You are Reading the Name of this Esolang ======================================== November 2007, Chris Pressey, Cat's Eye Technologies Introduction ------------ This programming language, called **You are Reading the Name of this Esolang**, is my first foray into the design space of programming languages whose programs contain undecidable elements. In the case of You are Reading the Name of this Esolang, these elements are the instructions themselves — or rather, the symbols that the instructions are composed of. Before we begin, some lexical notes. The name of this language is not pronounced exactly how it looks; rather, it is pronounced as an English speaker would pronounce the phrase "you are hearing the name of this esolang." In addition, it is strongly discouraged to refer to this language by the name "Yartnote", whether spoken or in writing, and in any capitalization scheme. After all, there may actually be a completely unrelated esolang called Yartnote one day, Zeus willing. A similar logic applies to the taboo on calling it "YRNE". Program structure ----------------- A You are Reading the Name of this Esolang program is a string of symbols drawn from the alphabet `0`, `1`, `[`, and `]`. `0` and `1` are interpreted as they are in the programming language Spoon, or rather, the slightly more clearly-specified version of Spoon that follows. The Spoon tape is considered unbounded in both directions. Each cell of the Spoon tape may contain any non-negative integer (again, unbounded.) In addition, attempting to decrement a cell below 0 results in an immediate termination of the program. Oh, and there's no way to change what the symbols are. But that's the extent of the difference, I think. For your convenience, the Spoon instructions are repeated here (taken from the public-domain [Spoon](http://esolangs.org/wiki/Spoon) entry on the [Esolang wiki](http://esolangs.org/wiki/)): 1 Increment the memory cell under the pointer 000 Decrement the memory cell under the pointer 010 Move the pointer to the right 011 Move the pointer to the left 0011 Jump back to the matching 00100 00100 Jump past the matching 0011 if the cell under the pointer is zero 001010 Output the character signified by the cell at the pointer 0010110 Input a character and store it at the cell in the pointer 00101110 Output the entire memory array 00101111 Immediately terminate program execution Each `[` must be matched with a `]`; between them lies a subprogram with the same structure as a general You are Reading the Name of this Esolang program. The meaning of this subprogram is determined from its structure as follows. The subprogram is considered to be given the same input as the entire program. If the subprogram halts on this input, it is reduced to a `1`, and if it loops forever on this input, it is reduced to a `0`. These reduced instructions are interpreted as they are in Spoon, as described above. Any output produced by the subprogram is simply discarded. Subprograms may themselves contain subprograms, nested to arbitrary depth, in which case the reduction above is recursively applied, from the inside out, until a string of only `0` and `1` symbols remains. This is then executed as if it were a Spoon program. Any syntactically ill-formed program or subprogram is considered to halt immediately, producing no output. Note however that, as a consequence of this, a subprogram can be syntactically ill-formed (for example consisting of a single `0`) while the parent program can still be syntactically OK (the subprogram just reduces to a `1` in the parent program.) Implementation notes -------------------- An implementation will determine if a particular subprogram halts, or not, if it can. Implementations may vary in the power of their proof methods used for this, but at a minimum must be able to recognize at least one subprogram that halts on any input, and one subprogram that loops forever on any input. This implementation-dependence should not strike anyone as too bizarre, I don't think — it is quite similar to how different implementations of a traditional systems-programming language can, for example, provide different levels of support for sizes of numerical data types like integers. Recall that the problem of telling if an arbitrary program in some given Turing-complete language halts on some input is *undecidable*, or equivalently, that set of all programs that halt on some input is *recursively enumerable*. The set of all programs that loop forever on some given input is the complement of this set, and it is called *co-recursively enumerable* (*co-r.e.