Integral Transforms - Lecture 8 The Fourier Inversion Theorem. Oxford Maths 2nd Yr Student Lecture

The Opinion Lighthouse
32 followers
2 Video Views·Nov 7, 2022
Integral Transforms - Lecture 8 The Fourier Inversion Theorem. Oxford Maths 2nd Yr Student Lecture
There you are. Welcome back for probably the highlight of the entire course,
this, the Fourier inversion theorem. We're going to work up to this gently.
We're going to look at some examples which motivate an absolutely key result
that tells us once we have the Fourier transform of a function,
how do we recover the function itself? So this is the analogue, of course,
of the Laplace inversion theorem, which tells us how to get a function back from its Laplace
transform. And it's going to be no surprise that the result is in the form of an integral.
But whereas the Laplace integral was a slightly unexpected looking contour integral,
we should see that the Fourier inversion looks very much like the Fourier formula itself.
And we're going to tie it all back into the theory of Fourier series,
which you studied earlier on in your course. Start with two examples, though. We've seen
them before. If you take the transform of e to the minus a times the modulus of x,
a double sided exponential distribution with a graph that looks sort of like that,
goes up to a point at the origin, then we know that the transform is 2 a over s squared plus a
squared. Now, I'm going to rewrite that slightly as two pi times a over pi times s squared plus
a squared, so that two is there. I put a pi on the top and the bottom.
And then the point is that this thing integrates to one. Now, what happened as I let a go to zero?
Well, we've seen this example a couple of times before. This thing tends to delta of s,
because when s is nonzero, when we let a tend to zero, we get zero. When s is zero,
we get a over a squared pi, which tends to infinity. And the integral is one.
So what happens to the left hand side, the function we took the transform of,
e to the minus a mod x? Well, that tends to one, because when a tends to zero, a times mod x
tends to one for all x. So the left hand side tends to one, the right hand side tends to delta of s.
Another example with the same conclusion. e to the minus a squared x squared has the Fourier
transform square root of pi over a, e to the minus s squared over four a squared.
Worked that one out last time. And again, I rewrite that so that it's two pi times something
whose integral is one. And this, of course, is the PDF of a normal variable with mean zero
and variance two a squared. We get four a squared there, that's twice the variance,
two sigma squared. So again, the right hand, the transform tends to two pi delta of s,
and the thing we were transforming tends to one because this is
what you get for a normal random variable when the density tends to infinity.
So, not density, the variance tends to infinity. So two examples where a function tends to one
and its transform tends to two pi delta of s.
So these very strongly suggest what you can only say is an amazing result,
so amazing that I worked out how to put it in a fancy box like this.
The Fourier transform of one is two pi delta of s. That's a formula you couldn't have conceived of
getting your head round before we started this course.
The Fourier transform of one is two pi delta of s.
And then there's a very quick corollary, which is that the Fourier transform of e to the i a x
is two pi delta of s minus a, which is just the standard formula for the Fourier
transform of e to the i a x f of x is f hat of s minus a.
So these formulas, you really do have to think about these formulas quite hard.
What they are not saying is that a function like e to the i a x is in any way integral
in the normal sense. It's not saying that one is integrable in the normal sense.
Because if you were to try to interpret this as an integral, you would say, well,
you multiply one by e to the minus s x and then you integrate in x.
Well, that would be asking you to integrate e to the minus i s x,
which of course you can't do. It's not an integrable function.
Nonetheless, we've obtained this as a sensible limiting process of functions for which the
Fourier transform was perfectly well defined, and for which all of the integrals involved
did exist in the real, the ordinary, classical sense.
So we don't have the proper framework to do this, but I hope you can believe
that we can take the framework of distributions that we introduced in the beginning of this course
and you can extend it a little bit to allow you to take the Fourier transforms of distributions.
The real technical obstacle, which we're not going to go into,
is that you need the Fourier transform of a test function to be a test function itself
for all the theory to work. And that's not true for the test functions that we
broke down earlier. So you have to have a slightly different kind of test function.
And then you just have to go through the same processes again.
We're not going to do that. That's too much for this course.
