How to work with Experimental`NumericalFunction?

Question

This question is intimately connected with previous one: "How to create internally optimized expression for computing with high WorkingPrecision?"

Oleksandr R. correctly states in the comment:

A good answer will hopefully discuss the Experimental`NumericalFunction, which is a structure produced by FindMinimum from the function and its Jacobian that is optimized for fast numerical evaluation. I know almost nothing about these objects or how to create/use them, but I would like to find out.

From other questions on this site it is apparent that other built-in numerical functions also use Experimental`NumericalFunction which seems to be the standard way to optimize numerical evaluation. It is not documented but it is placed in the Experimental context which "contains functions that are being considered for official inclusion in future versions of Mathematica". This undocumented function is of crucial importance because currently it is the only way to optimize arbitrary precision calculations. For MachinePrecision we have Compile but nothing documented for arbitrary precision. How to work with Experimental`NumericalFunction?

Answer 1

To create Experimental`NumericalFunction, one needs to evaluate Experimental`CreateNumericalFunction[vars, expr, dims] where vars is a list of arguments, expr - the expression from which the numerical function will be created, dims - the dimensions of the output matrix produced by this expression. If the output is scalar, then dims should be set to {}.

It also accepts an optional fourth argument, a list of input types corresponding to vars, such as {_Real, _Real}, as can be seen in this code that creates a new NIntegrate rule. If the fourth argument is not specified, the resulting function expects a single argument: a list of the vars values. If the fourth argument is specified, the resulting function expects a sequence of arguments instead of a single list.

The Experimental`CreateNumericalFunction has several options, including WorkingPrecision, EvaluationMonitor, StepMonitor, Compiled, Hessian, Gradient and Jacobian which work in the same way as in FindMinimum.

Hessian, Gradient and Jacobian can be Automatic (the default), Symbolic or FiniteDifference. According to the Documentation,

One of the things that NumericalFunction does is to handle derivatives automatically. By default, symbolic derivatives are used if they can be found, and otherwise finite differences are used.

I tried to use the "Sparse" suboption of Hessian and found that it is ignored. It also has interesting options "ErrorReturn", "SampleArgument" and Message the purpose of which is unclear to me.

---

The created Experimental`NumericalFunction has several properties:

f = Experimental`CreateNumericalFunction[{x, y}, {Sin[x + y], x^2 y}, {2}];
f["Properties"]

{"ArgumentDimensions", "ArgumentNames", "ArgumentUnits", "CompiledFunction", 
"FunctionExpression", "InputIndexes", "InputTypes", "Properties", "ResultUnits", 
"SolutionDataComponents", "WorkingPrecision"}

The most useful is that you can directly compute Gradient/Jacobian or Hessian for expr for any numerical values of parameters:

f = Experimental`CreateNumericalFunction[{x, y}, {Sin[x + y], x^2 y}, {2}];

f["Hessian"[{1, 2}]]
f["Gradient"[{1, 2}]]
f["Jacobian"[{1, 2}]]

{{{-0.14112,-0.14112},{-0.14112,-0.14112}},{{4.,2.},{2.,0.}}}
{{-0.989992,-0.989992},{4.,1.}}
{{-0.989992,-0.989992},{4.,1.}}

When called for the first time, symbolic Hessian or Jacobian will be created (if the corresponding parameter is set to Automatic or Symbolic), further evaluation will be executed MUCH faster. For Experimental`NumericalFunction Gradient and Jacobian options seem to do the same.

There also is NDSolve`ValidNumericalFunctionQ which seemingly tests whether the created Experimental`NumericalFunction is valid.

There also are some useful Messages for Experimental`CreateNumericalFunction defined in the file "Messages.m".

Another mentionable feature of NumericalFunction is its first argument isn't necessarily a list of Symbol:

f = Experimental`CreateNumericalFunction[{Subscript[x, 1], y}, {Sin[Subscript[x, 1] + y],
     Subscript[x, 1]^2 y}, {2}];
f[{2, 3}]   
(* {-0.958924, 12.} *)

Answer 2

Here's a little addition related to the performance of NumericalFunction.

Suppose we have a function like

func = Cos[x + y^2] Sin[x^2 - y];

A NumericalFunction may be created as follows. The first variant nf0[{x0, y0}] returns a scalar. The second variant nf1[{x0, y0}] returns a 1-vector {z0} for a given input {x0, y0}, and the reason for including it will be seen in the timing of the "Hessian" further down.

nf0 = Experimental`CreateNumericalFunction[{x, y}, func, {}];
nf1 = Experimental`CreateNumericalFunction[{x, y}, {func}, {1}];

For a NumericalFunction nf, nf["Jacobian"@{x, y}] and nf["Jacobian"@{x, y}] compute the Jacobian and Hessian, respectively. I thought to explore how efficient these were compared to the symbolic D[func, {{x, y}}] and D[func, {{x, y}, 2}].

The Hessian of func above turns out to be fairly complicated. I thought to use Experimental`OptimizeExpression to improve its speed for the sake of comparison. This was done as follows:

hf1 = Experimental`OptimizeExpression[
    Simplify@D[func, {{x, y}, 2}]] /. 
   Experimental`OptimizedExpression[f_] :> 
    Function @@ Hold[{x, y}, f];

It's possible to optimize func and its Jacobian expression, too, which I did mainly because I was making a table.

I also include compiled versions and vectorized ones. Functions were compiled with a simple Compile[{x, y}, expr]. I wasn't trying to find the absolute fastest method of evaluating, just giving some basis for evaluating the performance of NumericalFunction. Vectorization was done on the y coordinate as follows, using both the plain and optimized expressions for the function, Jacobian, and Hessian:

Do[testf[x, Range[0., 4., dx]], {x, 0., 4., dx}] // RepeatedTiming

The other functions were timed as follows:

Do[testf[x, y], {x, 0., 4., dx}, {y, 0., 4., dx}] // RepeatedTiming

Unsurprisingly, the results show that the vectorized optimized form is fast and usually fastest. The NumericalFunction is faster than the symbolic expression when evaluated on single inputs. The most surprising outcome is the difference between the performances of the scalar nf0 and the vector nf1 on the Hessian. The vector form evaluates at least twice as fast as the scalar form. They seem computationally equivalent to me. If either is more expensive, I would think it would be the First of the vector form nf1, which in fact is the case when evaluating the function and Jacobian. The same difference in speed is observed if func is changed to the simpler x^2 y^3.

dx = 1./8	dx = 1./128