can't use the code generator anymore.
EDIT 2: The way I implemented similar functionality over n commutative, associative and transitive operators was this:
My terms were of the form A | B | C | D, while queries were of the form B | C | $X, where $X was permitted to match zero or more things. I pre-sorted these using lexographic ordering, so that variables always occurred in the last position.
First, you construct all pairwise matches, ignoring variables for now, and recording those that match according to your rules.
{ (B,B), (C,C) }
If you treat this as a bipartite graph, then you are essentially doing a perfect marriage problem. There exist fast algorithms for finding these.
Assuming you find one, then you gather up everything that does not appear on the left-hand side of your relation (in this example, A and D), and you stuff them into the variable $X, and your match is complete. Obviously you can fail at any stage here, but this will mostly happen if there is no variable free on the RHS, or if there exists a constructor on the LHS that is not matched by anything (preventing you from finding a perfect match).
Sorry if this is a bit muddled. It's been a while since I wrote this code, but I hope this helps you, even a little bit!
For the record, this might not be a good approach in all cases. I had very complex notions of 'match' on subterms (i.e., not simple equality), and so building sets or anything would not have worked. Maybe that'll work in your case though and you can compute disjoint unions directly.
The terms do appear to be defined differently, but I've always thought of one implying the other; I can't think of any case when an expression is referentially transparent but not pure, or vice-versa.
Wikipedia maintains separate articles for these concepts and says:
From Referential transparency:
If all functions involved in the
expression are pure functions, then
the expression is referentially
transparent. Also, some impure
functions can be included in the
expression if their values are
discarded and their side effects are
insignificant.
From Pure expressions:
Pure functions are required to
construct pure expressions. [...] Pure
expressions are often referred to as
being referentially transparent.
I find these statements confusing. If the side effects from a so-called "impure function" are insignificant enough to allow not performing them (i.e. replace a call to such a function with its value) without materially changing the program, it's the same as if it were pure in the first place, isn't it?
Is there a simpler way to understand the differences between a pure expression and a referentially transparent one, if any? If there is a difference, an example expression that clearly demonstrates it would be appreciated.
If I gather in one place any three theorists of my acquaintance, at least two of them disagree on the meaning of the term "referential transparency." And when I was a young student, a mentor of mine gave me a paper explaining that even if you consider only the professional literature, the phrase "referentially transparent" is used to mean at least three different things. (Unfortunately that paper is somewhere in a box of reprints that have yet to be scanned. I searched Google Scholar for it but I had no success.)
I cannot inform you, but I can advise you to give up: Because even the tiny cadre of pointy-headed language theorists can't agree on what it means, the term "referentially transparent" is not useful. So don't use it.
P.S. On any topic to do with the semantics of programming languages, Wikipedia is unreliable. I have given up trying to fix it; the Wikipedian process seems to regard change and popular voting over stability and accuracy.
All pure functions are necessarily referentially transparent. Since, by definition, they cannot access anything other than what they are passed, their result must be fully determined by their arguments.
However, it is possible to have referentially transparent functions which are not pure. I can write a function which is given an int i, then generates a random number r, subtracts r from itself and places it in s, then returns i - s. Clearly this function is impure, because it is generating random numbers. However, it is referentially transparent. In this case, the example is silly and contrived. However, in, e.g., Haskell, the id function is of type a - > a whereas my stupidId function would be of type a -> IO a indicating that it makes use of side effects. When a programmer can guarantee through means of an external proof that their function is actually referentially transparent, then they can use unsafePerformIO to strip the IO back away from the type.
I'm somewhat unsure of the answer I give here, but surely somebody will point us in some direction. :-)
"Purity" is generally considered to mean "lack of side-effects". An expression is said to be pure if its evaluation lacks side-effects. What's a side-effect then? In a purely functional language, side-effect is anything that doesn't go by the simple beta-rule (the rule that to evaluate function application is the same as to substitute actual parameter for all free occurrences of the formal parameter).
For example, in a functional language with linear (or uniqueness, this distinction shouldn't bother at this moment) types some (controlled) mutation is allowed.
