Related
I am trying to practice fitting with the Lsq-Fit-function in Julia.
The derivative of a Cauchy-distribution with parameters \gamma and x_0.
Following this manual I tried
f(x, x_0, γ) = -2*(x - x_0)*(π * γ^3 * (1 + ((x - x_0)/γ)^2)^2)^(-1)
x_0 = 3350
γ = 50
xarr = range(3000, length = 5000, stop = 4000)
yarr = [f(x, x_0, γ) for x in xarr]
using LsqFit
# p ≡ [x_0, γ]
model(x, p) = -2*(x - p[1])*(π * (p[2])^3 * (1 + ((x - p[1])/p[2])^2)^2)^(-1)
p0 = [3349, 49]
curve_fit(model, xarr, yarr, p0)
param = fit.param
... and it does not work, giving a MethodError: no method matching -(::StepRangeLen[...], leaving me confused.
Can please somebody tell me what I am doing wrong?
There are a few issues with what you've written:
the model function is meant to be called with its first argument (x) being the full vector of independent variables, not just one value. This is where the error you mention comes from:
julia> model(x, p) = -2*(x - p[1])*(π * (p[2])^3 * (1 + ((x - p[1])/p[2])^2)^2)^(-1);
julia> p0 = [3349, 49];
julia> model(xarr, p0);
ERROR: MethodError: no method matching -(::StepRangeLen{Float64,Base.TwicePrecision{Float64},Base.TwicePrecision{Float64}}, ::Float64)
One way to fix this is to use the dot notation to broadcast all operators so that they work elementwise:
julia> model(x, p) = -2*(x .- p[1]) ./ (π * (p[2])^3 * (1 .+ ((x .- p[1])/p[2]).^2).^2);
julia> model(xarr, p0); # => No error
but if this is too tedious you can let the #. macro do the work for you:
# just put #. in front of the expression to transform every
# occurrence of a-b into a.-b (and likewise for all operators)
# which means to compute the operation elementwise
julia> model(x, p) = #. -2*(x - p[1])*(π * (p[2])^3 * (1 + ((x - p[1])/p[2])^2)^2)^(-1);
julia> model(xarr, p0); # => No error
Another issue is that the parameters you're looking for are meant to be floating-point values. But your initial guess p0 is initialized with integers, which confuses curve_fit. There are two ways of fixing this. Either put floating-point values in p0:
julia> p0 = [3349.0, 49.0]
2-element Array{Float64,1}:
3349.0
49.0
or use a typed array initializer to specify explicitly the element type:
julia> p0 = Float64[3349, 49]
2-element Array{Float64,1}:
3349.0
49.0
This is not really an error, but I would find it more intuitive to compute a/b instead of a*b^(-1). Also, yarr can be computed with a simple broadcast using dot notation instead of a comprehension.
Wrapping all this together:
f(x, x_0, γ) = -2*(x - x_0)*(π * γ^3 * (1 + ((x - x_0)/γ)^2)^2)^(-1)
(x_0, γ) = (3350, 50)
xarr = range(3000, length = 5000, stop = 4000);
# use dot-notation to "broadcast" f and map it
# elementwise to elements of xarr
yarr = f.(xarr, x_0, γ);
using LsqFit
model(x, p) = #. -2*(x - p[1]) / (π * (p[2])^3 * (1 + ((x - p[1])/p[2])^2)^2)
p0 = Float64[3300, 10]
fit = curve_fit(model, xarr, yarr, p0)
yields:
julia> fit.param
2-element Array{Float64,1}:
3349.999986535933
49.99999203625603
I am trying to write a function that takes x and raises it to the power of n.
This code works if x and n are integers:
let rec pow x n =
if n == 0 then 1 else
if (n mod 2 = 0) then pow x (n/2) * pow x (n/2) else
x * pow x (n/2) * pow x (n/2);;
If I try to change the code to work if x is a float, it falls apart:
let rec float_pow x n =
if n == 0.0 then 1.0 else
if n mod_float 2.0 == 0.0 then float_pow x (n /. 2) *. float_pow x (n /. 2) else
x *. float_pow x (n /. 2) *. float_pow x (n /. 2);;
I get this error:
Error: This expression has type float
This is not a function; it cannot be applied.
What do I do?
The key problem, I think, is that mod is a keyword in OCaml, and is treated as an infix operator. But mod_float is just an ordinary function. You need to use it in prefix form.
So x mod n should be translated to mod_float x n.
You have another problem, which is that you're using the special-purpose == operator for equality comparison. You want to use = for equality comparisons in OCaml unless you need a "physical" comparison (which is not what you want here).
This isn't just stylistic--it really makes a difference. Note the following results:
# 0.0 == 0.0;;
- : bool = false
# 0.0 = 0.0;;
- : bool = true
I now would like to run tarai, which reads in Prolog as follows. A test case would be to run ?- tarai(12,6,0,X). This is quite a hard test case, for example GNU Prolog crashes with this test case.
tarai(X, Y, Z, R) :-
X > Y ->
X1 is max(0,X-1), tarai(X1, Y, Z, Rx),
Y1 is max(0,Y-1), tarai(Y1, Z, X, Ry),
Z1 is max(0,Z-1), tarai(Z1, X, Y, Rz),
tarai(Rx, Ry, Rz, R);
R = Y.
I am mostly interested whether the test case can be run over a fully declarative version of some miniKanren code for tarai. Optionally I would be interested in running some test cases backwards.
