Can chess be solved?

There are some chess positions in which one player can force a win. Here’s an extremely simple example:

White just moves their queen up to a8 and checkmates Black. Some winning positions are harder to see than this. Take a look at the following position. Can you find a guaranteed two-move checkmate for White?

And for fun, here’s a harder one, again a guaranteed two-move checkmate, this time by Black:

Notice that in this last one, the opponent had multiple possible moves to choose from. A forced mate does not necessarily mean restricting your opponent to exactly one move on each of their turns. It just means that no matter what they do, you can still guarantee a win. Forced wins can become arbitrarily complicated and difficult to see if you’re looking many moves down the line, as you have to consider all the possible responses your opponent has at each turn. The world record for the longest forced win is the following position:

It’s White’s move, and White does have a strategy for a forced win. It just takes 549 turns to actually do this! (This strategy does violate the 50-move rule, which says that after 50 turns with no pawn moves or capture the game is drawn.) At this link you can watch the entire 549 move game. Most of it is totally incomprehensible to human players, and apparently top chess players that look at this game have reported that the reasoning behind the first 400 moves is opaque to them. Interestingly, White gets a pawn promotion after six moves, and it promotes it to a knight instead of a queen! It turns out that promoting to a queen actually loses for White, and their only way to victory is the knight promotion!

This position is the longest forced win with 7 pieces on the board. There are a few others that are similarly long. All of them represent a glimpse at the perfect play we might expect to see if a hypercomputer could calculate the whole game tree for chess and select the optimal move.

A grandmaster wouldn’t be better at these endgames than someone who had learned chess yesterday. It’s a sort of chess that has nothing to do with chess, a chess that we could never have imagined without computers. The Stiller moves are awesome, almost scary, because you know they are the truth, God’s Algorithm – it’s like being revealed the Meaning of Life, but you don’t understand one word.

Tim Krabbe

With six pieces on the board, the longest mate takes 262 moves (you can play out this position here). For five pieces, it’s 127 moves, for four it’s 43 moves, and the longest 3-man mate takes 28 moves.

But now a natural question arises. We know that a win can be forced in some positions. But how about the opening position? That is, is there a guaranteed win for White (or for Black) starting in this position?

Said more prosaically: Can chess be solved?

Zermelo’s theorem, published in “On an Application of Set Theory to the Theory of the Game of Chess” (1913), was the first formal theorem in game theory. It predated the work of von Neumann (the so-called “father of game theory”) by 15 years. It proves that yes, it is in fact possible to solve chess. We don’t know what the solution is, but we know that either White can force a win, or Black can force a win, or the result will be a draw if both play perfectly.

Of course, the guarantee that in principle there is a solution to chess doesn’t tell us much in practice. The exponential blowup in the number of possible games is so enormous that humans will never find this solution. Nonetheless, I still find it fascinating to think that the mystery of chess is ultimately a product of computational limitations, and that in principle, if we had a hypercomputer, we could just find the unique best chess game and watch it play out, either to a win by one side or to a draw. That would be a game that I would love to see.

✯✯✯

Here’s another fun thing. There’s an extremely bizarre variant on chess called suicide chess (or anti-chess). The goal of suicide chess is to lose all your pieces. Of course, if all the rules of play were the same, it would be practically impossible to win (since your opponent could always just keep refusing to take a piece that you are offering them). To remedy this, in suicide chess, capturing is mandatory! And if you have multiple possible captures, then you can choose among them.

Suicide chess gameplay is extremely complicated and unusual looking, and evaluating who is winning at any given moment tends to be really difficult, as sudden turnarounds are commonplace compared to ordinary chess. But one simplifying factor is that it tends to be easier to restrict your opponents’ moves. In ordinary chess, you can only restrict your opponents’ moves by blocking off their pieces or threatening their king. But in suicide chess, your opponents’ moves are restricted ANY time you put one of your pieces in their line of fire! This feature of the gameplay makes the exponential blow up in possible games more manageable.

Given this, it probably won’t be much of a surprise that suicide chess is, just like ordinary chess, in principle solvable. But here’s the crazy part. Suicide chess is solved!!

That’s right: it was proven a triple of years ago that White can force a win by moving first with e3!

