4
$\begingroup$

I'm reading the paper by ANGELIKA ROHDE AND ALEXANDRE B. TSYBAKOV, ESTIMATION OF HIGH-DIMENSIONAL LOW-RANK MATRICES.

And in the paper, they provide an inequation of the Schatten-p (quasi-)norm, namely, for any tow matrices $A,B \in \mathbb{R}^{m\times T}$, $\forall \ 0<p\le1$, we have

$$ \lVert A+B \rVert_{S_p}^p \le \lVert A \rVert_{S_p}^p + \lVert B \rVert_{S_p}^p $$

The$\lVert \cdot \rVert_{S_p}$means the Schatten-p norm, $$ \lVert A \rVert_{S_p}=\left(\sum_{j=1}^{\min\{m,T\}}\sigma_j(A)^p\right)^{1/p} $$ And $\sigma_j(A)$ means the singular value of $A$.

The authors say that we could find the result from MCCARTHY, C. A. (1967). $C_p$ or ROTFELD, S. Y. (1969). The singular numbers of the sum of completely continuous operators. But both papers are about the some compact operators on $\textbf{a}$ Hilbert space, from my perspective, the results from the both papers above can show the inequality is right when $A,B$ are square matrices, but how can it hold for the rectangular matrices?

I'm very puzzled, and hope for the answer sincerely.

THANK YOU!

$\endgroup$
1
  • $\begingroup$ Crossposted: math.stackexchange.com/q/1234163/166535 (Lin Yinan: It is generally considered bad form to post the same question here and at MSE simultaneously. If you don't get any answers for a couple of days, then you can crosspost, but do give a link to the other post.) $\endgroup$ – Joonas Ilmavirta Apr 14 '15 at 12:32
3
$\begingroup$

One can deduce the non-rectangular case from the rectangular one as follows. Let $$P\colon \mathbb{R}^{m+T}\to \mathbb{R}^T,\, \iota\colon \mathbb{R}^m\to \mathbb{R}^{m+T}$$ be the obvious orthogonal projection and the isometric imbedding respectively. Then for any linear operator $F\colon \mathbb{R}^T\to \mathbb{R}^m$ one has equality of the Shatten norms $$||F||_{S_p}=||\iota\circ F\circ P||_{S_p}.$$ Hence replace your operators $A$ and $B$ with their compositions with $\iota$ and $P$ and apply to the compositions the result from the references.

$\endgroup$
2
  • $\begingroup$ Schatten $p$ quasi norm does not satisfy triangle inequality. However the inequality here is triangle. What do I miss? Is the Schatten norm convex for all $p>0$? $\endgroup$ – 1.. Sep 24 '18 at 3:16
  • $\begingroup$ @1.. You missed the "^p" at each summand. The normal triangle inequality is indeed not satisfied. $\endgroup$ – Paul Pfeiffer Oct 20 '20 at 15:47

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.