I am a little bit at loss. I managed to install guile, a scheme variant, and can run the miniKanren test cases successfully. But miniKanren has no integer numbers, so what could be done?
Note that this algorithm only has one solution if it's exists and is deterministic. You may pass scheme variables as x,y,z directly to the kanren tarai function, there is no unification for those hence implementing the logic for X,Y,Z can be done without kanren variables. The R values however needs to be logic variables and you need to get the bounded values in the tarai(Rx,Ry,Rz,R) form e.g. lookup the value of Rx,Ry,Rz and feed those into the tarai function. Also you need to make sure that this form is executed after the first three forms has finished (which is easy to do because there is no pure multiple choices) so that you know that Rx,Ry,Rz is bounded. Also note that this algorithm might depend on execution order in order to be executed efficiently but again the determinism means that this point is straightforward to satisfy. Note that A -> B ; C translates here simply to schemes (if A B C) because A = X > Y is deterministic. So the code could look something like this in pseudocode
(define (tarai x y z r)
(lambda ()
(fresh (rx ry rz)
(if (> x y)
(conda
( (conda ((tarai (- x 1) y z rx)
(tarai (- y 1) z x ry)
(tarai (- z 1) x y rz)))
(project (rx ry rz) (tarai rx ry rz r))))
(== r y)))))
To do the deterministic verison in guile-log scheme macro interface see links on this site you can implement it with memoization as (without memoization the solution blows the variable stack on guile-log)
(define tarai
(memo
(<lambda> (x y z r)
(if (> x y)
(<var> (rx ry rz)
(<and>
(tarai (max (- x 1) 0) y z rx)
(tarai (max (- y 1) 0) z x ry)
(tarai (max (- z 1) 0) x y rz)
(tarai (<lookup> rx) (<lookup> ry) (<lookup> rz) r)))
(<=> r y)))))
scheme#(guile-user)> ,time (<run> 1 (r) (tarai 12 6 0 r))
$13 = (12)
;; 0.293411s real time, 0.290711s run time. 0.000000s spent in GC.
scheme#(guile-user)>
I tried running for example tarai(12,X,0,12) with clpfd, but it's to complex for it. And memoizing does not work with attributed variables for swipl. So the best solution atm that I can find is to use the deterministic memoized tarai with something like
tarai2(X,Y,Z,W) :-
(var(X)->between(0,20,X);true),
(var(Y)->between(0,20,Y);true),
(var(z)->between(0,20,Z);true),
tarai(X,Y,Z,W).
then all solutions Y,Z in that range with tarai2(12,Y,Z,12) can easilly be found.
The question was refrased to ask for how to implement a more general version of the tarai function in the prolog spec that allows variables in x,y,z fields. The technique here can be implemented in prolog with e.g. clpfd finite domain solver and something similar is needed for kanren (se comment discussions above e.g. the reference to numbers.scm). The key thing is to nuke -> and use guards for all cases and we will assume that operators >o =o <=o is all defined for variables (e.g. for var X, X > 0 will constrain X to be values 1,2,3,...). Also we will asume that '-o' is defined for variables as well of this kind using interval arithmetics constraints via a special 'iso'. Using this we can define tarai as the code below (this can be simplified if max and min is also defined as constraints but here we implement those via inequalities and a dived of cases in stead).
(define (taray x y z w)
(lambda ()
(conde ((<o x y)
(fresh (rx ry rz)
(conde
((conde
((>o x 0)
(fresh (xx)
(conde
((iso xx (-o x 1))
(tarai xx y z rx)))))
((== x 0) (tarai 0 y z rx)))
(conde
((>o y 0)
(fresh (yy)
(conde
((iso yy (-o y 1))
(tarai yy z x ry)))))
((== y 0) (tarai 0 z x ry)))
(conde
((>o z 0)
(fresh (zz)
(conde
((iso zz (-o z 1))
(tarai zz x y rz)))))
((== z 0) (tarai 0 x y r<)))
(tarai rx ry rz r)))))
((>=o x y) (== x y))))))
I'm the developer of guile-log a logic programming environment on guile scheme that has both minikanren constructs and prolog constructs and one can mix them. It also has a clpfd library ported so here you can do the following (unfourtunately atm it does not work (a bug i'm working on)). Assume that clpfd is imported. (,, ;; is interleaving kanren like ops). Replacing ,, with , and ;; with ; you get code that can run on for example swi prolog using the clpfd library.
tarai(X,Y,Z,W) :-
(
X #> Y ,
(
(
((X #> 0 , XX #= X - 1, tarai(XX,Y,Z,RX)) ;;
(X = 0 , tarai(0,Y,Z,RX))) ,,
((Y #> 0 , YY #= Y - 1, tarai(YY,Z,X,RY)) ;;
(Y = 0 , tarai(0,Z,X,RY))) ,,
((Z #> 0 , ZZ #= Z - 1, tarai(ZZ,X,Y,RZ)) ;;
(Z = 0 , tarai(0,X,Y,RZ)))
) ,,
tarai(RX,TY,RZ,R)
)
) ;;
(X #=< Y, R=Y).
This is a slow implementation that is declarative and uses numbers.scm (as suggested by the author of the question) and minkanren.
