Algorithm characterizations


Algorithm characterizations are attempts to formalize the word algorithm. Algorithm does not have a generally accepted formal definition. Researchers are actively working on this problem. This article will present some of the "characterizations" of the notion of "algorithm" in more detail.

The problem of definition

Over the last 200 years the definition of algorithm has become more complicated and detailed as researchers have tried to pin down the term. Indeed, there may be more than one type of "algorithm". But most agree that algorithm has something to do with defining generalized processes for the creation of "output" integers from other "input" integers – "input parameters" arbitrary and infinite in extent, or limited in extent but still variable—by the manipulation of distinguishable symbols with finite collections of rules that a person can perform with paper and pencil.
The most common number-manipulation schemes—both in formal mathematics and in routine life—are: the recursive functions calculated by a person with paper and pencil, and the Turing machine or its Turing equivalents—the primitive register machine or "counter machine" model, the Random Access Machine model, the Random access stored program machine model and its functional equivalent "the computer".
When we are doing "arithmetic" we are really calculating by the use of "recursive functions" in the shorthand algorithms we learned in grade-school, for example, adding and subtracting.
The proofs that every "recursive function" we can calculate by hand we can compute by machine and vice versa—note the usage of the words calculate versus compute—is remarkable. But this equivalence together with the thesis that this includes every calculation/computation indicates why so much emphasis has been placed upon the use of Turing-equivalent machines in the definition of specific algorithms, and why the definition of "algorithm" itself often refers back to "the Turing machine". This is discussed in more detail under Stephen Kleene's characterization.
The following are summaries of the more famous characterizations together with those that introduce novel elements—elements that further expand the definition or contribute to a more precise definition.
A mathematical problem and its result can be considered as two points in a space, and the solution consists of a sequence of steps or a path linking them. Quality of the solution is a function of the path. There might be more than one attribute defined for the path, e.g. length, complexity of shape, an ease of generalizing, difficulty, and so on.

Chomsky hierarchy

There is more consensus on the "characterization" of the notion of "simple algorithm".
All algorithms need to be specified in a formal language, and the "simplicity notion" arises from the simplicity of the language. The Chomsky hierarchy is a containment hierarchy of classes of formal grammars that generate formal languages. It is used for classifying of programming languages and abstract machines.
From the Chomsky hierarchy perspective, if the algorithm can be specified on a simpler language, it can be characterized by this kind of language, else it is a typical "unrestricted algorithm".
Examples: a "general purpose" macro language, like M4 is unrestricted, but the C preprocessor macro language is not, so any algorithm expressed in C preprocessor is a "simple algorithm".
See also Relationships between complexity classes.

Features of a Good Algorithm

The following are the features of a good algorithm;
In early 1870 W. Stanley Jevons presented a "Logical Machine" for analyzing a syllogism or other logical form e.g. an argument reduced to a Boolean equation. By means of what Couturat called a "sort of logical piano ... the equalities which represent the premises... are "played" on a keyboard like that of a typewriter.... When all the premises have been "played", the panel shows only those constituents whose sum is equal to 1, that is,... its logical whole. This mechanical method has the advantage over VENN's geometrical method...".
For his part John Venn, a logician contemporary to Jevons, was less than thrilled, opining that "it does not seem to me that any contrivances at present known or likely to be discovered really deserve the name of logical machines". But of historical use to the developing notion of "algorithm" is his explanation for his negative reaction with respect to a machine that "may subserve a really valuable purpose by enabling us to avoid otherwise inevitable labor":
He concludes that "I cannot see that any machine can hope to help us except in the third of these steps; so that it seems very doubtful whether any thing of this sort really deserves the name of a logical engine.".

