# Fourier Series

This chapter is devoted to the most popular and extremely important in various applications of Fourier series and its generalizations. Joseph Fourier made in nineteen century a remarkable discovery of function's expansion defined on a prescribed finite interval into infinite series of trigonometric functions. Later on, it was recognized that Fourier series is just a particular case of more general topic (now called the harmonic analysis) of representing functions as a linear combination (generally speaking, with infinite terms) over an orthogonal set of eigenfunctions corresponding to some Sturm--Liouville problem. Fourier made a very important step in understanding series representation of functions because previously Taylor's series played the most dominated role. To some extend, the Fourier series (written in complex form) is a particular case of the Lauraunt series (which is a generalization of Taylor's series). However, Taylor's series are determined by infinitesimal behavior of a function at the center of expansion because its coefficients are defined through derivatives evaluated at one point. In contrast, Fourier coefficients depend on the behavior of a function on the whole interval.

Fourier series are strongly related to the Fourier integral transformations; however, due to space constraint, we present only one section in this topic and refer the reader to next chapter for some of its applications. Since Fourier coefficients are determined by integrals over a finite interval of interest, pointwise convergence is not very suitable to them. Therefore, we go in two directions. First we introduce the mean square convergence (also called in mathematics as 𝔏² or 𝔏2 convergence). Since pointwise convergence of Fourier series for piecewise functions leads to Gibbs phenomena, we suggest to use another type of convergence named after Italian mathematician Ernesto Cesàro. Its application to Fourier series is based on Fejér's theorem, named for Hungarian mathematician Lipót Fejér, who proved it in 1899.

Introduction to Linear Algebra with Mathematica

In the 19th century, Joseph Fourier wrote: “The study of Nature is the most productive source of mathematical discoveries. By offering a specific objective, it provides the advantage of excluding vague problems and unwieldy calculations. It is also a means to form the Mathematical Analysis, and isolate the most important aspects to know and to conserve. These fundamental elements are those which appear in all natural effects.”

A Fourier series is a particular case of a more general orthogonal expansion with respect to eigenfunctions of differential operators. In particular, its eigenfunctions $$\sin \frac{n\pi x}{\ell}$$ and $$\cos \frac{n\pi x}{\ell} ,$$ where $$n=0,1,2,\ldots ,$$ are solutions of the following Sturm--Liouville problem:

$y'' + \lambda \,y =0 , \qquad y(\theta ) = y(\theta + 2\ell ) .$

A Fourier series is a way to represent a function as the sum of simple sine waves. More formally, a Fourier series is a way to decompose a periodic function or periodic signal with a finite period $$2\ell$$ into an infinite sum of its projections onto an orthonormal basis that consists of trigonometric polynomials. Therefore, a Fourier series provides a periodic extension of a function initially defined on a finite interval of length $$2\ell .$$ They are named in honor of Jean-Baptiste Joseph Fourier (1768--1830), who used trigonometric series in representing solutions of partial differential equations, after preliminary investigations by Leonhard Euler (1707--1783), Jean le Rond d'Alembert (1717--1783), and Daniel Bernoulli (1700--1782).

# Biography of Joseph Fourier

A Fourier series is a way of representing a periodic function as a (possibly infinite) sum of sine and cosine functions. It is actually a particular case of Taylor series, which represents functions as possibly infinite sums of monomial terms, when a variable belongs to the unit circle. For functions that are not periodic and defined on a finite interval, its Fourier series provides a periodic expansion from this interval.

Since the harmonic differential equation $$y'' + \lambda \,y =0$$ has periodic solutions, subject that λ is a positive number, it is convenient to represent its solution in complex form:

$y(\theta ) = A\, e^{{\bf j} 2\pi\,\theta / \ell + {\bf j} \varphi} .$
Here A the strength of the circle size, φ is the delay, or starting point, and $$2\pi /\ell$$ is the speed or the rate of change over time. Here’s an animation taken shamelessly from Better Explained describing a circular path:
The above figure shows how the entire signal on the circle is “drawn” (decomposed) into a wave. Next is an animation (from the same web page) for other trajectories generated by circular motions with a variable radius (meaning the clock arm of the circle is varied):

A cycle sequence with just one 1 in the i-th position means a pure cycle of (i+1) Hz, i.e., it completes i+1 cycles per time interval.

