Nununiform grids

[albop]

The decision rule object has had the ability for a while to interpolate on nonuniform cartesian grids, as long as it is supported by interpolation.py (currently multilinear is supported, cubic upcoming). But there is no way to specify a nonuniform spline in the yaml file. A very simple syntax could be as follows:

domain:   # unchanged
    z: [-2*sig_z/(1-rho^2)^0.5,  2*sig_z/(1-rho^2)^0.5]
    k: [ kss*0.5, kss*1.5]

options:
    grid:
        z: 40
        k: [kss*0.5, kss*0.6, kss*0.8, kss*0.9, kss, kss*1.1, kss*1.2, kss*1.3, kss*1.4, kss*0.1.5]

The arguments of grid, would be represented as a DiscretizationScheme, which would then be applied to the domain object. Arguments could be either a float (as for z) or an array. To avoid complications, one possibility could be to introduce a new discretization option, meant to eventually supersede options:grid. So it would be:

options:
    discretization: !DiscretizationScheme
        z: 40
        k: [kss*0.5, kss*0.6, kss*0.8, kss*0.9, kss, kss*1.1, kss*1.2, kss*1.3, kss*1.4, kss*0.1.5]

At a later stage, we could introduce a language to construct arrays, with concatenation (easy) or comprehension (tricky but feasible).

[albop]

One afterthought. A complementary change, would be to have an interpolation option. It would become:

options:
    discretization: !DiscretizationScheme
        z: 40
        k: [kss*0.5, kss*0.6, kss*0.8, kss*0.9, kss, kss*1.1, kss*1.2, kss*1.3, kss*1.4, kss*0.1.5]
    interpolation: cubic

This would seem to make a lot of sense.

[albop]

Second afterthought: the nice part of this structure, is that, once the arrays are constructed, all logic can be handled by the python object. Also, it doesn't involve complicated structures. As a result, discretization and interpolation could become natural options of all global solution methods, without inolving complex dictionary inheritance at all.

[albop]

Another question is whether the discretization scheme should use absolute or relative coordinates and how to choose which to use. For instance, the above example could be reformulated as:

options:
    discretization: !DiscretizationScheme
        z: 40   # this is implicilty relative, it requires the domain to have finite bounds
        k:  !RScale # relative scale
           - [0.5, 0.6, 0.8, 0.9, 1.1, 1.2, 1.3, 1.4, 0.15]
    interpolation: cubic

Another, heavy option, would be to find a way to denote upper bound of variables. For instance ^k could be the minimum of k domain and k^ would be the maximum. So the scale would be:

options:
    discretization: !DiscretizationScheme
        z: 40   # this is implicilty relative, it requires the domain to have finite bounds
        k:   [^k+0.5*(k^-^k), ^k+0.6*(k^-^k)]

which is super explicit, transparent, and, ahem, annoying.

[albop]

options:
    discretization: !DiscretizationScheme
        z: 40   # this is implicilty relative, it requires the domain to have finite bounds
        k:  !RScale # relative scale
           - [0.5, 0.6, 0.8, 0.9, 1.1, 1.2, 1.3, 1.4, 0.15]
    interpolation: !Special
       type: cubic
       extrapolation:
           up: !LinearBound
                   intercept: 1
                   slope: (r-1)
           down: !VerticalBound
                  value: 0