# Generalised Fourier Series, Part 3: Comparing Series Expansions

In the second part of this series on series expansions (introduction), we found a way to make a series expansion using any basis of orthogonal polynomials. This is how the standard Fourier series is defined, and then we also used it to make a series expansion out of the Legendre polynomials. Now we’re going to compare how these behave with different numbers of terms.

This is finally the fun part of the seminar! My second-year physics students should generally have already known the previous two articles from first-year courses, so they started here, pretty much. To interact with these series expansions, you should load up my Jupyter notebook on either Colab or myBinder, and try plotting your own functions and seeing how the expansions behave.

In reality we don’t use the infinite form of series expansions, we only use the first few terms to get a good approximation of a function. The notebook is used for investigating how different series expansions work with comparably few terms, and which would be most appropriate in a given situation. The series there are the standard Fourier series, the Fourier–Legendre series, and the Taylor series about x=0x=0 (which is formed by a different method to the other two). If you need a reminder of how these series look like, the Taylor and Fourier series forms were given in the introduction, and the Legendre series was derived in part two.

I asked the students to consider a couple of the following questions, and discuss them in groups of around five:

• Which series would you call the “most accurate”? Why? Does it depend on the function?
• What sorts of functions are the different expansions best at approximating? Which are they bad at?
• The Legendre series often seems to “give up” in the middle of some shapes at low orders (e.g. sinusoids). Why is this, and are there any things the series is still useful for?
• The Taylor series almost invariably has the largest pointwise error. Why is this? What is the Taylor series useful for?

Let’s look at some examples to illustrate what I hoped the students would discover, and then afterwards I’ll explain why all of these effects appear.

## Approximating a polynomial

How do the series behave if we attempt to approximate a polynomial function? I’m using a twelfth-degree polynomial, so

f(x)=c0+c1x+c2x2++c12x12, f(x) = c_0 + c_1 x + c_2 x^2 + \dotsb + c_{12} x^{12},

for some arbitrary coefficients that I picked to give a nice shape. Below I’ve plotted this function in the range x[1,1]x \in [-1,1] (in dashed black), and the Taylor, Legendre and Fourier series approximations with 13 terms each.

Since it’s a polynomial, given enough terms the Taylor series and Legendre series become exact. The Fourier series is not based on polynomials, though, so that remains an approximation even with this many terms.

What’s more interesting is how the series behave when there are only very few terms. This second graph is the exact same approximations, but with only the first five terms of each expansion.

Here we are starting to see what the series really care about. The Taylor series is a very poor approximation far from its central point of x=0x=0, but it is the only expansions which even seems to get any of the behaviour correct. The Fourier and Legendre series are approximately right on average across the whole function, but at any given point, they don’t really look like it at all.

## Approximating a Lorentzian function

Let’s move away from “nice” polynomials. This next function is a Lorentzian if you’re a physicist, or a scaled Cauchy distribution if you’re into statistics. As a probability distribution this has the form

f(x)=1πΓ[1+(xx0Γ)2], f(x) = \frac1{\pi\varGamma\Bigl[1 + {\bigl(\frac{x-x_0}{\varGamma}\bigr)}^2\Bigr]},

for some width parameter Γ\varGamma.

As a probability distribution, this can actually be a really tricky function to work with—we call it “pathological” because while it’s clearly symmetrical and has an obvious mean, the proper mathematical definition

x=xf(x) dx \langle x\rangle = \int_{-\infty}^\infty x f(x)\,\mathrm dx

doesn’t exist! The integral doesn’t converge, nor do higher-order statistical moments like the variance. It’s still a function, though, and we can approximate it with our expansions.

At low orders, we see the same behaviour of the Taylor series; it is the only approximation that even seems to make any effort to be close near the top of the peak, but then the polynomial approximation completely gives up. The Legendre and Fourier expansions here look fairly similar.

Adding many terms here doesn’t really change the Taylor series much, but the Fourier and Legendre series become significantly better. These approximations notionally have 13 terms in them, just like for the polynomial above, but really we can see from the function and the basis vectors that this won’t truly be the case; the function is clearly perfectly even, so no odd functions in the bases will contribute. In effect, this means that the Lorentzian is only being approximated by three and seven terms, though even if you use the attached notebook to add many more terms yourself, you’ll struggle to get the Taylor series to look good.

Functions that are asymptotically flat are generally a problem for polynomial expansions; an nnth-degree polynomial naturally goes as xnx^n for large nn. When only considering a small domain, as we are here, expansions can give quite a lot of insight.

## Approximating a logistic function

Finally, before we look at why these particular series expansions have these traits, let’s approximate the machine-learning enthusiast’s favourite type of the function: the logistic map. This particular one is

f(x)=11+e5x, f(x) = \frac1{1 + e^{-5x}},

and you can see its characteristic sigmoid (“s”) shape. Again, this tends to become flat, so we expect the Taylor series to struggle to find any convergence at the edges.

On the face of it, the logistic function is neither even nor odd, since it doesn’t pass through zero. However, given a single constant offset of one half, which is the lowest-order term in all of these expansions, it is actually odd. Similarly to the Lorentzian, then, only half of the subsequent terms will actually contribute to the approximations.

The most striking feature in this high-order approximation plot, perhaps, is the divergence of the Fourier series from the Legendre one at the edges, and the wiggles the Fourier approximation has towards these edges. Previously these two have generally been rather similar. The Legendre series is much better at approximating the logistic map than the Lorentzian, likely due to much smaller magnitude second derivatives of the function.

## Understanding what we’ve seen

So far we haven’t attempted to explain why we have seen things, we’ve only commented on what is present. Let’s go through and explain the parts now.

The Taylor series is only good close to x=0x=0. This is straightforward; the Taylor series is defined by successively making a better approximation around a given point. The series is extended by adding information from higher and higher order derivatives, and uses no knowledge of far-away points.

The Fourier and Legendre series are good on average. This is again because of how the series are defined. In the previous part, we saw that we found the series coefficients by minimising r,r\langle r,r\rangle, where r(x)=f(x)Af(x)r(x) = f(x) - A_f(x) was a “residual” or “error” function describing the difference between the true function ff and the approximation AfA_f. Since these inner products are integral, this is a minimisation of the root-mean-square error over the domain. In simpler words, these series expansions are designed to be good when averaged over every point in the interval we are considering, and the integral forms of their coefficients ensure that the expansions make use of data from everywhere relevant to achieve this. This property makes these series, and particularly the Legendre series, excellent for producing numerical integration rules of very high orders.

The Taylor and Legendre series were perfect when given a polynomial. Since they use polynomials as their basis functions, these series will naturally become perfect once they have enough terms. It’s easy enough to see that there is a possible solution using Legendre polynomials up to and including the degree of the polynomial; choose the coefficient of highest-degree Legendre polynomial to match the highest-degree term in the function, then do the same for the next-highest-degree terms taking into account what’s already present in the expansion, and so on. This gives a perfect result with no higher-order terms, and consequently also has zero error, since it’s the same function. Now we can also know that our method must reproduce this solution we’ve shown exists, because the minimum possible root-mean-square error of any function is zero, and our coefficients by definition minimise the root-mean-square error. The Fourier series does not use polynomials, and so never becomes exact for such functions; you must sum an infinite number of cosine waves of differing frequencies to approximate an x2x^2 term.

The Fourier series deviates when the function values on the left- and right-hand sides of the domain are not equal. This is actually one of the nice properties of Fourier series, though it doesn’t look like it here. As the Fourier series is based on the periodic trigonometric functions, the Fourier approximations are always also periodic, and the large divergence at the very end is the approximation keeping itself continuous between periods. I’ve only plotted the approximations between 1-1 and 11, but if we extended it beyond that, we’d see the Fourier series repeat again and again, while the Legendre expansions would shoot off to infinity in a similar manner to the Taylor series. Of course for our purposes, this is totally fine; these approximations were only meant to work within this region, and extrapolation is usually a sin in physics.

The Fourier series has large oscillations near a jump. This “ringing” effect of the Fourier series is called the Gibbs phenomenon, and it’s a rather funny behaviour that often appears near discontinuities. As the number of terms goes to infinity, the Fourier series will converge for every point of a square-integrable function (most sensible functions—very roughly these are functions that aren’t fractals and don’t have an infinite asymptote inside the boundaries). However, the maximum point-wise absolute divergence actually increases as you add more terms; the “ears” get thinner and closer to the jump, but they also get taller and taller. Each point converges individually, but in every Fourier approximation the worst point (which moves over time) gets worse and worse. This happens because the series must remain analytic—all its derivatives exist—everywhere, but it also needs to approximate an infinite gradient. The only way to do that is by these wiggles.

## Wrapping up

This series of posts has covered the maths behind making new series expansions using a generalised version of the Fourier-series method, based on a seminar I wrote for the whole second-year physics undergraduate cohort at Imperial College London. These series have many interesting properties, depending on which basis set of orthogonal polynomials are used. If you didn’t take the time at the start to do so, please do have a play with my example Jupyter notebook on Colab or myBinder to see how these series work with different functions!