www.digitalmars.com         C & C++   DMDScript  

digitalmars.D - A little challenge...

reply Norbert Nemec <Norbert Nemec-online.de> writes:
Hi everybody,

thinking about array expressions, I have stumbled over an interesting 
challenge for which I still have no idea:

Consider the mathematical sum notation:

	\sum_i a_i*b_i

here, the variable i is defined only at the scope inside the expression.

A analogous D syntax could be something like

	sum!(i)(a[i]*b[i])

where sum would have to be some kind of template that takes i as a name 
parameter and then defines it as variable inside the scope of the second 
expression.

(Of course, the scalar product is trivial enough to do it in a zillion 
of other ways, but for more complex sum expressions with potentially 
more than one variable, this would really be neat to have in an array 
library.)

Does anyone have an idea how this could be realized in D? Allowing a 
template to define additional symbols only for the scope of its 
arguments? Is it possible at all with the current language definion? 
Does anyone have an idea for an elegant language extension to allows 
this? It would mean taking lazy evaluation even one step further, all 
the way to a lazy symbol binding. Seems straighforward enough to me to 
implement in a compiler. Just the question how exactly to define the 
semantics.

Greetings,
Norbert
Feb 25 2010
next sibling parent reply Jason House <jason.james.house gmail.com> writes:
Norbert Nemec Wrote:


 Hi everybody,
 
 thinking about array expressions, I have stumbled over an interesting 
 challenge for which I still have no idea:
 
 Consider the mathematical sum notation:
 
 	\sum_i a_i*b_i
 
 here, the variable i is defined only at the scope inside the expression.
 
 A analogous D syntax could be something like
 
 	sum!(i)(a[i]*b[i])


Would sum!( "i", "a[i]*b[i]" ) be acceptable? That should be achievable with a
template mixin that does string mixins under the hood.
Feb 25 2010
next sibling parent reply Norbert Nemec <Norbert Nemec-online.de> writes:
Jason House wrote:

 Would sum!( "i", "a[i]*b[i]" ) be acceptable? That should be achievable with a
template mixin that does string mixins under the hood.


It is indeed a solution for the problem, but I still don't like it too 
much. For one, writing expressions as strings always feels awkward. Even 
though D can handle these strings at compile time, it just doesn't feel 
like writing native D code.

Beyond this "gut feeling" I also see two technical problems:

* code editors do not recognize the string content as code, so they 
cannot offer syntax highlighting or more advanced language tools

* the syntax does not allow nesting:

	sum(i)( a[i] * sum(j)(b[i,j]*c[j]) )

Greetings,
Norbert
Feb 26 2010
next sibling parent Don <nospam nospam.com> writes:
Norbert Nemec wrote:

 Jason House wrote:

 Would sum!( "i", "a[i]*b[i]" ) be acceptable? That should be 
 achievable with a template mixin that does string mixins under the hood.


 
 It is indeed a solution for the problem, but I still don't like it too 
 much. For one, writing expressions as strings always feels awkward. Even 
 though D can handle these strings at compile time, it just doesn't feel 
 like writing native D code.
 
 Beyond this "gut feeling" I also see two technical problems:
 
 * code editors do not recognize the string content as code, so they 
 cannot offer syntax highlighting or more advanced language tools


You can use the q{ } string syntax.


 * the syntax does not allow nesting:
 
     sum(i)( a[i] * sum(j)(b[i,j]*c[j]) )


This is a much bigger problem. It's not too difficult if you allow only 
a set of built-in operations, but if you allow user-defined operations, 
it's tough.

Template alias parameters have got much more powerful since I wrote 
BLADE, so maybe it's more feasible now.
Feb 26 2010
prev sibling parent Philippe Sigaud <philippe.sigaud gmail.com> writes:
On Fri, Feb 26, 2010 at 10:07, Norbert Nemec <Norbert nemec-online.de>wrote:


 Jason House wrote:


 Would sum!( "i", "a[i]*b[i]" ) be acceptable? That should be achievable
 with a template mixin that does string mixins under the hood.