(load "minkanren.scm)
(load "numbers.scm")
(define one (build-num 1))
(define zero (build-num 0))
(define (tarai x y z r)
(conde
((<=o x y) (== r y))
((<o y x)
(fresh (rx ry rz)
(conde
((fresh (xx)
(conde
((<o zero x) (minuso x one xx) (tarai xx y z rx))
((== zero x) (tarai zero y z rx))))
(fresh (yy)
(conde
((<o zero y) (minuso y one yy) (tarai yy z x ry))
((== zero y) (tarai zero z x ry))))
(fresh (zz)
(conde
((<o zero z) (minuso z one zz) (tarai zz x y rz))
((== zero z) (tarai zero x y rz))))
(tarai rx ry rz r)))))))
If the derivation direction is reversed (so wee add numbers) we can still solve due to the declarative style,
tarai(4,2,Z,4):, (you need to make binaries of the number) will lead to
Y = [0, 1],
W = [0, 0, 1],
Z = [0, 0, 1],
X = [0, 0, 1].
And supposedly if tabling is added:
tarai(12,6,0,W):
Z = (),
X = [0, 0, 1, 1],
Y = [0, 1, 1],
W = [0, 0, 1, 1].
Is it safe to replace a/(b*c) with a/b/c when using integer-division on positive integers a,b,c, or am I at risk losing information?
I did some random tests and couldn't find an example of a/(b*c) != a/b/c, so I'm pretty sure it's safe but not quite sure how to prove it.
Thank you.
Mathematics
As mathematical expressions, ⌊a/(bc)⌋ and ⌊⌊a/b⌋/c⌋ are equivalent whenever b is nonzero and c is a positive integer (and in particular for positive integers a, b, c). The standard reference for these sorts of things is the delightful book Concrete Mathematics: A Foundation for Computer Science by Graham, Knuth and Patashnik. In it, Chapter 3 is mostly on floors and ceilings, and this is proved on page 71 as a part of a far more general result:
In the 3.10 above, you can define x = a/b (mathematical, i.e. real division), and f(x) = x/c (exact division again), and plug those into the result on the left ⌊f(x)⌋ = ⌊f(⌊x⌋)⌋ (after verifying that the conditions on f hold here) to get ⌊a/(bc)⌋ on the LHS equal to ⌊⌊a/b⌋/c⌋ on the RHS.
If we don't want to rely on a reference in a book, we can prove ⌊a/(bc)⌋ = ⌊⌊a/b⌋/c⌋ directly using their methods. Note that with x = a/b (the real number), what we're trying to prove is that ⌊x/c⌋ = ⌊⌊x⌋/c⌋. So:
if x is an integer, then there is nothing to prove, as x = ⌊x⌋.
Otherwise, ⌊x⌋ < x, so ⌊x⌋/c < x/c which means that ⌊⌊x⌋/c⌋ ≤ ⌊x/c⌋. (We want to show it's equal.) Suppose, for the sake of contradiction, that ⌊⌊x⌋/c⌋ < ⌊x/c⌋ then there must be a number y such that ⌊x⌋ < y ≤ x and y/c = ⌊x/c⌋. (As we increase a number from ⌊x⌋ to x and consider division by c, somewhere we must hit the exact value ⌊x/c⌋.) But this means that y = c*⌊x/c⌋ is an integer between ⌊x⌋ and x, which is a contradiction!
This proves the result.
Programming
#include <stdio.h>
int main() {
unsigned int a = 142857;
unsigned int b = 65537;
unsigned int c = 65537;
printf("a/(b*c) = %d\n", a/(b*c));
printf("a/b/c = %d\n", a/b/c);
}
prints (with 32-bit integers),
a/(b*c) = 1
a/b/c = 0
(I used unsigned integers as overflow behaviour for them is well-defined, so the above output is guaranteed. With signed integers, overflow is undefined behaviour, so the program can in fact print (or do) anything, which only reinforces the point that the results can be different.)
But if you don't have overflow, then the values you get in your program are equal to their mathematical values (that is, a/(b*c) in your code is equal to the mathematical value ⌊a/(bc)⌋, and a/b/c in code is equal to the mathematical value ⌊⌊a/b⌋/c⌋), which we've proved are equal. So it is safe to replace a/(b*c) in code by a/b/c when b*c is small enough not to overflow.
While b*c could overflow (in C) for the original computation, a/b/c can't overflow, so we don't need to worry about overflow for the forward replacement a/(b*c) -> a/b/c. We would need to worry about it the other way around, though.
Let x = a/b/c. Then a/b == x*c + y for some y < c, and a == (x*c + y)*b + z for some z < b.
Thus, a == x*b*c + y*b + z. y*b + z is at most b*c-1, so x*b*c <= a <= (x+1)*b*c, and a/(b*c) == x.
Thus, a/b/c == a/(b*c), and replacing a/(b*c) by a/b/c is safe.
Nested floor division can be reordered as long as you keep track of your divisors and dividends.
#python3.x
x // m // n = x // (m * n)
#python2.x
x / m / n = x / (m * n)
Proof (sucks without LaTeX :( ) in python3.x:
Let k = x // m
then k - 1 < x / m <= k
and (k - 1) / n < x / (m * n) <= k / n
In addition, (x // m) // n = k // n
and because x // m <= x / m and (x // m) // n <= (x / m) // n
k // n <= x // (m * n)
Now, if k // n < x // (m * n)
then k / n < x / (m * n)
and this contradicts the above statement that x / (m * n) <= k / n
so if k // n <= x // (m * n) and k // n !< x // (m * n)
then k // n = x // (m * n)
and (x // m) // n = x // (m * n)
https://en.wikipedia.org/wiki/Floor_and_ceiling_functions#Nested_divisions
Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 4 years ago.
