12

This is a question from Linear Algebra Done Right.

Suppose $e_1, \dots, e_n$ is an orthonormal basis of $V$ and $v_1, \dots, v_n$ are vectors in $V$ such that $$ \| e_j - v_j \| < \frac{1} {\sqrt{n}} $$ for each $j$. Prove that $v_1, \dots, v_n$ is a basis of $V$. I tried to show this if $V$ is a two dimensional space by contradiction. However, it seems too messy to use the same idea for a general case.

Does anyone have hints for this problem?

user1101010
  • 3,638
  • 1
  • 17
  • 40
  • Duplicate of http://math.stackexchange.com/questions/1823716/orthonormal-basis-with-specific-norm/1823731#1823731. – levap Nov 23 '16 at 01:14
  • 2
    Despite the duplicate, I think this question should stay open. It stands to become the more useful resource. – Ben Grossmann Nov 23 '16 at 01:18
  • If you are interested in approaching this using the fact that the $\ell^1$ norm is $\le$ $\sqrt{2}$ times the $\ell^2$ norm , see my answer at the end. – Mkanders Nov 02 '23 at 02:28

10 Answers10

13

Thanks to one of the students in my group, I actually solved it a few days ago. The idea is similar to one of the comments. I put my solution here since it is more on linear algebra side without touching matrix.

We prove by contradiction. Assume otherwise, i.e. the set $\{v_i\}$ is not a basis for $V$. Then there exits some $w \in V$, such that $\langle w, v_i\rangle = 0$ for every $i = 1, \dots, n$ where $w \neq 0$. Such $w$ can be chosen from the annihilator of $\text{span}\{v_1, \dots, v_n\}$. Since we know $\dim( \text{span}\{v_1, \dots, v_n\}) < n$, $w$ always exists. Then \begin{align*} \|e_i - v_i\|^2 \|w\|^2 &\ge | \langle e_i-v_i, w\rangle |^2 && (\text{Cauchy-Schwarz}) \\ &= | \langle e_i, w\rangle - \langle v_i, w\rangle |^2 \\ &= | \langle e_i, w\rangle|^2 \end{align*} Now summing over $i = 1, \dots, n$, we get \begin{align*} \sum_{i=1}^n \| e_i - v_i\|^2 \|w\|^2 &\ge \sum_{i=1}^n (\langle e_i, w\rangle)^2 \\ &= \|w\|^2 \end{align*} This implies $\sum_{i=1}^n \|e_i - v_i\|^2 \ge 1$. On the other hand, by assumption, $\sum_{i=1}^n \|e_i-v_i\|^2 < 1$. This is a contradiction.

Hence, $\{v_1, \dots, v_n\}$ must be a basis.

Ben Grossmann
  • 234,171
  • 12
  • 184
  • 355
user1101010
  • 3,638
  • 1
  • 17
  • 40
6

Define $V$ as a matrix whose columns are $v_i$, taken componentwise in the $e$-basis, and a matrix $Δ = I - V$. Then the elements of $Δ^TΔ$ are of the form of scalar products between all pairs of $(e_i - v_i)$, about which we know that

$$(e_i - v_i)\cdot(e_i - v_i) = \|e_i - v_i\|^2 < \frac1n,$$ $$|(e_i - v_i)\cdot(e_j - v_j)| \le \|e_i - v_i\|\cdot\|e_j - v_j\| < \frac1n.$$

That's interesting so let's use it!

We know our matrix $Δ^TΔ$ has elements which are never $\frac1n$ or more, and want to prove that $I-Δ$ is regular. A contradiction would be to find a normalized vector $x$ such that

$$x = Δ x.$$

If such vector exists then $x^TΔ^TΔx = \|Δx\|^2 = \|x\|^2 = 1$. Now proceed as

$$1 = \sum_{j,k}|x_j(Δ^TΔ)_{jk}x_k| \le \sum_{jk} |x_j| |(Δ^TΔ)_{jk}| |x_k| < \frac1n \sum_{jk} |x_j| |x_k|$$

(here I used triangle inequality and the fact we know about elements of $Δ^TΔ$, notice the strict $<$ between the two latter terms),

$$\frac1n \sum_{jk} |x_j| |x_k| = n\left(\frac{\sum_j |x_j|}{n}\right)^2 \le n \frac{\sum_j |x_j|^2}{n}$$

(inequality between arithmetic and quadratic mean, probably Cauchy-Schwartz if you rephrase it in a clever way), thus

$$1 < \sum_j |x_j|^2 = 1.$$

That's a contradiction, so that's our proof done. It also shows that the bound is tight: for $\|e_j - v_j\| \le 1/\sqrt n$ you could find a counterexample which saturates each of the $≤$'s (namely, $(e_j - v_j)_k = 1/n, \forall j \forall k$).

The Vee
  • 3,151
6

Let $a_i \in \mathbb{F}$ such that $\sum_{i=1}^n a_i v_i = 0$. Then we have

$$ 0 = \left \| \sum_{i=1}^n a_i v_i \right \| = \left \| \sum_{i=1}^n a_i (e_i + (v_i - e_i)) \right \| = \left \| \sum_{i=1}^n a_i e_i + \sum_{i=1}^n a_i (v_i - e_i) \right \| \geq \left| \left \| \sum_{i=1}^n a_i e_i \right \| - \left \| \sum_{i=1}^n a_i (v_i - e_i) \right \|\right| $$

which implies that

$$ \left( \sum_{i=1}^n |a_i|^2 \right)^{\frac{1}{2}} = \left \| \sum_{i=1}^n a_i e_i \right \| = \left \| \sum_{i=1}^n a_i (v_i - e_i) \right \| \leq \sum_{i=1}^n |a_i| \| v_i - e_i \|. $$

If $a_i \neq 0$ for some $1 \leq i \leq n$ then using Cauchy-Schwartz we have

$$ \sum_{i=1}^n |a_i| \| v_i - e_i \| < \sum_{i=1}^n |a_i| \frac{1}{\sqrt{n}} \leq \left( \sum_{i=1}^n |a_i|^2 \right)^{\frac{1}{2}} \left( \sum_{i=1}^n \frac{1}{n} \right)^{\frac{1}{2}} = \left( \sum_{i=1}^n |a_i|^2 \right)^{\frac{1}{2}}$$

and we have arrived to a contradiction.

levap
  • 67,610
  • As straightforward as they come. Well done – Ben Grossmann Nov 23 '16 at 01:08
  • 1
    @Omnomnomnom: Thanks. I wasn't sure if I answered this question once already and it seems I did (http://math.stackexchange.com/questions/1823716/orthonormal-basis-with-specific-norm/1823731#1823731). The solution I wrote there is basically the same but it doesn't proceed by contradiction and uses only the regular triangle inequality (and not the reverse one) so it is slightly more elegant in my taste. – levap Nov 23 '16 at 01:12
5

Though if you want a neater proof...* ;-)

*) but don't try that at an exam.

  1. Draw this picture:

enter image description here

  1. Draw this picture:

enter image description here

  1. Generalize.

How it works

The $n$ vectors are linearly independent (and thus form a base) if there's no hyperplane that contains them all. If we draw an $n$-ball about the tip of each of the basis vectors, we mark all possible points that are compatible with the deviation $\|e_j - v_j\| < r$ (or $≤$, depending on whether we include the surface) from the elements of the orthonormal base, $e_j$. The hyperplane defined by the equation

$$\sum_j x_j = 0$$

has the property that for any other hyperplane the distance of at least one of the centers will be equal of higher (the equal ones are those which put a minus in front of some of the coordinates). Therefore it's the optimal candidate for one containing an overlap with all the spheres. If there's no overlap, any $n$-tuple of vectors pointing somewhere one into each of the spheres will be linearly independent.

The distance of each of the centers from this hyperplane is then the upper bound on the radius we can allow without breaking the guarantee of linear independence, and can be easily computed to be $1/\sqrt n$ by any suitable algebraic method.

