str2func and anonymous functions

Dear all,

I posed a problem earlier and it was not answered. I guess I did not pose the problem well. I try again as my understanding of the issues has improved.

Consider the following function

function test=mytest(a,b,c,ss,anonym)
if nargin<5
    anonym=true;
end

if anonym
    test=@(x)a(3,:)*x^2+b(4)*x+c(1)-ss;
else
    test=str2func('@(x)a(3,:)*x^2+b(4)*x+c(1)-ss');
end

end

It has two modes. When the anonym flag is set to true, the function handle returned is constructed through the anonymous function directly. When the flag is false, the function handle is constructed through a string.

I was hoping those two formulations would be equivalent. They are not!!!

try running

a=rand(10,1);b=rand(10,1);c=rand(10,1); ss=rand;

f1=mytest(a,b,c,ss,true)

f2=mytest(a,b,c,ss,false)

and then

f1(1)

f2(1)

The first one works, the second does not. Why?

While all the relevant variables appear in the workspace of f1, only "a" appears in the workspace of f2. Somehow, matlab treats "b", "c" and "ss" as functions I guess. On the other hand, "a" is in the workspace of f2, I guess because it is indexed as a matrix...

But, although it works, the first formulation also has problems of its own. The workspace of f1, which I access via tmp=functions(f1) is multiple. tmp.workspace{2}.test is identical to f1 itself and so it has the same number of workspaces and this goes on and on...

If the anonymous function is created outside a function, it has only one workspace.

I would like to use the second formulation (str2func), which only has one workspace, as it is the only alternative I have. The most direct route would be to declared all the inputs but then f2 will have to be called as f2(x,a,b,c,ss), which I am trying to avoid. Has anyone faced a similar problem or does anyone know the solution? Is there anyway to force some variables into the workspace of f2?

Thanks,

[Geoff Hayes]

Patrick - why are you concerned that the first function f1 has multiple workspaces? The functions command is (from the documentation) provided for querying and debugging purposes. Because its behavior may change in subsequent releases, you should not rely upon it for programming purposes. What makes you think that using f1 is less reliable than using f2?

[Patrick Mboma]

Hi Geoff,

It is not a matter of reliability, it is a matter of practicality. The f1 option implies huge memory costs depending on the problem at hand. Plus I do not understand why the behavior should be different depending on whether the anonymous function is created inside or outside a function.

I was hoping to use the nice property that you can call a function with one argument only instead of carrying around other arguments that never change...

[Geoff Hayes]

Patrick - without knowing how you are using this function it is difficult to provide a solution. Why not just nest your test function within that function that defines the a, b, c, and ss?

[Patrick Mboma]

Hi Geoff,

This is my current implementation and I was hoping I could improve it in terms of performance. It is called memoization, I think.

Answer 1

[Titus Edelhofer]

Hi,

the str2func does not work in this case, as you observed. But the documentation does not really suggest, that it should work... Nvertheless, you can achieve something similar by evaluating the string instead of using str2func:

    test=eval('@(x)a(3,:)*x^2+b(4)*x+c(1)-ss');

instead of the line with str2func should work as you expect it to work. Although generally speaking, I'm no fan of eval, if you really have to create an anonymous function by creating some string in advance, I guess this is the only way.

Titus

[Patrick Mboma]

Hi Titus,

The eval solution gives the same result as constructing the anonymous function directly. i.e. the function created has two workspace and the second of those workspaces contains the same original function handle, implying also two workspaces,...

[Titus Edelhofer]

Hi Patrick,

I'm pretty sure, that the two workspaces "look" as if they have doubled the memory, but as long as within mytest you don't change a, b etc, they do point to the same variable in memory. Please give it a try and make the variables a, b etc. rather large and follow the memory consumption ...

Titus

[Patrick Mboma]

Hi Titus,

The problem is that there is no reason for the second workspace to be there in the first place. The information is useless and not only that, it uses memory as discussed here .

That creating an anonymous function inside a function or outside results in different behavior seems strange to me. But that is the current implementation I have now. And in all likelihood, it seems that I might have to stick to it for now.