As discussed on page 753, an important indication of computational sophistication in a system is for its ultimate behavior to be undecidable, in the sense that a limited number of steps in a standard universal system cannot determine in general what the system will do after an infinite number of steps, and whether, for example, it will ever in some sense halt. Such undecidability is inevitable in any system that is universal. But what about other systems? So long as one only ever looks at the original input and final output it turns out that one can construct a system that exhibits undecidability but is not universal. One trivial way to do so is to take a universal system but modify it so that if it ever halts its output is discarded and, say, replaced by its original input. The lack of meaningful output prevents such a system from being universal, but the question of whether the system halts is still undecidable. Nevertheless, the pattern of this undecidability is just the same as for the underlying universal system. So one can then ask whether it is possible to have a system which exhibits undecidability, but with a pattern that does not correspond to that of any universal system.
As I discuss on page 1137, almost all known proofs of undecidability in practice work by reduction to the halting problem for some universal system—this is, by showing that if one could resolve whatever is supposed to be undecidable then one could also solve the halting problem for a universal system. But in 1956 Richard Friedberg and Albert Muchnik both gave an intricate and abstract construction of a system that has a halting problem which is undecidable but is not reducible to the halting problem of any universal system.
The pictures at the top of the facing page show successive steps in the evolution of an analog of their system. The input is an integer that gives a position in either of the two rows of cells at the bottom of each picture. All these cells are initially white, but some eventually become black—and the system is considered to halt for a particular input if the corresponding cell ever becomes black.
The rules for the system are quite complicated, and in essence work by progressively implementing a generalization of a diagonal argument of the kind discussed on page 1128. Note first that the configuration of cells in the rows at the bottom of each picture can be thought of as successive finite approximations to tables for an oracle (see page 1126) which gives the solution to the halting problem for each possible input to the system. To set up the generalized diagonal argument one needs a way to list all possible programs. Any type of program that supports universality can be used for this purpose; the pictures shown use essentially the register machines from page 97. Each row above the bottom one corresponds in effect to a successive register machine—and shows, if relevant, its output when given as input the integer corresponding to that position in the row, together with the complete bottom row of cells found so far. (A dot indicates that the register machine does not halt.) The way the system works is to put down new black cells in the bottom row in just such a way as to arrange that for any register machine at least the output shown will ultimately not agree with the cells in the bottom row. As indicated by vertical gray lines, there is sometimes temporary agreement, but this is always removed within a finite number of steps.
The fact that no register machine can ever ultimately give output that agrees everywhere with the bottom row of cells then demonstrates that the halting problem for the system—whose results appear in the bottom row—must be undecidable. Yet if this halting problem were reducible to a halting problem for a universal system, then by using its results one should ultimately be able to solve the halting problem for any system. However, even using the complete bottom row of cells on the left it turns out that the construction is such that no register machine can ever yield results after any finite number of steps that agree everywhere with the row of cells on the right—thus demonstrating that the halting problem for the system is not reducible to the halting problem for a universal system.
Note however that this result is extremely specific to looking only at what is considered output from the system, and that inside the system there are all sorts of components that are definitely universal.