* for short.) Despite this, there are many methods, ranging from simplistic to sophisticated, that can be used to prove in *specific* circumstances that a given program, on a given input, will either halt or fail to halt. These methods can be used in a You are Reading the Name of this Esolang implementation. The simplest method for proving that a subprogram halts is probably just to simulate it on the given input indefinately, returning `1` if it halts. If the subprogram does indeed halt, this technique will (eventually) reveal that fact. The simulation can be peformed concurrently with other subprograms, so that if no proof of halting for one subprogram is ever found, this will not prevent other subprograms from being checked. The simplest method for proving that a subprogram loops forever is probably to check it against a library of subprograms known to loop forever. For example, it can check if the program is `0010000000111001000011` (in Brainfuck: `[-]+[]`. You can readily assure yourself that this program loops forever, on any input.) This technique of course limits the number of recognizably looping subprograms that the implementation can handle to a finite number. Checking the general case, that is, recognizing an infinite number of forever-looping programs, is more difficult, however. Such an implementation will require techniques such as automatically finding a proof by induction, or abstract interpretation. However, ultimately, we know from the Halting Problem that there is no perfectly general technique that will recognize *every* program that loops forever. Computability class ------------------- You are Reading the Name of this Esolang can be trivially shown to be as powerful as Spoon, since valid Spoon programs are valid You are Reading the Name of this Esolang programs. (Modulo the absence of negative-valued tape cells and the other little variations mentioned above. Since these issues have been dealt with extensively in the Brainfuck "literature", such as it is, I'm not going to worry about them.) What about the other way around? Well, take as a starting point the fact (in classical logic, at least) that every Spoon program either halts at some point or loops forever. (Whether we can *discover* which of these is the case is a different story.) This means that every You are Reading the Name of this Esolang program has a "canonical" form consisting of just `1`'s and `0`'s — again, whether we have an interpreter powerful enough to discover it or not. At this level, Spoon is as powerful as "canonical" You are Reading the Name of this Esolang, because "canonical" You are Reading the Name of this Esolang programs are valid Spoon programs. Now we can go in the other direction. We can start with any "canonical" You are Reading the Name of this Esolang program, and replace each `1` with any You are Reading the Name of this Esolang subprogram that always halts. Since even the simple method, described above, of proving that a subprogram halts will always resolve to `1` if the subprogram does indeed halt, this subset of You are Reading the Name of this Esolang programs is still executable in Spoon (or any other Turing-complete language). We only need to add the halting proof mechanism to rewrite the program into "canonical" form, before executing or simulating it. This extends recursively to any arbitrary level of nesting, too: we can replace each `1` in each subprogram with subprograms that always halt, with no bound. We only have to test these subprograms recursively, from the inside out, to eventually recover a "canonical" program. However, something strange happens when we turn our attention to `0`'s. If we replace even one `0` with a subprogram that loops forever on some input, there is always a possibility that: a) the You are Reading the Name of this Esolang program will be run with that input, and that b) the interpreter cannot prove that the subprogram loops forever with that input. Because the set of programs that loop forever is co-r.e., there is no Turing machine (or Spoon program) that can look at any one Spoon program and say, yep, I'm certain that this here program loops forever on this input, so it should darn well be rewritten into a `0` symbol. Thus it seems that there are You are Reading the Name of this Esolang programs which no Spoon interpreter — indeed, not any interpreter for any Turing-complete language — is able to interpret. Since the case of `0`'s seems to mirror the case of `1`'s when it comes to expanding them into subprograms, we may conjecture that, instead of remaining basically the same as we consider more and more deeply nested subprograms (as it was with expanding `1`'s,) maybe the problem becomes more intractable the deeper we go in expanding `0`'s. Perhaps we are climbing up the arithmetic hierarchy? At any rate, *I* certainly initially conjectured that that was the case, but it appears to be off the mark. Say you have a Spoon interpreter that's equipped with an oracle. You can feed an input string and a Spoon program into the oracle, and the oracle tells you whether or not that program halts on that input. You could then use that oracle to resolve a given You are Reading the