Week 10 Discussion Notes

Table of Contents

Real Analytic Functions

Definition

Let URU \subseteq \R be an open set and f ⁣:UR\func{f}{U}{\R} be a function. We say that ff is real analytic on UU if for every aUa \in U, there exists r>0r > 0 such that

f(x)=n=0ca,n(xa)nfor xB(a,r).f\p{x} = \sum_{n=0}^\infty c_{a,n}\p{x - a}^n \quad\text{for } x \in B\p{a, r}.
Remark.

One thing important to keep in mind is that the coefficients can be different depending on the center aa. For example, if f=0f = 0 on (0,1)\p{0, 1} and f=1f = 1 on (2,3)\p{2, 3}, then ff is real analytic, but

f(x)=0+0+on (0,1)f(x)=1+0+on (2,3).\begin{aligned} f\p{x} &= 0 + 0 + \cdots \quad\text{on } \p{0, 1} \\ f\p{x} &= 1 + 0 + \cdots \quad\text{on } \p{2, 3}. \end{aligned}
Example 1.

11x=n=0xn\frac{1}{1 - x} = \sum_{n=0}^\infty x^n is real analytic on (1,1)\p{-1, 1}. (It's actually real analytic on (,1)\p{-\infty, 1}, but of course, the power series expansion will be different from the above when a<1a < -1.)

Theorem

Let ff be a real analytic function on an interval IRI \subseteq \R. Assume that the set E={x|f(x)=0}E = \set{x \st f\p{x} = 0} is non-empty and that there exists a0Ea_0 \in E and a sequence akE{a0}a_k \in E \setminus \set{a_0} such that aka0a_k \to a_0 (a0a_0 is called an accumulation point of EE). Then f=0f = 0 on II.

Proof.

Claim: f=0f = 0 in an open interval around a0a_0.

Since ff is real analytic at a0a_0, there exists r>0r > 0 such that

f(x)=n=0cn(xa0)nfor xB(a0,r).f\p{x} = \sum_{n=0}^\infty c_n\p{x - a_0}^n \quad\text{for } x \in B\p{a_0, r}.

Suppose f0f \neq 0 on B(a0,r)B\p{a_0, r}. Then at least one of the coefficients cnc_n must be non-zero or else the series expansion is just 00. Let n0n_0 be the first number such that cn00c_{n_0} \neq 0. Then

f(x)=n=n0cn(xa0)n=(xa0)n0n=n0cn(xa0)nn0g(x).f\p{x} = \sum_{n=n_0}^\infty c_n\p{x - a_0}^n = \p{x - a_0}^{n_0} \underbrace{\sum_{n=n_0}^\infty c_n\p{x - a_0}^{n-n_0}}_{\eqqcolon\,g\p{x}}.

Notice that

limncn(xa0)nn01n=xa0limncn1n\lim_{n\to\infty} \abs{c_n\p{x - a_0}^{n-n_0}}^{\frac{1}{n}} = \abs{x - a_0}\lim_{n\to\infty} \abs{c_n}^{\frac{1}{n}}

(the second limit exists since it's the reciprocal of the radius of convergence of ff at aa, which is at least rr). Thus, gg is also real analytic at aa, and in particular, gg is continuous.

Since akEa_k \in E,

0=f(ak)=(xa0)n0g(ak).0 = f\p{a_k} = \p{x - a_0}^{n_0} g\p{a_k}.

Moreover, because aka0a_k \neq a_0 for any kk, we can divide to get

g(ak)=0    g(a0)=limkg(ak)=0g\p{a_k} = 0 \implies g\p{a_0} = \lim_{k\to\infty} g\p{a_k} = 0

by continuity of gg. But

g(a0)=0    cn0=0,g\p{a_0} = 0 \implies c_{n_0} = 0,

a contradiction. Thus, cn=0c_n = 0 for all nn, so ff is 00 on B(a0,r)B\p{a_0, r}.

Claim: The set A={a|f(x)=0 in a neighborhood of a}A = \set{a \st f\p{x} = 0 \text{ in a neighborhood of } a} is non-empty, open, and closed.

From the first claim, we already have a0Aa_0 \in A, so AA is non-empty.

To see that AA is open, let aAa \in A, so by definition, there exists r>0r > 0 such that f=0f = 0 on B(a,r)B\p{a, r}. Since open balls are open, for any aB(a,r)a' \in B\p{a, r}, there exists r>0r' > 0 such that B(a,r)B(a,r)B\p{a', r'} \subseteq B\p{a, r}. But this means that f=0f = 0 on B(a,r)B\p{a', r'}. Since aa' was arbitrary, this shows B(a,r)AB\p{a, r} \subseteq A, so AA is open.

