This question is in part related to a question that I have already posed.

Say I have two symmetric positive definite matrices and their respective Cholesky decompositions $\mathbf{A} = \mathbf{L}_A \mathbf{L}_A^{\top}$ and $\mathbf{B} = \mathbf{L}_B \mathbf{L}_B^{\top}$, where $\mathbf{L}_A$ and $\mathbf{L}_B$ have positive diagonal entries. Furthermore, let $\sqrt{\mathbf{A}}$ and $\sqrt{\mathbf{B}}$ denote the unique symmetric positive definite square roots of $\mathbf{A}$ and $\mathbf{B}$, respectively, i.e., $\sqrt{\mathbf{A}}^{\top} = \sqrt{\mathbf{A}}$ and $\sqrt{\mathbf{A}}\sqrt{\mathbf{A}} = \mathbf{A}$.

I would like to know whether there is an inequality relating the operator norms of $\mathbf{L}_A - \mathbf{L}_B$ and $\sqrt{\mathbf{A}}- \sqrt{\mathbf{B}}$ up to a positive constant $C>0$, i.e., \begin{equation} \max_{\mathbf{x}} \frac{\Vert (\mathbf{L}_A - \mathbf{L}_B)\mathbf{x}\Vert}{\Vert\mathbf{x}\Vert} \leq C \max_{\mathbf{x}} \frac{\Vert(\sqrt{\mathbf{A}} - \sqrt{\mathbf{B}})\mathbf{x}\Vert}{\Vert\mathbf{x}\Vert}. \end{equation}

Here $\Vert \cdot \Vert$ denotes the Euclidean norm.

Edit: As pointed out by Denis Serre, the inequality does not hold in general. However, I can also include the assumption that the entries of $\mathbf{A}$ and $\mathbf{B}$ are bounded (and therefore also their square-roots and Cholesky decompositions). From Denis' answer, it seems to me that Lipschitz continuity from $\sqrt{\mathbf{A}}$ to $\mathbf{L}_A$ would hold, since the derivative of $\mathbf{L}_A$ with respect to the entries of $\sqrt{\mathbf{A}}$ cannot become unbounded.

  • $\begingroup$ Thanks for the suggestion. Equality does not hold in general. I have realized that I actually require inequality instead of equality, as specified in the new edit. $\endgroup$
    – Heinrich A
    Apr 19 '21 at 11:56

The answer is negative, and this happens as soon as $n=2$. The question is whether the composition $X\mapsto L:=L_{X^2}$ is globally Lipschitz over ${\bf SPD}_n$. Let $x_j\in{\mathbb R}^n$ denote the $j$th column of $X$. Then we have the following formulae $$\ell_{11}=\|x_1\|,\quad \ell_{j1}=\frac{\langle x_1,x_j\rangle}{\|x_1\|}\,,$$ and so on, the expression getting more and more complicated as the column index increases.

The first entry $l_{11}$ is obviously globally Lipschitz. But this is not the case for the next ones. We have $$d\ell_{j1}=\frac{x_1}{\|x_1\|}\cdot dx_j+x_j\cdot d\frac{x_1}{\|x_1\|}.$$ Because the entries $x_{kj}$ are independent from $x_1$ if $k\ge2$, the global Lipschitz property can hold only if $$d\frac{x_1}{\|x_1\|}\parallel \vec e_1,$$ which is not true in general.

Edit. The obstacle discussed above occurs at infinity, because there is no a priori bound of $x_{22}$, hence $d\ell_{21}$ is not uniformly bounded as $x_{22}\rightarrow+\infty$. But because $X\mapsto L$ is homogeneous of degree $1$, this has the counterpart that the Lipschitz property also fails near the origin. Actually the norm of $$d\frac{x_1}{\|x_1\|}=\frac1{\|x_1\|}\left(I_n-\frac{x_1\otimes x_1}{\|x_1\|^2}\right)dx_1$$ is $1/\|x_1\|$, which can be arbitrarily large. Thus even if we impose an a priori bound on $X$, the map $X\mapsto L$ is not Lipschitzian.

  • $\begingroup$ Thanks very much for the helpful answer! I was wondering whether the composition would be globally Lipschitz if I werer to introduce the assumption that the columns entries of the matrices $\mathbf{A}$, $\mathbf{B}$ (and therefore also their square roots and Cholesky decompositions) are bounded. $\endgroup$
    – Heinrich A
    Apr 19 '21 at 15:07
  • 1
    $\begingroup$ @HeinrichA. See my edit: even if we impose an a priori bound, the map $X\mapsto L$ is not Lipschitz. $\endgroup$ Apr 20 '21 at 6:44

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.