Why does this distribution of polynomial roots resemble a collection of affine IFS fractals?

Question

Consider the following spectacular image, created by Sam Derbyshire and described in John Baez's article "The Beauty of Roots":

In this image are plotted all the complex roots of all polynomials of degree $\le 24$ with coefficients drawn from the set $\{-1, 1\}$. There is some amazing structure in there, about which many questions can be asked! The symmetries about the $x$-axis, the $y$-axis, and the unit circle are easily explained, though: it is not hard to show that if $z$ is a root of a polynomial with coefficients in $\{-1,1\}$, then so are $\bar z$, $-z$, and $z^{-1}$.

Please do look at John Baez's aforementioned article for more zoomed images showing the beautiful detail in this distribution of roots. In particular, towards the interior of the unit circle, the distribution does not fall off smoothly, but seems to form fractal patterns, such as near $z = 4/5$, near $z = 4i/5$, and near $z = e^{i/5}/2$. In fact, these patterns are extremely reminiscent of the images produced by iterated function systems of affine transformations. With a little experimentation using David Eck's Chaos Game applet, I found that to produce fractals similar to these patterns, one only needs to use two affine maps, both of which are the composition of a scaling by about 70% and a rotation by $\arg z$, and differ by a translation. To illustrate:

I'm tempted to say that the whole distribution looks like a catalogue of such affine IFSs, the way the Mandelbrot set is like a catalogue of Julia sets.

My question is simple: Why is this so? What is the reason that the roots of $\{-1,1\}$ polynomials form such a fractal distribution, that behaves locally like an affine IFS?

Answer 1

I recently noticed that at some point since I posted this question, the article by John Baez that inspired it has been updated. It now includes an explanation of the phenomenon in terms of compositions of the one-parameter families of functions $f_+(x) = 1 + zx$ and $f_-(x) = 1 - zx$, which are evidently the two affine maps I was speculating about in my question. The relationship between the set of roots and the compositions of $f_+$ and $f_-$ apparently is indeed analogous to the Mandelbrot-Julia correspondence. I don't entirely follow the argument in the updated article yet, but it seems clear that the answer to my question is in there, so I'm posting a link to it as this community wiki answer.

(Since this is CW, if anyone wants to edit in a concise restatement of the linked argument, go right ahead.)

Answer 2

I can't give a complete answer, but I can make a few observations in the general direction of an answer. Note that most of this is based on memory from spending several hours looking at the behavior of these plots and observing how they relate to roots for individual polynomials. This is all "empirical math" (assuming that's even a thing, heh) but the patterns are pretty clear and while I doubt I have the mathematical background to make most of it rigorous I doubt it would be too difficult.

• It's not limited to coefficients from $\{-1,1\}$; almost any collection of polynomials with coefficients chosen from a very limited set will produce similar patterns.

• The character of the pattern is directly tied to the location on the complex plane, specifically the effect of multiplication. So I suspect that it catalogs the IFSs only insofar as the complex plane catalogs (some) affine transformations. Note that the Julia set formula, in contrast, lets the transformation vary during iteration with only an additive constant to alter the shape, which is why it has the more elaborate chaotic behavior vs. the IFS-like structure here.

• Each fragment of a pattern roughly resembles a Cantor set, with fragments overlapping heavily except at the outer fringes. Each fragment is twisted according to the nature of complex multiplication, and the combination creates the similarity you noticed.

The latter two points are fairly intuitive given the nature of complex numbers and mostly apparent from inspection of the plots, but don't really do much to explain why, which I imagine is why Baez didn't remark on it.

Also unexplained are the gaps, such as those around the roots of unity. I don't recall the exact placements, but I recall it also being obvious why the locations of the gaps were relevant, but not why root density drops off so sharply around them. I could probably make some guesses, but only with copious handwaving involved.

On the other hand, except for the gaps, I believe the thick ring around the unit circle simply consists of more variations on the same patterns, denser and heavily overlapped, until detail is no longer visible.

---

Regarding variations on plots like this, the same approximate pattern appears regardless of polynomial degree, and for any set of coefficients of the form $\{-N, -(N - 1) ... -1, 1 ... N - 1, N\}$. The density of roots around the unit circle, the prominence of the gaps around certain points, and the character of the fringe inside the gaps all vary with degree; this is hard to see in Derbyshire's plot rather than one limited to polynomials of a single degree.

The shape of the overall plot changes if the coefficients are not chosen as above, being either of different magnitudes or chosen from more than one set. Note the plots on this page, particularly those near the bottom. The coloring indicates sensitivity of the points to changes in coefficients, which indicates which parts of the plot may distort asymmetrically.

My observations, in general, were:

• Large differences in magnitude between coefficients makes the patterns tighter and denser, without changing the symmetry if each individual coefficient is chosen from a set like the original plot uses.

• Large differences in magnitude between the elements in each set makes the plot as a whole more asymmetric, while making the density variations more prominent and removing the fringe patterns around the edge.

Doing both in moderation creates... odd effects: