Mr Wu, author of the Singapore Maths Tuition blog, gave me a good nomination for this week’s topic: the j-function of number theory. Unfortunately I concluded I didn’t understand the function well enough to write about it. So I went to a topic of my own choosing instead.
The Jacobi Polynomials discussed here are named for Carl Gustav Jacob Jacobi. Jacobi lived in Prussia in the first half of the 19th century. Though his career was short, it was influential. I’ve already discussed the Jacobian, which describes how changes of variables change volume. He has a host of other things named for him, most of them in matrices or mathematical physics. He was also a pioneer in those elliptic curves you hear so much about these days.
Jacobi Polynomials are a family of functions. Polynomials, it happens; this is a happy case where the name makes sense. “Family” is the name mathematicians give to a bunch of functions that have some similarity. This often means there’s a parameter, and each possible value of the parameter describes a different function in the family. For example, we talk about the family of sine functions, . For every integer n we have the function where z is a real number between -π and π.
We like a family because every function in it gives us some nice property. Often, the functions play nice together, too. This is often something like mutual orthogonality. This means two different representatives of the family are orthogonal to one another. “Orthogonal” means “perpendicular”. We can talk about functions being perpendicular to one another through a neat mechanism. It comes from vectors. It’s easy to use vectors to represent how to get from one point in space to another. From vectors we define a dot product, a way of multiplying them together. A dot product has to meet a couple rules that are pretty easy to do. And if you don’t do anything weird? Then the dot product between two vectors is the cosine of the angle made by the end of the first vector, the origin, and the end of the second vector.
Functions, it turns out, meet all the rules for a vector space. (There are not many rules to make a vector space.) And we can define something that works like a dot product for two functions. Take the integral, over the whole domain, of the first function times the second. This meets all the rules for a dot product. (There are not many rules to make a dot product.) Did you notice me palm that card? When I did not say “the dot product is take the integral …”? That card will come back. That’s for later. For now: we have a vector space, we have a dot product, we can take arc-cosines, so why not define the angle between functions?
Mostly we don’t because we don’t care. Where we do care? We do like functions that are at right angles to one another. As with most things mathematicians do, it’s because it makes life easier. We’ll often want to describe properties of a function we don’t yet know. We can describe the function we don’t yet know as the sum of coefficients — some fixed real number — times basis functions that we do know. And then our problem of finding the function changes to one of finding the coefficients. If we picked a set of basis functions that are all orthogonal to one another, the finding of these coefficients gets easier. Analytically and numerically: we can often turn each coefficient into its own separate problem. Let a different computer, or at least computer process, work on each coefficient and get the full answer much faster.
The Jacobi Polynomials have three coefficients. I see them most often labelled α, β, and n. Likely you imagine this means it’s a huge family. It is huger than that. A zoologist would call this a superfamily, at least. Probably an order, possibly a class.
It turns out different relationships of these coefficients give you families of functions. Many of these families are noteworthy enough to have their own names. For example, if α and β are both zero, then the Jacobi functions are a family also known as the Legendre Polynomials. This is a great set of orthogonal polynomials. And the roots of the Legendre Polynomials give you information needed for Gaussian quadrature. Gaussian quadrature is a neat trick for numerically integrating a function. Take a weighted sum of the function you’re integrating evaluated at a set of points. This can get a very good — maybe even perfect — numerical estimate of the integral. The points to use, and the weights to use, come from a Legendre polynomial.
If α and β are both then the Jacobi Polynomials are the Chebyshev Polynomials of the first kind. (There’s also a second kind.) These are handy in approximation theory, describing ways to better interpolate a polynomial from a set of data. They also have a neat, peculiar relationship to the multiple-cosine formulas. Like, . And the second Chebyshev polynomial is . Imagine sliding between x and and you see the relationship. and . And so on.
Chebyshev Polynomials have some superpowers. One that’s most amazing is accelerating convergence. Often a numerical process, such as finding the solution of an equation, is an iterative process. You can’t find the answer all at once. You instead find an approximation and do something that improves it. Each time you do the process, you get a little closer to the true answer. This can be fine. But, if the problem you’re working on allows it, you can use the first couple iterations of the solution to figure out where this is going. The result is that you can get very good answers using the same amount of computer time you needed to just get decent answers. The trade, of course, is that you need to understand Chebyshev Polynomials and accelerated convergence. We always have to make trades like that.