 It is indeed a solution for the problem, but I still don't like it too
 much. For one, writing expressions as strings always feels awkward. Even
 though D can handle these strings at compile time, it just doesn't feel like
 writing native D code.

 Beyond this "gut feeling" I also see two technical problems:

 * code editors do not recognize the string content as code, so they cannot
 offer syntax highlighting or more advanced language tools

 * the syntax does not allow nesting:

        sum(i)( a[i] * sum(j)(b[i,j]*c[j]) )


You could use anonymous functions, like this:

sum!( (i,a,b) { return a[i]+b[i];})(array1, array2);

Here I consider a function of an index and some arrays, but you could use
functions on elements:

sum!( (a,b) { return a+b*a;})(array1, array2);

Editors can recognize these functions and they nest (though it's a bit
heavy):

sum!( (a,b) { return a + sum!((b) { return b*b;})(array2) * a)(array1);

A possible implementation for one array can be :

T sum(alias op, T)(T[] array) /* also, RandomAccessRange*/
{
    T result;
    foreach(i, elem; array) result += op(i, array);
    return result;
}

It's quite simplified: I consider op returns a T...

Now, what's interesting is generalizing it somewhat : any number of inputs,
and inputs can be input ranges and not only arrays:

auto sum(alias op, R...)(R ranges) if(allSatisfy!(isInputRange,R))
{
    alias staticMap!(ElementType,R) FrontType;
    FrontType f;
    typeof(op(f)) theSum;
    while(!ranges[0].empty)
    {
        foreach(i, Type; FrontType) { // extracting the fronts
            f[i] = ranges[i].front;
            ranges[i].popFront;
        }
        theSum += op(f);
    }
    return theSum;
}

usage:

int[] ar1 = [0,1,2,3];
int[] ar2 = [4,5,6,7];
auto s = sum!((a,b,c) {return a+b*c;})(ar1,ar2,ar1);
writeln(s); // 44

Note that we could sum ranges with any element type, as long as the .init
value can be summed. float/double being initialiazed to NaN, it's a bit
tricky.

We could adopt std.algorithm 'string functions' trick, to get this syntax,
which I quite like:

auto s = sum!"a+b*c"(ar1, ar2, ar3);

Maybe you could be interested by looking at some modules with these ideas
at:
http://www.dsource.org/projects/dranges

(addendum: another generalization is of course to have two operations: the
'mapping' operation and the 'reducing' operation
like this:
auto sum = mapReduce!("a*b*c+1", "a+b")(range1, range2, range3);
The first function is applied on the elements, the second on the successive
results of the first fun.

Philippe
Feb 26 2010
prev sibling parent "Igor Lesik" <curoles yahoo.com> writes:
"Jason House" <jason.james.house gmail.com> wrote in message 
news:hm75nn$1a2q$1 digitalmars.com...

 Norbert Nemec Wrote:


 Hi everybody,

 thinking about array expressions, I have stumbled over an interesting
 challenge for which I still have no idea:

 Consider the mathematical sum notation:

 \sum_i a_i*b_i

 here, the variable i is defined only at the scope inside the expression.

 A analogous D syntax could be something like

 sum!(i)(a[i]*b[i])


 Would sum!( "i", "a[i]*b[i]" ) be acceptable? That should be achievable 
 with a template mixin that does string mixins under the hood.


I think "i" could be implicit. Here is how I was able to implement it:

void main() {
    int[] a = [1,2,3];
    int[] b = [4,5,6];
    int[] c = [5,4,3];

    int IamLazy(){ writeln("am I?"); return 1; }

    writeln( sum!("a[i]*b[i]")(a,b) ); // 4 + 10 + 18 = 32
    writeln( sum!("b[i]/a[i]")(a,b) ); // 4 + 2 + 2 = 8
    writeln( sum!("(a[i]+b[i])%2 + c[i]")(a,b,c) ); // 6 5 4 = 15
    writeln( sum!("a[i]*b[i]")(a,b,IamLazy()) );

    // fun
    writeln( mixin( Msum!("a[i]*b[i]") ) );
    writeln( mixin( Msum!("(a[i]+b[i])%2 + c[i]") ) );
}