Improve this question
A Y-combinator is a computer science concept from the “functional” side of things. Most programmers don't know much at all about combinators, if they've even heard about them.
What is a Y-combinator?
How do combinators work?
What are they good for?
Are they useful in procedural languages?
A Y-combinator is a "functional" (a function that operates on other functions) that enables recursion, when you can't refer to the function from within itself. In computer-science theory, it generalizes recursion, abstracting its implementation, and thereby separating it from the actual work of the function in question. The benefit of not needing a compile-time name for the recursive function is sort of a bonus. =)
This is applicable in languages that support lambda functions. The expression-based nature of lambdas usually means that they cannot refer to themselves by name. And working around this by way of declaring the variable, refering to it, then assigning the lambda to it, to complete the self-reference loop, is brittle. The lambda variable can be copied, and the original variable re-assigned, which breaks the self-reference.
Y-combinators are cumbersome to implement, and often to use, in static-typed languages (which procedural languages often are), because usually typing restrictions require the number of arguments for the function in question to be known at compile time. This means that a y-combinator must be written for any argument count that one needs to use.
Below is an example of how the usage and working of a Y-Combinator, in C#.
Using a Y-combinator involves an "unusual" way of constructing a recursive function. First you must write your function as a piece of code that calls a pre-existing function, rather than itself:
// Factorial, if func does the same thing as this bit of code...
x == 0 ? 1: x * func(x - 1);
Then you turn that into a function that takes a function to call, and returns a function that does so. This is called a functional, because it takes one function, and performs an operation with it that results in another function.
// A function that creates a factorial, but only if you pass in
// a function that does what the inner function is doing.
Func<Func<Double, Double>, Func<Double, Double>> fact =
(recurs) =>
(x) =>
x == 0 ? 1 : x * recurs(x - 1);
Now you have a function that takes a function, and returns another function that sort of looks like a factorial, but instead of calling itself, it calls the argument passed into the outer function. How do you make this the factorial? Pass the inner function to itself. The Y-Combinator does that, by being a function with a permanent name, which can introduce the recursion.
// One-argument Y-Combinator.
public static Func<T, TResult> Y<T, TResult>(Func<Func<T, TResult>, Func<T, TResult>> F)
{
return
t => // A function that...
F( // Calls the factorial creator, passing in...
Y(F) // The result of this same Y-combinator function call...
// (Here is where the recursion is introduced.)
)
(t); // And passes the argument into the work function.
}
Rather than the factorial calling itself, what happens is that the factorial calls the factorial generator (returned by the recursive call to Y-Combinator). And depending on the current value of t the function returned from the generator will either call the generator again, with t - 1, or just return 1, terminating the recursion.
It's complicated and cryptic, but it all shakes out at run-time, and the key to its working is "deferred execution", and the breaking up of the recursion to span two functions. The inner F is passed as an argument, to be called in the next iteration, only if necessary.
If you're ready for a long read, Mike Vanier has a great explanation. Long story short, it allows you to implement recursion in a language that doesn't necessarily support it natively.
I've lifted this from http://www.mail-archive.com/boston-pm#mail.pm.org/msg02716.html which is an explanation I wrote several years ago.
I'll use JavaScript in this example, but many other languages will work as well.
Our goal is to be able to write a recursive function of 1
variable using only functions of 1 variables and no
assignments, defining things by name, etc. (Why this is our
goal is another question, let's just take this as the
challenge that we're given.) Seems impossible, huh? As
an example, let's implement factorial.
Well step 1 is to say that we could do this easily if we
cheated a little. Using functions of 2 variables and
assignment we can at least avoid having to use
assignment to set up the recursion.
// Here's the function that we want to recurse.
X = function (recurse, n) {
if (0 == n)
return 1;
else
return n * recurse(recurse, n - 1);
};
// This will get X to recurse.
Y = function (builder, n) {
return builder(builder, n);
};
// Here it is in action.
Y(
X,
5
);
Now let's see if we can cheat less. Well firstly we're using
assignment, but we don't need to. We can just write X and
Y inline.
// No assignment this time.
function (builder, n) {
return builder(builder, n);
}(
function (recurse, n) {
if (0 == n)
return 1;
else
return n * recurse(recurse, n - 1);
},
5
);
But we're using functions of 2 variables to get a function of 1
variable. Can we fix that? Well a smart guy by the name of
Haskell Curry has a neat trick, if you have good higher order
functions then you only need functions of 1 variable. The
proof is that you can get from functions of 2 (or more in the
general case) variables to 1 variable with a purely
mechanical text transformation like this:
// Original
F = function (i, j) {
...
};
F(i,j);
// Transformed
F = function (i) { return function (j) {
...
}};
F(i)(j);
where ... remains exactly the same. (This trick is called
"currying" after its inventor. The language Haskell is also
named for Haskell Curry. File that under useless trivia.)
Now just apply this transformation everywhere and we get
our final version.
// The dreaded Y-combinator in action!
function (builder) { return function (n) {
return builder(builder)(n);
}}(
function (recurse) { return function (n) {
if (0 == n)
return 1;
else
return n * recurse(recurse)(n - 1);
}})(
5
);
Feel free to try it. alert() that return, tie it to a button, whatever.
That code calculates factorials, recursively, without using
assignment, declarations, or functions of 2 variables. (But
trying to trace how it works is likely to make your head spin.
