Generalised Fourier Series, Part 1: Linear Algebra Basics

In the introduction to this series of articles, we introduced the idea of a series expansion, and saw the explicit forms of the Taylor and Fourier series. We are building towards a generalised series expansions based on “orthogonal” polynomials, but first we have to define some concepts from linear algebra. This article is roughly at the level of early first-year physics undergradutes.

Students at this level have come across the word “orthogonal” before when talking about Euclidean (“normal”) vectors. Here it means the same thing as “perpendicular” or “at right angles”, at least while you have three or fewer dimensions. Once you have more than that, or you’re dealing with some other type of vector, the definition is a little more abstract.

What is a vector?

What other types of vector are there? In pre-university physics, we usually describe a vector as a quantity which has both a magnitude and a direction. This is a Euclidean vector. Mathematically, though, the definition of a vector is more general than this; a vector is an element of a set of objects that have some defined operations between them (a “vector space”). There’s also a requirement that the vectors are “over a scalar field F\mathcal F”, but for our purposes we will always just be using real numbers for this.

To be a valid set of vectors V\mathcal V over the real numbers, we just need to have two defined operations: an abstract “addition” and “scalar multiplication”.

Addition can be any operation that satisfies the rules

  • x+(y+z)=(x+y)+z\bm x + (\bm y + \bm z) = (\bm x + \bm y) + \bm z,
  • x+y=y+x\bm x + \bm y = \bm y + \bm x,

for all vectors x\bm x, y\bm y and z\bm z in V\mathcal V, and the set of elements needs to contain a special vector 0\bm 0 that doesn’t do anything when added to any other vector. The collection of vectors also needs to contain an “inverse” vector for every member under addition, so that for every vector x\bm x in V\mathcal V there is another vector y\bm y that satisfies x+y=0\bm x + \bm y = \bm 0. You can see that the way you were taught to add Euclidean vectors together trivially satisfies all these rules.

The scalar multiplication operation has to satisfy some more rules in conjunction with the vector addition ones:

  • associativity: α(βx)=(αβ)x\alpha (\beta\bm x) = (\alpha\beta) \bm x,
  • multiplicative identity: 1x=x1\bm x = \bm x,
  • vector addition distributivity: α(x+y)=αx+αy\alpha (\bm x + \bm y) = \alpha\bm x + \alpha \bm y,
  • scalar addition distributivity: (α+β)x=αx+βx(\alpha + \beta)\bm x = \alpha\bm x + \beta\bm x,

for all scalars α\alpha and β\beta and vectors x\bm x and y\bm y. Again, this is familiar with Euclidean vectors.

Now, since these rules are quite abstract, there are lots of things that can also be valid vector spaces. These are separate to Euclidean vectors; you can’t add vectors from different spaces together. Here, we’re most interested in looking at functions defined in certain intervals (different intervals would be different vector spaces). It’s also interesting to note that even scalar numbers satisfy this definition of a vector space!

The notions of addition and scalar multiplication apply fairly straightforwardly to functions which return real numbers. If I have a function ff, then 2f2f is the function that does the same thing as ff, then multiplies the result by two. Similarly, the result of f+gf + g is a function which adds together the outputs of ff and gg. You can check that all the other rules are satisfied too; perhaps the least obvious one is the existence of the “zero” vector. In this case, that’s just a function that takes any input and returns the scalar zero.

The reason we do things like this mathematically is to apply one result to many different systems. There are lots of results that can be proved simply from the abstract rules I presented above. Now if you can show that a new system you’ve just come up with satisfies these rules, you get a whole lot of results for free.

Inner products and orthogonality

So far we haven’t said anything about orthogonality. That’s actually because the base definition of a mathematical vector doesn’t include it; your vector space needs to have some extra structure in the form of a new operation for it to be defined. Think about how you check if two Euclidean vectors are perpendicular—you see if the “dot product” is zero. The dot product is an example of this new operation, which in its abstract form we call an “inner product”. It takes two vectors as arguments and returns a scalar.

When we call it an inner product we denote the operation as x,y\langle \bm x, \bm y\rangle. This might be unfamiliar, since the dot product is usually denoted with (unsurprisingly) a dot.