It’s also possible to combine! A sequence 1 2 3 means that the 0 Hz cycle has strength 1, the 1 Hz cycle has strength 2, and the 2 Hz cycle has strength 3:

One can try to see other graphs by changing the input numbers. All the measurements are now on the circle in two dimensions. We call these “vectors”. Summing the vectors together give the final “power” of the frequency. This occurs when all the vectors line up and point in the same direction, creating high values that represent the “power” of the frequency.

Fourier series have a tremendous number of applications, of which we mention speech recognition and its applications in smart-phone communications due to its relevance today. Andrew Ng has long predicted that as speech recognition goes from 95% accurate to 99% accurate, it will become a primary way that we interact with computers. Contemporary smart phones constantly monitor sounds from the microphone, and when the owner says "hey Siri" or "ok google" followed by a question, it will respond to the owner. It only recognizes the owner's voice and not someone else’s so it uses sampling and comparison (with the Fourier series) to make sure your voice matches up with the samples it obtained when you first “trained” it.

The sounds we hear come from vibrations in air pressure. Humans do not hear all sounds: for example, we cannot hear the sound a dog whistle makes, but dogs can hear that sound. Marine animals can often hear sounds in a much higher frequency range than humans. What sound vibrations we can hear depends on the frequency and intensity of the air oscillations and one's individual hearing sensitivity. Frequency is the rate of repetition of a regular event. The number of cycles of a wave per second is expressed in units of hertz (Hz), in honor of the German physicist Heinrich Rudolf Hertz (1857--1894), who was the first to conclusively prove the existence of electromagnetic waves. Intensity is the average amount of sound power (sound energy per unit time) transmitted through a unit area in a specified direction; therefore, the unit of intensity is measured in watts per square meter. It is important to note that the sound intensity that scientists measure is not the same as loudness. Loudness describes how people perceive sound. Humans can hear sounds at frequencies from about 20 Hz to 20,000 Hz (or 20 kHz), though we hear sounds best at around 3,000 to 4,000 Hz, where human speech is centered. Scientists often specify sound intensity as a ratio, in decibels---the unit for intensity is the bel (dB), named in honor of the eminent British scientist Alexander Graham Bell (1847--1922), the inventor of the telephone---which is defined as 10 times the logarithm of the ratio of the intensity of a sound wave to a reference intensity.

When we detect sounds, or noise, our body is changing the energy in sound waves into nerve impulses which the brain then interprets. Such transmission occurs in human ears where sound waves cause the eardrum to vibrate. The vibrations pass through 3 connected bones in the middle ear, which causes fluid motion in the inner ear. Moving fluid bends thousands of delicate hair-like cells which convert the vibrations into nerve impulses that are carried to the brain by the auditory nerve; however, how the brain converts these electrical signals into what we hear'' is still unknown.

Sounds consist of vibrations in the air caused by its compression and decompression. Air is a gas containing atoms or molecules that can move freely in an arbitrary direction. The average velocity of air molecules at room temperature, under normal conditions, is around 450--500 meters per second. The mean free paths of air molecules before they collide with other air molecules is about 6 × 10-8 meters. So air consists of a large number of molecules in close proximity that collide with each other on average 1010 times per second, which is perceived as air pressure. When an object vibrates, it causes waves of increased and decreased pressure in the air that travel at about 340 meters per second.

The signals we use in the real world, such as our voices, are called "analog" signals. To process these signals in computers, we need to convert the signals to "digital" form. While an analog signal is continuous in both time and amplitude, a digital signal is discrete. To convert a continuous signal in time into a discrete one, a process, called sampling. When the continuous analog signal is sampled at a frequency f, the resulting discrete signal has more frequency components than the analog signal did. Sampling frequency is the parameter that controls the sampling process. The number of bits per sample is the parameter that controls the bit resolution. The intelligibility, which is the amount of information and also the perceptual quality of speech, depends on these two parameters.