And handing it, without the derivation, just slightly reformatted
will result in code that is sure to baffle and confuse.)
You can replace the 4 lines that recursively define factorial with
any other recursive function that you want.
I wonder if there's any use in attempting to build this from the ground up. Let's see. Here's a basic, recursive factorial function:
function factorial(n) {
return n == 0 ? 1 : n * factorial(n - 1);
}
Let's refactor and create a new function called fact that returns an anonymous factorial-computing function instead of performing the calculation itself:
function fact() {
return function(n) {
return n == 0 ? 1 : n * fact()(n - 1);
};
}
var factorial = fact();
That's a little weird, but there's nothing wrong with it. We're just generating a new factorial function at each step.
The recursion at this stage is still fairly explicit. The fact function needs to be aware of its own name. Let's parameterize the recursive call:
function fact(recurse) {
return function(n) {
return n == 0 ? 1 : n * recurse(n - 1);
};
}
function recurser(x) {
return fact(recurser)(x);
}
var factorial = fact(recurser);
That's great, but recurser still needs to know its own name. Let's parameterize that, too:
function recurser(f) {
return fact(function(x) {
return f(f)(x);
});
}
var factorial = recurser(recurser);
Now, instead of calling recurser(recurser) directly, let's create a wrapper function that returns its result:
function Y() {
return (function(f) {
return f(f);
})(recurser);
}
var factorial = Y();
We can now get rid of the recurser name altogether; it's just an argument to Y's inner function, which can be replaced with the function itself:
function Y() {
return (function(f) {
return f(f);
})(function(f) {
return fact(function(x) {
return f(f)(x);
});
});
}
var factorial = Y();
The only external name still referenced is fact, but it should be clear by now that that's easily parameterized, too, creating the complete, generic, solution:
function Y(le) {
return (function(f) {
return f(f);
})(function(f) {
return le(function(x) {
return f(f)(x);
});
});
}
var factorial = Y(function(recurse) {
return function(n) {
return n == 0 ? 1 : n * recurse(n - 1);
};
});
Most of the answers above describe what the Y-combinator is but not what it is for.
Fixed point combinators are used to show that lambda calculus is turing complete. This is a very important result in the theory of computation and provides a theoretical foundation for functional programming.
Studying fixed point combinators has also helped me really understand functional programming. I have never found any use for them in actual programming though.
For programmers who haven't encountered functional programming in depth, and don't care to start now, but are mildly curious:
The Y combinator is a formula which lets you implement recursion in a situation where functions can't have names but can be passed around as arguments, used as return values, and defined within other functions.
It works by passing the function to itself as an argument, so it can call itself.
It's part of the lambda calculus, which is really maths but is effectively a programming language, and is pretty fundamental to computer science and especially to functional programming.
The day to day practical value of the Y combinator is limited, since programming languages tend to let you name functions.
In case you need to identify it in a police lineup, it looks like this:
Y = λf.(λx.f (x x)) (λx.f (x x))
You can usually spot it because of the repeated (λx.f (x x)).
The λ symbols are the Greek letter lambda, which gives the lambda calculus its name, and there's a lot of (λx.t) style terms because that's what the lambda calculus looks like.
y-combinator in JavaScript:
var Y = function(f) {
return (function(g) {
return g(g);
})(function(h) {
return function() {
return f(h(h)).apply(null, arguments);
};
});
};
var factorial = Y(function(recurse) {
return function(x) {
return x == 0 ? 1 : x * recurse(x-1);
};
});
factorial(5) // -> 120
Edit:
I learn a lot from looking at code, but this one is a bit tough to swallow without some background - sorry about that. With some general knowledge presented by other answers, you can begin to pick apart what is happening.
The Y function is the "y-combinator". Now take a look at the var factorial line where Y is used. Notice you pass a function to it that has a parameter (in this example, recurse) that is also used later on in the inner function. The parameter name basically becomes the name of the inner function allowing it to perform a recursive call (since it uses recurse() in it's definition.) The y-combinator performs the magic of associating the otherwise anonymous inner function with the parameter name of the function passed to Y.
For the full explanation of how Y does the magic, checked out the linked article (not by me btw.)
Anonymous recursion
A fixed-point combinator is a higher-order function fix that by definition satisfies the equivalence
forall f. fix f = f (fix f)
fix f represents a solution x to the fixed-point equation
x = f x
The factorial of a natural number can be proved by
fact 0 = 1
fact n = n * fact (n - 1)
Using fix, arbitrary constructive proofs over general/μ-recursive functions can be derived without nonymous self-referentiality.
fact n = (fix fact') n
where
fact' rec n = if n == 0
then 1
else n * rec (n - 1)
such that
fact 3
= (fix fact') 3
= fact' (fix fact') 3
= if 3 == 0 then 1 else 3 * (fix fact') (3 - 1)
= 3 * (fix fact') 2
= 3 * fact' (fix fact') 2
= 3 * if 2 == 0 then 1 else 2 * (fix fact') (2 - 1)
= 3 * 2 * (fix fact') 1
= 3 * 2 * fact' (fix fact') 1
= 3 * 2 * if 1 == 0 then 1 else 1 * (fix fact') (1 - 1)
= 3 * 2 * 1 * (fix fact') 0
= 3 * 2 * 1 * fact' (fix fact') 0
= 3 * 2 * 1 * if 0 == 0 then 1 else 0 * (fix fact') (0 - 1)
= 3 * 2 * 1 * 1
= 6
This formal proof that
fact 3 = 6
methodically uses the fixed-point combinator equivalence for rewrites
fix fact' -> fact' (fix fact')
Lambda calculus
The untyped lambda calculus formalism consists in a context-free grammar
E ::= v Variable
| λ v. E Abstraction
| E E Application
where v ranges over variables, together with the beta and eta reduction rules
(λ x. B) E -> B[x := E] Beta
λ x. E x -> E if x doesn’t occur free in E Eta
Beta reduction substitutes all free occurrences of the variable x in the abstraction (“function”) body B by the expression (“argument”) E. Eta reduction eliminates redundant abstraction. It is sometimes omitted from the formalism. An irreducible expression, to which no reduction rule applies, is in normal or canonical form.
