Busy beaver

The busy beaver function quantifies the maximum score attainable by a Busy Beaver on a given measure. This is a noncomputable function. Also, a busy beaver function can be shown to have a greater growth rate[2] asymptotically than any computable function.

The busy beaver function, Σ: N → N, is defined such that Σ(n) is the maximum attainable score (the maximum number of 1s finally on the tape) among all halting 2-symbol n-state Turing machines of the above-described type, when started on a blank tape.

It is clear that Σ is a well-defined function: for every n, there are at most finitely many n-state Turing machines as above, up to isomorphism, hence at most finitely many possible running times.

This infinite sequence Σ is the busy beaver function, and any n-state 2-symbol Turing machine M for which σ(M) = Σ(n) (i.e., which attains the maximum score) is called a busy beaver. Note that for each n, there exist at least four n-state busy beavers (because, given any n-state busy beaver, another is obtained by merely changing the shift direction in a halting transition, another by shifting all direction changes to their opposite, and the final by shifting the halt direction of the all-swapped busy beaver).

Radó's 1962 paper proved that if f: ℕ → ℕ is any computable function, then Σ(n) > f(n) for all sufficiently large n, and hence that Σ is not a computable function.

Moreover, this implies that it is undecidable by a general algorithm whether an arbitrary Turing machine is a busy beaver. (Such an algorithm cannot exist, because its existence would allow Σ to be computed, which is a proven impossibility. In particular, such an algorithm could be used to construct another algorithm that would compute Σ as follows: for any given n, each of the finitely many n-state 2-symbol Turing machines would be tested until an n-state busy beaver is found; this busy beaver machine would then be simulated to determine its score, which is by definition Σ(n).)

Even though Σ(n) is an uncomputable function, there are some small n for which it is possible to obtain its values and prove that they are correct. It is not hard to show that Σ(0) = 0, Σ(1) = 1, Σ(2) = 4, and with progressively more difficulty it can be shown that Σ(3) = 6 and Σ(4) = 13 (sequence A028444 in the OEIS). Σ(n) has not yet been determined for any instance of n > 4, although lower bounds have been established (see the Known values section below).

In 2016, Adam Yedidia and Scott Aaronson obtained the first (explicit) upper bound on the minimum n for which Σ(n) is unprovable in ZFC. To do so they constructed a 7910-state[3] Turing machine whose behavior cannot be proven based on the usual axioms of set theory (Zermelo–Fraenkel set theory with the axiom of choice), under reasonable consistency hypotheses (stationary Ramsey property).[4][5] This was later reduced to 1919 states, with the dependency on the stationary Ramsey property eliminated,[6][7] and later to 748 states.[8]

A variant of Kolmogorov complexity is defined as follows [cf. Boolos, Burgess & Jeffrey, 2007]: The complexity of a number n is the smallest number of states needed for a BB-class Turing machine that halts with a single block of n consecutive 1s on an initially blank tape. The corresponding variant of Chaitin's incompleteness theorem states that, in the context of a given axiomatic system for the natural numbers, there exists a number k such that no specific number can be proved to have complexity greater than k, and hence that no specific upper bound can be proven for Σ(k) (the latter is because "the complexity of n is greater than k" would be proved if "n > Σ(k)" were proved). As mentioned in the cited reference, for any axiomatic system of "ordinary mathematics" the least value k for which this is true is far less than 10↑↑10; consequently, in the context of ordinary mathematics, neither the value nor any upper-bound of Σ(10 ↑↑ 10) can be proven. (Gödel's first incompleteness theorem is illustrated by this result: in an axiomatic system of ordinary mathematics, there is a true-but-unprovable sentence of the form "Σ(10 ↑↑ 10) = n", and there are infinitely many true-but-unprovable sentences of the form "Σ(10 ↑↑ 10) < n".)

In addition to the function Σ, Radó [1962] introduced another extreme function for the BB-class of Turing machines, the maximum shifts function, S, defined as follows:

• s(M) = the number of shifts M makes before halting, for any M in E_n,
• S(n) = max{ s(M) | M ∈ E_n } = the largest number of shifts made by any halting n-state 2-symbol Turing machine.

Because these Turing machines are required to have a shift in each and every transition or "step" (including any transition to a Halt state), the max-shifts function is at the same time a max-steps function.

Radó showed that S is noncomputable for the same reason that Σ is noncomputable — it grows faster than any computable function. He proved this simply by noting that for each n, S(n) ≥ Σ(n). Each shift may write a 0 or a 1 on the tape, while Σ counts a subset of the shifts that wrote a 1, namely the ones that hadn't been overwritten by the time the Turing machine halted; consequently, S grows at least as fast as Σ, which had already been proved to grow faster than any computable function.