R _sum(string OP, R : R[], T...)(lazy T op) {
    mixin Rename!(T.length);
    R ret = 0;
    foreach (i; 0 .. op[0].length) {
        ret += mixin(OP);
    }
    return ret;
}

auto sum(string OP, T...)(lazy T op) if (T.length > 0)
{
    return _sum!(OP, T[0], T)(op);
}
Feb 26 2010
prev sibling next sibling parent reply Daniel Keep <daniel.keep.lists gmail.com> writes:
BLADE may also be of some interest to you.

http://www.dsource.org/projects/mathextra/browser/trunk/blade
Feb 25 2010
parent Norbert Nemec <Norbert Nemec-online.de> writes:
Daniel Keep wrote:

 BLADE may also be of some interest to you.
 
 http://www.dsource.org/projects/mathextra/browser/trunk/blade


Thanks for the hint! I had seen BLADE sometimes before, but I realize 
that I should really study it in more detail.
Feb 26 2010
prev sibling next sibling parent reply "Robert Jacques" <sandford jhu.edu> writes:
On Thu, 25 Feb 2010 18:57:32 -0500, Norbert Nemec  
<Norbert nemec-online.de> wrote:

 Hi everybody,

 thinking about array expressions, I have stumbled over an interesting  
 challenge for which I still have no idea:

 Consider the mathematical sum notation:

 	\sum_i a_i*b_i

 here, the variable i is defined only at the scope inside the expression.

 A analogous D syntax could be something like

 	sum!(i)(a[i]*b[i])

 where sum would have to be some kind of template that takes i as a name  
 parameter and then defines it as variable inside the scope of the second  
 expression.

 (Of course, the scalar product is trivial enough to do it in a zillion  
 of other ways, but for more complex sum expressions with potentially  
 more than one variable, this would really be neat to have in an array  
 library.)

 Does anyone have an idea how this could be realized in D? Allowing a  
 template to define additional symbols only for the scope of its  
 arguments? Is it possible at all with the current language definion?  
 Does anyone have an idea for an elegant language extension to allows  
 this? It would mean taking lazy evaluation even one step further, all  
 the way to a lazy symbol binding. Seems straighforward enough to me to  
 implement in a compiler. Just the question how exactly to define the  
 semantics.

 Greetings,
 Norbert


Well there is std.algorithm's map and reduce. Somehow, I feel slicing,  
ranges/generators and array expressions should be able to handle this. For  
example: \sum_i i*a_i*b_i-1 => sum(iota * a * b[1..$]) But then again I'm  
having trouble thinking of real examples off the top of my head.
Feb 25 2010
parent reply Norbert Nemec <Norbert Nemec-online.de> writes:
Robert Jacques wrote:

 Well there is std.algorithm's map and reduce. Somehow, I feel slicing, 
 ranges/generators and array expressions should be able to handle this. 
 For example: \sum_i i*a_i*b_i-1 => sum(iota * a * b[1..$]) But then 
 again I'm having trouble thinking of real examples off the top of my head.


In fact, that is the way you would do it in e.g. Python/NumPy. It works 
fine for many common cases but does not scale up to more complex 
situations. The mathematical sum notation scales up arbitrarily and 
remains clear.

I would want to offer both options and leave it to the user to choose 
the more appropriate notation.
Feb 26 2010
parent reply "Robert Jacques" <sandford jhu.edu> writes:
On Fri, 26 Feb 2010 03:58:59 -0500, Norbert Nemec  
<Norbert nemec-online.de> wrote:


 Robert Jacques wrote:

 Well there is std.algorithm's map and reduce. Somehow, I feel slicing,  
 ranges/generators and array expressions should be able to handle this.  
 For example: \sum_i i*a_i*b_i-1 => sum(iota * a * b[1..$]) But then  
 again I'm having trouble thinking of real examples off the top of my  
 head.


 In fact, that is the way you would do it in e.g. Python/NumPy. It works  
 fine for many common cases but does not scale up to more complex  
 situations. The mathematical sum notation scales up arbitrarily and  
 remains clear.

