This post explains the surprising link between math and juggling. Juggling theory is a beautiful example of how we can get more insight in a seemingly non-mathematical area by building a theory and proving stuff.

Most of this post doesn't assume any specialized mathematical background. Towards the end, basics of graph theory and number theory are used.

## What is juggling?

Juggling is a recreational activity where a person manipulates multiple objects (usually balls, rings or clubs) by repeatedly throwing and catching them. If you haven't seen professional juggling before, I highly recommend watching this amazing performance by Anthony Gatto.

There are different patterns in juggling. The animations below depict three well-known patterns: cascade, shower and fountain.

The fundamental problem in juggling theory is how to succinctly describe a juggling pattern. I used GIF animations above to explain to you what cascade, shower and fountain mean, but we are interested in a way of representing juggling patterns that is much smaller (just a few characters long), reveals the structure of the pattern, and can be read and analyzed by machines. In 1981, Paul Klimek invented a way of associating a sequence of integers to each juggling pattern. This sequence of integers is called a siteswap, and it captures important properties of a jugging pattern. In this post, I'll show how to succinctly represent juggling patterns, and siteswaps form the most important part of this representation.

An amazing by-product of having such a representation is the ability to automate the process of inventing new patterns by brute-force enumeration of integer sequences. Previously, only non-beginner jugglers could invent patterns using hit-and-trial.

Another advantage of siteswaps is that they can be used as input to programs that create animations of juggling patterns. In fact, this is exactly how I created all the animations in this post, using Juggling Lab.

## The features of a juggling pattern

There are some assumptions that we'll impose on juggling patterns. These set of assumptions define our model. Here we will look at a model that is powerful enough to capture most of the well-known patterns and at the same time is simple enough to analyze.

Our first assumption is that we'll only look at patterns in the steady state, i.e. we won't worry about how the juggler started juggling. Alternatively, we can assume that the juggler started juggling at time $t=-\infty$ and will continue to juggle till time $t=\infty$.

A juggling pattern consists of several features that are orthogonal, i.e. the features can be varied independently of each other. We'll discuss some of the important features here, one of which is siteswap. We'll start with features that are easy to understand, and then move on to more complex but more important features.

### Speed

Every pattern can be made slower by throwing the objects higher.

Assumption: all throws are made at integral units of time.

A unit of time is called a tick or a beat. The speed of a pattern is defined by the duration of a tick in seconds.

### Dwell

The amount of time the juggler holds an object in his/her hand is called dwell. Dwell is measured in ticks. Using a large dwell is called lazy style and using a small dwell is called hot-potato style.

For simplicity of presentation, assume that the dwell is the same for all throws/catches.

### Hand position

A pattern can be modified by changing the positions of hands during throws and catches. Hand positions can be specified using a coordinate system.

Reverse cascade is similar to the normal cascade, except that instead of throwing from the inside and catching on the outside, we throw from the outside and catch on the inside. (A careful observation will reveal that the reverse cascade is actually a time-reversed version of the normal cascade.)

By using more complicated hand-positions, we can get awesome variations of the cascade, that don't really look like the cascade.

As stated before, hand positioning is independent of other features of the pattern. As an example, here are some hand positions applied to the fountain pattern:

### Siteswap

If you look at the first three animations (cascade, shower, fountain), they have the same hand-positioning and can be made to have the same speed and dwell. Yet these patterns look very different. The feature that differentiates them is the order in which objects are thrown and caught. This information is captured by siteswap. This is where things start getting mathematical!

Assumption: hands throw alternately, i.e. at even ticks, the left hand throws and at odd ticks, the right hand throws. This assumption is called asynchronicity.

Assumption: a hand can only hold 1 object at a time. This restriction is called non-multiplexing. (Most non-jugglers assume that non-multiplexing is mandatory for juggling, but that's not always true. We will, however, enforce non-multiplexing in our model.)

If all the other features (ticks per second, dwell, hand position) are fixed, then a juggling pattern can be uniquely determined by specifying two things for each throw: how high should the object be thrown, and whether the object is thrown to the same hand or to the other hand. (In the fountain pattern, all throws are made to the same hand, and in cascade all throws are made to the other hand.)

To specify how high an object is thrown, we associate a number with the throw, called order (also called throw-order). The order of a throw is defined to be the flight-time of a throw in ticks, plus the dwell. So if an object is thrown at time $t$ with order $x$, that object will be thrown next at time $t+x$. This implies that throw-orders must be integers.