Name of this Esolang subprogram into a `0` or `1`. But, you could also do this recursively, resolving them from the inside outward. A Spoon interpreter with such an oracle would be able to simulate any You are Reading the Name of this Esolang program — no more powerful oracle is needed, no matter how deep the subprograms are nested. Discussion ---------- I started designing You are Reading the Name of this Esolang shortly after reading about the programming language Gravity, while trying to determine the sense in which it is "non-computable." In particular, I noticed that these two statements (taken from the [Gravity](http://esolangs.org/wiki/Gravity) entry on the [Esolang wiki](http://esolangs.org/wiki/),) on which the claim of Gravity's "non-computability" apparently rests, have ready analogies to problems the world of Turing machines: - "Although [Gravity's] behavior is well-defined and deterministic, the evolution of its space is in general non-computable [...]" The evolution of the state-space (set of successive configurations) of a universal Turing machine is also in general non-computable (there's no Turing machine that can tell you that some given configuration will never be reached.) - "It can be shown that a Turing machine cannot compute, in the general case, whether even a single collision [in a given Gravity program] will ever happen." It can also be shown that a Turing machine cannot compute, in the general case, whether or not even a single given state of another Turing machine's finite control will ever be reached. (Just make that state a halt state, and you have the Halting Problem right there.) Because of this, I am skeptical that Gravity is any more "non-computable" than a universal Turing machine. (I am, however, far from an expert on the computability of differential equations; it could be that the rather nonspecific term "non-computable", as used in that subfield, means something stronger than simply "undecidable".) At any rate, the idea interested me enough to spur me into designing a language that I *could* be reasonably certain was non-computable, in some sense that I could explain. The name You are Reading the Name of this Esolang was drifting around nearby in the æther at that moment, and seemed fitting enough for this monstrosity. The general approach was to simply force the language interpreter to decide — that is, to reduce to either a `0` or a `1` — some problem that is undecidable. This led to looking for something that needed specifically either a `0` or a `1` to specify something necessary, and that in turn led to the choice of Spoon as a base language. (Of course, I could have picked just about any language which is "its own binary Gödel numbering"; there are plenty to choose from there, but Spoon had a cool name. What can I say — I like The Tick.) The obvious choice of undecidable problem was whether another program halts or not. Making the subject of this problem a *subprogram* with the same structure as the general program let me examine the case of unbounded recursive descent. This turned out to be not quite as interesting as I hoped, but perhaps still somewhat illuminating. (Just what *would* it take, to require that a Spoon interpreter have a more powerful oracle than HP, to run every You are Reading the Name of this Esolang program? Perhaps [Banana Scheme](http://esolangs.org/wiki/Banana_Scheme) could provide some inspiration, here. *It* certainly seems to be climbing the arithmetic hierarchy, although I can't quite say how far. Possibly "damned far.") I suppose one or two other things can be said about You are Reading the Name of this Esolang. Unlike both Gravity and Banana Scheme, You are Reading the Name of this Esolang has a recursively enumerable *syntax*: the problem of whether or not a given string over the alphabet `0`, `1`, `[`, and `]` is even a well-formed You are Reading the Name of this Esolang program is undecidable! It's not entirely clear how to interpret the instruction `00101110`, "Output the entire memory array," in the context of having a tape unbounded in both directions. I suppose I ought to stipulate that we are to just output the portion of the tape that has been "touched", i.e. that the tape head has moved over. But really, it's not so important for the goals of You are Reading the Name of this Esolang, so maybe I should just leave it undefined, for kicks. The fact that every subprogram takes the *same* input (same as the main program) might lead to some interesting programs — programs which are unduly sensitive to changes in the input, I imagine. Of course, this doesn't affect the undecidability of subprograms, since they are always free to ignore input completely. There is no implementation, yet, but constructing an efficient one would be a good exercise in static program analysis. Happy undeciding! -Chris Pressey November 5, 2007 Chicago, Illinois, USA