The periodic audio signals are known as (music) tones. A simple tone, or pure tone, has a sinusoidal waveform of amplitude a > 0, frequency ω0 > 0, and phase angle φ: $$x(t) = a\,\cos \left( \omega_0 t + \varphi \right) .$$ The frequency ω0 has units of radians/second = 1/seconds, with t in seconds, and phase angle φ is in radians. An alternative is to express the frequency in units of Hertz, abbreviated Hz, given by f0 = ω0/(2π). The perceived loudness of a pure tone is proportional to a0.6. The pitch of a pure tone is logarithmically related to the frequency.

Sound is recorded through regular samples. These samples are taken at a specified rate, named the sampling frequency or sampling rate f given in Hz (a hertz is the unit of frequency in the International System of Units (SI) and is defined as one cycle per second) or kHz (the kilohertz is equal to one thousand hertz or 1,000 Hz). The most common rate is 44,100 Hz = 44.1 kHz but lower rates can be used for low frequency sound (e.g. 22.05 kHz) or higher rates can be used for high frequency sound (up to 192 kHz). The speech signal has frequency components up to 8 kHz and hence 16 kHz is the optimal sampling frequency. However, when telephone communication started, with bandwidth being a precious resource, the speech signal was passed through an anti-aliasing low pass filter with cutoff frequency of 3.3 kHz and sampled at 8 kHz sampling frequency. Thus the speech signal collected over telephone networks will have a message bandwidth of about 4 kHz.

The information beyond 4 kHz is eliminated in the case of telephone bandwidth speech. Even though the sampling frequency seems to be fine for sounds like a and aa, it severely affects other sounds like s and sh. However, information up to 4 kHz bandwidth seem to be sufficient for intelligible speech. If the sampling frequency is further decreased from 8 kHz, then intelligibility of speech degrades significantly. Hence 8 kHz was chosen as the sampling frequency for telephone communication. The speech signal sampled at 8 kHz is termed as narrowband speech and the speech signal sampled at 16 kHz is termed as wideband speech.

Another important parameter of digitization is the process of quantization that consists of assigning a numerical value to each sample according to its amplitude. These numerical values are attributed according to a bit scale. A quantization of 8 bits will assign amplitude values along a scale of 28 = 256 states around 0 (zero). Most recording systems use a 216 = 65536 bit system. Quantization can be seen as a rounding process. A high bit quantization will produce values close to reality, i. e., values rounded to a high number of significant digits, whereas a low bit quantization will produce values far from reality, i.e., values rounded to a low number of significant digits. Low quantization can lead to impaired quality of the signal.

Sampling is a process of converting a signal (for example, a function of continuous time and/or space) into a numeric sequence (a function of discrete time and/or space). An analog-to-digital converter (ADC) can be modeled as two processes: sampling and quantization. Sampling converts a time-varying voltage signal into a discrete-time signal, a sequence of real numbers. Quantization replaces each real number with an approximation from a finite set of discrete values.

Fourier series applications are based on the fundamental sampling theorem that establishes a bridge between continuous-time signals (often called "analog signals") and discrete-time signals (often called "digital signals"). Strictly speaking, the theorem is only applied to a class of mathematical functions having a Fourier transform that is zero outside of a finite region of frequencies.

Note: The name Nyquist--Shannon sampling theorem honors Harry Nyquist (1889--1976) and Claude Shannon (1916--2001), known as "the father of information theory". The theorem was also discovered independently by Edmund Taylor Whittaker (1873--1956), by Vladimir Kotelnikov (1908--2005), and by others. Kotelnikov is mostly known for having discovered the sampling theorem, before all others, and he was the first to write down a precise statement of this theorem in relation to signal transmission in 1933. It is thus also known by the names Nyquist--Shannon--Kotelnikov, Whittaker--Shannon--Kotelnikov, Whittaker--Nyquist--Kotelnikov--Shannon, and cardinal theorem of interpolation.    ▣
Nyquist--Shannon--Kotelnikov Theorem: If a function f(t) contains no frequencies higher than B hertz, it is completely determined by giving its ordinates at a series of points spaced 1/(2B) seconds apart.

A sufficient sample-rate is therefore 2B samples/second, or anything larger. Equivalently, for a given sample rate fs, perfect reconstruction is deemed possible for a bandlimit B < fs/2. Therefore, the sampling theorem states that a signal can be exactly reproduced if it is sampled at a frequency greater than twice the maximum frequency in the signal.