Just like before, the inner product has a few rules that go along with it. These are rather more complicated, but that’s to be expected from an operation that provides us with so much more structure in our space. The rules when the scalars are the real numbers are

  • linearity in the first argument: αx+βy,z=αx,z+βy,z\langle \alpha\bm x + \beta\bm y, \bm z\rangle = \alpha\langle\bm x,\bm z\rangle + \beta\langle\bm y,\bm z\rangle,
  • symmetry: x,y=y,x\langle \bm x, \bm y\rangle = \langle \bm y, \bm x\rangle,
  • positive-semidefiniteness: x,x0\langle\bm x, \bm x\rangle \ge 0 with equality if and only if x=0\bm x = \bm0.

These lead to some really interesting abstract interpretations. The last point leads straight into the idea of “length”, or lets us define a distance between two vectors. Technically this is called a “norm”, though we won’t worry about all the additional rules that go along with that. You can see by comparison to Euclidean vectors that x,x\sqrt{\langle \bm x, \bm x\rangle} produces a quantity that we could call the “magnitude” of a vector; it’s always a positive number, and it’s zero only if the vector is the zero vector we defined earlier.

It also leads to the idea of measuring “how similar” two vectors are. You might have called this “the projection of one vector along another” when you learned about Euclidean vectors. You can show that

x,y2x,xy,y {\langle \bm x, \bm y\rangle}^2 \le \langle\bm x,\bm x\rangle \langle\bm y,\bm y\rangle

for all vectors x\bm x and y\bm y in any vector space that satisfy the rules we set out above, where the equality happens only if x=αy\bm x = \alpha\bm y. This last relation is the generalised form of two vectors being “parallel”. Another way of looking at this is to define “being parallel” as two vectors being equal when you divide them by their magnitudes. Scalar division by α\alpha is technically not defined yet, but we can rephrase it as scalar multiplication by 1/α1/\alpha. The two vectors are “more similar” when they are close to parallel, in the sense that the inequality is tighter.

Taking this further, just like we now have an abstract concept of “parallel”, we can also have an abstract concept of “perpendicular” or “orthogonal” by considering when the inequality is as far away as it can be; in other words, when x,y=0\langle\bm x, \bm y\rangle = 0. This relation defines “orthogonality”.

Geometrically in Euclidean vectors, the distance you have gone along one vector does not affect how far you have gone along another vector which is orthogonal to it. No matter how far up the yy-axis you go, your xx-position doesn’t change, but the closest point to the line y=xy = x (which is not at right angles to the yy-axis) does. Similarly, when we come to Fourier expansions, we will see how the amount of one function you use for your expansion does not affect how much of the orthogonal functions you use.

Let’s consider finite, continuous functions defined on the interval [a,a][-a, a]. There are several possibilities for an inner product, but the most useful for us uses integration:

f,g=aaf(x)g(x) dx. \langle f,g\rangle = \int_{-a}^a f(x) g(x)\,\mathrm dx.

You might be able to spot this type of operation in the definition of the Fourier series above!

Orthogonal bases

This last topic isn’t really new, it’s just a combination of what we’ve already defined. A “basis” (plural “bases” pronounced “bay-sees”) is a subset of the vectors in a vector space that can be used to create any other vector by addition and scalar multiplication, but you can’t make any of the basis vectors by a similar combination of the others. A basis is not unique. A familiar basis from 3D Euclidean vectors is the set {i, j, k}\{\bm i,\,\bm j,\,\bm k\}.

The property that basis vectors can’t be combinations of the other basis vectors is called “linear independence”, and it implies that when you have a vector space with a finite number of dimensions (like the Euclidean vectors), every possible basis contains the same number of elements. If we have a basis whose elements are called {x1, x2, , xn}\{\bm x_1,\,\bm x_2,\,\dotsc,\,\bm x_n\}, then any other vector y\bm y can be written as some

y=ncnxn, \bm y = \sum_n c_n \bm x_n,

for scalars cnc_n.

It is general useful to consider only bases where every element is orthogonal to every other element. The {i, j, k}\{\bm i,\,\bm j,\,\bm k\} basis above is also an example of this in its vector space. These bases are called (imaginatively) “orthogonal bases”.

These concepts of “orthogonality” and a vector-space “basis” are all we need to define a whole family of series expansions. The next article in this series will cover that, and then we’ll moved on to comparing the different series expansions we find.

This article is the first part of a series. You can find all of the rest of the articles in this series here: