Two iterative formulas of largest and smallest singular value of nonsingular matrices

$\Rightarrow:$ If $a=\sigma_{n}$, then

$$\sigma_{n}^{2}\left(\|A\|_{F}^{2}-\sigma_{n}^{2}\right)^{n-1}=|\operatorname{det}A|^{2}(n-1)^{n-1}.$$

Since $\|A\|_{F}^{2}=\sigma_{1}^{2}+\sigma_{2}^{2}+\cdots+\sigma_{n}^{2}$ and $|\operatorname{det}A|^{2}=\sigma_{1}^{2}\sigma_{2}^{2}\cdots\sigma_{n}^{2}$, we have

$$\begin{gathered}\sigma_{n}^{2}\left(\sigma_{1}^{2}+\sigma_{2}^{2}+\cdots+\sigma_{n-1}^{2}\right)^{n-1}=\sigma_{1}^{2}\sigma_{2}^{2}\cdots\sigma_{n}^{2}(n-1)^{n-1}.\\ \left(\frac{\sigma_{1}^{2}+\sigma_{2}^{2}+\cdots+\sigma_{n-1}^{2}}{n-1}\right)^{n-1}=\sigma_{1}^{2}\sigma_{2}^{2}\cdots\sigma_{n-1}^{2}\end{gathered}$$

By the arithmetic-geometric mean inequality, the above equation holds if and only if

$$\sigma_{1}^{2}=\sigma_{2}^{2}=\cdots=\sigma_{n-1}^{2}$$

if and only if

$$\sigma_{1}=\sigma_{2}=\cdots=\sigma_{n-1}$$

$\Leftarrow:$ Let

$$\sigma_{1}^{2}=\sigma_{2}^{2}=\cdots=\sigma_{n-1}^{2}=c^{2}(c>0).$$

Then $a$ is the smallest positive solution to the equation

$$x^{2}\left(c^{2}(n-1)+\sigma_{n}^{2}-x^{2}\right)^{n-1}=c^{2(n-1)}\sigma_{n}^{2}(n-1)^{n-1}$$

if and only if $a$ is the smallest positive zero point of

$$f(x)=x^{2}\left(1+\frac{\sigma_{n}^{2}-x^{2}}{c^{2}(n-1)}\right)^{n-1}-\sigma_{n}^{2}.$$

Next, we proof $a=\sigma_{n}$. Obviously, we can see that $f\left(\sigma_{n}\right)=0$ and $f(0)=-\sigma_{n}^{2}<0$. Next, we prove that $f(x)$ is an strictly increasing function on $\left[0,\sigma_{n}\right]$. Taking the derivative of $f(x)$, we can get

$$f^{\prime}(x)=\left(1+\frac{\sigma_{n}^{2}-x^{2}}{c^{2}(n-1)}\right)^{n-2}\frac{2x}{c^{2}(n-1)}\left(c^{2}(n-1)+\sigma_{n}^{2}-nx^{2}\right).$$

$\forall x_{0}\in\left(0,\sigma_{n}\right)$, we have

$$c^{2}(n-1)+\sigma_{n}^{2}-nx_{0}^{2}\geqslant n\sigma_{n}^{2}-nx_{0}^{2}>0.$$

We have $f^{\prime}\left(x_{0}\right)>0$. Therefore $f(x)$ is an strictly increasing function on $\left[0,\sigma_{n}\right]$ and $\sigma_{n}$ is the smallest positive zero point of

$$f(x)=x^{2}\left(1+\frac{\sigma_{n}^{2}-x^{2}}{c^{2}(n-1)}\right)^{n-1}-\sigma_{n}^{2}$$

Therefore, $\sigma_{n}$ is the smallest positive solution to the equation

$$x^{2}\left(c^{2}(n-1)+\sigma_{n}^{2}-x^{2}\right)^{n-1}=c^{2(n-1)}\sigma_{n}^{2}(n-1)^{n-1}.$$