In telephoning, the usable voice frequency band ranges from approximately 300 Hz to 3400 Hz. The bandwidth allocated for a single voice-frequency transmission channel is usually 4 kHz, including guard bands, allowing a sampling rate of 8 kHz to be used as the basis of the pulse code modulation system used for the digital PSTN (Short-time Fourier Transform). This gives a 64 kbit/s digital signal known as DS0, which is subject to compression encoding using either μ-law (mu-law) PCM (North America and Japan) or A-law PCM (Europe and most of the rest of the world). Pulse-code modulation (PSM) ) is a method used to digitally represent sampled analog signals. According to the Nyquist--Shannon--Kotelnikov sampling theorem, the sampling frequency (8 kHz) must be at least twice the highest component of the voice frequency via appropriate filtering prior to sampling at discrete times (4 kHz) for effective reconstruction of the voice signal. The voiced speech of a typical adult male will have a fundamental frequency from 85 to 180 Hz, and that of a typical adult female from 165 to 255 Hz.

Note: In the United States, the National Inventors Hall of Fame has honored Bernard M. Oliver (1916--1995) and Claude Shannon as the inventors of PCM, as described in "Communication System Employing Pulse Code Modulation", U.S. Patent 2,801,281 filed in 1946 and 1952, granted in 1956. Another patent by the same title was filed by John R. Pierce in 1945, and issued in 1948: U.S. Patent 2,437,707. The three of them published "The Philosophy of PCM" in 1948. John Robinson Pierce (1910--2002), was an American engineer and author. He worked extensively in the fields of radio communication, microwave technology, computer music, psychoacoustics, and science fiction. As a sideline to his professional career, he wrote science fiction for many years under various names. The compact disc (CD) brought PCM to consumer audio applications with its introduction in 1982. The CD uses a 44,100 Hz sampling frequency and 16-bit resolution.    ▣

Now we need to define the Fourier transform that decomposes a function of time (a signal) into the frequencies that make it up, in a way similar to how a musical chord can be expressed as the frequencies (or pitches) of its constituent notes. The Fourier transform of the function f is traditionally denoted by adding a circumflex:

$\hat{f} (\xi ) = \int_{-\infty}^{\infty} f(x)\, e^{-2\pi\,{\bf j} x\xi} \,{\text d} x ,$
where $$f\,:\, \mathbb{R} \to \mathbb{C}$$ is an integrable function, and j is a unit vector in the positive vertical direction on the complex plane. The function f is restored from its Fourier transform via the inverse transform:
$f (x ) = \int_{-\infty}^{\infty} \hat{f}(\xi )\, e^{2\pi\,{\bf j} x\xi} \,{\text d} x , \qquad x\in \mathbb{R}.$
For speech recognition, we do not need all frequencies, but only a finite portion of them. Therefore, integration over the infinite interval should be replaced with a finite one, which leads to the discrete Fourier transform:
$X_k = \sum_{n=0}^{N-1} x_n \,e^{-2\pi\,{\bf j} kn/N} = \sum_{n=0}^{N-1} x_n \left( \cos \frac{2\pi\,kn}{N} - {\bf j} \,\sin \frac{2\pi\,kn}{N} \right)$
that maps the sequence $$x_0 , x_1 , \ldots , x_{N-1}$$ into another sequence of complex numbers $$X_0 , X_1 , \ldots , X_{N-1} .$$ Here N is number of samples, n is the current sample, k is current frequency, between 0 Hz to N-1 Hz, the ratio 1/N is not necessary but it gives the actual sizes of the time spikes, 2πk is speed in radians/second, and $$e^{-{\bf j}x}$$ is the backwards-moving circular path. The above sum is actually the truncated Fourier series, which we are going to study shortly.

1. Peters, Terry M. editor, The Fourier Transform in Biomedical Engineering, Boston, MA : Birkhäuser, 1998.
2. Serov, V., Fourier Series, Fourier Transform and Their Applications to Mathematical Physics, Springer International Publishing, 2017.
3. Zaanen, A.C., Continuity, Integration and Fourier Theory, Springer Berlin Heidelberg, 1989.