The following connection between Σ and S was used by Lin & Radó [Computer Studies of Turing Machine Problems, 1965] to prove that Σ(3) = 6: For a given n, if S(n) is known then all n-state Turing machines can (in principle) be run for up to S(n) steps, at which point any machine that hasn't yet halted will never halt. At that point, by observing which machines have halted with the most 1s on the tape (i.e., the busy beavers), one obtains from their tapes the value of Σ(n). The approach used by Lin & Radó for the case of n = 3 was to conjecture that S(3) = 21, then to simulate all the essentially different 3-state machines for up to 21 steps. By analyzing the behavior of the machines that had not halted within 21 steps, they succeeded in showing that none of those machines would ever halt, thus proving the conjecture that S(3) = 21, and determining that Σ(3) = 6 by the procedure just described.

Inequalities relating Σ and S include the following (from [Ben-Amram, et al., 1996]), which are valid for all n ≥ 1:

${\displaystyle {\begin{aligned}S(n)&\geq \Sigma (n)\\S(n)&\leq (2n-1)\Sigma (3n+3)\\S(n)&<\Sigma (3n+6);\end{aligned}}}$

and an asymptotically improved bound (from [Ben-Amram, Petersen, 2002]): there exists a constant c, such that for all n ≥ 2,

${\displaystyle S(n)\leq \Sigma \left(n+\left\lceil {\frac {8n}{\log _{2}n}}\right\rceil +c\right).}$

${\displaystyle S(n)}$ tends to be close to the square of ${\displaystyle \Sigma (n)}$, and in fact many machines give ${\displaystyle S(n)}$ less than ${\displaystyle \Sigma ^{2}(n)}$.

As of 2016 the function values for Σ(n) and S(n) are only known exactly for n < 5.[5]

The current (as of 2018) 5-state busy beaver champion produces 4098 1s, using 47176870 steps (discovered by Heiner Marxen and Jürgen Buntrock in 1989), but there remain 18 or 19 (possibly under 10, see below) machines with non-regular behavior which are believed to never halt, but which have not been proven to run infinitely. Skelet lists 42 or 43 unproven machines, but 24 are already proven.[9] The remaining machines have been simulated to 81.8 billion steps, but none halted. Daniel Briggs also proved some machines.[10] Another source says 98 machines remain unproven. There is an analysis of holdouts.[11] So, it is likely that Σ(5) = 4098 and S(5) = 47176870, but this remains unproven, and it is unknown if there are any holdouts left (as of 2018). At the moment the record 6-state champion produces over 3.515×10¹⁸²⁶⁷ 1s (exactly (25*4³⁰³⁴¹+23)/9), using over 7.412×10³⁶⁵³⁴ steps (found by Pavel Kropitz in 2010). As noted above, these are 2-symbol Turing machines.

A simple extension of the 6-state machine leads to a 7-state machine which will write more than 10^{10101018705353} 1s to the tape, but there are undoubtedly much busier 7-state machines. However, other busy beaver hunters have different sets of machines.

Milton Green, in his 1964 paper "A Lower Bound on Rado's Sigma Function for Binary Turing Machines", constructed a set of Turing machines demonstrating that

${\displaystyle \Sigma (2k)>3\uparrow ^{k-2}3>A(k-2,k-2)\qquad {\mbox{for }}k\geq 2,}$

where ↑ is Knuth up-arrow notation and A is Ackermann's function.

Thus

${\displaystyle \Sigma (10)>3\uparrow \uparrow \uparrow 3=3\uparrow \uparrow 3^{3^{3}}=3^{3^{3^{.^{.^{.^{3}}}}}}}$

(with 3³³ = 7625597484987 terms in the exponential tower), and

${\displaystyle \Sigma (12)>3\uparrow \uparrow \uparrow \uparrow 3=g_{1},}$

where the number g₁ is the enormous starting value in the sequence that defines Graham's number.

In 1964 Milton Green developed a lower bound for the Busy Beaver function that was published in the proceedings of the 1964 IEEE symposium on switching circuit theory and logical design. Heiner Marxen and Jürgen Buntrock described it as "a non-trivial (not primitive recursive) lower bound". This lower bound can be calculated but is too complex to state as a single expression in terms of n. When n=8 the method gives Σ(8) ≥ 3 × (7 × 3⁹² - 1) / 2 ≈ 8.248×10⁴⁴.

It can be derived from current lower bounds that:

${\displaystyle \Sigma (2k+1)>3\uparrow ^{k-2}31>A(k-2,k-2)\qquad {\mbox{for }}k\geq 2,}$

In contrast, the best current (as of 2018) lower bound on Σ(6) is 10¹⁸²⁶⁷, which is greater than the lower bound given by Green's formula, 3³ = 27 (which is tiny in comparison). In fact, it is much greater than the lower bound: 3 ↑↑ 3 = 3³³ = 7625597484987, which is Green's first lower bound for Σ(8), and also much greater than the second lower bound: 3*(7*3⁹²-1)/2.

