Login  Register

Re: LexicographicalView class help!!

Posted by Luigi Ballabio-4 on Mar 19, 2002; 1:33pm
URL: http://quantlib.414.s1.nabble.com/LexicographicalView-class-help-tp1903p1904.html

Re: [Quantlib-users] LexicographicalView class help!!

Hi Toyin,
        sorry for the delay---I pushed your mail in the stack and it was a while before I could pop it...

At 7:42 PM +0000 3/14/02, Toyin Akin wrote:
Your doc refers to the possibility of implementing 2D Finite differences using the Lexicographical
class. It also states that no 2D operator is currently implemented apart from the identity operator.
(Page 20 of the pdf, version 0.3.0a5)
 
Could you point me to the location of where this 2D identity operator is implemented?

Ouch, that was a leftover from code that isn't there anymore. There's no 2D identity operator.

This would give me ideas of how the other 2D operators can be implemented with respect to the
Lexicographical class.

Hmm, I probably wrote that in a misleading way. Let's see if I can write this  straight. Apologies if I'm stating the obvious.

Representing a discretized 1-D function f(x) is straightforward---one just chooses a series of x_i, i in 0..N-1, and the function is then the array
[f_0,...,f_{N-1}] where f_i = f(x_i). 1-D differential operators are discretized just as easily. Let's take the x_i are evenly spaced with x_{i+1} - x_i = h.
The discrete derivative in x_i is f'_i = (x_{i+1}-x_{i-1})/2h which translates in the differential operator
    |? ?                 |
    |-1  0  1            |
  1 |   -1  0  1         |
  - |       .....        |
 2h |         -1  0  1   |
    |            -1  0  1|
    |                ?  ?|
where the ? depend on which boundary conditions are chosen.

2-D functions and operators are a different breed of cats.
Given a function f(x,y), one way to start is to discretize the x and y domain as [x_0,...,x_{N-1}] and [y_0,...,y_{M-1}]. We end up with a collection of points (x_i,y_j) and with the corresponding f(x_i,y_j). Hic sunt leones. We want to put these points in an array [f_0,...,f_{N*M-1}] (note: not a matrix!) so that we can use linear algebra to apply a differential operator D to the function. However, there is no obvious ordering of the points (or rather, there are more than one), or put differently, there is no single mapping of the (i,j) into the indexes of the array f.

Let's say we have [x_0,x_1,x_2] and [y_0,...,y_3]. The corresponding points are
f_00 f_10 f_20
f_01 f_11 f_21
f_02 f_12 f_22
f_03 f_13 f_23

Lexicographical mappings give the array
[f_00 f_10 f_20 f_01 f_11 f_21 f_02 f_12 f_22 f_03 f_13 f_23];

reverse lexicographical mapping gives
[f_00 f_01 f_02 f_03 f_10 f_11 f_12 f_13 f_20 f_21 f_22 f_23];

red-black mapping gives
[f_00 f_20 f_11 f_02 f_22 f_13 f_10 f_01 f_21 f_12 f_03 f_23]
(consider the points above as on a chessboard; order lexicographically all the red points first, then the black).

There are many other which have different applications. For instance, lexicographical orderings seem most natural. However, red-black ordering has the advantage that in a 5-star differential formula (that is, a derivative in one point depends on the points immediately above, below and besides it but not the ones touching its corners only) all points in the formula except the central one have the same color. Thus, the resulting differential operator has tighter bands than the one resulting from lexicographical ordering.

Now, as for your question: I'm afraid I have only half an answer.

Mapping classes such as LexicographicalView have the purpose of making it easier to manage discretized 2-D functions. Its use is straightforward enough: given [x_0,x_{N-1}] and [y_0,y_{M-1}], you can write:

Array f(N*M);  // discretized function

LexicographicalView view(f.begin(),f.end(),N);   // put a "view" over f
for (Size i=0; i<N; i++)
    for (Size j=0; j<M; j++)
        view[i][j] = f(x[i],y[j]);   // view will put this value at the
                                     // right position in f

had you used RedBlackView (which doesn't exist yet, alas) you would have written exactly the same for loop; however, the discrete values would have ended up at different positions in the f array corresponding to a red-black ordering. Thus, these classes save you the hassle of calculating such indexes yourself.

As for implementing 2-D differential operators, it is clear that, say, the first derivative with respect to x will have a different matricial form depending on the mapping used for storing the function into its array (I would give some examples, but they would be too painful for me to write and for you read in plain ASCII).
Therefore, there is some coupling between constructing the array and applying the operator: they both must be aware of the mapping used (note that this is not a problem introduced by the LexicographicalView class; the problem would be the same had we build the array by hand).

I admit I don't have a design clear in my mind right now.

Of course, one possibility would be to let the user shoot his own foot (which is actually kind of appealing since it would promote natural selection :) but I will discard it for the time being. Other possibilities all involve carrying view information with the array.

We could pass the view to the operator and let it operate on the array through the view, but this would prevent composing operators unless we jump through some major hoops.

Another possibility I like better is to declare the operators as templates and specialize them for the view classes. This would allow us to write code such as:

template <class View>
result_type doSomething(whatever) {
    Array f(N*M);
    View view(f.begin(),f.end(),N);
    ... // fill the array
    Operator L = Dx<View>() + 2.0*Dy<View>()   // simple matricial algebra
    L.applyTo(f);
    ... // etc.
}

I will have to think about this and put together a QuEP.

Time to go to bed now.

Bye,
        Luigi











--