Here’s the paper. The proof amounts to basically running a program that looks at all possible responses to e3 and expands out the game tree, ultimately showing that all branches can be terminated with White losing all pieces and winning the game.

Not only do we know that by starting with e3, White is guaranteed a win, we also know that Black can force a win if White starts with any of the following moves: a3, b4, c3, d3, d4, e4, f3, f4, h3, h4, Nc3, Nf3. As far as I was able to tell, there are only six opening moves remaining for which we don’t know if White wins, Black wins, or they draw: a4, b3, c4, e3, g3, and g4.

✯✯✯

Alright, final chess variant trivia. Infinite chess is just chess, but played on an infinite board.

There’s a mind-blowing connection between infinite chess and mathematical logic. As a refresher, a little while back I discussed the first-order theory of Peano arithmetic. This is the theory of natural numbers with addition and multiplication. If you recall, we found that Peano arithmetic was incomplete (in that not all first-order sentences about the natural numbers can be proven from its axioms). First order PA is also undecidable, in that there exists no algorithm that takes in a first order sentence and returns whether it is provable from the axioms. (In fact, first order logic in general is undecidable! To get decidability, you have to go to a weaker fragment of first order logic known as monadic predicate calculus, in which predicates take only one argument and there are no functions. As soon as you introduce a single binary predicate, you lose decidability.)

Okay, so first order PA (the theory of natural numbers with addition and multiplication) is incomplete and undecidable. But there are weaker fragments of first order PA that are decidable! Take away multiplication, and you have Presburger arithmetic, the theory of natural numbers with addition. Take away addition, and you have Skolem arithmetic, the theory of natural numbers with multiplication. Both of these fragments are significantly weaker than Peano arithmetic (each is unable to prove general statements about the missing operation, like that multiplication is commutative for Presburger arithmetic). But in exchange for this weakness, you get both decidability and completeness!

How does all this relate to infinite chess? Well, consider the problem of determining whether there exists a checkmate in n turns from a given starting position. This seems like a really hard problem, because unlike in ordinary chess, now it’s possible for there to be literally infinite possible moves for a given player from a position. (For instance, a queen on an empty diagonal can move to any of the infinite locations on this diagonal.) So apparently, the game tree for infinite chess, in general, branches infinitely. Given this, we might expect that this problem is not decidable.

Well, it turns out that any instance of this problem (any particular board setup, with the question of whether there’s a mate-in-n for one of the players) can be translated into a sentence in Presburger arithmetic. You do this by translating a position into a fixed length sequence of natural numbers, where each piece is given a sequence of numbers indicating its type and location. The possibility of attacks can be represented as equations about these numbers. And since the distance pieces (bishops, rooks, and queens – those that have in general an infinite number of available moves) all move in straight lines, there are simple equations expressible in Presburger arithmetic that describe whether these pieces can attack other pieces! From the attack relations, you can build up more complicated relations, including the mate-in-n relation.

So we have a translation from the mate-in-n problem to a sentence in Presburger arithmetic. But Presburger arithmetic is decidable! So there must also be a decision procedure for the mate-in-n problem in infinite chess. And not only is there a decision procedure for the mate-in-n problem, but there’s an algorithm that gives the precise strategy that achieves the win in the fewest number of moves!

Here’s the paper in which all of this is proven. It’s pretty wild. Many other infinite chess problems can be proven to be decidable by the same method (demonstrating interpretability of the problem in Presburger arithmetic). But interestingly, not all of them! This has a lot to do with the limitations of first-order logic. The question of whether, in general, there is a forced win from a given position can not be shown to be decidable in this way. (This relates to the general impossibility in first-order logic of expressing infinitely long statements. Determining whether a given position is a winning position for a given player requires looking at the mate-in-n problem, but without any upper bound on what this n is – on how many moves the win may take.) It’s not even clear whether the winning-position problem can be phrased in first-order arithmetic, or whether it requires going to second-order!

The paper takes this one step further. This proof of the decidability of the mate-in-n problem for infinite chess doesn’t crucially rest upon the two-dimensionality of the chess board. We could easily translate the proof to a three-dimensional board, just by changing the way we code positions! So in fact, we have a proof that the mate-in-n problem for k-dimensional infinite chess is decidable!

