Let $0 < \sigma_i \in \mathbb{R}$ ($i=1 \dots d$). Is it true that

$$ c \sum_{i=1}^d (\sigma_i-1)^2 \le \sqrt{\sum_{i=1}^d (\sigma_i^2-1)^2} \tag{1}$$

for some $c>0$ which does not depend on the $\sigma_i$.

(I actually suspect this holds for $c=1$).


This inequality is equivalent to the following

$$c \|\sqrt{A^TA}-I\|^2 \le \|A^TA-I\|, \tag{2}$$

where $A$ is a $d \times d$ real matrix. (The equivalence is obtained by considering SVD).

(Inequality $(2)$ comes from comparing different ways to measure deviation of a linear transformation from being an isometry).


Isn't this just a direct application of Cauchy-Schwarz inequality? We have \begin{align*} \sum_{i=1}^d (\sigma_i-1)^2 &= \left\langle\left((\sigma_1-1)^2,\ldots,(\sigma_d-1)^2\right),\ (1,\ldots,1)\right\rangle\\ &\le \left\|\left((\sigma_1-1)^2,\ldots,(\sigma_d-1)^2\right)\right\|\cdot \|(1,\ldots,1)\|\\ &= \sqrt{d}\sqrt{\sum_i(\sigma_i-1)^4}\\ &= \sqrt{d}\sqrt{\sum_i\left[(\sigma_i-1)^4-(\sigma_i^2-1)^2+(\sigma_i^2-1)^2\right]}\\ &= \sqrt{d}\sqrt{\sum_i\left[-4\sigma_i^2(\sigma_i-1)^2+(\sigma_i^2-1)^2\right]}\\ &\le \sqrt{d}\sqrt{\sum_i(\sigma_i^2-1)^2}. \end{align*} Equality holds in the first inequality iff all $|\sigma_i-1|$s are equal, and equality holds in the second inequality iff each $\sigma_i$ is $0$ or $1$. Hence ties occur in both inequalities iff all $\sigma_i$ are equal to zero or all $\sigma_i$ are equal to one, i.e. iff $A=0$ or $A$ is real orthogonal.

  • $\begingroup$ Well spotted!${}$ $\endgroup$ – Omnomnomnom Mar 10 '17 at 4:28

Sorry for all the edits...

The inequality holds for $c=1/d$. Indeed, $$ \sum(\sigma_i-1)^2 \le d \max_i (\sigma_i-1)^2 \le d \max_i |\sigma_i^2-1|\le d\sqrt{\sum (\sigma_i^2-1)^2}, $$

The inequality does not hold for any $c>1/\sqrt{d}$: Assume $\sigma_i = n+1$ for all $i$. Then the inequality reads $$ c\,d\,n^2 \le \sqrt{d(n^2+2n)^2} = \sqrt{d}(n^2+2n),$$ and by taking $n\to \infty$ you get that $c\le 1/\sqrt{d}$.

In fact, if $\sum \sigma_i \ge d$, then inequality holds with $c=1/\sqrt{d}$. Indeed, using Hölder's inequality, we have $$ \sum(\sigma_i-1)^2 = \sum(\sigma_i^2-1) - 2\sum(\sigma_i-1) \le \sqrt{d}\sqrt{\sum (\sigma_i^2-1)^2} - 2\sum(\sigma_i-1). $$ I suspect $c=1/\sqrt{d}$ is the best constant.


It's clear that in the case of $\sum_i(\sigma_i-1)^2 \leq 1$, we have $$ \sqrt{\sum_{i=1}^d (\sigma_i^2-1)^2} = \sqrt{\sum_{i=1}^d (\sigma_i-1)^2(\sigma_i+1)^2} \geq \sqrt{\sum_{i=1}^d (\sigma_i-1)^2} \geq \sum_{i=1}^d (\sigma_i-1)^2 $$ I'm not sure about the general case, though.

Consider the case of $\sigma_1 = x+1 >\sigma_2 = y+1 > 1$ $\sigma_3 = \cdots = \sigma_d = 1$. We define $$ f(x,y) = \sqrt{\sum_{i=1}^d (\sigma_i^2-1)^2} = \sqrt{[(x+1)^2 - 1]^2 + [(y+1)^2-1]^2}\\ = \sqrt{(x^2 + 2x)^2 + (y^2 + 2y)^2}\\ g(x,y) = \sum_{i=1}^d(\sigma_i - 1)^2 = x^2 + y^2 $$ Now, if we select $x = \alpha y^2$ for $\alpha > 0$, I think we'll notice something. I have a hunch that $f(x,y)/g(x,y)$ has no lower bound.

  • $\begingroup$ Thanks, this is nice. I still hope for a solution of the general case. $\endgroup$ – Asaf Shachar Mar 9 '17 at 17:56
  • $\begingroup$ Added a little something to consider. $\endgroup$ – Omnomnomnom Mar 9 '17 at 19:49

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.