30 May 2021

What is imperative programming?

In this post, I explain why you should care about Turing machines and what they are. In the next parts of this series, I will contrast Turing machines with lambda calculus and expand on why understanding both models is useful for everyday programming.

Functional v. object-oriented

Over the past fifteen years or so, fear of the flattening of Moore's law has driven renewed interest in functional programming. Many programmers are still unclear on what functional programming really is, and how it relates to their current approaches, commonly designated as "object-oriented" programming.

This is, however, not a useful question. Functional code does not contrast with object-oriented code, because they sit alongside different axes. They are orthogonal to each other and, to a large extent, do not apply at the same level.

In the next few posts, I'll try to explain what functional programming is, and what it should be contrasted to. One of the biggest hurdles for such a discussion is that programming, as a nascent discipline, does not have a widely-agreed-upon jargon1 yet. For the purposes of this blog post, I assume "object-oriented" is one of two things, which in my experience are at the heart of the most common meanings for the term:

  • An organization of code whereby data and the functions to manipulate that data are colocated under a common reference. This is the nuts-and-bolt "objects and classes" language-level set of features.
  • An architectural technique whereby systems are divided into separate parts that communicate through message-passing. This can be done in almost any language, or across languages if the messages are passed through some sort of network or queue.

Models of computation

Functional programming, however, is a model of computation based on Church's "lambda calculus". It is code of a different nature, not a different way to organize code. It has pretty much nothing to do with either of the above concepts. If a comparison needs to be made, it should be to imperative programming, which is a model of computation based on Turing's "Turing machine", and is inarguably the dominant programming paradigm of the last 200 hundred years.2

So what's a model of computation? At the end of the 19th century, Cantor and Russel broke mathematics.3 That lead mathematicians to question the validity of their entire field, and thus to try and rigorously redefine mathematics from the ground up in order to prove Cantor and Russel wrong. The most important mathematical consequence from that effort is Gödel's incompleteness theorems, which proved that mathematics was, and would remain, broken. Mathematicians have since then made their peace with that state of affairs. That thread of enquiry also led a few mathematicians to ponder the notion of computation: Cantor's argument made a pretty good case for the existence of real numbers that exist and can be precisely defined, yet cannot be computed.

This was a bit unsettling, and thus prompted a handful of mathematicians to ponder the deeper meaning of computation: what does it mean, in a precise, rigorous mathematical sense, to compute a number? How can we tell the difference between "this number cannot be computed" and "we don't know how to compute that number yet"? One of the first proposition for such a formal definition of a model for computation was Church's lambda calculus. Having a model does not really answer the question, though. All you can say with a model is "this number can be computed within this model", but what if the model itself is incomplete?

A few years later, Turing proposed his idea of a Turing machine, but more importantly he offered a proof that the set of functions that can be described with his Turing machine is exactly the same as the set of functions that can be described with Church's lambda calculus. Since then, a fairly large number of alternative models of computation have been proposed, and so far none of them has been proven to be able to compute anything the Turing machine cannot compute. In other words, the Turing machine is exactly as powerful as the lambda calculus, and no other model we know of is more powerful (i.e. "can compute more things") than either.

Note that we do not have a proof that there is not a more powerful model of computation available. All we know is that so far we have not found one. The idea that the Turing machine is the most powerful model of computation (as opposed to "the most powerful one we have found so far) is known as "the Church-Turing hypothesis".

Why am I telling you this? Mostly because I like to ramble, but also because lambda calculus and Turing machines turn out to be really important for modern computing, and because understanding them, even at a superficial level, can give you deep insights into computer programming.