λ x y. E
is shorthand for
λ x. λ y. E
(abstraction multiarity),
E F G
is shorthand for
(E F) G
(application left-associativity),
λ x. x
and
λ y. y
are alpha-equivalent.
Abstraction and application are the two only “language primitives” of the lambda calculus, but they allow encoding of arbitrarily complex data and operations.
The Church numerals are an encoding of the natural numbers similar to the Peano-axiomatic naturals.
0 = λ f x. x No application
1 = λ f x. f x One application
2 = λ f x. f (f x) Twofold
3 = λ f x. f (f (f x)) Threefold
. . .
SUCC = λ n f x. f (n f x) Successor
ADD = λ n m f x. n f (m f x) Addition
MULT = λ n m f x. n (m f) x Multiplication
. . .
A formal proof that
1 + 2 = 3
using the rewrite rule of beta reduction:
ADD 1 2
= (λ n m f x. n f (m f x)) (λ g y. g y) (λ h z. h (h z))
= (λ m f x. (λ g y. g y) f (m f x)) (λ h z. h (h z))
= (λ m f x. (λ y. f y) (m f x)) (λ h z. h (h z))
= (λ m f x. f (m f x)) (λ h z. h (h z))
= λ f x. f ((λ h z. h (h z)) f x)
= λ f x. f ((λ z. f (f z)) x)
= λ f x. f (f (f x)) Normal form
= 3
Combinators
In lambda calculus, combinators are abstractions that contain no free variables. Most simply: I, the identity combinator
λ x. x
isomorphic to the identity function
id x = x
Such combinators are the primitive operators of combinator calculi like the SKI system.
S = λ x y z. x z (y z)
K = λ x y. x
I = λ x. x
Beta reduction is not strongly normalizing; not all reducible expressions, “redexes”, converge to normal form under beta reduction. A simple example is divergent application of the omega ω combinator
λ x. x x
to itself:
(λ x. x x) (λ y. y y)
= (λ y. y y) (λ y. y y)
. . .
= _|_ Bottom
Reduction of leftmost subexpressions (“heads”) is prioritized. Applicative order normalizes arguments before substitution, normal order does not. The two strategies are analogous to eager evaluation, e.g. C, and lazy evaluation, e.g. Haskell.
K (I a) (ω ω)
= (λ k l. k) ((λ i. i) a) ((λ x. x x) (λ y. y y))
diverges under eager applicative-order beta reduction
= (λ k l. k) a ((λ x. x x) (λ y. y y))
= (λ l. a) ((λ x. x x) (λ y. y y))
= (λ l. a) ((λ y. y y) (λ y. y y))
. . .
= _|_
since in strict semantics
forall f. f _|_ = _|_
but converges under lazy normal-order beta reduction
= (λ l. ((λ i. i) a)) ((λ x. x x) (λ y. y y))
= (λ l. a) ((λ x. x x) (λ y. y y))
= a
If an expression has a normal form, normal-order beta reduction will find it.
Y
The essential property of the Y fixed-point combinator
λ f. (λ x. f (x x)) (λ x. f (x x))
is given by
Y g
= (λ f. (λ x. f (x x)) (λ x. f (x x))) g
= (λ x. g (x x)) (λ x. g (x x)) = Y g
= g ((λ x. g (x x)) (λ x. g (x x))) = g (Y g)
= g (g ((λ x. g (x x)) (λ x. g (x x)))) = g (g (Y g))
. . . . . .
The equivalence
Y g = g (Y g)
is isomorphic to
fix f = f (fix f)
The untyped lambda calculus can encode arbitrary constructive proofs over general/μ-recursive functions.
FACT = λ n. Y FACT' n
FACT' = λ rec n. if n == 0 then 1 else n * rec (n - 1)
FACT 3
= (λ n. Y FACT' n) 3
= Y FACT' 3
= FACT' (Y FACT') 3
= if 3 == 0 then 1 else 3 * (Y FACT') (3 - 1)
= 3 * (Y FACT') (3 - 1)
= 3 * FACT' (Y FACT') 2
= 3 * if 2 == 0 then 1 else 2 * (Y FACT') (2 - 1)
= 3 * 2 * (Y FACT') 1
= 3 * 2 * FACT' (Y FACT') 1
= 3 * 2 * if 1 == 0 then 1 else 1 * (Y FACT') (1 - 1)
= 3 * 2 * 1 * (Y FACT') 0
= 3 * 2 * 1 * FACT' (Y FACT') 0
= 3 * 2 * 1 * if 0 == 0 then 1 else 0 * (Y FACT') (0 - 1)
= 3 * 2 * 1 * 1
= 6
(Multiplication delayed, confluence)
For Churchian untyped lambda calculus, there has been shown to exist a recursively enumerable infinity of fixed-point combinators besides Y.