Σ(7) is by the same way much, much greater than current common lower bound 3³¹ (nearly 618 trillion), so the second lower bound is also very, very weak.

Suppose that S(n) is a computable function and let EvalS denote a TM, evaluating S(n). Given a tape with n 1s it will produce S(n) 1s on the tape and then halt. Let Clean denote a Turing machine cleaning the sequence of 1s initially written on the tape. Let Double denote a Turing machine evaluating function n + n. Given a tape with n 1s it will produce 2n 1s on the tape and then halt. Let us create the composition Double | EvalS | Clean and let n₀ be the number of states of this machine. Let Create_n₀ denote a Turing machine creating n₀ 1s on an initially blank tape. This machine may be constructed in a trivial manner to have n₀ states (the state i writes 1, moves the head right and switches to state i + 1, except the state n₀, which halts). Let N denote the sum n₀ + n₀.

Let BadS denote the composition Create_n₀ | Double | EvalS | Clean. Notice that this machine has N states. Starting with an initially blank tape it first creates a sequence of n₀ 1s and then doubles it, producing a sequence of N 1s. Then BadS will produce S(N) 1s on tape, and at last it will clear all 1s and then halt. But the phase of cleaning will continue at least S(N) steps, so the time of working of BadS is strictly greater than S(N), which contradicts to the definition of the function S(n).

The uncomputability of Σ(n) may be proved in a similar way. In the above proof, one must exchange the machine EvalS with EvalΣ and Clean with Increment — a simple TM, searching for a first 0 on the tape and replacing it with 1.

The uncomputability of S(n) can also be established by reference to the blank tape halting problem. The blank tape halting problem is the problem of deciding for any Turing machine whether or not it will halt when started on an empty tape. The blank tape halting problem is equivalent to the standard halting problem and so it is also uncomputable. If S(n) was computable, then we could solve the blank tape halting problem simply by running any given Turing machine with n states for S(n) steps; if it has still not halted, it never will. So, since the blank tape halting problem is not computable, it follows that S(n) must likewise be uncomputable.

For any model of computation there exist simple analogs of the busy beaver. For example, the generalization to Turing machines with n states and m symbols defines the following generalized busy beaver functions:

1. Σ(n, m): the largest number of non-zeros printable by an n-state, m-symbol machine started on an initially blank tape before halting, and
2. S(n, m): the largest number of steps taken by an n-state, m-symbol machine started on an initially blank tape before halting.

For example, the longest-running 3-state 3-symbol machine found so far runs 119112334170342540 steps before halting. The longest running 6-state, 2-symbol machine which has the additional property of reversing the tape value at each step produces 6147 1s after 47339970 steps. So S_RTM(6) ≥ 47339970 and Σ_RTM(6) ≥ 6147.

It is possible to further generalize the busy beaver function by extending to more than one dimension.

Likewise we could define an analog to the Σ function for register machines as the largest number which can be present in any register on halting, for a given number of instructions.

The following table lists the exact values and some known lower bounds for S(n, m) and Σ(n, m) for the generalized busy beaver problems. Note: entries listed as "?" are bounded from below by the maximum of all entries to left and above. These machines either haven't been investigated or were subsequently surpassed by a smaller machine.

The Turing machines that achieve these values are available on Pascal Michel's webpage. Each of these websites also contains some analysis of the Turing machines and references to the proofs of the exact values.

Values of S(n, m)
n m	2-state	3-state	4-state	5-state	6-state	7-state
2-symbol	6	21	107	47176870 ?	> 7.4×1036534	> 1010101018705353
3-symbol	38	≥ 119112334170342540	> 1.0×1014072	?	?	?
4-symbol	≥ 3932964	> 5.2×1013036	?	?	?	?
5-symbol	> 1.9×10704	?	?	?	?	?
6-symbol	> 2.4×109866	?	?	?	?	?
Values of Σ(n, m)
n m	2-state	3-state	4-state	5-state	6-state	7-state
2-symbol	4	6	13	4098 ?	> 3.5×1018267	> 1010101018705353
3-symbol	9	≥ 374676383	> 1.3×107036	?	?	?
4-symbol	≥ 2050	> 3.7×106518	?	?	?	?
5-symbol	> 1.7×10352	?	?	?	?	?
6-symbol	> 1.9×104933	?	?	?	?	?

A 748-state binary Turing machine has been constructed that halts iff ZFC is inconsistent.[14] A 744-state Turing machine has been constructed that halts iff the Riemann hypothesis is false.[6][14] A 43-state Turing machine has been constructed that halts iff Goldbach's conjecture is false, and a 27-state machine for that conjecture has been proposed but not yet verified.[6][14]