I’ll leave you with this infinite chess puzzle:

It’s White’s turn. Can they guarantee a checkmate in 12 moves or less?

For Loops and Bounded Quantifiers in Lambda Calculus

Lambda calculus is an extremely minimal language that is powerful enough to express every computation that a Turing machine can do. Here is a fantastic video explaining the basic “rules of the game”:

Figuring out how to translate programs into lambda calculus can be a challenging puzzle. I recently discovered a nice way to use lambda calculus to generate for loops and bounded quantifiers, thus allowing recursion without requiring the Y Combinator. I’m sure similar things have been discovered by others, but things feel cooler than they are when you’ve found them, so I’ll present this function here in this post. 🙂

First, let me lay out some of the most important basic functions in lambda calculus. (A lot of this will be rehashing bits of the video lecture.)

Useful Tools

Propositional Logic

To give you a sense of how proving things works in lambda calculus, here’s a quick practice proof that ¬¬T is the same thing as T. We’ll prove this by reducing ¬¬T to T, which will show us that the two functions are extensionally equivalent (they return the same values on all inputs). Each step, we will either be substituting in the definition of some symbol, or evaluating a function by substituting its inputs in.

Another more convoluted approach is to show that the function (¬¬T ↔ T) reduces to the function T, which is to say that the two are equivalent. Notice that we’re not saying that the expression (¬¬T ↔ T) “has the truth value True”. There is no such thing as expressions with truth values in lambda calculus, there are just functions, and some functions are equivalent to the True function.

To shorten this proof, I’ll start out by proving the equivalence of ¬T and F, as well as the equivalence of ¬F and T. This will allow me to translate between these functions in one step.

Got the hang of it? Ok, on to the natural numbers!

Natural Numbers

In lambda calculus, numbers are adverbs. 1 is not “one”, it’s “once”. 1 is a function that takes in a function f and tells you to apply it one time. 2 is “twice”; it’s a function that tells you to apply f twice. And so on.

Defining things this way allows us to have beautifully simple definitions of addition, multiplication, and exponentiation. For instance, the successor of n is just what you get by one more application of f on top of n applications of f. And n plus m is just n S m, because this means that S (the successor function) should be applied n times to m. See if you can see for yourself why the definitions of multiplication and exponentiation make sense (and how they’re different! It’s not just the order of the n and m!).

Let’s prove that 1 + 1 = 2 using lambda calculus!

One more: Here’s a proof that 0 + k = k.

Interestingly, formally proving that k + 0 = k is significantly more convoluted, and there’s no obvious way to do it in general for all k (as opposed to producing a separate proof for each possible value of k). In addition, the proof length will end up being the size of k.

Okay, let’s dive deeper by looking at our first non-trivial lambda calculus data structure, the pair.

Pair Data Structure

The pair function can be fed two values (a and b) which can then be referenced at a later time by feeding it one more function. Feeding the function either of T or F will just select either the first or the second element in the pair. It might not be immediately obvious, but having this ability to, as it were, store functions “in memory” for later reference gives us enormous power. In fact, at this point we have everything we need to introduce the magical function which opens the door to quantification and recursion:

Magic Function

All that Φ does is take a pair of functions (a, b) to a new pair (f(a, b), g(a, b)). Here it’s written in more familiar notation:

Φ: (a, b)  (f(a, b), g(a, b))

Different choices of f and g give us very different types of behavior when we repeatedly apply Φ. First of all, let’s use Φ to subtract and compare numbers.

Subtraction and Comparison

φ1 simply chooses f and g to get the following behavior:

φ1: (a, b)  (b, b+1)

Here’s what happens with repeated application of φ1 to the pair (0, 0):

(0, 0) → (0, 1) → (1, 2) → (2, 3) → …

Looking at this, you can see how n applications of φ1 to (0, 0) gives the pair (n – 1, n), which justifies our definition of the predecessor function.

Our power is enhancing every minute; now, we create for loops!

For Loops and Quantifiers

B S F is the composition of successor and True, so it takes two inputs, fetches the value of the first input, and then adds one to it. Since B S T is the first input to Φ, it will be the function that determines the value of the first member of the output pair. In other words, φ2 gives the following mapping:

φ2: (a, b)  (a + 1, f(a, b))

To get a for loop, we now just iterate this function the appropriate number of times!