X = λ f. (λ x. x x) (λ x. f (x x))
Y' = (λ x y. x y x) (λ y x. y (x y x))
Z = λ f. (λ x. f (λ v. x x v)) (λ x. f (λ v. x x v))
Θ = (λ x y. y (x x y)) (λ x y. y (x x y))
. . .
Normal-order beta reduction makes the unextended untyped lambda calculus a Turing-complete rewrite system.
In Haskell, the fixed-point combinator can be elegantly implemented
fix :: forall t. (t -> t) -> t
fix f = f (fix f)
Haskell’s laziness normalizes to a finity before all subexpressions have been evaluated.
primes :: Integral t => [t]
primes = sieve [2 ..]
where
sieve = fix (\ rec (p : ns) ->
p : rec [n | n <- ns
, n `rem` p /= 0])
David Turner: Church's Thesis and Functional Programming
Alonzo Church: An Unsolvable Problem of Elementary Number Theory
Lambda calculus
Church–Rosser theorem
Other answers provide pretty concise answer to this, without one important fact: You don't need to implement fixed point combinator in any practical language in this convoluted way and doing so serves no practical purpose (except "look, I know what Y-combinator is"). It's important theoretical concept, but of little practical value.
Here is a JavaScript implementation of the Y-Combinator and the Factorial function (from Douglas Crockford's article, available at: http://javascript.crockford.com/little.html).
function Y(le) {
return (function (f) {
return f(f);
}(function (f) {
return le(function (x) {
return f(f)(x);
});
}));
}
var factorial = Y(function (fac) {
return function (n) {
return n <= 2 ? n : n * fac(n - 1);
};
});
var number120 = factorial(5);
A Y-Combinator is another name for a flux capacitor.
I have written a sort of "idiots guide" to the Y-Combinator in both Clojure and Scheme in order to help myself come to grips with it. They are influenced by material in "The Little Schemer"
In Scheme:
https://gist.github.com/z5h/238891
or Clojure:
https://gist.github.com/z5h/5102747
Both tutorials are code interspersed with comments and should be cut & pastable into your favourite editor.
As a newbie to combinators, I found Mike Vanier's article (thanks Nicholas Mancuso) to be really helpful. I would like to write a summary, besides documenting my understanding, if it could be of help to some others I would be very glad.
From Crappy to Less Crappy
Using factorial as an example, we use the following almost-factorial function to calculate factorial of number x:
def almost-factorial f x = if iszero x
then 1
else * x (f (- x 1))
In the pseudo-code above, almost-factorial takes in function f and number x (almost-factorial is curried, so it can be seen as taking in function f and returning a 1-arity function).
When almost-factorial calculates factorial for x, it delegates the calculation of factorial for x - 1 to function f and accumulates that result with x (in this case, it multiplies the result of (x - 1) with x).
It can be seen as almost-factorial takes in a crappy version of factorial function (which can only calculate till number x - 1) and returns a less-crappy version of factorial (which calculates till number x). As in this form:
almost-factorial crappy-f = less-crappy-f
If we repeatedly pass the less-crappy version of factorial to almost-factorial, we will eventually get our desired factorial function f. Where it can be considered as:
almost-factorial f = f
Fix-point
The fact that almost-factorial f = f means f is the fix-point of function almost-factorial.
This was a really interesting way of seeing the relationships of the functions above and it was an aha moment for me. (please read Mike's post on fix-point if you haven't)
Three functions
To generalize, we have a non-recursive function fn (like our almost-factorial), we have its fix-point function fr (like our f), then what Y does is when you give Y fn, Y returns the fix-point function of fn.
So in summary (simplified by assuming fr takes only one parameter; x degenerates to x - 1, x - 2... in recursion):
We define the core calculations as fn: def fn fr x = ...accumulate x with result from (fr (- x 1)), this is the almost-useful function - although we cannot use fn directly on x, it will be useful very soon. This non-recursive fn uses a function fr to calculate its result
fn fr = fr, fr is the fix-point of fn, fr is the useful funciton, we can use fr on x to get our result
Y fn = fr, Y returns the fix-point of a function, Y turns our almost-useful function fn into useful fr
Deriving Y (not included)
I will skip the derivation of Y and go to understanding Y. Mike Vainer's post has a lot of details.
The form of Y
Y is defined as (in lambda calculus format):
Y f = λs.(f (s s)) λs.(f (s s))
If we replace the variable s in the left of the functions, we get
Y f = λs.(f (s s)) λs.(f (s s))
=> f (λs.(f (s s)) λs.(f (s s)))
=> f (Y f)
So indeed, the result of (Y f) is the fix-point of f.
Why does (Y f) work?
Depending the signature of f, (Y f) can be a function of any arity, to simplify, let's assume (Y f) only takes one parameter, like our factorial function.
def fn fr x = accumulate x (fr (- x 1))
since fn fr = fr, we continue
=> accumulate x (fn fr (- x 1))
=> accumulate x (accumulate (- x 1) (fr (- x 2)))
=> accumulate x (accumulate (- x 1) (accumulate (- x 2) ... (fn fr 1)))
the recursive calculation terminates when the inner-most (fn fr 1) is the base case and fn doesn't use fr in the calculation.
