1. Introduction to Extrapolation with Cost Functional
The problem of non-uniqueness of solutions arises often while applying some general-type highly nonlinear optimal conditions. In particular, it is typical to various extrapolations, when the solutions to a various complicated problems are given in the form of a truncated asymptotic expansion of the real function of real variable
as
[
1,
2]
From the truncation (
1) we need to construct the behavior of
at
under assumption that as
the function behaves as a power-law
with unknown
B and
. Assume also that from some additional arguments, one can find the (critical) index
. Thus, by extrapolating (
1) to the form (
2), we have to find “only” the critical amplitude
B [
3].
The latter problem can be attacked by means of the iterative Borel summation
defined following to [
1,
4] to generalize celebrated Borel summation [
5,
6,
7,
8]. Instead of the original series (
1) we are going to work with the series
with the control parameter
p. The parameter
p can be real or discrete in two different realizations of the Borel iterative summation suggested in [
4].
The summation procedure for a Borel transform
is realized analytically by means of self-similar iterated root approximants [
4,
9],
with all parameters
defined in a closed form by the accuracy-through order procedure.
Strikingly, the large-variable behaviour of the function (
4) is a power-law with the index
at infinity. The sought amplitude at infinity
is found analytically; and the marginal amplitude
derives from
. It is given analytically as well, dependent on the parameter
p.
In order to find
p one can consider local optimization conditions in application to the large-variable amplitude
and real control parameter
p. By analogy to [
10,
11] we can write down the minimal difference condition,
and the minimal derivative condition,
Obviously, there are at least two possible solutions to optimization. In fact, there could be more than two solutions. However, due to a strictly one-parameter nature of the iterated Borel summation, the solutions can be relatively easily enumerated. Solving separately both these equations, we can obtain several solutions for the control parameters, which are enumerated as
, with
. The solutions to the two optimizations could be considered all together and on equal footing. Then, we calculate their average and variance of all solutions. In such a manner the non-uniqueness is getting addressed.
In addition to finding all solutions with a subsequent averaging, there is an alternative approach. As suggested in [
10,
12,
13,
14,
15], a penalty term should be introduced into optimization problems to avoid a non-uniqueness of solutions altogether. In the latter approach, one can adapt classical Tikhonov regularization schemes to the extrapolation problems [
10,
14,
15]. In such an approach, it is necessary to construct a cost functional to be minimized directly with respect to the control parameter considered as a variable. In addition, a group of lasso measures [
12,
13], can be adapted to the extrapolation problems [
10].
Firstly, following [
10], we introduce a cost functional
which combines the minimal-difference and minimal-derivative conditions, complemented with addition of a ridge-type penalty term. The cost functional
is written down by analogy to [
10]. The penalty is introduced in such a way that it selects the unique solution which deviates minimally from the solutions found at different
, so that
with
. We set above in the expression (
8),
(
, or
);
(
, or
;
(
, or
).
employs non-transformed iterated roots as a reference state, while
employs a pure Borel-transform as reference state.
Computations are performed variationally, for the Borel solutions with amplitudes dependent on
p. The optimal control parameter minimizes the cost functional (
8), so that
Secondly, we consider for sake of completeness the “No-Penalty” functional
Thirdly, in the discrete case of a non-negative integer
p, we consider for extrapolation the simplest sequences introduced in [
4],
with
. It could be easily extended to larger integers
p if necessary. Formula (
11) will be called for brevity “Discr.”.
Fourthly, let us demand that the results of a continuous iterative summation with arbitrary real
p should deviate minimally from the results (
11) of the discrete iterations. Technically such objective is achieved through adding a corresponding penalty term to
which takes such deviations into account, so that a new cost functional
is introduced. The cost-functional (
12) will be called for brevity “Collar”.
The optimal control parameter minimizes the cost functional (
12), so that
The penalty term
included into the cost functional (
12), is inspired by a martingale collar from the theory of random processes [
16]. For martingales the expected values at future times could not be better than the last actually available value of the time series. Imposing martingale conditions on random processes significantly restricts the class of random processes considered in finance.
Simple methods respecting martingale constraints are even considered by respectful Makridakis Competitions as a naive benchmark for actual time series extrapolations; i.e., the last value in time known from the time series is used for extrapolation to the future. Surprisingly, it is not easy to outperform the method of extrapolation based on such a naive assumption of martingale.