Back to the Jacobi Polynomials family. If α and β are the same number, then the Jacobi functions are a family called the Gegenbauer Polynomials. These are great in mathematical physics, in potential theory. You can turn the gravitational or electrical potential function — that one-over-the-distance-squared force — into a sum of better-behaved functions. And they also describe zonal spherical harmonics. These let you represent functions on the surface of a sphere as the sum of coefficients times basis functions. They work in much the way the terms of a Fourier series do.
If β is zero and there’s a particular relationship between α and n that I don’t want to get into? The Jacobi Polynomials become the Zernike Polynomials, which I never heard of before this paragraph either. I read they are the tools you need to understand optics, and particularly how lenses will alter the light passing through.
Since the Jacobi Polynomials have a greater variety of form than even poison ivy has, you’ll forgive me not trying to list them. Or even listing a representative sample. You might also ask how they’re related at all.
Well, they all solve the same differential equation, for one. Not literally a single differential equation. A family of differential equations, where α and β and n turn up in the coefficients. The formula using these coefficients is the same in all these differential equations. That’s a good reason to see a relationship. Or we can write the Jacobi Polynomials as a series, a function made up of the sum of terms. The coefficients for each of the terms depends on α and β and n, always in the same way. I’ll give you that formula. You won’t like it and won’t ever use it. The Jacobi Polynomial for a particular α, β, and n is the polynomial
Its domain, by the way, is the real numbers from -1 to 1. We need something for the domain. It turns out there’s nothing you can do on the real numbers that you can’t fit into the domain from -1 to 1 anyway. (If you have to do something on, say, the interval from 10 to 54? Do a change of variable, scaling things down and moving them, and use -1 to 1. Then undo that change when you’re done.) The range is the real numbers, as you’d expect.
(You maybe noticed I used ‘z’ for the independent variable there, rather than ‘x’. Usually using ‘z’ means we expect this to be a complex number. But ‘z’ here is definitely a real number. This is because we can also get to the Jacobi Polynomials through the hypergeometric series, a function I don’t want to get into. But for the hypergeometric series we are open to the variable being a complex number. So many references carry that ‘z’ back into Jacobi Polynomials.)
Another thing which links these many functions is recurrence. If you know the Jacobi Polynomial for one set of parameters — and you do; — you can find others. You do this in a way rather like how you find new terms in the Fibonacci series by adding together terms you already know. These formulas can be long. Still, if you know and for the same α and β? Then you can calculate with nothing more than pen, paper, and determination. If it helps,
and this is true for any α and β. You’ll never do anything with that. This is fine.
There is another way that all these many polynomials are related. It goes back to their being orthogonal. We measured orthogonality by a dot product. Back when I palmed that card I told you was the integral of the two functions multiplied together. This is indeed a dot product. We can define others. We make those others by taking a weighted integral of the product of these two functions. That is, integrate the two functions times a third, a weight function. Of course there’s reasons to do this; they amount to deciding that some parts of the domain are more important than others. The weight function can be anything that meets a few rules. If you want to get the Jacobi Polynomials out of them, you start with the function and the weight function
As I say, though, you’ll never use that. If you’re eager and ready to leap into this work you can use this to build a couple Legendre Polynomials. Or Chebyshev Polynomials. For the full Jacobi Polynomials, though? Use, like, the command JacobiP[n, a, b, z] in Mathematica, or jacobiP(n, a, b, z) in Matlab. Other people have programmed this for you. Enjoy their labor.
In my work I have not used the full set of Jacobi Polynomials much. There’s more of them than I need. I do rely on the Legendre Polynomials, and the Chebyshev Polynomials. Other mathematicians use other slices regularly. It is stunning to sometimes look and realize that these many functions, different as they look, are reflections of one another, though. Mathematicians like to generalize, and find one case that covers as many things as possible. It’s rare that we are this successful.
I thank you for reading this. All of this year’s A-to-Z essays should be available at this link. The essays from every A-to-Z sequence going back to 2015 should be at this link. And I’m already looking ahead to the M, N, and O essays that I’ll be writing the day before publication instead of the week before like I want! I appreciate any nominations you have, even ones I can’t cover fairly.