The for function I’ve written takes in four parameters (n, m, f, a), and is equivalent to the following program:

The  function is equivalent to the following:

𝑥 ∈ [n, m-1] θ(𝑥)

In code, this is:

And the  function is equivalent to:

𝑥 ∈ [n, m-1] θ(𝑥)

With these powerful tools in hand, we can now define more interesting functions, like one that detects if an input number is prime!

Let’s use lambda calculus to see if 2 is prime!

That got a little messy, but it’s nothing compared to the computation of whether 3 is prime!

This might sound strange, but I actually find it amazing how short and simple this is. To be fair, I skipped steps if all that they did was substitute in the definition of a function, instead opting to just immediately apply the definition and cut the number of steps in half. The full proof would actually be twice as long!

But nonetheless, try writing a full proof of the primality of 3 using only ZFC set theory in anywhere near as few lines! It’s interesting to me that a language as minimal and bare-bones as that of lambda calculus somehow manages to produce such concise proofs of interesting and complicated statements.

There’s a problem with infinity

Last post I described the Ross-Littlewood paradox, in which an ever-expanding quantity of numbered billiard balls are placed into a cardboard box in such a way that after an infinite number of steps the box ends up empty. Here’s a version of this paradox:

Process 1
Step 1: Put 1 through 9 into the box.
Step 2: Take out 1, then put 10 through 19 into the box.
Step 3: Take out 2, then put 20 through 29 into the box.
Step 4: Take out 3, then put 30 through 39 into the box.
And so on.

Box contents after each step
Step 1: 1 through 9
Step 2: 2 through 19
Step 3: 3 through 29
Step 4: 4 through 39
And so on.

Now take a look at a similar process, where instead of removing balls from the box, we just change the number that labels them (so, for example, we paint a 0 after the 1 to turn “Ball 1” to “Ball 10″).

Process 2
Step 1: Put 1 through 9 into the box
Step 2: Change 1 to 10, then put 11 through 19 into the box.
Step 3: Change 2 to 20, then put 21 through 29 in.
Step 3: Change 3 to 30, then put 31 through 39 in.
And so on.

Box contents after each step
Step 1: 1 through 9
Step 2: 2 through 19
Step 3: 3 through 29
Step 4: 4 through 39
And so on.

Notice that the box contents are identical after each step. If that’s all that you are looking at (and you are not looking at what the person is doing during the step), then the two processes are indistinguishable. And yet, Process 1 ends with an empty box, and Process 2 ends with infinitely many balls in the box!

Why does Process 2 end with an infinite number of balls in it, you ask?

Process 2 ends with infinitely many balls in the box, because no balls are ever taken out. 1 becomes 10, which later becomes 100 becomes 1000, and so on forever. At infinity you have all the natural numbers, but with each one appended an infinite number of zeros.

So apparently the method you use matters, even when two methods provably get you identical results! There’s some sort of epistemic independence principle being violated here. The outputs of an agent’s actions should be all that matters, not the specific way in which the agent obtains those outputs! Something like that.

Somebody might respond to this: “But the outputs of the actions aren’t the same! In Process 1, each step ten are added and one removed, whereas in Process 2, each step nine are added. This is the same with respect to the box, but not with respect to the rest of the universe! After all, those balls being removed in Process 1 have to go somewhere. So somewhere in the universe there’s going to be a big pile of discarded balls, which will not be there in Process 2.

This responds holds water as long as our fictional universe doesn’t violate conservation of information, as if not, these balls can just vanish into thin air, leaving no trace of their existence. But that rebuttal feels cheap. Instead, let’s consider another variant that gets at the same underlying problem of “relevance of things that should be irrelevant”, but avoids this problem.

Process 1 (same as before)
Step 1: Put 1 through 9 into the box.
Step 2: Take out 1, then put 10 through 19 into the box.
Step 3: Take out 2, then put 20 through 29 into the box.
Step 4: Take out 3, then put 30 through 39 into the box.
And so on.

Box contents after each step
Step 1: 1 through 9
Step 2: 2 through 19
Step 3: 3 through 29
Step 4: 4 through 39
And so on.