Finally, to see that AA is closed, assume there exists aAAa \in \cl{A} \setminus A. By definition, there exists a sequence bkA{a}b_k \in A \setminus \set{a} such that bkab_k \to a. But notice that a,bkEa, b_k \in E by definition, so by the first claim, this means that aAa \in A, a contradiction. Thus, A=AA = \cl{A}, so AA is closed.

Claim: A=IA = I, i.e., f=0f = 0 on all of II.

AA is open, and because AA is closed, AcA^\comp is also open. Moreover, AAc=IA \cup A^\comp = I. Thus, because II is connected and AA is non-empty, it follows that A=IA = I. (Otherwise, if AcA^\comp \neq \emptyset, then II is the union of two disjoint, non-empty, and open sets, which would imply that II is disconnected.) \square

Inner Products

Definition

Let VV be a real or complex vector space. Then an inner product ,\inner{\,\cdot\,, \,\cdot\,} is a function V×V(R or C)V \times V \to \p{\R \text{ or } \C} which is:

  1. (conjugate symmetric) x,y=y,x\inner{x, y} = \overline{\inner{y, x}}
  2. (linear in the first argument) x+λy,z=x,y+λy,z\inner{x + \lambda y, z} = \inner{x, y} + \lambda \inner{y, z}
  3. (positive definite) If x0x \neq 0, then x,x(0,)\inner{x, x} \in \p{0, \infty}
Example 2.

The dot product on Rn\R^n, i.e.,

xy=i=1nxiyix \cdot y = \sum_{i=1}^n x_iy_i

is an inner product on Rn\R^n. Recall that

xy=xycosθ,x \cdot y = \norm{x} \norm{y} \cos{\theta},

so intuitively, inner products allow us to talk about angles in more general vector spaces.

Lemma

The function

f,g=01f(x)g(x)dx\inner{f, g} = \int_0^1 f\p{x} \overline{g\p{x}} \,\diff{x}

defines an inner product on C([0,1])C\p{\br{0, 1}}. Thus,

f2f,f=(01f(x)2dx)12\norm{f}_2 \coloneqq \sqrt{\inner{f, f}} = \p{\int_0^1 \abs{f\p{x}}^2 \,\diff{x}}^{\frac{1}{2}}

defines a norm on C([0,1])C\p{\br{0, 1}} called the L2L^2-norm.

Remark.

C([0,1])C\p{\br{0, 1}} is not complete with respect to the L2L^2-norm.

Example 3.

Use Gram-Schmidt to find an orthonormal basis for V=span{1,x}V = \span\set{1, x}, where the span is taken over R\R.

Solution.

We can take e1=1e_1 = 1 since 11 is already normalized. Then

u2=xproje1(x)=xx,e1e1.u_2 = x - \proj_{e_1}\p{x} = x - \inner{x, e_1}e_1.

Here,

x,e1=01x1dx=12,\inner{x, e_1} = \int_0^1 x \cdot 1 \,\diff{x} = \frac{1}{2},

so u2=x12u_2 = x - \frac{1}{2}. Now we just need to normalize it:

u22=(01(x12)2dx)12=123    e2=u2u22=2x33.\norm{u_2}_2 = \p{\int_0^1 \p{x - \frac{1}{2}}^2 \,\diff{x}}^{\frac{1}{2}} = \frac{1}{2\sqrt{3}} \implies e_2 = \frac{u_2}{\norm{u_2}_2} = 2x\sqrt{3} - \sqrt{3}.

Thus,

{1,2x33}\boxed{\set{1, 2x\sqrt{3} - \sqrt{3}}}

is an orthonormal basis for VV.

Example 4.

Find the function in V=span{1,x}V = \span\set{1, x} that minimizes the L2L^2-distance to x2x^2.

Solution.

Recall that the minimizer is the orthogonal projection of x2x^2 to VV. This is

projV(x2)=x2,e1e1+x2,e2e2,\proj_V\p{x^2} = \inner{x^2, e_1}e_1 + \inner{x^2, e_2}e_2,

where {e1,e2}\set{e_1, e_2} is the orthonormal basis from the previous problem. Thus,

x2,e1=01x21dx=13x2,e2=01x2(2x33)dx=3233,\begin{gathered} \inner{x^2, e_1} = \int_0^1 x^2 \cdot 1 \,\diff{x} = \frac{1}{3} \\ \inner{x^2, e_2} = \int_0^1 x^2\p{2x\sqrt{3} - \sqrt{3}} \,\diff{x} = \frac{\sqrt{3}}{2} - \frac{\sqrt{3}}{3}, \end{gathered}

so the minimizer is

13+(3233)(2x33).\boxed{\frac{1}{3} + \p{\frac{\sqrt{3}}{2} - \frac{\sqrt{3}}{3}}\p{2x\sqrt{3} - \sqrt{3}}}.