We get $a=\sigma_{n}.$ ∎

Let $0\leqslant\lambda<\sigma_{n}^{2}$, by the arithmetic-geometric mean inequality, we have

$$\left(\sigma_{1}^{2}-\lambda\right)\left(\sigma_{2}^{2}-\lambda\right)\cdots\left(\sigma_{n-1}^{2}-\lambda\right)\leqslant\left(\frac{\sigma_{1}^{2}+\cdots+\sigma_{n-1}^{2}-(n-1)\lambda}{n-1}\right)^{n-1}$$

Since

	$\displaystyle\left(\sigma_{1}^{2}-\lambda\right)\left(\sigma_{2}^{2}-\lambda\right)\cdots\left(\sigma_{n-1}^{2}-\lambda\right)$	$\displaystyle=\frac{\left(\sigma_{1}^{2}-\lambda\right)\left(\sigma_{2}^{2}-\lambda\right)\cdots\left(\sigma_{n}^{2}-\lambda\right)}{\sigma_{n}^{2}-\lambda}$
		$\displaystyle=\frac{\left|\operatorname{det}\left(\lambda I_{n}-A^{H}A\right)\right|}{\sigma_{n}^{2}-\lambda}$

we have

$$\frac{\left|\operatorname{det}\left(\lambda I_{n}-A^{H}A\right)\right|}{\sigma_{n}^{2}-\lambda}\leqslant\left(\frac{\sigma_{1}^{2}+\cdots+\sigma_{n-1}^{2}-(n-1)\lambda}{n-1}\right)^{n-1}$$

$$\begin{gathered}\sigma_{n}^{2}\geqslant\lambda+\left|\operatorname{det}\left(\lambda I_{n}-A^{H}A\right)\right|\left(\frac{n-1}{\sigma_{1}^{2}+\cdots+\sigma_{n-1}^{2}-(n-1)\lambda}\right)^{n-1}\\ \sigma_{n}\geqslant\left(\lambda+\left|\operatorname{det}\left(\lambda I_{n}-A^{H}A\right)\right|\left(\frac{n-1}{\sigma_{1}^{2}+\cdots+\sigma_{n-1}^{2}+\sigma_{n}^{2}-(n-1)\lambda}\right)^{n-1}\right)^{1/2}\\ \sigma_{n}\geqslant\left(\lambda+\left|\operatorname{det}\left(\lambda I_{n}-A^{H}A\right)\right|\left(\frac{n-1}{\|A\|_{F}^{2}-(n-1)\lambda}\right)^{n-1}\right)^{1/2}\end{gathered}$$

Let $\lambda\rightarrow\sigma_{n}^{2-}\left(\lambda\right.$ tends to $\sigma_{n}^{2}$ from the left). We get that the above inequality is also true for $\lambda=\sigma_{n}^{2}$. Therefore, for $0\leqslant\lambda\leqslant\sigma_{n}^{2}$, we have

$$\sigma_{n}\geqslant\left(\lambda+\left|\operatorname{det}\left(\lambda I_{n}-A^{H}A\right)\right|\left(\frac{n-1}{\|A\|_{F}^{2}-(n-1)\lambda}\right)^{n-1}\right)^{1/2}$$

(1)

We show by induction on $k$ that

$$\sigma_{n}\geqslant a_{k+1}\geqslant a_{k}>0$$

We have $\sigma_{n}\geqslant a_{1}>0$. In equation 1, let $\lambda=a_{1}^{2}$, we have

$$\sigma_{n}\geqslant a_{2}=\left(a_{1}^{2}+\left|\operatorname{det}\left(a_{1}^{2}I_{n}-A^{H}A\right)\right|\left(\frac{n-1}{\|A\|_{F}^{2}-(n-1)a_{1}^{2}}\right)^{n-1}\right)^{1/2}\geqslant a_{1}>0$$

Assume that our claim is true for $k=m$, that is $\sigma_{n}\geqslant a_{m+1}\geqslant a_{m}>0$. Now we consider the case when $k=m+1$. In equation 1, let $\lambda=a_{m+1}^{2}$, we have

$$\sigma_{n}\geqslant a_{m+2}=\left(a_{m+1}^{2}+\left|\operatorname{det}\left(a_{m+1}^{2}I_{n}-A^{H}A\right)\right|\left(\frac{n-1}{\|A\|_{F}^{2}-(n-1)a_{m+1}^{2}}\right)^{n-1}\right)^{1/2}\geqslant a_{m+1}>0$$

Hence $\sigma_{n}\geqslant a_{m+2}\geqslant a_{m+1}>0$. This proves $\sigma_{n}\geqslant a_{k+1}\geqslant a_{k}>0(k=1,2,\cdots)$. By the well known monotone convergence theorem, $\lim_{k\rightarrow\infty}a_{k}$ exists. Let $\lim_{k\rightarrow\infty}a_{k}=$ $\sigma$, then

$$\sigma=\left(\sigma^{2}+\left|\operatorname{det}\left(\sigma^{2}I_{n}-A^{H}A\right)\right|\left(\frac{n-1}{\|A\|_{F}^{2}-(n-1)\sigma^{2}}\right)^{n-1}\right)^{1/2},k=1,2,\cdots$$

We have

$$\left|\operatorname{det}\left(\sigma^{2}I_{n}-A^{H}A\right)\right|\left(\frac{n-1}{\|A\|_{F}^{2}-(n-1)\sigma^{2}}\right)^{n-1}=0$$

Since

$$\frac{n-1}{\|A\|_{F}^{2}-(n-1)\sigma^{2}}\neq 0$$

we have

$$\operatorname{det}\left(\sigma^{2}I_{n}-A^{H}A\right)=0.$$

We get that $\sigma^{2}$ is the eigenvalue of $A^{H}A$. Since $\sigma_{n}^{2}$ is the smallest eigenvalue of $A^{H}A$, we have $\sigma^{2}\geqslant\sigma_{n}^{2}$. According to the definition of $\sigma$, we have $\sigma\leqslant\sigma_{n}$. Therefore, $\sigma^{2}\leqslant\sigma_{n}^{2}$ and we get $\sigma=\sigma_{n}$. Hence $\lim_{k\rightarrow+\infty}a_{k}=\sigma_{n}$. ∎

From [3], we have $\sigma_{n}\geqslant a>0$. We know Corollary 1 is a special case of Theorem 1. ∎

Set $\lambda>\sigma_{1}^{2}$, according to the arithmetic geometric mean inequality, we can get

$$\left(\lambda-\sigma_{2}^{2}\right)\left(\lambda-\sigma_{3}^{2}\right)\cdots\left(\lambda-\sigma_{n}^{2}\right)\leqslant\left(\frac{(n-1)\lambda-(\sigma_{2}^{2}+\cdots+\sigma_{n}^{2})}{n-1}\right)^{n-1}.$$

Since

$$\begin{aligned} \left(\lambda-\sigma_{2}^{2}\right)\left(\lambda-\sigma_{3}^{2}\right)\cdots\left(\lambda-\sigma_{n}^{2}\right)&=\frac{\left(\lambda-\sigma_{1}^{2}\right)\left(\lambda-\sigma_{2}^{2}\right)\cdots\left(\lambda-\sigma_{n}^{2}\right)}{\lambda-\sigma_{1}^{2}}\\ &=\frac{|\det(\lambda I_{n}-A^{H}A)|}{\lambda-\sigma_{1}^{2}}\end{aligned},$$

we can get

$$\frac{|\det(\lambda I_{n}-A^{H}A)|}{\lambda-\sigma_{1}^{2}}\leqslant\left(\frac{(n-1)\lambda-(\sigma_{2}^{2}+\cdots+\sigma_{n}^{2})}{n-1}\right)^{n-1}$$

