Creative Commons License


  1. Home

  2. 1. Epicycles

  3. 2. From Circles to Squares

  4. 3. Ticking of the Clock

  5. 4. Mixing it up

So an online friend of mine saw a tweet about a secret society of scholars who plan to create a universe where everything is square.

The difference between circles and squares is something I've been thinking about of late because a world in which circles are squares is a world without Pythagoras' Theorem, and that's been on my mind a bit recently.

More precisely, if we take either the 1 or norm on 2 then the "unit circle" is actually a square.

But that's not replacing circles with squares. A circle or a square can be defined without regard to the ambient norm. So to actually replace circles with squares, we need something with some circles in it where replacing them by squares makes an actual difference.

1 Epicycles

Enter Epicycles. Every now and again I see a post on epicycles. Every now and again I get annoyed by a post on epicycles.

I'm not a historian of mathematics, but my reading of the story of epicycles goes something like this. Circles were viewed as a form of perfection, and the heavens (by which I really mean everything outside the atmosphere) were perfect, so therefore the heavenly bodies must move on circles. They clearly don't. So they must move on some complicated path that involves circles moving along circles. These are epicycles.

Along came Brahe, Copernicus, Kepler, and ultimately Newton who pointed out that actually everything moved on ellipses and so epicycles faded back into the aether. Until Fourier, that is, who showed that everything (for some definition of "everything") was actually made up of circles providing you used an infinite number of them. In other words, along came Fourier series.

Cue lots of posts on the internet showing how to take a shape and draw it by lots of cute circles rotating about each other at different speeds. These tend to irritate me a little, mainly because they gloss over some detail or other. For a while I had in mind writing about those details, but I couldn't think of an "in" other than that all the other posts I'd read irritated me and that didn't feel like a good basis for a post.

The theme of replacing circles with squares, however, does feel like a good basis for a post.

I do need to explain one thing about what irritates me about the way epicycles are presented. There seems to be some "wow" factor about drawing a picture with a lot of circles. To me, it is more impressive to draw a picture with as few circles as possible.

I did write a little program with the idea of fixing the number of circles and playing around with the various pictures one could get by varying the coefficients in the resulting Fourier series. But that was as far as I got until the idea of squares and circles reared its head.

2 From Circles to Squares

The key idea of replacing circles by squares is very simple: take a finite Fourier series and replace the circle functions (either sin(2kπt) and cos(2kπt) or e2kπit depending on whether you are enlightened or not) by their equivalent square functions.

Let's unpack that a bit. We can think of the function tcos(2kπt) as the function which tracks the x–coordinate of a point moving around the unit circle at such a speed as to go around k times in 1 unit of time. Similarly tsin(2kπt) does the same for the y–coordinate. We want to replace these by functions that track the x and y–coordinates of a point that moves around the unit square.

Just as for the circle, we start from (1,0) and travel anticlockwise. The graphs of the coordinate functions look like Figure 1.

Figure 1: The Square Functions

Their definitions look a little tricksey, but aren't all that complicated.


Implementing them in code is straightforward, so I could easily add them to my little program.

My proposal, therefore, for "replacing circles with squares" is to take a finite Fourier series and replace the circle functions with their square equivalents.

It would be nice to do this for a general Fourier series, but there's a slight hitch. The square functions do not form an orthonormal basis for the appropriate space of functions. So there's no guarantee that the result of swapping the circle functions for square ones in a general Fourier series will result in a well-defined square integrable function. It may be that there is some obvious way to change the norm to make the square functions orthogonal, but I'll leave that question open for now.

3 Ticking of the Clock

The replacement idea is still not enough. I need a ready source of curves to apply this to. I quite like the idea of "a curve a day", so I started thinking about how to change a date into coefficients.

To give enough scope for variation, I considered series of the form:


with each cj. I skipped c0 as that would only shift the resulting shape.

So I wanted a way to convert a date into a set of values for the cj. The first step is to discretise the possibilities, so rather than cj we consider cj+i, namely complex numbers with integer real and imaginary parts (so-called Gaussian integers).

There are a few transformations that essentially preserve the resulting shape. Multiplying every coefficient by the same complex number scales and rotates the shape, while swapping t-t reflects the shape in the x–axis. Also, replacing t by kt means that we go round the shape k times.

To ensure that we don't get essentially repeated shapes, we want to allow only two possibilities for c1, namely 1 or 1+i. If we insist that c-1 have real and imaginary components one of -1, 0, or 1 then we get 9 choices for c-1 for each choice of c1, but some of these should be viewed as equivalent. We regard two as equivalent if they are related by a rotation, reflection, or interchanging the pair (or any combination of the two). This leaves the following choices for c-1:

This gives 10 choices in all.

That accounts for c1 and c-1, leaving us with four complex parameters to play with. Ideally, I'd like to say that there is a fixed set of choices and each is drawn from that set. Given a set of size k, we therefore have a total parameter space of 10k4. With 365 days, this gives a set of size a little under 3, which is not very big and therefore doesn't give us much freedom in generating curves. I wanted something a little more to play with.

Calendars and clocks go hand in hand, so let's try making it a time and date concept. There are 3600 seconds in an hour and 24 hours in a day. So the number of seconds in a year is roughly


Now it just so happens that 192=361 which is very close to both 360 and 365. In fact, 194=130321 and 360×365=131400, a discrepancy of a little over a thousand which is almost nothing in this context. So if we fix on a set of 19 Gaussian integers and draw c±2 and c±3 from this set, we can use one lot to account for days and the other to account for 10s of seconds (so the "tick" for our clock will be every 10 seconds). It gets a bit confusing to refer to "10s of seconds" with all the numbers flying around, so I'm going to use the term "decasecond".