The Vee
  • 3,151
3

Associate to each list $u = (u_1, \dots, u_n)$ of $n$ vectors in $V$ a linear map \begin{align*} T_u : \mathbb C^n &\to V \\ T_ua &= a_1 u_1 + \cdots + a_n u_n. \end{align*} By the definition of linear independence, $u$ is linearly independent iff $T_u a \neq 0$ for all $a \neq 0$. The idea is, for $a \neq 0$, to bound the distance $T_v a$ from $0$ by (1) bounding it's distance from $T_e a$, and (2) bounding the distance from $T_e a$ to $0$, to show that $\|T_v a - 0\| > 0$, hence $T_v a \neq 0$. Since $a$ was arbitrary, it follows that $v$ is linearly independent, and since $v$ has $n = \dim V$ vectors, it's a basis.

For (1) we use the given condition and the triangle inequality: $$ \|T_e a - T_v a\| = \left\| \sum_{i=1}^n a_i e_i - \sum_{i=1}^n a_i v_i \right\| \leq \sum_{i=1}^n |a_i| \|e_i - v_i\| < \frac{1}{\sqrt{n}} \| a \|_1. $$

For (2), $$ \|T_e a - 0 \| = \left\| \sum_{i=1}^n a_i e_i \right\| = \sqrt{|a_1|^2 + \cdots + |a_n|^2} = \|a\|_2 $$ by the Pythagorean theorem.

It's a consequence of Cauchy-Schwarz that $\frac{1}{\sqrt{n}} \|a\|_1 \leq \|a\|_2$: $$ \|a\|_1 = \langle (|a_1|, \dots, |a_n|), (1, \dots, 1) \rangle \leq \|(|a_1|, \dots, |a_n|)\|_2 \|(1, \dots, 1)\|_2 = \|a\|_2 \sqrt{n}. $$ Then by the reverse triangle inequality $\|u - w\| \geq \big| \|u\| - \|w\| \big|$, it follows that $T_v a$ is a positive distance from $0$: $$ 0 \leq \|a\|_2 - \frac{1}{\sqrt{n}} \|a\|_1 < \|Te_a - 0\| - \|T_e a - T_v a\| \leq \|T_v a - 0\|. $$

In fact, my answer to Regarding linear independence on a normed-linear space given a condition proves essentially the same result with more of a topology/analysis bent.

Jon Warneke
  • 5,017
3

For each $i,$ let $v_i=e_i-u_i.$ Suppose there are scalars $s_1,.., s_n,$ not all $0,$ such that $\sum_{i=1}^ns_iv_i=0.$ $$\text {Let }\quad e= \sum_{i=1}^ns_ie_i,\quad u=\sum_{i=1}^ns_iu_i.$$ Then $e=u.$ And $\|e\|^2=\sum_{i=1}^n|s_i|^2.$ We have $$\|u\|^2\leq \left(\sum_{j=1}^n\|s_iu_i\|\right)^2<\left(\sum_{i=1}^n|s_i|/\sqrt n\right)^2.$$ The second inequality is strict because not all $s_i$ are $0$, and because $\|u_i\|<1\sqrt n$ for every $i$.

Consider the vectors $s=(|s_1|,..,|s_n|)$ and $t=(t_1,..,t_n)$ where $t_i=1/\sqrt n$ for every $i.$ We have $$\|u\|^2<\left(\sum_{i=1}^n|s_i|/\sqrt n\right)^2=(s\cdot t)^2\leq \|s\|^2\; \|t\|^2=\|s\|^2=\sum_{i=1}^n|s_i|^2=\|e\|^2=\|u\|^2,$$ a contradiction.

Remarks: I thought of this before reading the other answers. It seems to be sufficiently different, at least in its style. I think the one from jing007 is the best (including mine)... At one point I wrote $|s_i|^2$ rather than $s_i^2$, so that it is valid for $\mathbb C^n$ as well as for $\mathbb R^n.$

1

Say $v_i=e_i+x_i$ and suppose there exist $a_i$s such that $\sum_1^n a_i v_i=0$

Then $$ \left \langle \sum_1^n a_i e_i,\sum_1^n a_i v_i\right \rangle =0\\ \sum_1^n a_i^2\|e_i\|^2+\left \langle \sum_1^n a_i e_i,\sum_1^n a_i x_i\right \rangle=0\\ $$ Which is impossible since $$ \|\left \langle \sum_1^n a_i e_i,\sum_1^n a_i x_i\right \rangle\|<\\ \sqrt{\sum_1^n a_i^2}\frac{\sum_1^n |a_i|}{\sqrt n}=\\ \sqrt{\sum_1^n a_i^2}\sqrt n \frac{\sum_1^n |a_i|}{n}\le\\ \sqrt{\sum_1^n a_i^2}\sqrt n \sqrt{\frac{\sum_1^n |a_i|^2}{n}}=\\ \sum_1^n a_i^2 $$