And…

Process 3
Step 1: Put 1 through 9 into the box.
Step 2: Take out 9, then put 10 through 19 into the box.
Step 3: Take out 19, then put 20 through 29 into the box.
Step 4: Take out 29, then put 30 through 39 into the box.
And so on.

Box contents after each step
Step 1: 1 through 9
Step 2: 1 to 8, 10 to 19
Step 3: 1 to 8, 10 to 18, 20 to 29
Step 4: 1 to 8, 10 to 18, 20 to 28, 30 to 39
And so on

Okay, so as I’ve written it, the contents of each box after each step are different in Processes 1 and 3. Just one last thing we need to do: erase the labels on the balls. The labels will now just be stored safely inside our minds as we look over the balls, which will be indistinguishable from one another except in their positions.

Now we have two processes that look identical at each step with respect to the box, AND with respect to the external world. And yet, the second process ends with an infinite number of balls in the box, and the first with none! (Every number that’s not one less than a multiple of ten will be in there.) It appears that you have to admit that the means used to obtain an end really do matter.

But it’s worse than this. You can arrange things so that you can’t tell any difference between the two processes, even when observing exactly what happens in each step. How? Well, if the labelling is all in your heads, then you can switch around the labels you’ve applied without doing any harm to the logic of the thought experiment. So let’s rewrite Process 3, but fill in both the order of the balls in the box and the mental labelling being used:

Process 3
1 2 3 4 5 6 7 8 9
Mentally rotate labels to the right:
9 1 2 3 4 5 6 7 8
Remove the furthest left ball:
1 2 3 4 5 6 7 8
Add the next ten balls to the right in increasing order:
1 2 3 4 5 6 7 8 10 11 12 13 14 15 16 17 18 19
Repeat!

Compare this to Process 1, supposing that it’s done without any relabelling:

Process 1
1 2 3 4 5 6 7 8 9
Remove the furthest left ball:
2 3 4 5 6 7 8 9
Add the next tell balls to the right in increasing order:
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 18 19
Repeat!

If the labels are all in your head, then these two processes are literally identical except for how a human being is thinking about them.

But looking at Process 3, you can prove that after Step 1 there will always be a ball labelled 1 in the box. Same with 2, 3, 4, and all other numbers that are not a multiple of 10 minus one. Even though we remove an infinity of balls, there are ball numbers that are never removed. And if we look at the pile of discarded balls, we’ll see that it consists of 9, 19, 29, 39, and so on, but none of the others. Unless some ball numbers vanish in the process (which they never do!), all the remainders must still be sitting in the box!

So we have two identical-in-every-relevant-way processes, one of which ends with an infinite number of balls in the box and the other with zero. Do you find this troubling? I find this very troubling. If we add some basic assumption that an objective reality exists independent of our thoughts about it, then we’ve obtained a straightforward contradiction.

✯✯✯

Notice that it’s not enough to say “Well, in our universe this process could never be completed.” This is for two reasons:

First of all, it’s actually not obvious that supertasks (tasks involving the completion of an infinite number of steps in a finite amount of time) cannot be performed in our universe. In fact, if space and time are continuous, then every time you wave your hand you are completing a sort of supertask.

You can even construct fairly physically plausible versions of some of the famous paradoxical supertasks. Take the light bulb that blinks on and off at intervals that get shorter and shorter, such that after some finite duration it has blinked an infinity of times. We can’t say that the bulb is on at the end (as that would seem to imply that the sequence 0101010… had a last number) or that it is off (for much the same reason). But these are the only two allowed states of the bulb! (Assume the bulb is robust against bursting and all the other clever ways you can distract from the point of the thought experiment.)

Now, here’s a variant that seems fairly physically reasonable:

A ball is dropped onto a conductive plate that is attached by wire to a light bulb. The ball is also wired to the bulb, so that when the ball contacts the plate, a circuit is completed that switches the light bulb on. Each bounce, the ball loses some energy to friction, cutting its velocity exactly in half. This means that after each bounce, the ball hangs in the air for half as long as it did the previous bounce.

Suppose the time between the first and second bounce was 1 second. Then the time between the second and third will be .5 seconds. And next will be .25 seconds. And so on. At 2 seconds, the ball will have bounced an infinite number of times. So at 2 seconds, the light bulb will have switched on and off an infinite number of times.