	$\displaystyle\sigma_{1}^{2}$	$\displaystyle\leqslant\lambda-|\operatorname{det}\left(\lambda-A^{H}A\right)\mid\left(\frac{n-1}{(n-1)\lambda+\sigma_{1}^{2}-\|A\|_{F}^{2}}\right)^{n-1}$
		$\displaystyle\leqslant\lambda-\left|\operatorname{det}\left(\lambda-A^{H}A\right)\right|\left(\frac{n-1}{(n+1)\lambda-\|A\|_{F}^{2}}\right)^{n-1}$

$$\sigma_{1}\leqslant\left(\lambda-\left|\operatorname{det}\left(\lambda-A^{H}A\right)\right|\left(\frac{n-1}{(n+1)\lambda-\|A\|_{F}^{2}}\right)^{n-1}\right)^{1/2}$$

Let $\lambda\to{\sigma_{1}^{2}}^{+}$($\lambda$ tend to $\sigma_{1}^{2}$ from the right). We get that the above equation is also true for $\lambda=\sigma_{1}^{2}$. So, for $\lambda\geqslant\sigma_{1}^{2}$, we have

$$\sigma_{1}\leqslant\left(\lambda-\left|\operatorname{det}\left(\lambda-A^{H}A\right)\right|\left(\frac{n-1}{(n+1)\lambda-\|A\|_{F}^{2}}\right)^{n-1}\right)^{1/2}$$

(2)

We use induction on $k$ to prove that:

$$\sigma_{1}\leqslant a_{k+1}\leqslant a_{k}$$

For $k=1$, $\sigma_{1}\leqslant a_{1}$ can be obtained from the condition. In the equation (2), taking $\lambda=a_{1}^{2}$, we can get

$$\sigma_{1}\leqslant a_{2}=\left(a_{1}^{2}-\left|\operatorname{det}\left(a_{1}^{2}-A^{H}A\right)\right|\left(\frac{n-1}{(n+1)a_{1}^{2}-\|A\|_{F}^{2}}\right)^{n-1}\right)^{1/2}\leqslant a_{1}$$

Suppose our conclusion holds for $k=m$, that is, $\sigma_{1}\leqslant a_{m+1}\leqslant a_{m}$. Now let’s consider the case of $k=m+1$. In equation (2), let $\lambda=a_{m+1}^{2}$, we can get

$$\sigma_{1}\leqslant a_{m+2}=\left(a_{m+1}^{2}-\left|\operatorname{det}\left(a_{m+1}^{2}-A^{H}A\right)\right|\left(\frac{n-1}{(n+1)a_{m+1}^{2}-\|A\|_{F}^{2}}\right)^{n-1}\right)^{1/2}\leqslant a_{m+1}$$

So $\sigma_{1}\leqslant a_{k+1}\leqslant a_{k}$. This proves that $\sigma_{1}\leqslant a_{k+1}\leqslant a_{k}(k=1,2,\cdots)$. From the monotone convergence theorem, $\lim_{k\to\infty}a_{k}$ exists. Let $\lim_{k\to\infty}a_{k}=\sigma$, then

$$\sigma=\left(\sigma^{2}-|\det(\sigma^{2}I_{n}-A^{H}A)|\left(\frac{n-1}{(n+1)\sigma^{2}-\|A\|_{F}^{2}}\right)^{n-1}\right)^{1/2}$$

We have

$$|\det(\sigma^{2}I_{n}-A^{H}A)|\left(\frac{n-1}{(n+1)\sigma^{2}-\|A\|_{F}^{2}}\right)^{n-1}=0.$$

because

$$\frac{n-1}{(n+1)\sigma^{2}-\|A\|_{F}^{2}}\not=0,$$

we can get

$$\det(\sigma^{2}I_{n}-A^{H}A)=0.$$

So $\sigma^{2}$ is the eigenvalue of $A^{H}A$. Because $\sigma_{1}^{2}$ is the largest eigenvalue of $A^{H}A$, there is $\sigma^{2}\leqslant\sigma_{1}^{2}$. According to the definition of $\sigma$, we have $\sigma\geqslant\sigma_{1}$, so $\sigma^{2}\geqslant\sigma_{1}^{2}$. We get $\sigma=\sigma_{1}$, so $\lim_{k\to+\infty}a_{k}=\sigma_{1}$. ∎