But we do believe that the Fourier transform of one is two pi delta of s
on the basis of limiting arguments like these two here.
Here is a couple of asides, a couple of comments on that result before we go and apply it
to prove the Fourier inversion theorem. The first one is that another example that
we've already seen, we took the Fourier transform of the interval,
the indicator function of the interval minus a a to get two sine of a s over s.
And if you let a tend to infinity, you should be able to get the result that one hat is two
pi delta of s again by letting a tend to infinity in this indicator function.
So it just gets wider and wider.
And then you have to do the appropriate thing on the Fourier transform.
But you will find that it works. Now, this is a little bit more,
this second talking point is a little bit more out there.
Remember, when we did Laplace transforms, we took the Laplace transform of delta of x minus a,
and we didn't worry about that too much. We just said we multiply by e to the minus p x
and integrate. And because e to the minus p x is continuous, we get e to the minus p a.
Evaluate e to the minus p x at x equals a.
Now, if we believe the Laplace inversion theorem, we ought to have that delta of x minus a,
the thing we started with, is one over two pi i, the integral of sigma minus infinity to sigma
plus i infinity of e to the minus p a, the transform e to the p x d p.
And if I just collect together the exponential, this is the integral of e to the p times x minus a
d p. Now, if we were to try to do that by contour integration, which, of course,
is the obvious thing to do, if x is less than a, then x minus a is negative, real and negative.
So that means that e to the p times x minus a decays as p goes to plus infinity.
So that means you can take your semicircle, your inversion contour, which goes vertically up like
that, you can close it with a semicircle in the right hand half plane. You can show that the
integral around the semicircle goes to zero. And so by Cauchy's theorem, because this is an entire
function with no singularities in the finite complex plane, you know that the integral is zero.
Which, of course, you expect because delta of x minus a is zero if x is less than a.
In the footnote, I've just put down a little note here about the technical manoeuvres you need
to make sure that the integral around the semicircular arc goes to zero.
The point is you can't quite use Jordan's lemma as it stands in your notes because
you need the maximum of the modulus of the thing you're integrating, in this case,
e to the p times x minus a. That has to go to zero as fast as one over r as the radius goes
to infinity. And that's not true because the maximum of this is equal to one at the end,
at the top and bottom of the semicircle. But you can do a thing a little bit like we did
when we proved that a Laplace transform vanishes as infinity,
and you can combine it with the Riemann-Lebesgue lemma.
I'm not going to go through those details, but it's not a particularly difficult proof.
You might like to have a go at it yourselves.
So what we've shown is that if x is less than a,
you close the contour with a semicircle on the right, you integrate and you get zero.
If x is greater than a, you do the same thing on the left, because then this is positive.
So as p goes to minus infinity, then you'll get exponential decay.
And so again, if x is greater than a, you will also get zero.
And if x is equal to a, then remember, you're integrating e to the p times x minus a.
So if x is equal to a, this is just one. You put p is sigma plus i s in the integral.
Remember, you're integrating from sigma minus i infinity to sigma plus i infinity.
So you put p is sigma plus i s and you have, well, it's not an integral.
It's a thing you might write like this, one over two pi times the integral of one d s.
Well, this is exactly what you saw in the formula one hat over two pi is equal to delta,
or one hat is two pi delta.
So although this isn't a proof, this is another piece of evidence that says
this all hangs together as a mathematical framework.
It's not a proof, not least because delta of x minus a is nowhere near satisfying the
conditions you need on the function for the Laplace inversion theorem to hold.
So again, we're taking it on trust that all of this works.
We're going to come back to the Fourier Laplace transform relationship a little bit later on,
but we'll move on now and we'll state the famous Fourier inversion theorem.
And here it is. Theorem 84. f is continuous and integrable,
then f of x can be recovered from its Fourier transform.
So here's the Fourier transform, f hat of s is the integral of f e to the minus i s x d x.
f of x is one over two pi, there's a difference, one over two pi,
the integral of the transform times e to the plus i x s d s.
So there's the difference between the function,
it's given with a plus sign in the exponent and the transform has a minus.
And then there's one over two pi here.