So I guess we have sorted out what "purity" and "side-effects" might be.
Referential transparency (according to the Wikipedia article you cited) means that variable can be replaced by the expression it denotes (abbreviates, stands for) without changing the meaning of the program at hand (btw, this is also a hard question to tackle, and I won't attempt to do so here). So, "purity" and "referential transparency" are indeed different things: "purity" is a property of some expression roughly means "doesn't produce side-effects when executed" whereas "referential transparency" is a property relating variable and expression that it stands for and means "variable can be replaced with what it denotes".
Hopefully this helps.
These slides from one ACCU2015 talk have a great summary on the topic of referential transparency.
From one of the slides:
A language is referentially transparent if (a)
every subexpression can be replaced by any other
that’s equal to it in value and (b) all occurrences of
an expression within a given context yield the
same value.
You can have, for instance, a function that logs its computation to the program standard output (so, it won't be a pure function), but you can replace calls for this function by a similar function that doesn't log its computation. Therefore, this function have the referential transparency property. But... the above definition is about languages, not expressions, as the slides emphasize.
[...] it's the same as if it were pure in the first place, isn't it?
From the definitions we have, no, it is not.
Is there a simpler way to understand the differences between a pure expression and a referentially transparent one, if any?
Try the slides I mentioned above.
The nice thing about standards is that there are so many of them to choose
from.
Andrew S. Tanenbaum.
...along with definitions of referential transparency:
from page 176 of Functional programming with Miranda by Ian Holyer:
8.1 Values and Behaviours
The most important property of the semantics of a pure functional language is that the declarative and operational views of the language coincide exactly, in the following way:
Every expression denotes a value, and there are valuescorresponding to all possible program behaviours. Thebehaviour produced by an expression in any context is completely determined by its value, and vice versa.
This principle, which is usually rather opaquely called referential transparency, can also be pictured in the following way:
and from Nondeterminism with Referential Transparency in Functional Programming Languages by F. Warren Burton:
[...] the property that an expression always has the same value in the same environment [...]
...for various other definitions, see Referential Transparency, Definiteness and Unfoldability by Harald Søndergaard and Peter Sestoft.
Instead, we'll begin with the concept of "purity". For the three of you who didn't know it already, the computer or device you're reading this on is a solid-state Turing machine, a model of computing intrinsically connected with effects. So every program, functional or otherwise, needs to use those effects To Get Things DoneTM.
What does this mean for purity? At the assembly-language level, which is the domain of the CPU, all programs are impure. If you're writing a program in assembly language, you're the one who is micro-managing the interplay between all those effects - and it's really tedious!
Most of the time, you're just instructing the CPU to move data around in the computer's memory, which only changes the contents of individual memory locations - nothing to see there! It's only when your instructions direct the CPU to e.g. write to video memory, that you observe a visible change (text appearing on the screen).
For our purposes here, we'll split effects into two coarse categories:
those involving I/O devices like screens, speakers, printers, VR-headsets, keyboards, mice, etc; commonly known as observable effects.
and the rest, which only ever change the contents of memory.
In this situation, purity just means the absence of those observable effects, the ones which cause a visible change to the environment of the running program, maybe even its host computer. It is definitely not the absence of all effects, otherwise we would have to replace our solid-state Turing machines!
Now for the question of 42 life, the Universe and everything what exactly is meant by the term "referential transparency" - instead of herding cats trying to bring theorists into agreement, let's just try to find the original meaning given to the term. Fortunately for us, the term frequently appears in the context of I/O in Haskell - we only need a relevant article...here's one: from the first page of Owen Stephen's Approaches to Functional I/O:
Referential transparency refers to the ability to replace a sub-expression with one of equal value, without changing the value of the outer expression. Originating from Quine the term was introduced to Computer Science by Strachey.
Following the references:
From page 9 of 39 in Christopher Strachey's Fundamental Concepts in Programming Languages:
One of the most useful properties of expressions is that called by Quine referential transparency. In essence this means that if we wish to find the value of an expression which contains a sub-expression, the only thing we need to know about the sub-expression is its value. Any other features of the sub-expression, such as its internal structure, the number and nature of its components, the order in which they are evaluated or the colour of the ink in which they are written, are irrelevant to the value of the main expression.