1943, 1952 Stephen Kleene's characterization

This section is longer and more detailed than the others because of its importance to the topic: Kleene was the first to propose that all calculations/computations—of every sort, the totality of—can equivalently be calculated by use of five "primitive recursive operators" plus one special operator called the mu-operator, or be computed by the actions of a Turing machine or an equivalent model.
Furthermore, he opined that either of these would stand as a definition of algorithm.
A reader first confronting the words that follow may well be confused, so a brief explanation is in order. Calculation means done by hand, computation means done by Turing machine. . A "function" can be thought of as an "input-output box" into which a person puts natural numbers called "arguments" or "parameters" and gets out a single nonnegative integer. Think of the "function-box" as a little man either calculating by hand using "general recursion" or computing by Turing machine.
"Effectively calculable/computable" is more generic and means "calculable/computable by some procedure, method, technique... whatever...". "General recursive" was Kleene's way of writing what today is called just "recursion"; however, "primitive recursion"—calculation by use of the five recursive operators—is a lesser form of recursion that lacks access to the sixth, additional, mu-operator that is needed only in rare instances. Thus most of life goes on requiring only the "primitive recursive functions."

1943 "Thesis I", 1952 "Church's Thesis"

In 1943 Kleene proposed what has come to be known as Church's thesis:
In a nutshell: to calculate any function the only operations a person needs are the 6 primitive operators of "general" recursion.
Kleene's first statement of this was under the section title "12. Algorithmic theories". He would later amplify it in his text as follows:
This is not as daunting as it may sound – "general" recursion is just a way of making our everyday arithmetic operations from the five "operators" of the primitive recursive functions together with the additional mu-operator as needed. Indeed, Kleene gives 13 examples of primitive recursive functions and Boolos–Burgess–Jeffrey add some more, most of which will be familiar to the reader—e.g. addition, subtraction, multiplication and division, exponentiation, the CASE function, concatenation, etc., etc.; for a list see Some common primitive recursive functions.
Why general-recursive functions rather than primitive-recursive functions?
Kleene et al. had to add a sixth recursion operator called the minimization-operator because Ackermann produced a hugely growing function—the Ackermann function—and Rózsa Péter produced a general method of creating recursive functions using Cantor's diagonal argument, neither of which could be described by the 5 primitive-recursive-function operators. With respect to the Ackermann function:
But the need for the mu-operator is a rarity. As indicated above by Kleene's list of common calculations, a person goes about their life happily computing primitive recursive functions without fear of encountering the monster numbers created by Ackermann's function.

1952 "Turing's thesis"

hypothesizes the computability of "all computable functions" by the Turing machine model and its equivalents.
To do this in an effective manner, Kleene extended the notion of "computable" by casting the net wider—by allowing into the notion of "functions" both "total functions" and "partial functions". A total function is one that is defined for all natural numbers. A partial function is defined for some natural numbers but not all—the specification of "some" has to come "up front". Thus the inclusion of "partial function" extends the notion of function to "less-perfect" functions. Total- and partial-functions may either be calculated by hand or computed by machine.
We now observe Kleene's definition of "computable" in a formal sense:
Thus we have arrived at Turing's Thesis:
Although Kleene did not give examples of "computable functions" others have. For example, Davis gives Turing tables for the Constant, Successor and Identity functions, three of the five operators of the primitive recursive functions:
Boolos–Burgess–Jeffrey give the following as prose descriptions of Turing machines for:
With regards to the counter machine, an abstract machine model equivalent to the Turing machine:
Demonstrations of computability by abacus machine and by counter machine :
The fact that the abacus/counter machine models can simulate the recursive functions provides the proof that: If a function is "machine computable" then it is "hand-calculable by partial recursion". Kleene's Theorem XXIX :
The converse appears as his Theorem XXVIII. Together these form the proof of their equivalence, Kleene's Theorem XXX.

1952 Church–Turing Thesis

With his Theorem XXX Kleene proves the equivalence of the two "Theses"—the Church Thesis and the Turing Thesis. :
Thus by Kleene's Theorem XXX: either method of making numbers from input-numbers—recursive functions calculated by hand or computated by Turing-machine or equivalent—results in an "effectively calculable/computable function". If we accept the hypothesis that every calculation/computation can be done by either method equivalently we have accepted both Kleene's Theorem XXX and the Church–Turing Thesis.