Ben Grossmann
  • 234,171
  • 12
  • 184
  • 355
  • I've changed your answer in order to put it into more standard notation. You'll have to clarify why $|\sum x_i| < 1$, though. Also, note that you've been using $u_i$ instead of $e_i$. – Ben Grossmann Nov 22 '16 at 20:09
  • @Omnomnomnom oh, I didn't see it is $e$, I really thought it was $u$. Also thought it was $\frac 1 n$ not $\frac 1 {\sqrt n}$. Wow, I must be super tired... – Djura Marinkov Nov 22 '16 at 20:28
  • Well, in any case, I appreciate the effort. I won't decide on who gets the bounty until the end of the week, if you decide you want to take another shot at it. – Ben Grossmann Nov 22 '16 at 20:29
1

Lemma: given $0 \le a_1 , .. a_n$ and $s = \sum a_i$, we have $\sum a_i^2 \ge \frac{s^2}{n}$.

Proof of lemma: Weierstrass :)

$\{(a_1,\dots,a_n) \mid 0 \le a_1, \dots, a_n, \sum a_i = s\}$ is a compact set in $\Bbb R^n \implies \sum a_i^2$ achieves its minimum.

Let $i,j$ be such that $a_i \ne a_j$. Change $a_i$ and $a_j$ to $\frac{a_i +a_j}{2}$ and get the solution with less value of $\sum a_i^2$.

So, minimum is achieved if $a_i = \frac{s}{n}$ for all i $\implies$ minimum is $\frac{s^2}{n}$.
QED

For each $i$, write $v_i = e_i + c_i$ where $\|e_i\| = 1$ and $\|c_i\| < \frac{1}{\sqrt{n}}$.

Assume (for the sake of contradiction) that $\{v_i\}$ is not a basis.

$$ \sum b_iv_i = 0 \implies \sum b_ie_i + \sum b_ic_i = 0 $$

Let $a_i = |b_i|, \sum a_i = s > 0$, then $\| \sum b_ie_i \| \ge \frac{s}{\sqrt{n}}$ by the lemma. However, we also have

$$ \| \sum b_ic_i \| \leq \sum |b_i|\, \|c_i\| < \frac 1{\sqrt{n}} \sum |b_i| = \frac{s}{\sqrt{n}} $$ which is absurd.

Ben Grossmann
  • 234,171
  • 12
  • 184
  • 355
kotomord
  • 1,900
  • 1
    Not sure what you mean by Weierstrass. Normally, I see that lemma proven using the Cauchy-Schwarz inequality. – Ben Grossmann Nov 22 '16 at 20:58
  • Yes, Cauchy-Schwarz inequality is more fast way to prove; – kotomord Nov 22 '16 at 21:02
  • 1
    What exactly is "Weierstass" here, though? I hope you don't mind that I've rewritten your post a bit to make things clearer. – Ben Grossmann Nov 22 '16 at 21:07
  • sorry, fixed ^) Thanks! – kotomord Nov 22 '16 at 21:08
  • I think that lemma shows that the $\ell^1$ norm of an n-dimensional vector is less than or equal to the $\sqrt{n}$ times the $\ell^2$ norm. $||v||_1 \le \sqrt{n} \cdot ||v||_2$. I learned this from Strang, Linear Algebra and Learning from Data, Problem Set 1.11, #3. – Mkanders Oct 29 '23 at 17:44
0

Hint: Normalize $v_i$ for simplicity. Write $v_i = (v_i, a_i) a_i + b_i$ and observe that $(a_i, b_i)= 0$ and $|(v_i, a_i)| >> ||bi||$, estimate each side, then assume the set {${v_j}$} is linearly dependent and derive a contradiction.

Burrrrb
  • 1,566
0

Let $\mathbf{E}$ and $\mathbf{V}$ be the $n$-by-$n$ matrices with columns $\mathbf{e_1},\ldots,\mathbf{e_n}$ and $\mathbf{v_1},\ldots,\mathbf{v_n}$.