And of course, at 2 seconds the ball is at rest on the plate, completing the circuit. So at 2 seconds, upon the completion of the supertask, the light will be on.

Notice that there are no infinite velocities here, or infinite quantities of energy. Just ordinary classical mechanics applied to a bouncing ball and a light bulb. What about infinite accelerations? Well even that is not strictly speaking necessary; we just imagine that each velocity reversal takes some amount of time, which shrinks to zero as the velocity shrinks to zero in such a way as to keep all accelerations finite and sum to a finite total duration.

All this is just to say that we shouldn’t be too hasty in dismissing the real-world possibility of apparently paradoxical supertasks.

But secondly, and more importantly, physical possibility is not the appropriate barometer of whether we should take a thought experiment seriously. Don’t be the person that argues that the fat man wouldn’t be sufficient to stop a trolley’s momentum. When we find that some intuitive conceptual assumptions lead us into trouble, the takeaway is that we need to closely examine and potentially revise our concepts!

Think about Russell’s paradox, which showed that some of our most central intuitions about the concept of a set lead us to contradiction. Whether or not the sets that Bertie was discussing can be pointed to in the physical world is completely immaterial to the argument. Thinking otherwise would have slowed down progress in axiomatic set theory immensely!

These thought experiments are a problem if you believe that it is logically possible for there to be a physical universe in which these setups are instantiated. That’s apparently all that’s required to get a paradox, not that the universe we live in happens to be that one.

So it appears that we have to conclude some limited step in the direction of finitism, in which we rule out a priori the possibility of a universe that allows these types of supertasks. I’m quite uncomfortable with this conclusion, for what it’s worth, but I don’t currently see a better option.

The Puzzle

You have in front of you an empty box. You also have on hand an infinite source of billiard balls, numbered 0, 1, 2, 3, 4, and so on forever.

At time zero, you place balls 0 and 1 in the box.

In thirty minutes, you remove ball 0 from the box, and place in two new balls (2 and 3).

Fifteen minutes after that, you remove ball 1 from the box, and place in two new balls (4 and 5).

7.5 minutes after that, you remove ball 2 and place in balls 6 and 7.

And so on.

After an hour, you will have taken an infinite number of steps. How many billiard balls will be in the box?

✯✯✯

At time zero, the box contains two balls (0 and 1). After thirty minutes, it contains three (1, 2, and 3). After 45 minutes, it contains four (2, 3, 4, and 5). You can see where this is going…

Naively taking the limit of this process, we arrive at the conclusion that the box will contain an infinity of balls.

But hold on. Ask yourself the following question: If you think that the box contains an infinity of balls, name one ball that’s in there. Go ahead! Give me a single number such that at the end of this process, the ball with that number is sitting in the box.

The problem is that you cannot do this. Every single ball that is put in at some step is removed at some later step. So for any number you tell me, I can point you to the exact time at which that ball was removed from the box, never to be returned to it!

But if any ball that you can name can be proven to not be in the box.. and every ball you put in there was named… then there must be zero balls in the box at the end!

In other words, as time passes and you get closer and closer to the one-hour mark, the number of balls in the box appears to be growing, more and more quickly each moment, until you hit the one-hour mark. At that exact moment, the box suddenly becomes completely empty. Spooky, right??

Let’s make it weirder.

What if at each step, you didn’t just put in two new balls, but one MILLION? So you start out at time zero by putting balls 0, 1, 2, 3, and so on up to 1 million into the empty box. After thirty minutes, you take out ball 1, but replace it with the next 1 million numbered balls. And at the 45-minute mark, you take out ball 2 and add the next 1 million.

What’ll happen now?

Well, the exact same argument we gave initially applies here! Any ball that is put in the box at any point, is also removed at a later point. So you literally cannot name any ball that will still be in the box after the hour is up, because there are no balls left in the box! The magic of infinity doesn’t care about how many more balls you’ve put in than removed at any given time, it still delivers you an empty box at the end!

Now, here’s a final variant. What if, instead of removing the smallest numbered ball each step, you removed the largest numbered ball?

So, for instance, at the beginning you put in balls 0 and 1. Then at thirty minutes you take out ball 1, and put in balls 2 and 3. At 45 minutes, you take out ball 3, and put in balls 4 and 5. And so on, until you hit the one hour mark. Now how many balls are there in the box?