From page 163 of 314 in Willard Van Ormond Quine's Word and Object:
[...] Quotation, which thus interrupts the referential force of a term, may be said to fail of referential transparency2. [...] I call a mode of confinement Φ referentially transparent if, whenever an occurrence of a singular term t is purely referential in a term or sentence ψ(t), it is purely referential also in the containing term or sentence Φ(ψ(t)).
with the footnote:
2 The term is from Whitehead and Russell, 2d ed., vol. 1, p. 665.
Following that reference:
From page 709 of 719 in Principa Mathematica by Alfred North Whitehead and Bertrand Russell:
When an assertion occurs, it is made by means of a particular fact, which is an instance of the proposition asserted. But this particular fact is, so to speak, "transparent"; nothing is said about it, bit by means of it something is said about something else. It is the "transparent" quality which belongs to propositions as they occur in truth-functions.
Let's try to bring all that together:
Whitehead and Russell introduce the term "transparent";
Quine then defines the qualified term "referential transparency";
Strachey then adapts Quine's definition in defining the basics of programming languages.
So it's a choice between Quine's original or Strachey's adapted definition. You can try translating Quine's definition for yourself if you like - everyone who's ever contested the definition of "purely functional" might even enjoy the chance to debate something different like what "mode of containment" and "purely referential" really means...have fun! The rest of us will just accept that Strachey's definition is a little vague ("In essence [...]") and continue on:
One useful property of expressions is referential transparency. In essence this means that if we wish to find the value of an expression which contains a sub-expression,
the only thing we need to know about the sub-expression is its value. Any other features of the sub-expression, such as its internal structure, the number and nature of
its components, the order in which they are evaluated or the colour of the ink in which they are written, are irrelevant to the value of the main expression.
(emphasis by me.)
Regarding that description ("that if we wish to find the value of [...]"), a similar, but more concise statement is given by Peter Landin in The Next 700 Programming Languages:
the thing an expression denotes, i.e., its "value", depends only on the values of its sub-expressions, not on other properties of them.
Thus:
One useful property of expressions is referential transparency. In essence this means the thing an expression denotes, i.e., its "value", depends only on the values of its sub-expressions, not on other properties of them.
Strachey provides some examples:
(page 12 of 39)
We tend to assume automatically that the symbol x in an expression such as 3x2 + 2x + 17 stands for the same thing (or has the same value) on each occasion it occurs. This is the most important consequence of referential transparency and it is only in virtue of this property that we can use the where-clauses or λ-expressions described in the last section.
(and on page 16)
When the function is used (or called or applied) we write f[ε] where ε can be an expression. If we are using a referentially transparent language all we require to know about the expression ε in order to evaluate f[ε] is its value.
So referential transparency, by Strachey's original definition, implies purity - in the absence of an order of evaluation, observable and other effects are practically useless...
I'll quote John Mitchell Concept in programming language. He defines pure functional language has to pass declarative language test which is free from side-effects or lack of side effects.
"Within the scope of specific deceleration of x1,...,xn , all occurrence of an expression e containing only variables x1,...,xn have the same value."
In linguistics a name or noun phrase is considered referentially transparent if it may be replaced with the another noun phrase with same referent without changing the meaning of the sentence it contains.
Which in 1st case holds but in 2nd case it gets too weird.
Case 1:
"I saw Walter get into his new car."
And if Walter own a Centro then we could replace that in the given sentence as:
"I saw Walter get into his Centro"
Contrary to first :
Case #2 : He was called William Rufus because of his read beard.
Rufus means somewhat red and reference was to William IV of England.
"He was called William IV because of his read beard." looks too awkward.
Traditional way to say is, a language is referentially transparent if we may replace one expression with another of equal value anywhere in the program without changing the meaning of the program.
So, referential transparency is a property of pure functional language.
And if your program is free from side effects then this property will hold.
So give it up is awesome advice but get it on might also look good in this context.
Pure functions are those that return the same value on every call, and do not have side effects.
Referential transparency means that you can replace a bound variable with its value and still receive the same output.
Both pure and referentially transparent:
def f1(x):
t1 = 3 * x
t2 = 6
return t1 + t2
Why is this pure?
Because it is a function of only the input x and has no side-effects.
Why is this referentially transparent?
You could replace t1 and t2 in f1 with their respective right hand sides in the return statement, as follows
def f2(x):
return 3 * x + 6
and f2 will still always return the same result as f1 in every case.
Pure, but not referentially transparent:
Let's modify f1 as follows:
def f3(x):
t1 = 3 * x
t2 = 6
x = 10
return t1 + t2
Let us try the same trick again by replacing t1 and t2 with their right hand sides, and see if it is an equivalent definition of f3.
def f4(x):
x = 10
return 3 * x + 6
We can easily observe that f3 and f4 are not equivalent on replacing variables with their right hand sides / values. f3(1) would return 9 and f4(1) would return 36.
Referentially transparent, but not pure:
Simply modifying f1 to receive a non-local value of x, as follows:
def f5:
global x
t1 = 3 * x
t2 = 6
return t1 + t2
Performing the same replacement exercise from before shows that f5 is still referentially transparent. However, it is not pure because it is not a function of only the arguments passed to it.
Observing carefully, the reason we lose referential transparency moving from f3 to f4 is that x is modified. In the general case, making a variable final (or those familiar with Scala, using vals instead of vars) and using immutable objects can help keep a function referentially transparent. This makes them more like variables in the algebraic or mathematical sense, thus lending themselves better to formal verification.
Reading this question got me thinking: For a given function f, how can we know that a loop of this form:
while (x > 2)
x = f(x)
will stop for any value x? Is there some simple criterion?
(The fact that f(x) < x for x > 2 doesn't seem to help since the series may converge).
Specifically, can we prove this for sqrt and for log?
For these functions, a proof that ceil(f(x))<x for x > 2 would suffice. You could do one iteration -- to arrive at an integer number, and then proceed by simple induction.
For the general case, probably the best idea is to use well-founded induction to prove this property. However, as Moron pointed out in the comments, this could be impossible in the general case and the right ordering is, in many cases, quite hard to find.
Edit, in reply to Amnon's comment:
If you wanted to use well-founded induction, you would have to define another strict order, that would be well-founded. In case of the functions you mentioned this is not hard: you can take x << y if and only if ceil(x) < ceil(y), where << is a symbol for this new order. This order is of course well-founded on numbers greater then 2, and both sqrt and log are decreasing with respect to it -- so you can apply well-founded induction.
Of course, in general case such an order is much more difficult to find. This is also related, in some way, to total correctness assertions in Hoare logic, where you need to guarantee similar obligations on each loop construct.
There's a general theorem for when then sequence of iterations will converge. (A convergent sequence may not stop in a finite number of steps, but it is getting closer to a target. You can get as close to the target as you like by going far enough out in the sequence.)
The sequence x, f(x), f(f(x)), ... will converge if f is a contraction mapping. That is, there exists a positive constant k < 1 such that for all x and y, |f(x) - f(y)| <= k |x-y|.
(The fact that f(x) < x for x > 2 doesn't seem to help since the series may converge).
If we're talking about floats here, that's not true. If for all x > n f(x) is strictly less than x, it will reach n at some point (because there's only a limited number of floating point values between any two numbers).
Of course this means you need to prove that f(x) is actually less than x using floating point arithmetic (i.e. proving it is less than x mathematically does not suffice, because then f(x) = x may still be true with floats when the difference is not enough).
There is no general algorithm to determine whether a function f and a variable x will end or not in that loop. The Halting problem is reducible to that problem.
For sqrt and log, we could safely do that because we happen to know the mathematical properties of those functions. Say, sqrt approaches 1, log eventually goes negative. So the condition x < 2 has to be false at some point.
Hope that helps.
In the general case, all that can be said is that the loop will terminate when it encounters xi≤2. That doesn't mean that the sequence will converge, nor does it even mean that it is bounded below 2. It only means that the sequence contains a value that is not greater than 2.
That said, any sequence containing a subsequence that converges to a value strictly less than two will (eventually) halt. That is the case for the sequence xi+1 = sqrt(xi), since x converges to 1. In the case of yi+1 = log(yi), it will contain a value less than 2 before becoming undefined for elements of R (though it is well defined on the extended complex plane, C*, but I don't think it will, in general converge except at any stable points that may exist (i.e. where z = log(z)). Ultimately what this means is that you need to perform some upfront analysis on the sequence to better understand its behavior.
The standard test for convergence of a sequence xi to a point z is that give ε > 0, there is an n such that for all i > n, |xi - z| < ε.
As an aside, consider the Mandelbrot Set, M. The test for a particular point c in C for an element in M is whether the sequence zi+1 = zi2 + c is unbounded, which occurs whenever there is a |zi| > 2. Some elements of M may converge (such as 0), but many do not (such as -1).
Sure. For all positive numbers x, the following inequality holds:
log(x) <= x - 1
(this is a pretty basic result from real analysis; it suffices to observe that the second derivative of log is always negative for all positive x, so the function is concave down, and that x-1 is tangent to the function at x = 1). From this it follows essentially immediately that your while loop must terminate within the first ceil(x) - 2 steps -- though in actuality it terminates much, much faster than that.
A similar argument will establish your result for f(x) = sqrt(x); specifically, you can use the fact that:
sqrt(x) <= x/(2 sqrt(2)) + 1/sqrt(2)
for all positive x.
If you're asking whether this result holds for actual programs, instead of mathematically, the answer is a little bit more nuanced, but not much. Basically, many languages don't actually have hard accuracy requirements for the log function, so if your particular language implementation had an absolutely terrible math library this property might fail to hold. That said, it would need to be a really, really terrible library; this property will hold for any reasonable implementation of log.
I suggest reading this wikipedia entry which provides useful pointers. Without additional knowledge about f, nothing can be said.
I'd like to write a program that lets users draw points, lines, and circles as though with a straightedge and compass. Then I want to be able to answer the question, "are these three points collinear?" To answer correctly, I need to avoid rounding error when calculating the points.
Is this possible? How can I represent the points in memory?
(I looked into some unusual numeric libraries, but I didn't find anything that claimed to offer both exact arithmetic and exact comparisons that are guaranteed to terminate.)
Yes.
I highly recommend Introduction to constructions, which is a good basic guide.
Basically you need to be able to compute with constructible numbers - numbers that are either rational, or of the form a + b sqrt(c) where a,b,c were previously created (see page 6 on that PDF). This could be done with algebraic data type (e.g. data C = Rational Integer Integer | Root C C C in Haskell, where Root a b c = a + b sqrt(c)). However, I don't know how to perform tests with that representation.
Two possible approaches are:
Constructible numbers are a subset of algebraic numbers, so you can use algebraic numbers.
All algebraic numbers can be represented using polynomials of whose they are roots. The operations are computable, so if you represent a number a with polynomial p and b with polynomial q (p(a) = q(b) = 0), then it is possible to find a polynomial r such that r(a+b) = 0. This is done in some CASes like Mathematica, example. See also: Computional algebraic number theory - chapter 4
Use Tarski's test and represent numbers. It is slow (doubly exponential or so), but works :) Example: to represent sqrt(2), use the formula x^2 - 2 && x > 0. You can write equations for lines there, check if points are colinear etc. See A suite of logic programs, including Tarski's test
If you turn to computable numbers, then equality, colinearity etc. get undecidable.
I think the only way this would be possible is if you used a symbolic representation,
as opposed to trying to represent coordinate values directly -- so you would have
to avoid trying to coerce values like sqrt(2) into some numerical format. You will
be dealing with irrational numbers that are not finitely representable in binary,
decimal, or any other positional notation.
To expand on Jim Lewis's answer slightly, if you want to operate on points that are constructible from the integers with exact arithmetic, you will need to be able to operate on representations of the form:
a + b sqrt(c)
where a, b, and c are either rational numbers, or representations in the form given above. Wikipedia has a pretty decent article on the subject of what points are constructible.