Actually, we don't need to specify whether an object is thrown to the same hand or not, because that can be inferred from the throw order.

Lemma 1: An object is thrown to the same hand iff the throw-order is even.

Proof. Define the parity of an integer $y$ to be 1 if $y$ is odd and 0 if $y$ is even. Since hands throw alternately, the parity of the throwing time decides whether the object is thrown from the left hand or the right hand. Suppose an object is thrown at time $t$ with order $x$. Then it is thrown to the same hand iff the parities of $t$ and $t+x$ are the same, which happens iff $x$ is even. $\Box$

Assumption: patterns are periodic. So, we only need to state the smallest repeating subsequence of the infinite sequence of throw orders. This finite-sized subsequence is called a siteswap. The length of this subsequence is called the period of the siteswap.

Looking at a pattern and figuring out its siteswap can be tricky at first, especially if you're new to juggling. I'll show some examples so that you get the hang of it.

## Siteswap Examples

Firstly, note that all throws are identical. So all throws have the same throw-order.

Suppose the blue ball is thrown at time $t$. At time $t+1$ the red ball is thrown, at time $t+2$ the green ball is thrown, and at time $t+3$ the blue ball is thrown again. Therefore, the time between consecutive throws of the blue ball is 3 ticks. Hence, all throws have order 3. This gives us the siteswap  for this pattern.

Let's now look at a pattern, which, as far as I know, doesn't have a common name. It's identified solely by its siteswap.

We'll repeat the same trick as last time. Just keep noting down the colors of the throws. Denote the colors gray, red, green, blue by characters _, r, g, b, respectively. Then the colors of the throws is the string _bg_rb_gr repeated indefinitely. Now for each throw, find the next throw of the same color, and compute the difference of their throw-times to the get the throw-order. This gives us the sequence [3, 4, 5] repeated indefinitely.

Therefore, our siteswap is [3, 4, 5]. Usually people just concatenate the numbers together and write 345 instead of [3, 4, 5]. This is not a problem, since throw orders above 9 rarely arise in practice, and even if they do, we can use a for 10, b for 11, and so on.

## When is a siteswap valid?

If we take an arbitrary sequence of integers, will it be the siteswap of some juggling pattern? As we'll see, the answer turns out to be "no". But why? Try to think about it before reading further. (Hint: Is 43 a valid siteswap? If yes, what does it look like? If no, why is it invalid?)

Turns out that to throw a ball, there must be a ball in your hand. If your throw-orders at odd ticks are odd and throw-orders at even ticks are even, balls will land in your hands only on even ticks, so at odd ticks, eventually you won't have any balls left to throw. But there's more. You also need to ensure that the number of balls landing in your hand at each tick is not more than 1. Otherwise, you'll have more balls than you can throw. Think of these two constraints as the equivalent of conservation of mass.

So how can we find out if a sequence of integers is a valid siteswap? I'll get straight to the point, like this:

Permutation Theorem: Let $a = [a_0, a_1, \ldots, a_{n-1}]$ be a sequence of $n$ non-negative integers. Let $\mathbb{Z}_n = \{0, 1, \ldots, n-1\}$. For each $i \in \mathbb{Z}_n$, let $b_i = (i + a_i) \bmod n$. Then $a$ is a valid siteswap iff $b$ is a permutation of $\mathbb{Z}_n$.

Wait, $$a$$ is a sequence of non-negative integers? What does a throw-order of 0 mean?

Good question! That's a technical detail that I conveniently brushed under the rug. I'll address this at the end of the blog post. For now, you can assume that $a$ only has positive integers and throw orders can only be positive. The proofs work for the general case that includes zero-order throws.

The permutation theorem gives us the following $O(n)$-time algorithm (python code ahead):

def is_valid(a):
n = len(a)
freq =  * n  # a list of n zeros
for i in range(n):  # i from 0 to n-1
b_i = (i + a[i]) % n
freq[b_i] += 1
for j in freq:
if j != 1:
return False
return True


We'll now try to prove the permutation theorem. To do that, we first need to get a good characterization of what a valid siteswap is. Essentially, we're trying to remove the juggling from the problem and reduce it to a pure math problem.

Let $a$ be a sequence of $n$ non-negative integers. We'll now define a function $f_a: \mathbb{Z} \mapsto \mathbb{Z},$ that takes as input a throw time and outputs the corresponding catch time. $f_a(x) = x + a[x \bmod n]$ Here $a[i] = a_i$ and $x \bmod n$ is the remainder obtained after dividing $x$ by $n$.