This leaves us with just the hours. We have a factor of 10 from the choices of c1 and c-1. If we could just leverage that up to 12 then we'd have the hours up to a.m. and p.m., with then a factor of 2 to finish with. We have 6 options when c1=1, but only 4 when c1=1+i. So if we add in two more for when c1=1+i we'll be home and dry. Adding in 2i and -2 provides a pleasing symmetry, and while the choice of, say, c1=1+i and c-1=2i is related to the choice of c1=1 and c-1=1+i by multiplication by 1+i, if we pick our 19 points for the other coefficients carefully we won't get any duplicated pictures.

Let's choose those 19 points. I have to admit, 19 is a strange number to be choosing as it affords very little symmetry. Figure 2 is what I ended up with. (Though if you count the points, you'll find that there are twenty – so for each of the four coefficients I remove one of the four innermost points.)

Figure 2: The nineteen points

The final step is the factor of 2 from the morning and afternoon split. But here's where the circles and squares comes to the rescue: one is circles, the other squares.

4 Mixing it up

The final step was to mix up the choices. The way I have laid it out above, the obvious course is to take a date and time and write it as a triple using days (of the year), hours (of the day), and decaseconds (of the hour).

At time of writing, we are at the 105th day of the year, 17th hour of the day, and approximately 315th decasecond of the hour. Then we take the 105 and the 315 and write them in base 19 as 510 and 1611, respectively. Finally, we write the 17 in base 12 as 15. We therefore have our numbers as:


So we can use the 5 and 10 to choose c2 and c-2, the 16 and 11 to choose the c3 and c-3, the 5 tells us which pair to pick for c1 and c-1, and the 1 tells us whether to use squares or circles.

The problem with this is that it is all rather too predictable. Only one number is guaranteed to change every decasecond and all morning pictures will use circles.

So let's mix it up.

I'll show what I do with the numbers given above: 105th day of the year, 17th hour of the day, and 315th decasecond of the hour.

The first thing is to use the parity (by which I mean "odd/even"ness) of the decasecond number to set the circle/square choice. Rather than swapping every decasecond, what I decided to do was to swap every odd decasecond, so the sequence goes "square", "circle", "circle", "square". This means that every decasecond we either swap between using squares and using circles, or we change our coefficients. I quite like the effect of this as it means that the change in type and change of coefficients happen at different steps.

Once we've used the parity of the decaseconds to determine whether we're using squares or circles, we need to remove it from consideration. This means that we round 315 down to the nearest even number, in this case 314.

Our decaseconds count now only has half the information it had before, so we augment it back up to full strength using the "morning/afternoon"ness of the hour. In this case, 17 is afternoon so we add 1 to the decaseconds and replace 17 by its equivalent on a 12 hour clock (with a slight modification that we use 0 instead of 12). So we're back to 315 decaseconds but now have 5 hours.

We'll use the 5 hours to set c1 and c-1. We start by using its "odd/even"ness to set c1: 5 is odd so we take c1=1+i. Then we divide 5 by 2 and throw away any remainder to get 2. We use this number to choose one of the 6 choices for c-1. The formula I used was to use its "odd/even"ness to choose the imaginary component (in this case 0), then again dividing by 2 and throwing away the remainder gives us one of 0, 1, or 2, in this case 1. Subtracting 1 maps this to -1, 0, or 1 which gives us the real part, in this case 0. So our resulting number is actually 0. Finally, we multiply that by c1 to get the correct choice for c-1.

Now we look at our two numbers for days of the year and decaseconds of the hour. We want to mix these together in a fashion that means that we still get a unique set of choices for the remaining ck (ignoring, that is, the minor issue of the discrepancy between 360×365 and 3612 – we basically ignore that).

Our numbers of days and decaseconds will be changing by one each time they change (apart from on the wrap around). So we make it so that a change by one in these numbers means a bigger change in the numbers we actually use. To do that we multiply them by some fixed number, and then take remainders on dividing by 361. To ensure that when we do that we don't lose any information, we need to pick our fixed number carefully – it needs to be invertible modulo 361, meaning that it must not have any factors in common with 361. We'd also like it to be quite big, somewhere in the middle of 361, to have the biggest disruptive effect. I picked 127.

So we multiply our numbers by 127 and then take remainders modulo 361. This mixes up the numbers without losing information. It doesn't, though, change the fact that one of the numbers only changes each day. My solution to that is to combine the number of decaseconds with the number of days in such a fashion that if we know the combination and the number of decaseconds then we could recover the number of days (this means that we don't lose any information). By combining this with multiplying by 127, we can mix up these numbers quite considerably without changing the amount of information contained in them.

The procedure I went for was the following. Start by multiplying the decaseconds by 127, in our example we get 40005. We'll take its remainder after dividing by 361 to get one of our numbers, which will be 295. This throws away the quotient, which is 110, so we'll add that to our other number, to get 215 in our example. Now multiply that by 127 and take its remainder on dividing by 361, in our case we get 230.

So from our original numbers of 315 and 105 we have generated two new numbers 295 and 230. Let's see what happens to those numbers as the seconds tick by.


And finally, we convert each of those numbers to a pair of numbers less than 19 to pick one of our chosen points as the coefficient.

The last step in designing the clock is to figure out what to do with the 10 seconds between each change. As we're drawing a shape, I use 8 of those seconds to trace round it and then the final 2 seconds to morph from one shape to the next.

The result is the animation at the top of the page. A more intricate version (showing the construction details) can be found on my code pages with source on github.