Let $\mathbf{x}$ be a non-zero vector with elements $x_1, \ldots, x_n$.

We want to show that $\mathbf{Vx} \neq \mathbf{0}$.

We will use the fact that the $\ell^1$ norm of an n-dimensional vector is less than or equal to the $\sqrt{n}$ times the $\ell^2$ norm.

\begin{align*} ||\mathbf{x}||_1 \le \sqrt{n} \cdot ||\mathbf{x}||_2\\ \frac{1}{\sqrt{n}} \cdot ||\mathbf{x}||_1 \le ||\mathbf{x}||_2 \end{align*}

I learned this from Strang, Linear Algebra and Learning from Data, Problem Set 1.11, #3. The proof is to choose a vector $\mathbf{w}$ consisting of all $1$s and $-1$s . The length of this vector is $\sqrt{n}$. If you choose the $1$ or $-1$ to correspond to the sign of the element $x_i$, then $\mathbf{w^Tx} = ||\mathbf{x}||_1$. \begin{align*} |\mathbf{w^Tx} | &\le ||\mathbf{w}||_2 \cdot ||\mathbf{x}||_2 && \text{by Cauchy-Schwartz}\\ ||\mathbf{x}||_1 &\le \sqrt{n} \cdot ||\mathbf{x}||_2 \end{align*} I will explain why below, but $||\mathbf{e_i}-\mathbf{v_i}||_2 < \frac{1}{\sqrt{n}}$ means that the $\ell^2$ norm of $(\mathbf{E} - \mathbf{V})\mathbf{x}$ is strictly less than $\frac{1}{\sqrt{n}}$ times the $\ell^1$ norm of $\mathbf{x}$: $$||(\mathbf{E} - \mathbf{V})\mathbf{x}||_2 < \frac{1}{\sqrt{n}} ||\mathbf{x}||_1$$ Now, let's use this to show that $\mathbf{Vx} \neq \mathbf{0}$. \begin{align*} ||(\mathbf{E} - \mathbf{V})\mathbf{x}||_2 &= ||(\mathbf{Ex} - \mathbf{Vx})||_2\\ &\ge |(||(\mathbf{Ex} ||_2 - ||\mathbf{Vx}||_2))| && \text{by the reverse triangle inequality}\\ &\ge |(||\mathbf{x} ||_2 - ||\mathbf{Vx}||_2)| && \text{because } \mathbf{E} \text{ is an orthonormal matrix}\\ & \ge |\frac{1}{\sqrt{n}}||\mathbf{x}||_1 - ||\mathbf{Vx}||_2)| && \text{ by the fact from Strang about the } \ell^1 \text {norm} \end{align*} But we are given that $||(\mathbf{E} - \mathbf{V})\mathbf{x}||_2 < \frac{1}{\sqrt{n}} ||\mathbf{x}||_1$, so $||\mathbf{Vx}||_2 > 0$ and $\mathbf{Vx} \neq \mathbf{0}$.

Were we really given that $||(\mathbf{E} - \mathbf{V})\mathbf{x}||_2 < \frac{1}{\sqrt{n}} ||\mathbf{x}||_1$? \begin{align*} ||(\mathbf{E} - \mathbf{V})\mathbf{x}||_2 &= ||\sum_{i=1}^n (\mathbf{e_i}-\mathbf{v_i})x_i||_2 && \text{by the definition of matrix multiplication} \\ &\le \sum_{i=1}^n ||\mathbf{e_i}-\mathbf{v_i}||_2 \cdot |x_i| && \text{by the triangle inequality}\\ &< \sum_{i=1}^n \frac{1}{\sqrt{n}} \cdot |x_i| &&\text{because } ||\mathbf{e_i}-\mathbf{v_i}||_2 < \frac{1}{\sqrt{n}}\\ &< \frac{1}{\sqrt{n}} \sum_{i=1}^n |x_i| \\ &< \frac{1}{\sqrt{n}} ||\mathbf{x}||_1 && \text{by definition of the } \ell^1 \text{norm} \end{align*}

Mkanders
  • 193
  • 9