Some people would put a square root of two pi on both sides to get a nice symmetrical formula,
but we're using the standard one over two pi here.
So how do we prove this? You are going to have to look through this proof a little carefully,
because there's one step in it that really you cannot sit and watch me do it.
You've got to sit and do it yourselves and think with a piece of paper.
But I'll come to that in a moment.
We start with the answer. We start with one over two pi,
the integral of f hat e to the i x s with one over two pi.
Then we substitute for f hat of s here in terms of its definition with f.
So there is f, which is what we're after, buried inside this double integral here.
We have the Fourier transform integral with e to the minus i s y.
And this y here is, of course, just a dummy variable of integration.
And then e to the i x s d s.
And that, of course, is another dummy variable of integration,
the s, and we can call those anything we like.
We can call them u or v or psi or Fred or whatever we want to call them.
So what we now do is we just rearrange this a little bit.
We take the f of y out here and we take the d y out the back.
So d s d y, not d y d s.
So, of course, we're playing a bit fast and loose here.
You can't just change the order of integration in the double integral,
but we're going to do it anyway.
And then in the middle, we have the integral from minus infinity to infinity d s.
That's the wad, the outer integral here.
e to the i x s, that one there, e to the minus i y s.
And that's this one here with s y turned into y s.
Now, this is where you have to think.
And listening to me explain it is not going to be as good as you going away
and thinking about this very hard.
But this is, in effect, the Fourier transform using s as the variable.
So when you take a Fourier transform like this one up here, you integrate over a variable.
Well, that variable is called s here.
And then when you take a Fourier transform like this one up here,
you have a variable s inside, which becomes the argument of the Fourier transform.
Well, that variable here is going to be called y.
And then x is just a parameter in this integral.
So as I said, you have to sit and think about this.
But the point is, this is just the Fourier transform of e to the i x s,
where x is just a constant and s is the variable you're taking the transform over.
And so the answer is two pi delta of y minus x.
This thing here is two pi delta of y minus x.
And that is using this result that the Fourier transform of e to the i a x is two pi delta of s
minus a. You're just permuting the names of the variables, two pi delta of y minus x.
But now this is easy, because when you multiply f of y by delta of y minus x and integrate,
you just get f of x, because you pick up the value of f where y minus x is equal to zero,
shifting property. So you get f of x.
Wow. I mean, wow. That is just an amazing proof, if you ask me.
It is, of course, not a totally rigorous proof, but the idea of it is right,
and the intuition of it is right.
And it shows you the power of this distributional framework using delta functions.
It shows you you can do extraordinary things.
And if you don't believe me, go and look in any book
for a classical proof of the Fourier inversion theorem,
and you will find exactly the same argument here, except that instead of this step here,
you'll have a limiting process which amounts to approximating the delta function.
So the classical proof will be by effectively smoothing out this delta function into something
that looks a bit like a delta function, but is integrable, and then taking the limit.
But the truth of it is in this calculation here.
So I've made a few, a couple of notes on this.
I've already commented we needed to change the order of integration,
and I've commented about the relabelling of the variables in the key step,
and I've written out for you how this works.
As I say, you need to go work with this yourselves.
I've noted that there are lots of classical proofs.
And you might be interested that if the function is discontinuous at a point,
is piecewise continuous, but has a jump in discontinuity at a point,
that comes along, goes up, say, and then goes along,
the Fourier transform gives you the average of the left and right limits,
which is exactly what you would have got for Fourier series.
So that's a reassuring sign that the Fourier transform is similar to the Fourier series.
So here's an example. Invert e to the minus a mod s.
Well, in fact, you could already do this because you can go and look up the answer
by taking the Fourier transform of e to the minus a mod x and substituting for s with minus x.
But we can do it equally well by direct integration.
We need one over two pi, the integral from minus infinity to infinity,
e to the minus a mod s times e to the i x s d s.
Split the range of integration in two for when s is positive, you have e to the a s,
sorry, when s is negative, you have e to the a s.
When s is positive, you have e to the minus a s from zero to infinity and from minus infinity to zero.
And then you add them up and you get one over two pi times two a over x squared plus a squared,
which is a over pi x squared plus a squared, as we saw before.