Infinity! Why not zero like before? Well, because now I can name you an infinity of numbers whose billiard balls are still guaranteed to be in the box when the hour’s up. Namely, 0, 2, 4, 6, and all the other even numbered balls are still going to be in there.

Take a moment to reflect on how bizarre this is. We removed the exact same number of balls each step as we did last time. All that changed is the label on the balls we removed! We could even imagine taking off all the labels so that all we have are identical plain billiard balls, and just labeling them purely in our minds. Now apparently the choice of whether to mentally label the balls in increasing or decreasing order will determine whether at the end of the hour the box is empty or packed infinitely full. What?!? It’s stuff like this that makes me sympathize with ultrafinitists.

One final twist: what happens if the ball that we remove each step is determined randomly? Then how many balls will there be once the hour is up? I’ll leave it to you all to puzzle over!

Can an irrational number raised to an irrational power be rational?

There’s a wonderful proof that yes, indeed, this is possible, and it goes as follows:

Let’s consider the number $\sqrt 2 ^ {\sqrt 2}$. This number is either rational or irrational. Let’s examine each case.

Case 1: $\sqrt 2 ^ {\sqrt 2}$$\sqrt 2 ^ {\sqrt 2}$ is rational.

Recall that $\sqrt 2$ is irrational. So if $\sqrt 2 ^ {\sqrt 2}$ is rational, then we have proven that it’s possible to raise an irrational number to an irrational power and get a rational value. Done!

Case 2: $\sqrt 2 ^ {\sqrt 2}$$\sqrt 2 ^ {\sqrt 2}$ is irrational.

In this case, $\sqrt 2 ^ {\sqrt 2}$ and $\sqrt 2$ are both irrational numbers. So what if we raise the $\sqrt 2 ^ {\sqrt 2}$ to the power of $\sqrt 2$?

$\left( \sqrt 2 ^ {\sqrt 2} \right) ^{\sqrt 2} = \sqrt 2 ^ {\sqrt 2 \cdot \sqrt 2} = \sqrt 2 ^ 2 = 2$

So in this case, we have again found a pair of irrational numbers such that one raised to the power of the other is a rational number! Proof complete!

✯✯✯

One thing that’s fun about this proof is that the result is pretty surprising. I would not have guessed a priori that you could get a rational by raising one irrational to another; it just seems like irrationality is the type of thing that would be closed under ordinary arithmetic operations.

But an even cooler thing is that it’s a non-constructive proof. By the end of the proof, we know for sure that there is a pair of irrational numbers such that one raised to the other gives us a rational number, but we have no idea whether it’s ($\sqrt 2$, $\sqrt 2$) or ($\sqrt 2 ^ {\sqrt 2}$,$\sqrt 2$).

(It turns out that it’s the second. The Gelfond–Schneider theorem tells us that for any two non-zero algebraic numbers a and b with a ≠ 1 and b irrational, the number ab is irrational. So $\sqrt 2 ^ {\sqrt 2}$ is in fact irrational.)

Now, most mathematicians are totally fine with non-constructive proofs, as long as they follow all the usual rules of proofs. But there is a branch of mathematics known as constructive mathematics that only accepts constructive proofs of existence. Within constructive mathematics, this proof is not valid!

Now, it so happens that you can prove the irrationality of $\sqrt 2 ^ {\sqrt 2}$ by purely constructive means, but that’s besides the point. To my eyes, the refusal to accept such an elegant and simple proof because it asserts a number’s existence without telling us exactly what it is just looks a little silly!

Along similar lines, here’s one more fun problem.

Are $\pi + e$$\pi + e$ and $\pi - e$$\pi - e$ transcendental?

We know that $\pi$ and $e$ are both transcendental numbers (i.e. they cannot be expressed as the roots of any polynomial with rational coefficients). But are $\pi + e$ and $\pi - e$ both transcendental?

It turns out that this amazingly simple sounding problem is unsolved to this day! But one thing that we do know is that it can’t be that neither of them are transcendental. Because if this was the case, then their sum $(\pi + e) + (\pi - e) = 2 \pi$ would also not be transcendental, which we know is false! So we know that at least one of them has to be true, using a proof that doesn’t guarantee the truth of either of them! Cool, right?