Answering the question of exact equality (as necessary to establish colinearity) with such representations is a rather tricky problem.
If you try to compare co-ordinates for your points, then you have a problem. Leaving aside co-linearity for a moment, how about just working out whether two points are the same or not?
Supposing that one has given co-ordinates, and the other is a compass-straightedge construction starting from certain other co-ordinates, you want to determine with certainty whether they're the same point or not. Either way is a theorem of Euclidean geometry, it's not something you can just measure. You can prove they aren't the same by spotting some difference in their co-ordinates (for example by computing decimal places of each until you encounter a difference). But in general to prove they are the same cannot be done by approximate methods. Compute as many decimal places as you like of some expansions of 1/sqrt(2) and sqrt(2)/2, and you can prove they're very close together but you won't ever prove they're equal. That takes algebra (or geometry).
Similarly, to show that three points are co-linear you will need theorem-proving software. Represent the points A, B, C by their constructions, and attempt to prove the theorem "A, B and C are colinear". This is very hard - your program will prove some theorems but not others. Much easier is to ask the user for a proof that they are co-linear, and then verify (or refute) that proof, but that's probably not what you want.
In general, constructable points may have an arbitrarily complex symbolic form, so you must use a symbolic representation to work them exactly. As Stephen Canon noted above, you often need numbers of the form a+b*sqrt(c), where a and b are rational and c is an integer. All numbers of this form form a closed set under arithmetic operations. I have written some C++ classes (see rational_radical1.h) to work with these numbers if that is all you need.
It is also possible to construct numbers which are sums of any number of terms of rational multiples of radicals. When dealing with more than a single radicand, the numbers are no longer closed under multiplication and division, so you will need to store them as variable length rational coefficient arrays. The time complexity of operations will then be quadratic in the number of terms.
To go even further, you can construct the square root of any given number, so you could potentially have nested square roots. Here, the representations must be tree-like structures to deal with root hierarchy. While difficult to implement, there is nothing in principle preventing you from working with these representations. I'm not sure just what additional numbers can be constructed, but beyond a certain point, your symbolic representation will be expressive enough to handle very large classes of numbers.
Addendum
Found this Google Books link.
If the grid axes are integer valued then the answer is fairly straight forward, the points are either exactly colinear or they are not.
Typically however, one works with real numbers (well, floating points) and then draws the rounded values on the screen which does exist in integer space. In this case you have no choice but to pick a tolerance and use it to determine colinearity. Keep it small and the users will never know the difference.
You seem to be asking, in effect, "Can the normal mathematics (integer or floating point) used by computers be made to represent real numbers perfectly, with no rounding errors?" And, of course, the answer to that is "No." If you want theoretical correctness, then you will be stuck with the much harder problem of symbolic manipulation and coding up the equivalent of the inferences that are done in geometry. (In short, I'm agreeing with Steve Jessop, above.)
Some thoughts in the hope that they might help.
The sort of constructions you're talking about will require multiplication and division, which means that to preserve exactness you'll have to use rational numbers, which are generally easy to implement on top of a suitable sort of big integer (i.e., of unbounded magnitude). (Common Lisp has these built-in, and there have to be other languages.)
Now, you need to represent square roots of arbitrary numbers, and these have to be mixed in.
Therefore, a number is one of: a rational number, a rational number multiplied by a square root of a rational number (or, alternately, just the square root of a rational), or a sum of numbers. In order to prove anything, you're going to have to get these numbers into some sort of canonical form, which for all I can figure offhand may be annoying and computationally expensive.
This of course means that the users will be restricted to rational points and cannot use arbitrary rotations, but that's probably not important.
I would recommend no to try to make it perfectly exact.
The first reason for this is what you are asking here, the rounding error and all that stuff that comes with floating point calculations.
The second one is that you have to round your input as the mouse and screen work with integers. So, initially all user input would be integers, and your output would be integers.
Beside, from a usability point of view, its easier to click in the neighborhood of another point (in a line for example) and that the interface consider you are clicking in the point itself.