Now, I mentioned earlier on that the Fourier transform
is really in the same spirit as the Fourier series.
So I want to spend a few minutes just talking about that.
This is for your general education and background, as it were,
rather than being directly relevant to exams.
Please don't switch off for that reason, though.
And because this is really one of the raison d'etre of the Fourier transform,
is that it's the idea of expanding a function in terms of an orthogonal,
in fact, an orthonormal basis.
So this is such a fundamental idea in linear mathematics,
and you've seen it many times already.
For example, in linear algebra, in a finite dimensional linear algebra,
and you might think of linear algebra in R n, for example, straightforward real linear algebra,
you take a basis of vectors, for example, the standard coordinate vectors in R n,
and they're orthogonal in the sense that the inner product, the dot product of e i
and e j is delta i j. This is the Kronecker delta.
And so this, by the way, is the matrix, this is the identity matrix, delta i j.
The i j entry of the identity matrix is delta i j.
So you've seen, then, you know how to expand any vector in terms of e i and what you do.
The coefficient of e i is just the inner product of your vector and that coordinate vector.
You've seen how to do this for a Fourier series.
So I'm writing the Fourier series in complex form rather than in real form.
And you use the complex exponentials e to the i n x as your basis functions.
And you know that those are orthogonal because the integral from minus pi to pi
of e to the i m x e to the minus i n x is two pi, two pi is the length of the interval,
delta m n, same Kronecker delta.
Notice I put a minus here. This is the Hermitian structure that anyone doing
quantum mechanics will have seen. And it's there to ensure that when m is equal to n,
we have something real. In this case, it will be one.
And now what you're doing is you're taking this hierarchy, if you like,
and you're extending it to an infinite setting, which is uncountable,
because this, of course, is countable.
The number of functions of this type is countably infinite.
Now we're going to expand in a function not on a finite interval, but on an infinite interval.
And that requires an uncountable basis set, the set of functions e to the i s x,
where s is any real number. And your orthogonality is no longer
an integral like with e to the i m x, e to the minus i n x, e to the i x s, e to the minus i y s.
So you see, s has become x and m has become x and n has become y.
And this integral is two pi delta of y minus x.
So there's a very clear correspondence between these two.
And it is, in fact, possible to get to the Fourier transform
by letting the length of this interval tend to infinity.
Now, obviously, there's a lot of theory needed to do this.
You need to know all sorts of questions, even for Fourier series.
The convergence of Fourier series is not straightforward.
You need to know what functions can be represented in this form.
And you need to know a bit more about what does this
really mean in the sense of, properly in the sense of distributions.
But this is the framework. This is what it's all about.
It's about expanding in an orthogonal basis.
And you might also ask, where does that orthogonal basis come from?
Well, it comes mostly from, or it comes in one sense, from differential equations.
And the trig functions are eigenfunctions of the operator d two by dx squared
on the interval minus pi to pi.
These functions are eigenfunctions of the operator d two by dx squared
on the interval minus infinity to infinity.
But that's going a bit too far for this course.
But you ought to know that there is that motivation.
Going back, we can compare the Fourier transform and the Fourier series.
And here they are, one above the other.
If you have a Fourier series, you can write f of x is the sum of
complex coefficients c n times e to the i n x,
and c n are given by one over two pi, one over the length of the interval,
multiplied by the integral of minus pi to pi f of x e to the minus i n x dx.
So this is the standard way of working out Fourier series.
When you first did it, of course, you did it in terms of sines and cosines,
but they are entirely equivalent.
And this is obviously a much more compact and elegant way of writing it,
although it's maybe a bit less practical
when it comes to working out the details in a given problem.
So there you are. The function is its Fourier series,
a sum, and the coefficients are given by integrating the function.
Now look at your Fourier transform pair.
The transform is the integral of a function multiplied by the complex conjugate of
the e to the—the complex conjugate of the e to the i n x is now going to be
the complex conjugate of the e to the i x s, e to the minus i s x.
So that's the Fourier transform.
And instead of a sum, we have an integral because we don't have
a countably infinite set of n, we have an uncountably infinite set of s's.