Turing machines

Formally, a Turing machine is a 7-tuple \(M=\langle Q, \Gamma, b, \Sigma, \delta, q_0, F\rangle\) where:

  • \(Q\) is a finite set of states (it has to be non-empty as it needs to contain at least \(q_0\)).
  • \(\Gamma\) is an alphabet, i.e. a set of symbols (it has to be non-empty as it needs to contain at least \(b\)).
  • \(b\in\Gamma\) is a special symbol of the alphabet, which can appear an infinite number of times on the tape (see below for what the tape is); all other symbols in \(\Gamma\) can only appear on the tape a finite number of times.
  • \(\Sigma \subseteq (\Gamma \setminus \{b\})\) is the set of symbols that are allowed to appear on the initial tape contents.
  • \(q_0\in Q\) is the initial state of the machine.
  • \(F\subseteq Q\) is the set of final states.
  • \(\delta: (Q\setminus F)\times\Gamma\to Q\times\Gamma\times\{L,R\}\) is a (partial) function4 called the transition function.

This may or may not look like a jumble of meaningless symbols, and quite frankly understanding the precise meaning of each of those is not very important to my point here. (I do enjoy looking at typeset mathematics, though.) These are just the components; a formal definition would also need to proceed with exactly how they are used (i.e. what their semantics are), but I prefer to drop the formality level a few notches at this point. For our purposes, it is enough to know that a formal definition exists.

Informally, one can visualize a Turing machine as having three components:

  • A ribbon, or tape, with things written on it. The ribbon has a starting point (generally assumed to be on the left), but it is infinite on the other side (generally the right side). It is one-dimensional, and is composed of cells, with exactly one symbol per cell. A given Turing machine defines the alphabet that can be used on the ribbon, \(\Gamma\), but not the contents of the ribbon. This make sense if you think of the definition of the Turing machine as the formula for a function, and the (initial) content of the ribbon as the input to that function. When the machine "starts", the ribbon will contain a description of the input using symbols of \(\Sigma\), followed by an infinite number of cells containing \(b\).
  • A head, which is pointing to a given cell on the ribbon. It starts at 0 (the "first" cell on the finite side of the ribbon).
  • A "processor", which can direct the head to move either one cell to the right or one cell to the left, after having written a new symbol on the ribbon at its current position.

The operation of the processor is the core of the Turing machine, and is mostly described by the \(\delta\) function. The Turing machine can be thought of as performing a number of steps, where each step consists of the following operations:

  1. Check the current state. If the current state is in \(F\), the machine has finished running. Stop here. Exactly what it means for the machine to have finished depends on how the machine was defined, but in most cases it means the output of the function is a combination of the current state of the machine and what is currently written on the ribbon.
  2. Read the symbol on the ribbon, at the current position of the head. Apply the \(\delta\) function, using as arguments the symbol that was just read and the current state of the machine. The result gives us a new state for the processor, a new symbol that we write on the ribbon (possibly the same one we just read), and a direction to move the head into.
  3. GOTO 1.

Note that, while there is no "internal structure" to the states of the processor in a Turing machine, this is not a limitation as far as mathematicians care: any internal structure can just be expanded to a set of individual states by enumerating all of the possible values of all internal variables. For example, if you wanted your Turing machine state to be "any JSON value representable using 30kb", that would be OK: this is a well-defined, finite set of numbers5. You could even decide to set the name of each state to "the string representing that JSON value". This is not cheating. I swear.

Additionally, while the formal definition only allows the head to move one cell at a time, it is trivial to implement larger moves by simply adding intermediate states that do nothing but rewrite what they just read from it on the ribbon and keep moving, perhaps decrementing a counter in their non-existent-but-still-useable internal state representation.

As a concrete example, a Turing machine that does 2-bit addition could be defined by:

  • \(Q := \{B, B_0, B_1, B_2, B_3, H_0, H_1, H_2, H_3, H\}\)
  • \(\Gamma := \{0, 1, 2, 3, x\}\)
  • \(b := x\)
  • \(\Sigma := \{0, 1, 2, 3\}\)
  • \(q_0 := BB\)
  • \(F := \{H\}\)
  • \(\delta\) is defined by the following table:

StateSymboltoNew stateWritten symbolMove

There are missing entries in that table. For example, if the machine ever reads a \(2\) symbol on the ribbon while in the \(H_0\) state, the \(\delta\) function is undefined. In such cases, we say that the computation failed. This is ok: \(\delta\) is defined as a partial function to start with. If this bothers you, though, you can simply add another state to \(F\) (and thus to \(Q\)) representing failure, and add all the missing combinations of \(\Gamma\) and \(Q\) to the table, mapping all of them to that new state along with a movement to the right (which is always safe).

