Trigonometric Series

Default

In this article we create a concept that has no important math meaning, only for the convenience of narration.

F-integrable: Riemann-integrable or in defect integral square-integrable.

Brief

Trigonometric series can be seen as a linear combination of basic trigonometric system $1,\sin x,\cos x,\sin 2x,\cdots$, which is orthogonal.

There is a distinction between trigonometric series and Fourier series, according to my professor.

When we say Fourier, we assume there exists a function $f$ matching this series.

Cesàro sum a Fourier series may turn it to a non-Fourier trigonometric series.

Definition: Trigonometric Series of a Function

In the Real Field

Definition:
$$\begin{gathered} f\sim \frac{a_0}{2}+\sum _{k=1}^{+\infty }{a}_k\cos (kx)+b_k\sin (kx). \\ {a}_k:=\frac{1}{\pi }{\int }_{-\pi }^{\pi }f(x)\cos (kx)\text{d}x,b_k=\frac{1}{\pi }{\int }_{-\pi }^{\pi }f(x)\sin (kx)\text{d}x. \end{gathered}$$
Why we define $a_k,b_k$ as such?

Suppose $f=\frac{a_0}{2}+{\sum }_{k=1}^{+\infty }{a}_k\cos (kx)+b_k\sin (kx)$ and the right part is uniformly convergent.

$⟹f\in C[-\pi ,\pi ].⟹f\in R[-\pi ,\pi ].$

Multiply $\cos kx$ at both sides and integrate.

(That is a conclusion about orthogonal basis!)

Then we have:
$${\int }_{-\pi }^{\pi }f(x)\cos (kx)=\pi a_k.$$

Then you know why we add $\frac{a_0}{2}$ here.

Furthermore, it fits into other formulas.

$$\begin{gathered} \sigma :R[-\pi ,\pi ]\to T, \\ f\mapsto \sum a_n\cos (nx)+b_n\sin (nx). \end{gathered}$$
$\sigma {|}_{C[-\pi ,\pi ]}$ is an injective.$⟺(a_n=b_n=0.⟹f\equiv 0.).$

But $\sigma$ is not an injective.

A Fourier-series converges to $f$.

$f^{\ast }$ and $f$ is different at a zero measurement set.

then F-series won’t converges to $f^{\ast }$.

And if you guarantee $f^{\ast }$'s integrability, $\sigma (f^{\ast })=$this F-series.

All $\sin ,\cos$ here are $2\pi -$periodic, and according to the definition, this expansion has nothing to do with $f$ out of $[-\pi ,\pi ]$.

Thus we require $f$ to be a $2\pi -$periodic function. or pick $f_1:=f{|}_{[-\pi ,\pi ]}$, also you can choose $[t,t+2\pi ]$, as long as $f(t)=f(t+2\pi )$.

If $f$ is on $[-T,T]$:
$${\int }_{-T}^{T}f(x)\cos (k\frac{\pi }{T}x)=T{a}_k.$$

In the Complex Field

$$\begin{gathered} f\sim \sum _{n=1}^{+\infty }{c}_n{e}^{inx}. \\ {c}_n:=\frac{1}{2\pi }{\int }_{-\pi }^{\pi }f(x)e^{-inx}\text{d}x. \end{gathered}$$
For the sake of convenience, we can rephrase real one as in complex:
$$c_n:=\{\begin{array}{l}{a}_n/2+b_n/(2i),n\in {\mathbb{Z}}^{+}.\\ a_0/2,n=0.\\ (a-n)/2-(b-n)/2i,n\in {\mathbb{Z}}^{-}.\end{array}$$
$\hat{f}(n):=c_n$.(Actually this is the notation for Fourier transformation and $c_n$ is its value at $n$. But here we just take it as a notation now.)

Figure out Fourier Series

Extension:

odd extension

even extension

Black one is the original function. Red one is gained by odd extension. Green one is gained by even extension.

Even extension helps to construct a cosine series, while odd one helps to construct sine series.
If it is equal to these two, then also their linear combinations.