Looking at Y again:
fr = Y fn = λs.(fn (s s)) λs.(fn (s s))
=> fn (λs.(fn (s s)) λs.(fn (s s)))
So
fr x = Y fn x = fn (λs.(fn (s s)) λs.(fn (s s))) x
To me, the magical parts of this setup are:
fn and fr interdepend on each other: fr 'wraps' fn inside, every time fr is used to calculate x, it 'spawns' ('lifts'?) an fn and delegates the calculation to that fn (passing in itself fr and x); on the other hand, fn depends on fr and uses fr to calculate result of a smaller problem x-1.
At the time fr is used to define fn (when fn uses fr in its operations), the real fr is not yet defined.
It's fn which defines the real business logic. Based on fn, Y creates fr - a helper function in a specific form - to facilitate the calculation for fn in a recursive manner.
It helped me understanding Y this way at the moment, hope it helps.
BTW, I also found the book An Introduction to Functional Programming Through Lambda Calculus very good, I'm only part through it and the fact that I couldn't get my head around Y in the book led me to this post.
Here are answers to the original questions, compiled from the article (which is TOTALY worth reading) mentioned in the answer by Nicholas Mancuso, as well as other answers:
What is a Y-combinator?
An Y-combinator is a "functional" (or a higher-order function — a function that operates on other functions) that takes a single argument, which is a function that isn't recursive, and returns a version of the function which is recursive.
Somewhat recursive =), but more in-depth definition:
A combinator — is just a lambda expression with no free variables.
Free variable — is a variable that is not a bound variable.
Bound variable — variable which is contained inside the body of a lambda expression that has that variable name as one of its arguments.
Another way to think about this is that combinator is such a lambda expression, in which you are able to replace the name of a combinator with its definition everywhere it is found and have everything still work (you will get into an infinite loop if combinator would contain reference to itself, inside the lambda body).
Y-combinator is a fixed-point combinator.
Fixed point of a function is an element of the function's domain that is mapped to itself by the function.
That is to say, c is a fixed point of the function f(x) if f(c) = c
This means f(f(...f(c)...)) = fn(c) = c
How do combinators work?
Examples below assume strong + dynamic typing:
Lazy (normal-order) Y-combinator:
This definition applies to languages with lazy (also: deferred, call-by-need) evaluation — evaluation strategy which delays the evaluation of an expression until its value is needed.
Y = λf.(λx.f(x x)) (λx.f(x x)) = λf.(λx.(x x)) (λx.f(x x))
What this means is that, for a given function f (which is a non-recursive function), the corresponding recursive function can be obtained first by computing λx.f(x x), and then applying this lambda expression to itself.
Strict (applicative-order) Y-combinator:
This definition applies to languages with strict (also: eager, greedy) evaluation — evaluation strategy in which an expression is evaluated as soon as it is bound to a variable.
Y = λf.(λx.f(λy.((x x) y))) (λx.f(λy.((x x) y))) = λf.(λx.(x x)) (λx.f(λy.((x x) y)))
It is same as lazy one in it's nature, it just has an extra λ wrappers to delay the lambda's body evaluation. I've asked another question, somewhat related to this topic.
What are they good for?
Stolen borrowed from answer by Chris Ammerman: Y-combinator generalizes recursion, abstracting its implementation, and thereby separating it from the actual work of the function in question.
Even though, Y-combinator has some practical applications, it is mainly a theoretical concept, understanding of which will expand your overall vision and will, likely, increase your analytical and developer skills.
Are they useful in procedural languages?
As stated by Mike Vanier: it is possible to define a Y combinator in many statically typed languages, but (at least in the examples I've seen) such definitions usually require some non-obvious type hackery, because the Y combinator itself doesn't have a straightforward static type. That's beyond the scope of this article, so I won't mention it further
And as mentioned by Chris Ammerman: most procedural languages has static-typing.
So answer to this one — not really.
A fixed point combinator (or fixed-point operator) is a higher-order function that computes a fixed point of other functions. This operation is relevant in programming language theory because it allows the implementation of recursion in the form of a rewrite rule, without explicit support from the language's runtime engine. (src Wikipedia)
The y-combinator implements anonymous recursion. So instead of
function fib( n ){ if( n<=1 ) return n; else return fib(n-1)+fib(n-2) }
you can do
function ( fib, n ){ if( n<=1 ) return n; else return fib(n-1)+fib(n-2) }
of course, the y-combinator only works in call-by-name languages. If you want to use this in any normal call-by-value language, then you will need the related z-combinator (y-combinator will diverge/infinite-loop).
The this-operator can simplify your life:
var Y = function(f) {
return (function(g) {
return g(g);
})(function(h) {
return function() {
return f.apply(h(h), arguments);
};
});
};
Then you avoid the extra function:
var fac = Y(function(n) {
return n == 0 ? 1 : n * this(n - 1);
});
Finally, you call fac(5).
I think the best way to answer this is to pick a language, like JavaScript:
function factorial(num)
{
// If the number is less than 0, reject it.
if (num < 0) {
return -1;
}
// If the number is 0, its factorial is 1.
else if (num == 0) {
return 1;
}
// Otherwise, call this recursive procedure again.
else {
return (num * factorial(num - 1));
}
}
Now rewrite it so that it doesn't use the name of the function inside the function, but still calls it recursively.
The only place the function name factorial should be seen is at the call site.
Hint: you can't use names of functions, but you can use names of parameters.
Work the problem. Don't look it up. Once you solve it, you will understand what problem the y-combinator solves.