 I would want to offer both options and leave it to the user to choose  
 the more appropriate notation.


That sounds sensible. However, extensive experience in Matlab has taught  
me that resorting to custom for-loop indicates you've failed to  
sufficiently think in arrays. :)
Take, for example, your composition example from the other thread:
sum(i)( a[i] * sum(j)(b[i,j]*c[j]) ) => sum(a.*(b'*c)) or  
a.*sum(b.*(c*ones(1,length(c)))) ,1)
or something like that. (As an aside, having a more efficient syntax for  
broadcasts, instead of having to use the outer product all the time, would  
be nice)

Is there some example of a complex case you can post? I think we'll all  
think of better solutions with a goal in sight.
Feb 26 2010
parent reply Norbert Nemec <Norbert Nemec-online.de> writes:
Robert Jacques wrote:

 That sounds sensible. However, extensive experience in Matlab has taught 
 me that resorting to custom for-loop indicates you've failed to 
 sufficiently think in arrays. :)


Indeed, most use cases are simple enough to be handled in array 
notation. I have worked with Matlab and Python and managed to come up 
with array notations in many non-trivial cases as well. However, once in 
a while, it just cannot be done. Typically, this happens when you have 
to handle non-linear terms or high order tensorial objects.

Of course, my examples were simple enough to permit alternative 
expressions, but I have encountered quite a number of cases where
I could not avoid a loop in Python. I is hard to spontaneously construct 
something useful that I can describe in a few lines.

Imagine a charge density in one dimension:
	rho[r]
and then compute the coulomb energy
	sum(r1,r2)(rho[r1]*rho[r2]/(r1-r2))

Or an expression containing function calls

	sum(i,j)(f(i)*g(j)*A[i,j]))

Ultimately, 'sum' and other reduction would actually be just one use 
case. One could even use the same mechanism to construct arrays from 
expressions.

	auto A = array(a=0:10,b=0:20)(2*a + b%3)

(Disregard the exact syntax here...)

I will think further about this and try to come up with more specific 
use cases.

Greetings,
Norbert
Feb 27 2010
parent "Robert Jacques" <sandford jhu.edu> writes:
On Sat, 27 Feb 2010 04:56:00 -0500, Norbert Nemec  
<Norbert nemec-online.de> wrote:


 Robert Jacques wrote:

 That sounds sensible. However, extensive experience in Matlab has  
 taught me that resorting to custom for-loop indicates you've failed to  
 sufficiently think in arrays. :)


 Indeed, most use cases are simple enough to be handled in array  
 notation. I have worked with Matlab and Python and managed to come up  
 with array notations in many non-trivial cases as well. However, once in  
 a while, it just cannot be done. Typically, this happens when you have  
 to handle non-linear terms or high order tensorial objects.

 Of course, my examples were simple enough to permit alternative  
 expressions, but I have encountered quite a number of cases where
 I could not avoid a loop in Python. I is hard to spontaneously construct  
 something useful that I can describe in a few lines.

 Imagine a charge density in one dimension:
 	rho[r]
 and then compute the coulomb energy
 	sum(r1,r2)(rho[r1]*rho[r2]/(r1-r2))

 Or an expression containing function calls

 	sum(i,j)(f(i)*g(j)*A[i,j]))

 Ultimately, 'sum' and other reduction would actually be just one use  
 case. One could even use the same mechanism to construct arrays from  
 expressions.

 	auto A = array(a=0:10,b=0:20)(2*a + b%3)

 (Disregard the exact syntax here...)

 I will think further about this and try to come up with more specific  
 use cases.