In our case we suggest an analog of such naive extrapolation with the penalty for any deviations from the value expected from the iterated roots constructed for original non-transformed truncation (
1) as
,
To find a naive extrapolation it suffices to replace
in the cost functional (
12) with
to arrive at a naive extrapolation with such defined cost functional
.
In our opinion it is desirable to have some general restrictions on the expected results of extrapolations considered in the current paper. Inspired by the martingale condition, we require the expected value of sought critical amplitude to be restricted by the results of a discrete Borel summation. To such an end the penalty term is written making the solution to optimization contingent on penalties for the deviations of the expected continuous value from the result of a discrete iterated Borel summation given above as . Indeed, the continuous approach to resummation is more general but more assumptive. While the discrete approach is more natural but more restrictive. In such a manner the discrete and continuous versions of iterations are getting reconciled.
The meaning of collaring achieved by imposing the novel penalty term becomes clear from the context, since the optimum is getting bounded from below and above by two competing terms arising from continuous no-penalty and penalty contributions.
Altogether, twenty different methods were advanced in the paper [
10] to tackle the problem of non-uniqueness. Such a number originates from application of Borel–Leroy, Mittag-Leffler, Fractional Riemann–Liouville and Fractional Integral Borel summations in conjunction with various cost functionals. Clearly, it is highly desirable to advance a single method with a single penalty function which would allow for a unique solution to the extrapolation problem while also maintaining a good accuracy competitive with the results of [
10].
Thus, the main objective of the current paper is to improve extrapolation accuracy and achieve uniqueness of the results using a new cost functional imposed on a single method of resummation.
3. Concluding Remarks
We consider iterated Borel-type summation with the number of iterations employed as control parameters associated with some optimization problem. For the optimization problem a new cost functional is suggested inspired by a martingale, with a penalty term written to penalize the optimal number of iterations for the deviations of expected value of critical amplitude from the results of a discrete iterated Borel summation. The optimization technique employs the penalty which by itself is expressed through the critical amplitude. The variational solution to the problem is accurate, robust and uniquely defined for a variety of extrapolation problems. Making the penalty dependent on expected value is novel, to the best of our knowledge.
The cost functional (
12) systematically outperforms, or stays close, to the naive cost functional with the penalty conditioned on deviations of the critical amplitude from the naive value (
15). Current approach is also much simpler and direct than also accurate two-parameter fractional summation techniques of [
38]. By applying the complementarity principle, it is possible to make the solution uniquely defined as well [
38].
The penalty term in the formula (
12) is written with the aim to penalize for any deviations of the expected value
with fractional
p from the result of a discrete iterated Borel summation
. In such a manner the discrete and continuous versions of iterated Borel summations are getting reconciled. The fractional approach is more general, since it employs real control parameters, but also is more assumptive in making the number of iterations to be fractional. While the discrete approach is more natural but restricted to a non-negative number of iterations.
The influence of the collar-penalty term (
14) could be very weak and rather strong, or else it can lead to a compromise solution depending on the situation. The goals here are not to spoil the results when a no-penalty cost-functional (
10) works well, and to produce a plausible solution emerging from the discrete version of Borel summation (
11) when no-penalty cost functionals fail.
The collar cost functionals lead to a good quality, robust extrapolants. They produce reasonable results for a wide variety of problems with rapidly growing, rapidly decaying or even irregular coefficients
. In such a sense the method based on a collar cost functionals works better than all other ridge cost-functionals based on Tikhonov’s ideas, advanced in the current paper and in [
10]. It is also more robust than a direct averaging over all solutions to the minimal-derivative and minimal-difference optimization problems.
It is possible to envisage several extensions to the cost functional (
12). For instance, in place of the simplest expected value given by the critical amplitude for given
k by itself, one can employ some average approximations arising from averaging over approximants with varying
k.
It would be also of a general interest to lift the property of differentiability pertinent to functional of the type (
12). Considering a non-differentiable cost functionals could lead to interesting properties of the optimum to be found. The optimum could become a catastrophe. Of course, the accuracy of the ensuing cost functionals should be checked against its differentiable counterparts.