Now we need to prove two things:

• At any time $y$, we have a ball to throw, i.e. $\forall y \in \mathbb{Z},$ $\exists x \in \mathbb{Z},$ $f_a(x) = y$.
• We don't have collisions, i.e. there shouldn't be multiple balls landing in your hand at the same time: $\forall x_1, x_2 \in \mathbb{Z},$ $(f_a(x_1) = f_a(x_2)$ $\implies x_1 = x_2)$.

Does this look familiar? This is exactly the definition of a bijection! The first conditions says that $f_a$ should be onto, and the second condition says that $f_a$ should be one-to-one. Therefore, we get that $a$ is a valid siteswap iff $f_a$ is a bijection.

Define the predicate $P$ as:
$P(a)$: For $b_i = (i + a[i]) \bmod n,$ $b$ is a permutation of $\mathbb{Z}_n$.
Now the permutation theorem reduces to this lemma:

Lemma 2: Let $a$ be a sequence of $n$ non-negative integers. Then $f_a$ is a bijection iff $P(a)$.

Lemma 2 has no reference to juggling. It's a purely mathematical fact. Now that we're in familiar territory, you should try proving it yourself before you read my proof below.

Mathematics shares one feature with juggling: for maximum enjoyment, you have to try it yourself.

Joe Buhler and Ron Graham

### Easy part of the proof

Lemma 3: If $f_a$ is one-to-one, then $P(a)$ is true.

Proof. Assume $P(a)$ is false. Then $\exists i_1, i_2 \in \mathbb{Z}_n$ such that $i_1 \neq i_2$ and $b[i_1] = b[i_2]$.

\begin{align} & b[i_1] = b[i_2] \\ &\implies (i_1 + a[i_1]) \bmod n = (i_2 + a[i_2]) \bmod n \\ &\implies \exists k \in \mathbb{Z}, i_1 + a[i_1] = i_2 + a[i_2] + kn \\ &\implies \exists k \in \mathbb{Z}, f_a(i_1) = f_a(i_2 + kn) \end{align}

This is a contradiction, since $f_a$ is one-to-one. Therefore, $P(a)$ is true. $\Box$

Lemma 4: $P(a)$ implies that $f_a$ is one-to-one.

Proof. Assume $f_a$ is not one-to-one. Then $\exists u_1 \neq u_2$ such that $f_a(u_1) = f_a(u_2)$. Let $i_1 = u_1 \bmod n$ and $i_2 = u_2 \bmod n$.

Case 1: $i_1 = i_2$.
$f_a(u_1) = f_a(u_2)$ $\implies u_1 + a[i_1] = u_2 + a[i_2]$ $\implies u_1 = u_2.$ This is a contradiction, since $u_1 \neq u_2$.

Case 2: $i_1 \neq i_2$. \begin{align} & f_a(u_1) = f_a(u_2) \\ &\implies u_1 + a[i_1] = u_2 + a[i_2] \\ &\implies i_1 + a[i_1] \equiv i_2 + a[i_2] \pmod{n} \\ &\implies b[i_1] = b[i_2] \end{align}

Since $b$ contains a duplicate entry, it cannot be a permutation of $\mathbb{Z}_n$. This contradicts $P(a)$. Therefore, $f_a$ is one-to-one. $\Box$

### Not-so-easy part of the proof

Lemma 5: If $f_a$ is one-to-one, then $f_a$ is onto.

Proof. There are 3 crucial insights in this proof. The first is that we can represent $f_a$ as a graph of infinite size. Formally, let $G_a$ be a graph whose vertex set is $\mathbb{Z}$. For each $x \in \mathbb{Z}$, there is an edge from $x$ to $f_a(x)$. Hence, the out-degree of each vertex is 1. $f_a$ is one-to-one means that the in-degree of each vertex is at most 1. We want to show that $f_a$ is onto, i.e. every vertex has in-degree at least 1. (Actually, there's a name for $G_a$ when $a$ is a valid siteswap: it's called a ladder diagram.)

It is easy to see that if $(u, v)$ is an edge in $G_a$, then for any $k \in \mathbb{Z}$, $(u - kn, v - kn)$ is also an edge in $G_a$. This is the second crucial idea of this proof.

