Is the Euclidean distance from $\text{SO}(n)$ approximately convex?

468 Views Asked by At

Let $M_n$ be the vector space of real $n \times n$ matrices.

Define $f:M_n \to \mathbb{R}$ by $f(A)=\text{dist}^2(A,\text{SO}(n))$, where the distance is measured using the Frobenius (Euclidean) norm. It is known that $f$ is not convex.

Question: Does there exist a $c>0$ such that $$ f(tA+(1-t)B)\le c \big(tf(A)+(1-t)f(B)\big)$$ for every $A,B \in M_n, t \in [0,1]$?

If so, can we say something about the optimal $c$?

($c$ certainly cannot be arbitrarily close to $1$, since this would imply $f$ is convex).

If it helps, here is an explicit description of $f(A)$ in terms of the singular values of $A$:

$$ f(A)=(\sigma_1'-1)^2 + \sum_{i=2}^n (\sigma_i-1)^2,$$

where $\sigma_1 \le \sigma_2 \le \dots \le \sigma_n$ are the singular values of $A$, and $\sigma_1'=\text{sign} (\det A) \sigma_1$.

In particular, if $\det A \ge 0$, then

$$ f(A)= \sum_{i=1}^n (\sigma_i-1)^2.$$

1

There are 1 best solutions below

0
On BEST ANSWER

As commented by Max, there is no such $c>0$. Suppose $n$ is even, and take $A=Id,B=-Id,t=\frac{1}{2}$. Then,

$$f(tA+(1-t)B)\le c \big(tf(A)+(1-t)f(B)\big)$$ becomes

$$f(0)\le \frac{1}{2}c \big(f(A)+f(B)\big)=0,$$

since $A,B \in \text{SO}(n)$. This implies $n=f(0) \le 0$ which is a contradiction.

When $n$ is odd, we can embed $\text{SO}(n-1)$ inside $\text{SO}(n)$ and use the same argument.