A note of dissent: "There's more to algorithm..." Blass and Gurevich (2003)

The notion of separating out Church's and Turing's theses from the "Church–Turing thesis" appears not only in Kleene but in Blass-Gurevich as well. But while there are agreements, there are disagreements too:

1954 A. A. Markov Jr.'s characterization

provided the following definition of algorithm:
He admitted that this definition "does not pretend to mathematical precision". His 1954 monograph was his attempt to define algorithm more accurately; he saw his resulting definition—his "normal" algorithm—as "equivalent to the concept of a recursive function". His definition included four major components :
In his Introduction Markov observed that "the entire significance for mathematics" of efforts to define algorithm more precisely would be "in connection with the problem of a constructive foundation for mathematics". Ian Stewart shares a similar belief: "...constructive analysis is very much in the same algorithmic spirit as computer science...". For more see constructive mathematics and Intuitionism.
Distinguishability and Locality: Both notions first appeared with Turing --
Locality appears prominently in the work of Gurevich and Gandy . Gandy's "Fourth Principle for Mechanisms" is "The Principle of Local Causality":

1936, 1963, 1964 Gödel's characterization

1936: A rather famous quote from Kurt Gödel appears in a "Remark added in proof in his paper "On the Length of Proofs" translated by Martin Davis appearing on pp. 82–83 of The Undecidable. A number of authors—Kleene, Gurevich, Gandy etc. -- have quoted the following:
1963: In a "Note" dated 28 August 1963 added to his famous paper On Formally Undecidable Propositions Gödel states his belief that "formal systems" have "the characteristic property that reasoning in them, in principle, can be completely replaced by mechanical devices". "... due to "A. M. Turing's work a precise and unquestionably adequate definition of the general notion of formal system can now be given a completely general version of Theorems VI and XI is now possible.". In a 1964 note to another work he expresses the same opinion more strongly and in more detail.
1964: In a Postscriptum, dated 1964, to a paper presented to the Institute for Advanced Study in spring 1934, Gödel amplified his conviction that "formal systems" are those that can be mechanized:
The * indicates a footnote in which Gödel cites the papers by Alan Turing and Emil Post and then goes on to make the following intriguing statement:
Church's definitions encompass so-called "recursion" and the "lambda calculus". His footnote 18 says that he discussed the relationship of "effective calculatibility" and "recursiveness" with Gödel but that he independently questioned "effectively calculability" and "λ-definability":
It would appear from this, and the following, that far as Gödel was concerned, the Turing machine was sufficient and the lambda calculus was "much less suitable." He goes on to make the point that, with regards to limitations on human reason, the jury is still out:

1967 Minsky's characterization

Minsky baldly asserts that "an algorithm is "an effective procedure" and declines to use the word "algorithm" further in his text; in fact his index makes it clear what he feels about "Algorithm, synonym for Effective procedure":
Other writers use the word "effective procedure". This leads one to wonder: What is Minsky's notion of "an effective procedure"? He starts off with:
But he recognizes that this is subject to a criticism:
His refinement? To "specify, along with the statement of the rules, the details of the mechanism that is to interpret them". To avoid the "cumbersome" process of "having to do this over again for each individual procedure" he hopes to identify a "reasonably uniform family of rule-obeying mechanisms". His "formulation":
In the end, though, he still worries that "there remains a subjective aspect to the matter. Different people may not agree on whether a certain procedure should be called effective"
But Minsky is undeterred. He immediately introduces "Turing's Analysis of Computation Process". He quotes what he calls "Turing's thesis"
After an analysis of "Turing's Argument"
he observes that "equivalence of many intuitive formulations" of Turing, Church, Kleene, Post, and Smullyan "...leads us to suppose that there is really here an 'objective' or 'absolute' notion. As Rogers put it:

