GENERALIZED FUNCTION

Generalized Function

Maths

This work will extend the classic function to the generalized one typified by Dirac function in order to deal with the convergence issue of Fourier transform.

Dilemma

One of the most stunning motivations to carry over to the generalized function is to confront the convergence of Fourier transform. There are two issues to be considered,

---- Convergence of integral defining Fourier transform

---- Fourier inversion

For instance, take the Fourier transform of a constant function f(t) = 1. There’s no way to make sense of

Likewise, for trigonometric functions, it cannot make sense of

Again, we can easily get the inverse Fourier transform of sinc(t) via duality. However, the following integral is intractable,

From the examples above, we can see the whole theory is collapsing, even for some common signals. Therefore, we need a more robust definition of Fourier transform to deal with common signals, ones for which the classic definition will not do.

There are actually two ways of dealing with this problem.

(1) Advanced special techniques to solve a particular problem, not a general way.

(2) Rework the foundations and definitions.

The work will focus on the latter way, the more general one. Analogous to Fourier series, a new conception of convergence is required. The idea is how to choose the basic phenomena which can be used to explain the others. So, back away from the specific problems, we are looking for the best situation.

Therefore, to identify the “best class” of functions, denoted by Schwartz, for Fourier transform, we demand for the following nice properties (*),

(1) Suppose f(t) is in Schwartz, then the Fourier transform of f is defined and also in Schwartz, i.e.

(2) Fourier inversion works, i.e.

A further property along with the above two, known as Parseval’s Identity, holds as,

which indicates that the total energy stays the same both in time domain and frequency domain. (See the proof in Appendix I)

How to Define Schwartz Class

What kind of class is the best class for Fourier transform? This problem was solved by Laurent Schwartz, an extraordinary French mathematician, who claims that Schwartz Class is a class of rapidly decreasing functions f(x) defined as,

(1) f(x) is infinitely differentiable.

(2) For any m, n >= 0,

Note that m and n are independent, (2) indicates that, any derivative of f(x) tends to zero faster than any power of x.

One might doubt whether there exist any such functions, due to these seemingly terribly restricted conditions. The fact is, however, any Gaussian is in Schwartz Class. (Check it yourself)

Why is the Schwartz Class the best for Fourier transform? Because it satisfies both of the properties we demand for above.

Apparently,

So, how can we be sure that we have not lost anything and will gain greater generality? It is incumbent to pick up the main subject of this work, the idea of “generalized functions”, also known as “distribution”, typified by Dirac function.

Dirac Function

Formally, Dirac function is defined as,

Dirac function is supposed to represent a function which is concentrated at a point. There are various ways of doing this, but it is always via a limiting process. (See Appendix II)

Definition of Generalized Function

A generalized function or distribution is defined as,

Given a test function phi (in Schwartz Class), a distribution T(.), associated with a test function, is linear functional on phi, i.e.,

Besides, T(.) satisfies the property of continuity, i.e.,

where phi_n is a sequence of test functions.

To capture this idea and include more, we need to change a point of view, that is, a distribution induced by a classic function is focus on the outcome rather than on the process. In this sense, a distribution is always paired with a test function, denoted by T(phi), or in a more common way,

Therefore, given a test function, it is necessary merely to know how a distribution operates on it.

Now, recover Dirac function in context of this definition. Operationally, the effect of Dirac function is to evaluate a function of the origin. Therefore, Dirac function formalized by a distribution can be alternatively defined in the following way.

It can be proved that Dirac function defined in this way satisfies all the properties of a generalized function or a distribution.

Considering the Dirac function shifted by a, we obtain a distribution as,

Likewise, constant functions, triangle functions and trigonometric functions come back in in the similar way, that is, they are defined as generalized functions or distributions paired with a test function in Schwartz Class. And the pairing is always by integration, just as Dirac function does.

Therefore, for “any” function, f(x), we can consider it as a generalized function or a distribution by defining the pairing,

Fourier Transform of a Distribution

A distribution (generalized function) is continuous linear functional on a test function. Particularly, when taking test functions to be ones in Schwartz Class, i.e., rapidly decreasing functions, we will obtain the nice properties (*) for Fourier transform. Correspondingly, this class of distributions is called tempered distributions. If T is a tempered distribution, we want to define its Fourier transform which is another tempered distribution defined by,

Like Dirac function conducts as a typical distribution, we might as well assume the pairing above is also by integration, i.e.,

Now, we can turn this into a definition, given a tempered distribution T, we define its Fourier transform by

Note that

always makes sense, since

Similarly, we can define the inverse Fourier transform by

therefore,

Armed with this special technique, we are now capable of recovering the unsolved problems mentioned at the beginning of this work.

First, what is the Fourier transform of Dirac function?

Due to the arbitrariness of the rapidly decreasing function phi, we have,

Since Dirac function is “infinitely concentrated” while its Fourier transform, 1, “uniformly spreads out”, this indicates that, concentration of time domain is spreading out in frequency domain.

What is the Fourier transform of shifted Dirac function?

What is the Fourier transform of complex exponential?

Considering a special case where a = 0, we have,

which indicates again the reciprocal relationship that stretching out in time domain means concentration in frequency domain.

Ultimately, sines and cosines do not have classical Fourier transform, neither. However, they have generalized Fourier transforms, since they make sense as generalized functions or distributions.

According to Euler’s formula,

 

we obtain its Fourier transform,

 

Likewise,

 

Derivative of a Distribution

Given a distribution T, T’ is also defined as a distribution by a pairing with a test function,

We can now turn this into a definition,

Here are some examples.

What is the derivative of unit step function?

Unit step function, u(x), can be defined as a distribution since

Therefore, u’ exists as a distribution defined by a pairing,

What is the derivative of signum function?

Again,

Moreover, the derivative theorem given by

can be applied to find Fourier transform.

For instance, what is the Fourier transform of signum function?

What about the Fourier transform of unit step function?

Noting that

we have

Distribution Multiplication

Multiplication of functions does not carry over to multiplication of distributions, that is, if S and T are distributions, then ST is generally not defined. What is defined, however, is fT when f is a function. Similarly, fT is defined in pairing,

only if fphi is a test function.

There is a special case where a function f multiplies Dirac function, which is also the sampling operation on f in mathematics, defined by

Therefore,

More generally,

which is called the sampling property of Dirac function.

Distribution Convolution

The convolution between distributions, S * T, is not always defined, either. However, we can define it by pairing as well but in extra conditions. On the other hand, f * T most often makes sense when f is a function. Note that the convolution here is not quite defined as integral as that of classic functions, which the work will not cover. But the convolution theorem somehow still holds as

There is also a special case that is worthy of noticing, where

where Dirac function serves as the identity for convolution.

More generally,

We can even convolve Dirac function with itself, i.e.,

Moreover, the scaling property of Dirac function can be defined by pairing

The similar argument goes for a < 0. In conclusion,

Appendix

I. Parseval’s Identity

Proof:

In general, suppose f(.) and g(.) are in Schwartz Class, we are to prove

Note that

All above is perfectly justified because Schwartz Class is the perfect class for all Fourier analysis.

When f(.) = g(.), Parseval’s Identity is properly proved.

II.The limiting process to Dirac function

Considering a family of rectangle functions,

Regardless of the value of epsilon, the area of the rectangle stays the same as 1.

As epsilon tends to zero, the height of the rectangle becomes infinitely large, i.e.,

Checking the definition of Dirac function,

moreover,

since phi(x) is infinitely differentiable.