# Staircase spectrum: is there a known solution for this problem?

• Let $A$ be a (real) dense (non-sparse) matrix with size $rp\times M$ ($M < rp$) with orthonormal columns (i.e. $A^TA = I_M$).

• Similarly, let $B$ be a dense matrix with size $rq\times N$ ($N < rq$) with orthonormal columns.

• Let $V=\bigoplus_{i=1}^qA.$ Equivalently, $V=I_q\bigotimes A,$ where $\bigoplus$ denotes the direct sum, and $\bigotimes$ denotes the Kronecker product.

• Let $W=\bigoplus_{j=1}^pB.$ Equivalently, $W=I_p\bigotimes B.$

• Let $R=V^TW,$ and therefore, $$R=\big(\bigoplus_{i=1}^qA^T\big)\big(\bigoplus_{j=1}^pB\big)$$ $$=\big(I_q\bigotimes A^T\big)\big(I_p\bigotimes B\big)$$

• Assume: $rpq > qM + pN$ (so that the augmented matrix $\left[\begin{array}{c|c}V & W\end{array}\right]$ is a tall matrix).

• Note: $p\neq q$, $p>1$ and $q>1$.

Prove that the singular values of $R$ come in $n$-tuples, with $$n=\text{GCD}(p,q),$$ where $\text{GCD}(p,q)$ is the greatest common divisor of $p$ and $q$. In other words, prove that a singular value of $R$ has a multiplicity of $\text{GCD}(p,q)$.

Note: I am ONLY interested in proving that the largest singular value of $R$ is distinct when $p$ and $q$ are relatively prime ($\text{GCD}(p,q) = 1$). However, I think giving the larger picture of the behavior of the singular values of $R$ is far more interesting, and probably even helpful in proving the specific case I'm interested in. Please refer to the Matlab code below to verify that $n=\text{GCD}(p,q)$.

Matlab code:

clear all; close all;

r = 3; % integer >= 1

p = 12; q = 16; %17 % for the singular values of R to be distinct, p and q must be relatively prime integers and larger than 1 (i.e. GCD(p,q) = 1 and p > 1 and q > 1)

M = 8; % # of columns of A
N = 14; % # of columns of B. M and N MUST satisfy 1 < M < rp, 1 < N < rq AND they must satisfy assumption that r*p*q/(q*M+p*N) > 1.

[A,~] = qr(rand(r*p,M),0); % size = rp X M (M < rp, A'A = I) % Note: you can orthonormalize (e.g. via Gram Schmidt or SVD) any arbitrary
% (full rank) data matrix other than noise, I'm only using
% noise as a data matrix because it's conevniently easy to generate.
[B,~] = qr(rand(r*q,N),0); % size = rq X N (N < rq, B'B = I)

V = kron(eye(q),A); % size = q(rp) X qM
W = kron(eye(p),B); % size = p(rq) X pN

R = V'*W; % i.e. R = direct sum (over q) of (A') X direct sum (over p) of (B)

% assumption
r*p*q/(q*M+p*N) % check is greater than 1 % This assumption is required for the augmented matrix [V W] to be tall.
gcd(p,q) % check for relative primeness of p and q. % This assumption is required for the singular values of R to be distinct.

figure;stem(svd(R));
title 'note how the singular values of R come in GCD(p,q)-tuples'


## 1 Answer

Let me write \begin{align} V &= I_q \otimes A = \text{blkdiag}(A, ..., A), \\ W &= I_p \otimes B = \text{blkdiag}(B, ..., B). \end{align} We can write $$V^TW = I_q \otimes (A^TB)=\text{blkdiag}(A^TB,...,A^TB)$$ only if $p=q$. But if $p \neq q$, we can still do the following: $$R=V^TW = I_n \otimes (V_{\text{sub}}^TW_{\text{sub}})=\text{blkdiag}(V_{\text{sub}}^TW_{\text{sub}},...,V_{\text{sub}}^TW_{\text{sub}}),$$ with \begin{align} n &= \text{GCD}(p, q), \\ V_{\text{sub}} &= I_{q/n} \otimes A = \text{blkdiag}(A,...,A), \\ W_{\text{sub}} &= I_{p/n} \otimes B = \text{blkdiag}(B,...,B). \end{align} So in the end, it turns out that $R$ is a block diagonal matrix with $n=\text{GCD}(p,q)$ identical matrices. To compute the singular values of a block diagonal matrix, it suffices to compute the singular values of the matrices. Hence, the singular values of $R$ are simply $n$ times the singular values of $V_{\text{sub}}^TW_{\text{sub}}$.

Note that the fact that $A$ and $B$ are orthogonal is not used for this derivation. It also holds in case $A$ and $B$ are non-orthogonal.

• First of all, thank you very much for your answer. But did you try to image $R$? For example, using the Matlab code above, add figure;imagesc(R); $R$ is not even square for it to be possibly a block-diagonal matrix. Yes, it does contain a submatrix (made up of 3 almost-diagonal blocks) repeated $n$ times, but overall $R$ is not block-diagonal. Looking forward for your feedback. (btw, what is $k$?). – seeker Nov 9 '17 at 10:15
• Yes I did. Actually, it was the first thing I did. I do not really see the problem. I think I nowhere mention that R had to be square – EdG Nov 9 '17 at 10:23
• Also, note that a diagonal matrix has to be square, BUT a block diagonal matrix does not need to be square – EdG Nov 9 '17 at 10:25
• Oh sorry for $k$. That is supposed to be $n$. I will fix that – EdG Nov 9 '17 at 10:27
• Ok, but do you agree that the submatrix: $V_{sub}^{T} W_{sub}$ is not block diagonal? – seeker Nov 9 '17 at 10:43