1967 Rogers' characterization

In his 1967 Theory of Recursive Functions and Effective Computability Hartley Rogers' characterizes "algorithm" roughly as "a clerical procedure... applied to... symbolic inputs and which will eventually yield, for each such input, a corresponding symbolic output". He then goes on to describe the notion "in approximate and intuitive terms" as having 10 "features", 5 of which he asserts that "virtually all mathematicians would agree ". The remaining 5 he asserts "are less obvious than *1 to *5 and about which we might find less general agreement".
The 5 "obvious" are:
The remaining 5 that he opens to debate, are:
has given a list of five properties that are widely accepted as requirements for an algorithm:
  1. Finiteness: "An algorithm must always terminate after a finite number of steps... a very finite number, a reasonable number"
  2. Definiteness: "Each step of an algorithm must be precisely defined; the actions to be carried out must be rigorously and unambiguously specified for each case"
  3. Input: "...quantities which are given to it initially before the algorithm begins. These inputs are taken from specified sets of objects"
  4. Output: "...quantities which have a specified relation to the inputs"
  5. Effectiveness: "... all of the operations to be performed in the algorithm must be sufficiently basic that they can in principle be done exactly and in a finite length of time by a man using paper and pencil"
Knuth offers as an example the Euclidean algorithm for determining the greatest common divisor of two natural numbers.
Knuth admits that, while his description of an algorithm may be intuitively clear, it lacks formal rigor, since it is not exactly clear what "precisely defined" means, or "rigorously and unambiguously specified" means, or "sufficiently basic", and so forth. He makes an effort in this direction in his first volume where he defines in detail what he calls the "machine language" for his "mythical MIX...the world's first polyunsaturated computer". Many of the algorithms in his books are written in the MIX language. He also uses tree diagrams, flow diagrams and state diagrams.
"Goodness" of an algorithm, "best" algorithms: Knuth states that "In practice, we not only want algorithms, we want good algorithms...." He suggests that some criteria of an algorithm's goodness are the number of steps to perform the algorithm, its "adaptability to computers, its simplicity and elegance, etc." Given a number of algorithms to perform the same computation, which one is "best"? He calls this sort of inquiry "algorithmic analysis: given an algorithm, to determine its performance characteristcis"

1972 Stone's characterization

Stone and Knuth were professors at Stanford University at the same time so it is not surprising if there are similarities in their definitions :
Stone is noteworthy because of his detailed discussion of what constitutes an “effective” rule – his robot, or person-acting-as-robot, must have some information and abilities within them, and if not the information and the ability must be provided in "the algorithm":
Furthermore, "...not all instructions are acceptable, because they may require the robot to have abilities beyond those that we consider reasonable.” He gives the example of a robot confronted with the question is “Henry VIII a King of England?” and to print 1 if yes and 0 if no, but the robot has not been previously provided with this information. And worse, if the robot is asked if Aristotle was a King of England and the robot only had been provided with five names, it would not know how to answer. Thus:
After providing us with his definition, Stone introduces the Turing machine model and states that the set of five-tuples that are the machine’s instructions are “an algorithm... known as a Turing machine program”. Immediately thereafter he goes on say that a “computation of a Turing machine is described by stating:
This precise prescription of what is required for "a computation" is in the spirit of what will follow in the work of Blass and Gurevich.

1995 Soare's characterization

2000 Berlinski's characterization

While a student at Princeton in the mid-1960s, David Berlinski was a student of Alonzo Church. His year-2000 book The Advent of the Algorithm: The 300-year Journey from an Idea to the Computer contains the following definition of algorithm:

2000, 2002 Gurevich's characterization

A careful reading of Gurevich 2000 leads one to conclude that he believes that "an algorithm" is actually "a Turing machine" or "a pointer machine" doing a computation. An "algorithm" is not just the symbol-table that guides the behavior of the machine, nor is it just one instance of a machine doing a computation given a particular set of input parameters, nor is it a suitably programmed machine with the power off; rather an algorithm is the machine actually doing any computation of which it is capable. Gurevich does not come right out and say this, so as worded above this conclusion is certainly open to debate:
In Blass and Gurevich 2002 the authors invoke a dialog between "Quisani" and "Authors", using Yiannis Moshovakis as a foil, where they come right out and flatly state:
This use of the word "implementation" cuts straight to the heart of the question. Early in the paper, Q states his reading of Moshovakis:
But the authors waffle here, saying "et's stick to "algorithm" and "machine", and the reader is left, again, confused. We have to wait until Dershowitz and Gurevich 2007 to get the following footnote comment:

