# What is… a numerical infinitesimal? and Humpty-Dumpty logic

A notion of a numerical infinitesimal is mentioned in a pair of recent articles:

Caldarola, Fabio; The Sierpinski curve viewed by numerical computations with infinities and infinitesimals. Appl. Math. Comput. 318 (2018), 321-328.

and

Gaudioso, Manlio; Giallombardo, Giovanni; Mukhametzhanov, Marat; Numerical infinitesimals in a variable metric method for convex nonsmooth optimization. Appl. Math. Comput. 318 (2018), 312-320.

I was naturally curious to find out what a numerical infinitesimal is and looked inside, but was unable to find a definition other than a reference to an as-yet unpublished article. Help, anyone?

Note 1. I found the two articles in MathSciNet (formerly known as Mathematical Reviews) in october 2017. I posted this question on 31 october 2017. Shortly afterwards the two articles were deleted from the MathSciNet database and currently cannot be found there.

Note 2. Here's another one:

Sergeyev, Yaroslav; Numerical infinities and infinitesimals: Methodology, applications, and repercussions on two Hilbert problems. EMS Surv. Math. Sci. 4 (2017), no. 2, 219–320.

Following the publication of the above article by Sergeyev in EMS Surveys in Mathematical Sciences, the editors published the following clarification:

Statement of the editorial board

We deeply regret that this article appears in this issue of the EMS Surveys in Mathematical Sciences.

It was a serious mistake to accept it for publication. Owing to an unfortunate error, the entire processing of the paper, including the decision to accept it, took place without the editorial board being aware of what was happening. The editorial board unanimously dissociates itself from this decision. It is not representative of the very high level that we expect to see in our journal, which can be assessed from all other papers that we have published.

Both editors-in-chief have assumed responsibility for these mistakes and resigned from their position. Having said that, we add that this journal would not exist without their dedication and years of hard work, and we wish to register our thanks to them.

Note. See related MSE thread here.

• See mathoverflow.net/questions/226277/what-is-a-grossone for a discussion of "gross-one" and the infinity computer. There are apparent similarities to Laugwitz' Omega calculus and automatic differentiation ideas. – LutzL Oct 31 '17 at 10:44
• @LutzL, I am familiar with Laugwitz's work but I don't recall anything about infinitesimals being numerical. – Mikhail Katz Oct 31 '17 at 11:22
• Hi, I did not make the connection that you wrote some of the answers in the linked post on MO. As was already formulated there, it is futile to speculate on non-existing quantities. On the one hand, the segment $[1,..,G]$ is considered to exist inside $\Bbb N$, on the other, $G=|\Bbb N|$, this makes no sense. -- The approximation of $\Bbb N$ by $[1,..,G]$ (and likewise of the reals by a finite set of inf. close numbers) is something E. Nelson did in his "radically elementary calculus". – LutzL Oct 31 '17 at 11:31
• @Wouter: In principle, there are similarities, the mockup of the "infinity calculator" was identified as (truncated) Puiseux series arithmetic. But for whatever reason, the originator of this concept says it is not automatic differentiation (where dual numbers belong to) or (truncated) Taylor- or Puiseux arithmetic. The idea seems to be that there is a kind of floating point format on some vapor ware device called "infinity computer" where 0.000...001 with an unconstructible (but in some way finite) numbers of zeros is validly representable. – LutzL Nov 4 '17 at 18:01
• @LutzL, would you mind posting your comments as an answer? – Mikhail Katz Nov 5 '17 at 9:47

I was asked to write down my comments on the insights of the thread https://mathoverflow.net/questions/226277/what-is-a-grossone, which is about the framework of these articles and my thoughts. This is not really an answer as statements on this "new" infinity calculus are contradictory, so the very existence of such entities "numerical (or numeral) infinitesimal" is in doubt.

## non-standard infinitely large integers

One very old introductory justification for non-standard calculus (from Reeb?) is the following: The Peano axioms define the set of natural numbers as the minimal set that contains $1,2,3$ and any other number obtained by counting up. Now there is an abstract decision to make: Do these "constructable" numbers fill up $\Bbb N$ or is there some additional unavoidable fluff that is larger than any of these "standard" numbers. With the first choice, nothing interesting happens. Formalizing the second variant gives a non-standard arithmetic in the flavour of the internal set theory of Nelson, or the ultra-filter construction of the $\Omega$-calculus of Laugwitz.

One could get the impression from http://www.theinfinitycomputer.com/The_second_paper_to_read_(Lagrange_Lecture).pdf that Sergeyevs "grossone" is obtained by taking one of these "i-large" natural numbers $N$ and setting $①=N!$. Then he restricts "his" natural numbers to the segment $[1,2,...,①]$, so that this restricted set indeed has $①$ elements, and calls all numbers $①+1,...,2①,...$ "extended natural numbers". This seems to be some kind of Humpty-Dumpty logic.

“When I use a word,” Humpty Dumpty said, in rather a scornful tone, “it means just what I choose it to mean—neither more nor less.” “The question is,” said Alice, “whether you can make words mean so many different things.” “The question is,” said Humpty Dumpty, “which is to be master—that’s all.” (Lewis Carroll: Through the looking glass)

## non-standard "algebraic" extensions of number sets

Another, more classical, way to introduce infinite and infinitesimal "numbers" is to add to the number system an external symbol, say $X$, and the rules $1<X$, $2<X$, $3<X$, ... The arithmetic closure of this extension is the field of rational functions in $X$ with the order relation based on the definition that a rational expression is positive if the leading coefficients in numerator and denominator have the same sign. The algebraic closure of this field is the field of Puiseux series $\sum_{k\in \Bbb Z}c_kX^{k/n}$. The arithmetic of Taylor series is part of this.

The "numeral system" Sergeyev introduces is more oriented on this algebraic approach to infinities and thus also infinitesimals. It consists of finite sums $\sum_{k\in I}c_k\,①^{p_k}$ where the pairs $(c_k,p_k)$ are represented by standard floating point numbers. Arithmetic is done following usual arithmetic rules and and function evaluation seems to follow Taylor series arithmetic using not really apparent truncation rules (see the screen shot in https://mathoverflow.net/a/227536).

Despite the apparent symbolic nature of $①$ in this calculus, Sergeyev insists that this is not symbolic calculus. Despite the fact that automatic differentiation contains truncated Taylor series arithmetic as one of its tools, he insists that his techniques are different and unrelated to automatic differentiation. Operation counts are done relative to the full numerals, independent of truncation depth.

"Numerical infinitesimals" appear as elements $c\,①^p$ with $p<0$ and their combinations, most often $①^{-1}$ as initial infinitesimal. The religiously repeated claim is then that the evaluation of $f(x_0+①^{-1})$ is in no way related to Taylor series or automatic differentiation, without going into details.

If one allows all these pre-existing techniques, the patented methods as in

could be replicated by standard CAS software like Magma, using G for $①$, E for the infinitesimal $①^{-1}$

P<E>:=PuiseuxSeriesRing(RealField(8),400);
ratio:=func<p|Truncate(p*100)/100>;
G:=E^(-1);
num:=G^14-2.8+3*G^(-ratio(0.3));
den:=5*G^(ratio(7.2))+2.4*G^(-ratio(3.1));
num/den;


with output

0.20000000*E^(-34/5) - 0.096000000*E^(7/2) - 0.56000000*E^(36/5) +
0.59999999*E^(15/2) + 0.046080000*E^(69/5) + 0.26880000*E^(35/2) -
0.28800000*E^(89/5) - 0.022118400*E^(241/10) - 0.12902400*E^(139/5) +
0.13824000*E^(281/10) + O(E^(166/5))


where one would only have to replace E with G^(-1), which is string manipulation, no further calculation.

• Hi Lutz, if there is anything to be added specifically from the CS perspective feel free to comment here. – Mikhail Katz Dec 24 '17 at 10:20