 Greetings,
 Norbert


Thank you. I understand the difficulty of finding good examples.  I did  
take a look at my own research, but didn't find any good examples of  
complex, single line expressions. Somehow, this feels like a narrow  
problem area; simple things are easy to express using array ops, really  
complex things take multiple lines to express and should really be done  
with loops in the first place. Finding the middle ground and then finding  
a concise way of correctly expressing it both seem like difficult  
problems, though definitely ones worth investigating. Anyways, I've taken  
a look at the new examples:

Let Map create an infinite array whose elements are generated via function  
calls. (I've run into the function call issue before)
Let Index = Map!"a" or something similar
Let .B!D denote broadcasting an array along the D dimension
Let .* denote the element wise multiply. Practically, you might want to  
use .A and .M to switch between element-wise array and matrix style math.
Practically, both Map and .B require the concept of one or more dimensions  
which are 'unbounded'. Unbounded dimensions become bounded when they  
interact with any operation that does bounds checking with a bounded  
dimension.

sum(r1,r2)(rho[r1]*rho[r2]/(r1-r2))      => sum( rho.B!1.*rho.B!0./(  
Index.B!1 - Index.B!0 ) )
sum(i,j)(f(i)*g(j)*A[i,j]))              => sum( Map!f.B!1.*Map!g.B!0.*A )
auto A = array(a=0:10,b=0:20)(2*a + b%3) => auto A = take([0,10],[0,20],  
2*Index.B!1 + Index.B!0 % 3 );
                                             auto A = take( Map!"2*a +  
b%3"[0..10,0..20] )

Although the ideal syntax is definitely shorter and a bit more readable  
than the array ops version, the array ops seems more concise than the  
delegate based alternative:

sum( Map!((int i, int j){ return rho(i)*rho(j)/(i-j);  
})[0..rho.length,0..rho.length] );

By the way, I'd recommend keeping all these examples for use in your  
documentation, as people don't naturally think in arrays and array ops.
Feb 27 2010
prev sibling parent reply Ary Borenszweig <ary esperanto.org.ar> writes:
Norbert Nemec wrote:

 Hi everybody,
 
 thinking about array expressions, I have stumbled over an interesting 
 challenge for which I still have no idea:
 
 Consider the mathematical sum notation:
 
     \sum_i a_i*b_i
 
 here, the variable i is defined only at the scope inside the expression.
 
 A analogous D syntax could be something like
 
     sum!(i)(a[i]*b[i])
 
 where sum would have to be some kind of template that takes i as a name 
 parameter and then defines it as variable inside the scope of the second 
 expression.


You are missing i's initial and ending values.

I think it should be something like:

sum!("i", 0, n)(a[i]*b[i])
Feb 26 2010
parent reply Norbert Nemec <Norbert Nemec-online.de> writes:
Ary Borenszweig wrote:

 Norbert Nemec wrote:

 Hi everybody,

 thinking about array expressions, I have stumbled over an interesting 
 challenge for which I still have no idea:

 Consider the mathematical sum notation:

     \sum_i a_i*b_i

 here, the variable i is defined only at the scope inside the expression.

 A analogous D syntax could be something like

     sum!(i)(a[i]*b[i])

 where sum would have to be some kind of template that takes i as a 
 name parameter and then defines it as variable inside the scope of the 
 second expression.


 
 You are missing i's initial and ending values.
 
 I think it should be something like:
 
 sum!("i", 0, n)(a[i]*b[i])


I assumed them to default to the array boundaries, but that does not 
really matter at this point.
Feb 26 2010
parent "Robert Jacques" <sandford jhu.edu> writes:
On Fri, 26 Feb 2010 10:49:30 -0500, Norbert Nemec  
<Norbert nemec-online.de> wrote:


 Ary Borenszweig wrote:

 Norbert Nemec wrote:

 Hi everybody,

 thinking about array expressions, I have stumbled over an interesting  
 challenge for which I still have no idea:

 Consider the mathematical sum notation:

     \sum_i a_i*b_i

 here, the variable i is defined only at the scope inside the  
 expression.

 A analogous D syntax could be something like

     sum!(i)(a[i]*b[i])

 where sum would have to be some kind of template that takes i as a  
 name parameter and then defines it as variable inside the scope of the  
 second expression.


  You are missing i's initial and ending values.
  I think it should be something like:
  sum!("i", 0, n)(a[i]*b[i])


 I assumed them to default to the array boundaries, but that does not  
 really matter at this point.


Detecting those array boundaries is not an easy task. For example, if you  
take the expression in as a delegate you can't detect which ranges are  
used in it.
Feb 26 2010