But you can see the parallelity is exact.
And again, we do need theory to establish all of this properly,
but this is what is going on.
And I can't resist then saying what happens if you put
a periodic function into the Fourier transform.
Well, what happens? You want to take the Fourier transform of a sum of c n e to the i n x.
Well, you know what the transform of e to the i n x is, it's delta of s minus n.
And so the Fourier transform is the sum of c n times delta of s minus n.
And any engineer would tell you that what you've got there
is the decomposition of a signal, f of x as a signal,
decomposed into its components, frequencies, which are multiples of the integers,
or are concentrated on the integers, and those are the amplitudes of those frequencies.
So that's Fourier series and Fourier transforms joined up.
As I say, that was an aside and it's not a formal part of the course,
but I do think it's a really important and interesting
way in which this course is joined up to what you've done before.
We're going to end, though, by proving the Laplace inversion theorem,
because if you remember, we did a terrible job of that before.
There is a proof in the lecture notes for a very particular case,
and I didn't go into any detail about that in the lectures because it's technical
and it doesn't even do a proper job of doing the Laplace inversion.
But with the help of the Fourier inversion, we can do this in a very few lines.
And this is a beautiful proof with which to end today's lecture.
So we're going to have a continuous function f of x,
and it's got a Laplace transform f bar of p.
Remember the Laplace transform, you multiply f of x by e to the minus p x
and integrate from zero to infinity.
That gives you f bar of p, and it exists for the real part of p greater than p zero.
If it exists at all, it will exist on a half plane like this, so we assume it does.
Then for positive x, which, of course, is where f of x is defined,
f of x is one over two pi i, the integral sigma minus i infinity to sigma plus i infinity,
f bar of p e to the x p dp.
And sigma is an arbitrary constant greater than p zero.
So we integrate in the p plane, we integrate up a vertical contour
to the right of all the singularities of f bar, and that gives f of x.
A rather amazing and slightly mysterious looking result,
because you wouldn't really have expected that.
Whereas the Fourier inversion theorem, you might have expected,
given what we know about Fourier series.
So here's the proof. We're going to put in this integral, we're going to put p equal to
sigma plus i s. And we're going to start by working out, slightly
unexpectedly, by working out what is the Laplace transform evaluated at sigma plus i s.
And sigma plus i s, as I say, is going to be a point on this contour.
So first of all, we'll work out what f bar of p is when p is sigma plus i s.
Well, that's the integral zero to infinity f of x e to the minus
sigma plus i s times x dx, definition of the Laplace transform.
Now, I just rearrange that very slightly. I take the e to the minus sigma x with the f,
leave the e to the minus i s x alone.
And then I say, hey, that e to the minus i s x is exactly what I have
in a Fourier transform. This is an integral from zero to infinity,
but I can make it an integral from minus infinity to infinity by just multiplying f of x
e to the minus sigma x by the Heaviside function.
So this is the Fourier transform, the hat there, the Fourier transform of f of x
e to the minus sigma x times the Heaviside of x.
So what we've shown is that f bar, anywhere on that inversion contour,
f bar can be written as a Fourier transform.
So that means I can invert. If I take any point where x is positive,
f of x e to the minus sigma x is the Fourier inverse of e to the i x s
times this f bar of sigma plus i s, because this is a Fourier transform of that,
and therefore that is the Fourier inverse of this.
One over two pi out front, and f bar of sigma plus i s is the Fourier transform.
Then what do I do? The e to the minus sigma x here, I just take this over to the other side,
and I multiply by i twice. I put a two pi i there because I know I've got to have that in the answer.
And I put an i there with the i d s.
And then I realise that sigma plus i s there and sigma plus i s there are both equal to p,
because that's where I started. So that tells me that f of x is one over two pi i,
the integral sigma minus i infinity to sigma plus i infinity.
Notice I changed the limits of integration, f bar of p e to the x p dp.
And isn't that a beautiful way of proving the Laplace inversion theorem? Absolutely amazing.
Right. Well, that's the end for today. We've got one more lecture.
So I will see you then. Goodbye.