2003 Blass and Gurevich's characterization

Blass and Gurevich describe their work as evolved from consideration of Turing machines and pointer machines, specifically Kolmogorov-Uspensky machines, Schönhage Storage Modification Machines, and linking automata as defined by Knuth. The work of Gandy and Markov are also described as influential precursors.
Gurevich offers a 'strong' definition of an algorithm :
The above phrase computation as an evolution of the state differs markedly from the definition of Knuth and Stone—the "algorithm" as a Turing machine program. Rather, it corresponds to what Turing called the complete configuration -- and includes both the current instruction and the status of the tape. .
In Algorithm examples we see the evolution of the state first-hand.

1995 – Daniel Dennett: evolution as an algorithmic process

Philosopher Daniel Dennett analyses the importance of evolution as an algorithmic process in his 1995 book Darwin's Dangerous Idea. Dennett identifies three key features of an algorithm:
It is on the basis of this analysis that Dennett concludes that "According to Darwin, evolution is an algorithmic process"..
However, in the previous page he has gone out on a much-further limb. In the context of his chapter titled "Processes as Algorithms", he states:
It is unclear from the above whether Dennett is stating that the physical world by itself and without observers is intrinsically algorithmic or whether a symbol-processing observer is what is adding "meaning" to the observations.

2002 John Searle adds a clarifying caveat to Dennett's characterization

is a proponent of strong artificial intelligence: the idea that the logical structure of an algorithm is sufficient to explain mind. John Searle, the creator of the Chinese room thought experiment, claims that "syntax is by itself not sufficient for semantic content ". In other words, the "meaning" of symbols is relative to the mind that is using them; an algorithm—a logical construct—by itself is insufficient for a mind.
Searle cautions those who claim that algorithmic processes are intrinsic to nature :

2002: Boolos-Burgess-Jeffrey specification of Turing machine calculation

An example in Boolos-Burgess-Jeffrey demonstrates the precision required in a complete specification of an algorithm, in this case to add two numbers: m+n. It is similar to the Stone requirements above.
They have discussed the role of "number format" in the computation and selected the "tally notation" to represent numbers:
At the outset of their example they specify the machine to be used in the computation as a Turing machine. They have previously specified that the Turing-machine will be of the 4-tuple, rather than 5-tuple, variety. For more on this convention see Turing machine.
Previously the authors have specified that the tape-head's position will be indicated by a subscript to the right of the scanned symbol. For more on this convention see Turing machine. :
This specification is incomplete: it requires the location of where the instructions are to be placed and their format in the machine--
This later point is important. Boolos-Burgess-Jeffrey give a demonstration that the predictability of the entries in the table allow one to "shrink" the table by putting the entries in sequence and omitting the input state and the symbol. Indeed, the example Turing machine computation required only the 4 columns as shown in the table below :
1BRH11R21RHR2
2BP321R22P3R2
3BL431R33L4R3
4BL541E44L5E4
5BRH51L55RHL5

2006: Sipser's assertion and his three levels of description

Sipser begins by defining '"algorithm" as follows:
Does Sipser mean that "algorithm" is just "instructions" for a Turing machine, or is the combination of "instructions + a Turing machine"? For example, he defines the two standard variants of his particular variant and goes on, in his Problems, to describe four more variants. In addition, he imposes some constraints. First, the input must be encoded as a string and says of numeric encodings in the context of complexity theory:
Van Emde Boas comments on a similar problem with respect to the random access machine abstract model of computation sometimes used in place of the Turing machine when doing "analysis of algorithms":
"The absence or presence of multiplicative and parallel bit manipulation operations is of relevance for the correct understanding of some results in the analysis of algorithms.
"... here hardly exists such as a thing as an "innocent" extension of the standard RAM model in the uniform time measures; either one only has additive arithmetic or one might as well include all reasonable multiplicative and/or bitwise Boolean instructions on small operands."
With regard to a "description language" for algorithms Sipser finishes the job that Stone and Boolos-Burgess-Jeffrey started. He offers us three levels of description of Turing machine algorithms :