The third crucial idea is to assign two labels to each number: a major label $M(x) = \lfloor x/n \rfloor$ and a minor label $m(x) = x \bmod n$. We then partition $\mathbb{Z}$ by the major label. Therefore, there are an infinite number of partitions and each partition has $n$ numbers.

For an arbitrary integer $r$, consider the set $S$ of integers having major label $r$. There are $n$ edges that come out of $S$, i.e. $n$ edges have their source vertex in $S$. We now ask: how many edges enter $S$ (i.e. how many edges have their target vertex in $S$)? We'll now show that at least $n$ edges enter $S$. Since $f_a$ is one-to-one, this would prove that the number of edges entering $S$ is equal to $n$ and each vertex in $S$ has in-degree exactly 1.

Let $S = \{u_0, u_1, \ldots, u_{n-1}\},$ where $u_i = rn + i$. Let $v_i = f_a(u_i) = u_i + a[i]$. Suppose the major label of $v_i$ is $s_i$. Then the major label of $v_i' = v_i - (s_i-r)n$ is $r$, so $v_i' \in S$. Let $u_i' = u_i - (s_i-r)n$. Then the minor label of $u_i'$ is $i$.

Since $(u_i, v_i)$ is an edge of $G_a$, $(u_i', v_i')$ is also an edge in $G_a$ (as per the second crucial idea). This edge enters $S$. Also, the integers $u_0', u_1', \ldots, u_{n-1}'$ are all distinct, since the minor label of $u_i'$ is $i$. Therefore, at least $n$ edges enter $S$. $\Box$

### Average theorem

Average Theorem: Let $a$ be a siteswap of period $n$. Then the number of objects being juggled is $(\sum_{i=1}^n a_i)/n$.

The proof of the average theorem is quite involved, so I'm not going to put it here. It uses the same ideas as in the proof of the permutation theorem. It also gives a quick sanity check of a siteswap's validity: the average should be an integer.

For the competitive programmers reading this who are looking for a challenge, you may want to solve the problem ICL1703 on Codechef, which is a generalization of the average theorem when multiplexing is allowed. I have abstracted out juggling terminology from this problem, so you don't need to know anything about juggling to understand it. I have also written an editorial for it.

### Special throw orders

We assumed that we will throw a ball on every tick. We can relax this assumption, without violating conservation of mass, by allowing instances when no ball falls into your hand and so you don't throw anything. This situation is called a zero-order throw. Here are some examples:

Sometimes you may want to keep holding on to a ball in your hand instead of throwing it. This can be achieved via order-2 throws (think why it makes sense).

### Extensions

We assumed that in each tick, only one hand throws and hands throw alternately. This assumption is called asynchronicity. This is the most restrictive assumption that we have made. There are many nice patterns that don't fit in this model. Fortunately, all of the theory we saw here can be ported to the synchronous model, and synchronous patterns have their own, somewhat different, siteswap notation.

Further generalizations include multiplexing, i.e. allowing multiple balls in a single hand, and passing, i.e. multiple people juggling together. Siteswap notation and the associated theory extends to these too.

### Uses and limitations of mathematical analysis

We can invent new juggling patterns by brute-force enumeration of integer sequences up to a certain length and then using the permutation theorem to filter out invalid sequences. Each siteswap output by this process corresponds to a valid juggling pattern. This is great, and rumors say that 441 was invented like this.

However, this approach has limitations. A lot of patterns look very similar to each other. Out of the many patterns output by this algorithm, only some may be aesthetically pleasing, that too after carefully choosing an appropriate hand positioning. So in practice, highly skilled humans outperform machines in inventing interesting patterns.

But that doesn't mean this theory is useless! There are other uses of juggling theory, like animation programs, transitioning across patterns, generating patterns under constraints, etc. Also, an understanding of juggling theory can help jugglers invent interesting patterns more easily. This is how I invented 615150 and 6051, and these patterns helped me learn to juggle 3 balls in 1 hand.

You can try inventing your own patterns, and then easily see what they look like using Juggling Lab or its GIF server.

## Want to learn how to juggle?

Juggling is a very enjoyable hobby. If you feel motivated, see my other blog post, Juggling 101, on how to learn juggling.

I started learning juggling a few months before I joined BITS Pilani as an undergraduate student. By the end of my third semester, I had learned around 30 different patterns.

1. Joe Buhler and Ron Graham. Juggling patterns, passing, and posets.
In Mathematical Adventures for Students and Amateurs (2004): 99–116.