In these picture we know
$$\frac{\pi }{4}-\frac{2}{\pi }\sum _{n=1}^{+\infty }\frac{\cos ((2n-1)x)}{(2n-1{)}^2}=\frac{x}{2}=\sum _{n=1}^{+\infty }\frac{(-1{)}^{n+1}}{n}\sin (nx).$$

And why in same term number, cosine series performs way better?

An interesting problem.

Convergence of Fourier Series

Point-Wise Convergence

Periodic function averging (actually you have to prove this under Lebesgue-integral system) and F-integrable guarantee that Fourier coefficients $\to 0$.

---

$$\phi (t):=f(x_0+t)+f(x_0-t)-2S.$$
Dini theorem:

If $f$ is an F-integrable function.

$\exists \delta >0s.t.{\int }_0^{\delta }\frac{|\varphi (t)|}{t}\text{d}t\exists .$

Then the fourier series converges to S.

---

Holder continuous

If f is $\alpha$-Holder continuous, then its Fourier series converges to itself.

---

Dirichlet:

If $f$ has different monotonicity on $(x_0-\delta ,x_0),(x_0,x_0+\delta )$, then the Fourier series at $x_0$ converges to $\frac{1}{2}(f(x_0^{-})+f(x_0^{+}))$.

That says piece-wisely monotonic function’s trigonometric series converges to $\frac{f(x_0^{-})+f(x_0^{+})}{2}$.

Actually this also stands for piece-wisely differentiable function.

Mean-Square Convergence

Actually it implies probability convergence and it’s a special case of $L^p$ convergence. That is $E((f_n-f{)}^p)\to 0$.

In this special case:
$$\underset{n\to +\infty }{\lim }{\int }_a^b(f_n(x)-f(x){)}^2\text{d}x=0.$$
Then we pick any trigonometric series and consider ${\Delta }_n^2:=\int (f(x)-T_n(x){)}^2\text{d}x$.

After some violent calculation, we get following properties.

Trigonometric Best Approximation Theorem

Content:

Suppositions:

1, $f$ is an F-integrable function.

2, Note the sequence of its trigonometric series { S n } \{S_n\} {Sn​}.

Conclusions:

1, $\forall \text{ trigonometric polynomial }{T}_n,\mathrm{deg}{T}_n=n,E((f-S_n{)}^2)⩽E((f-T_n{)}^2)$, which is only equal when $T_n=S_n$.

---

And as $n$ goes bigger, the error only descends.

Through these two we reach this proposition:
$$E((S_n-f{)}^2)\to 0.$$
Proof:

Use continuous function to approximate $f$, then use trigonometric series to approximate continuous function.

---

勾股-Theorem in F-Integrable Function Space

Content:

If $f$ is F-integrable, then
$$\frac{1}{\pi }{\int }_{-\pi }^{\pi }{f}^2(x)\text{d}x=\frac{a_0^2}{2}+\sum _{n=1}^{+\infty }(a_n^2+b_n^2).$$

---

Proof:

Consider Trigonometric Best Approximation Theorem above.

---

Essence:

Usually it’s called Parsevel equality, but actually it’s an extrapolation of 勾股-theorem in more complex space.

Say, the left part of the equality is the length of a vector in $L^2$ space with an orthogonal basis ($1,\sin x,\cos x,\cdots$), and the right part is its component in each dimension.

---

Application:

That tells us some Fourier series are not the match of any square integrable function

$\sum \frac{\cos nx}{\sqrt{n}},\sum \frac{\sin nx}{\ln n}$.

Actually $\frac{\cos nx}{n^{\alpha }},\alpha \in (0,1)$.

---

Extrapolation:

Consider the 勾股-theorem of $f+g$ and $f-g$; then by subtraction we get
$$\frac{1}{\pi }\int fg=\frac{a_0{\alpha }_0}{2}+\sum _{n=1}^{+\infty }(a_n{\alpha }_n+b_n{\beta }_n).$$

---

Inequality: Bessel(you just consider part, not all, of the components.)

Uniform Convergence

we often use 勾股-Theorem in this topic.

If $f$ differentiable and $f^{\prime }$ integrable then its trigonometric series uniformly converges to $f$.

---

Proof:

Consider the trigonometric series of $f^{\prime }$.

By definition we know that $a_n^{\prime }=n{b}_n,b_n^{\prime }=-n{a}_n$.

That is the self-same form you may get from differentiation.

Then we prove this series is absolutely, uniformly convergent because it’s controled by $\sum |a_n|+|b_n|$.
$$\begin{gathered} \sum |a_n|+|b_n| \\ =\sum \frac{|a_n^{\prime }|+|b_n^{\prime }|}{n}⩽{\left[\sum _{n=1}^N(a_n^{\prime 2}+b_n^{\prime 2})\right]}^{\frac{1}{2}}{\left(\sum _{n=1}^N\frac{2}{n^2}\right)}^{\frac{1}{2}}\le \sqrt{\frac{{\pi }^2}{3}}{\left[\frac{1}{\pi }{\int }_{-\pi }^{\pi }(f^{\prime }(x){)}^2\text{d}x\right]}^{\frac{1}{2}}. \end{gathered}$$

Then we consider where it converges to.

Consider that piece-wise differentiable conditions. Here we have stronger condition.

Term-By-Term Differentiation

Conditions:

1, $f^{\prime \prime }\exists ,\in R.$ That is a very strong condition!

Just a higher order form of uniform convergence.

Term-By-Term Integration

Fourier series integrable just need $f$ integrable and that’s what Fourier series need to exist!

---

Proof:

It’s a constructive one.

Consider $g(t)=\{\begin{array}{l}\pi ,t\in (0,x).\\ 0,t\in (x,2\pi ).\\ \frac{\pi }{2},t=0,t=x.\end{array}$.

And 勾股-Theorem for $\int fg$.

Uniform Approximation of Continuous Function by Trigonometric Polynomials

Also called Weierstrass second approximation.

---

Content:

$\forall f\in C[-\pi ,\pi ],f(\pi )=f(-\pi ),\exists$ trigonometric series uniformly converges to $f$.

This approaching series is trigonometric but may not Fourier.

---

Proof:

Consider the Cesàro sum of Fourier series. (Fejér kernel)

Conclusions

1,

$\forall x\in [0,2\pi ),\forall N\in {\mathbb{Z}}^{+},|{\sum }_{n=1}^N\frac{\sin (nx)}{n}|⩽3\sqrt{\pi }.$

---

2,

{ a n } \{a_n\} {an​} is decreasing.

$\sum a_n\sin (nx)$ uniformly converges on $\mathbb{R}$.$⟺\lim n{a}_n=0.$

Left$⟹$Right:

Let $x=\frac{\pi }{2m},m>2n,|{\sum }_{i=n}^m{a}_i\sin (ix)|<\epsilon$, that is Cauchy criterion. Then $ix$ has a range, minify them to create $m{a}_n$.

Right$⟹$Left:

Respectively consider $x\in [0,\pi /m],[\pi /m,\pi /n],[\pi /n,\pi ]$.

1, Violent amplification.

3, Abel transformation.

2, Lo and Behold! $l:=[\pi /x],S_{n,m}=S_{n,l}+S_{l+1,m}$, then this is deduced to 1 and 3.

Here we use a scaling to make something in the mid be the sum of two ends.

---

3,

$f$ is integrable. $⟹$ It has trigonometric series.

There exists L-integrable function whose trigonometric series diverges at everywhere. (1927 Kolmogorov)

But even $f$ is continuous, its trigonometric series can have countably infinite divergent points.

Its trigonometric series can converge at its discontinuities.

The trigonometric series of an L-square-integrable function converges almost everywhere.

In mathematics, the Gibbs phenomenon, discovered by Henry Wilbraham (1848) and rediscovered by J. Willard Gibbs (1899), is the oscillatory behavior of the Fourier series of a piecewise continuously differentiable periodic function around a jump discontinuity. The function’s $N^{th}$ partial Fourier series (formed by summing its $N$ lowest constituent sinusoids) produces large peaks around the jump which overshoot and undershoot the function’s actual values. This approximation error approaches a limit of about 9% of the jump as more sinusoids are used, though the infinite Fourier series sum does eventually converge almost everywhere except the point of discontinuity.
From Wikipedia

Remaining Problems

Trigonometric (Fourier) Expansion

Series converge?

End-points continuity?