Also note how the set of states really represents two internal variables: where we are in the computation (not read anything yet, need to read one more variable, need to write result, done) and an integer accumulator to carry out the actual addition.

In order to compute 2 + 1 using this Turing machine, we would set up the ribbon with the following symbols:

\[2, 1, x, x, \dots\]

The entire state of the computation could be written:

\[B:\{[2], 1, x, x, \dots\}\]

where \(B\) represents the current state of the machine and \([]\) represents the current position of the head. Using this notation, the computation would proceed as follows:

  1. \(B:\{[2], 1, x, x, \dots\}\): \(B\) is not in \(F\), so we proceed. \(\delta(B, 2)\to\{B_2, x, R\}\), so the next state is:
  2. \(B_2:\{x, [1], x, x, \dots\}\): \(B_2\) is not in \(F\), so we proceed. \(\delta(B_2, 1)\to\{H_3, x, L\}\), so the next state is:
  3. \(H_3:\{[x], x, x, x, \dots\}\): \(H_3\) is not in \(F\), so we proceed. \(\delta(H_3, x)\to\{H, 3, R\}\), so the next state is:
  4. \(H:\{3,[x],x,x,\dots\}\): \(H\) is in \(F\), so we stop.

We can now read the ribbon, and we have computed \(2 + 1 = 3\).

Programmable machines

The example Turing machine above is running a specific program. This may seem a bit limiting, but, as programmers, we know that if you can write a program in a "Turing-complete" language, you can write a programmable program. This turns out to be true of Turing machines. They are, after all, the very definition of "Turing-complete".

That is, the model of a Turing machine is powerful enough to describe the model of a Turing machine. In other words, it is possible to define a specific Turing machine, generally called "the unversal Turing machine" (UTM), which can read a description of a Turing machine and an input for it, and simulate the execution of the described Turing machine on the given input. More precisely:

  • The alphabet of the UTM is rich enough that one can describe any other Turing machine using it. (For example, that alphabet could be \(\{0, 1, \epsilon\}\).)
  • That definition would include the specification of the input alphabet of the "simulated" Turing machine, or an encoding thereof.
  • A given run of the UTM would correspond to a given run of the simulated machine on a specific input for that machine; therefore the initial ribbon of the UTM can be described as containing two parts: first, a description of the simulated Turing machine, and second, the input for the simulated Turing machine.

The UTM can simulate itself simulating another machine, etc., though it obviously gets really tedious really fast.

Why is any of this relevant? If you squint a little bit, a Turing machine is a really good description of a programmer trying to understand imperative code. As an example, let's take a look at the following C code:

/* 1 */  int x = 1;
/* 2 */  int i = 1;
         do {
/* 3 */    x = x * i;
/* 4 */    i = i + 1;
/* 5 */  } while (i <= 4);
/* 6 */  printf("%d\n", x);

In order to predict what will be printed, a programmer would read the code one line at a time, updating their internal state of belief for what the value of each variable is at each step. Calling \(H\) the current line (for head), here is how such an evaluation would go:

  • \(\{x = ?, i = ?, H = 1\}\): x = 1 we need to set \(x\) to \(1\).
  • \(\{x = 1, i = ?, H = 2\}\): i = 1 we need to set \(i\) to \(1\).
  • \(\{x = 1, i = 1, H = 3\}\): x = x * i we need to set \(x\) to \(1\).
  • \(\{x = 1, i = 1, H = 4\}\): i = i + 1 we need to set \(i\) to \(2\).
  • \(\{x = 1, i = 2, H = 5\}\): i <= 4 is true so we set \(H\) to \(3\).
  • \(\{x = 1, i = 2, H = 3\}\): x = x * i we need to set \(x\) to \(2\).
  • \(\{x = 2, i = 2, H = 4\}\): i = i + 1 we need to set \(i\) to \(3\).
  • \(\{x = 2, i = 3, H = 5\}\): i <= 4 is true so we set \(H\) to \(3\).
  • \(\{x = 2, i = 3, H = 3\}\): x = x * i we need to set \(x\) to \(6\).
  • \(\{x = 6, i = 3, H = 4\}\): i = i + 1 we need to set \(i\) to \(4\).
  • \(\{x = 6, i = 4, H = 5\}\): i <= 4 is true so we set \(H\) to \(3\).
  • \(\{x = 6, i = 4, H = 3\}\): x = x * i we need to set \(x\) to \(24\).
  • \(\{x = 24, i = 4, H = 4\}\): i = i + 1 we need to set \(i\) to \(5\).
  • \(\{x = 24, i = 5, H = 5\}\): i <= 4 is false so we set \(H\) to \(6\).

And the program reaches its end. Note how closely this resembles the workings of a Turing machine: each numbered line of the program could be a symbol from the input alphabet, and the transition between each bullet point can easily be encoded as a \(\delta\) function. As described, the value of all variables can also be easily encoded into a state representation.


This correspondence between Turing machines and C code is no accident. The reason Turing machines are the most important model of computation we have is because they can be built. CPUs really are just physical implementations of a universal Turing machine (albeit with a finite ribbon of memory).

Like a Turing machine, imperative code is executed by reading instructions that tell:

  1. How to update the internal state of a processor, and
  2. Where to look for the next instruction.

Up next

This is why imperative code is the dominant programming paradigm: we started with a handful of models of computation, were able to build one of them, and then wrote programs for that one. Over time, we added layers of abstraction, but ultimately we mostly remained stuck within the confines of the underlying hardware. Importantly, note that this story does not say we're programming Turing machines because it's easy, or good. It simply was possible.

In the next post, I will describe another model of computation, lambda calculus, how it relates to functional programming, and why you should care.

  1. Jargon is usually seen as a bad thing to the uninitiated, but the development of jargon is actually a very important sign that a discipline is maturing. Jargon is what lets experts in a field have fruitful discussions with each other, where they can precisely refer to discipline-specific concepts and ideas and reduce ambiguity and misunderstandings. Computer programming is still a very young field of study; it has been pretty prolific in creating new words, or appropriating existing words, but it has not yet reached the point where those words truly have agreed-upon, precise definitions across the discipline. Even basic terms like "variable" or "type" can mean widely different things to different people. Object-oriented is definitely such a term, with at least one definition per programming language that claims to be it.

  2. The field of computer programming is generally regarded as starting in the 1950s, while the mathematical concept of a Turing machine was first described by Turing in 1936. Still, Babbage's analytical engine was, at its heart, a Turing machine, and thus imperative programming has been the dominant paradigm since 1837.

  3. This is obviously a very simplified version of history, with some inaccuracies thrown in for dramatic effect.

  4. It may seem recursive that we invented Turing machines as a way to express functions, yet we need a function in the description of a Turing machine. This is not actually a problem, because the purpose of a Turing machine is not to offer a way to define functions, but to offer a way to describe how to compute functions. "Simple" functions can be described as a set of pairs such that the first element is the input and the second element is the output (and there are no two pairs with the same first element). The functions that can be described in this way using a finite set of pairs do not need to be computed, as they can just be described with their values directly available. \(\delta\) is such a function. Some functions cannot be described using a finite set of pairs, but can still be described succinctly by a simple formula, such as the function "x squared". But some functions have an infinite set of inputs and no simple formula, such as square root.

  5. Because JSON is defined as using UTF-8 encoding, any JSON representation maps directly to a well-defined, unique sequence of bytes, which can be seen as just one big number. I'm using "representation" here instead of "value" because there can be multiple representations for the same value, as JSON authorizes the use of non-significant whitespace.

Tags: paradigms