2
$\begingroup$

Suppose that $a_j,b_j \in \mathbb C$ are complex numbers, $j=1,\dots,n$, with the property that $|a_j|,|b_j| \geq c > d >0$ where $c,d$ are positive real numbers. I'm interested in the stability of the product $$ \frac{\prod_{j=1}^n a_j}{\prod_{j=1}^n b_j} $$ under noise in the following sense: Let $$ \tilde{a}_j = a_j + \delta_j $$ and $$ \tilde{b}_j = a_j + \gamma_j $$ where $\delta$ and $\gamma$ is $\ell^2$ noise such that $\| \delta \|_{\ell^2},\| \gamma \|_{\ell^2} \leq d$. The assumption on the $\ell^2$ norm to be smaller than $d$ implies that the quotient $$ \frac{\prod_{j=1}^n \tilde a_j}{\prod_{j=1}^n \tilde b_j} $$ is well-defined. The difference $$ \left | \frac{\prod_{j=1}^n \tilde a_j}{\prod_{j=1}^n \tilde b_j} - \frac{\prod_{j=1}^n a_j}{\prod_{j=1}^n b_j} \right | $$ tends to zero as the noise tends to zero (w.r.t the $\ell^2$-norm). I'm searching for a (possibly sharp) estimate on this difference in terms of some norm of the noise and $n$. Is there a way to get such inequalities? This question seems to be quite elementary so maybe there exists some literature on this?

$\endgroup$

1 Answer 1

1
$\begingroup$

Perhaps the simplest approach is to consider the problem as the calculation of a functional derivative. Let's put $$ (a_1, \ldots, a_n, b_1, \ldots, b_n)\triangleq (\mathbf{a}, \mathbf{b})\in\Bbb C^{2n} $$ and, in an analogous fashion, $$ (\delta_1, \ldots, \delta_n, \gamma_1, \ldots, \gamma_n)\triangleq (\boldsymbol{\delta}, \boldsymbol{\gamma})\in\Bbb C^{2n} $$ Then, defining the complex valued functional $$ \mathfrak{F}\big[(\mathbf{a}, \mathbf{b})\big]= \frac{\prod_{j=1}^n a_j}{\prod_{j=1}^n b_j} $$ it is natural to write $$ \mathfrak{F}\big[(\mathbf{a}, \mathbf{b})+\epsilon(\boldsymbol{\delta}, \boldsymbol{\gamma})\big]= \frac{\prod_{j=1}^n (a_j+\epsilon\delta_j)}{\prod_{j=1}^n (b_j+\epsilon \gamma_j)}\quad \forall \epsilon >0 $$ and calculating the functional derivative as usual $$ {\delta \mathfrak{F}}\big[(\mathbf{a}, \mathbf{b})\big](\boldsymbol{\delta}, \boldsymbol{\gamma}) \triangleq \left.\frac{\mathrm{d}}{\mathrm{d}\epsilon}\mathfrak{F}\big[(\mathbf{a}, \mathbf{b})+\epsilon(\boldsymbol{\delta}, \boldsymbol{\gamma})\big]\right|_{\epsilon=0} $$ This procedure occasionally shows up in problems of sensitivity analysis in signal processing. This functional derivative is linear respect to the variation $(\boldsymbol{\delta}, \boldsymbol{\gamma})$: if it is a Fréchet derivative, it is possibly relatively easy to obtain a nice estimate in the form $$ \left|\mathfrak{F}\big[(\mathbf{a}, \mathbf{b})+(\boldsymbol{\delta}, \boldsymbol{\gamma})\big]-\mathfrak{F}\big[(\mathbf{a}, \mathbf{b})\big]\right|\le c\left|{\delta \mathfrak{F}}\big[(\mathbf{a}, \mathbf{b})\big](\boldsymbol{\delta}, \boldsymbol{\gamma})\right|\label{1}\tag{☆} $$ with a sharp constant $c$. However, even if it is a Gâteaux derivative (i.e. it depends on the structure of the variation $(\boldsymbol{\delta}, \boldsymbol{\gamma})$) you may able to obtain a sharp estimate.

Addendum. Without going in the details and assuming the (plausible) non degeneracy hypothesis $\left|{\delta \mathfrak{F}}\big[(\mathbf{a}, \mathbf{b})\big](\boldsymbol{\delta}, \boldsymbol{\gamma})\right| >0$ when $(\boldsymbol{\delta}, \boldsymbol{\gamma})\neq 0$, further fairly elementary steps for possibly arriving at \eqref{1} are the following ones: $$ \begin{split} \left|\mathfrak{F}\big[(\mathbf{a}, \mathbf{b})+(\boldsymbol{\delta}, \boldsymbol{\gamma})\big]-\mathfrak{F}\big[(\mathbf{a}, \mathbf{b})\big]\right| &= \int\limits_0^1\frac{\mathrm{d}}{\mathrm{d}\epsilon}\mathfrak{F}\big[(\mathbf{a}, \mathbf{b})+\epsilon(\boldsymbol{\delta}, \boldsymbol{\gamma})\big]\mathrm{d}\epsilon \\ & \le \int\limits_0^1\left| \frac{\mathrm{d}}{\mathrm{d}\epsilon}\mathfrak{F}\big[(\mathbf{a}, \mathbf{b})+\epsilon(\boldsymbol{\delta}, \boldsymbol{\gamma})\big] \right| \mathrm{d}\epsilon \\ & = \left| {\delta \mathfrak{F}}\big[(\mathbf{a}, \mathbf{b})\big](\boldsymbol{\delta}, \boldsymbol{\gamma})\right|\int\limits_0^1\left| \frac{\frac{\mathrm{d}}{\mathrm{d}\epsilon}\mathfrak{F}\big[(\mathbf{a}, \mathbf{b})+\epsilon(\boldsymbol{\delta}, \boldsymbol{\gamma})\big]}{\left.\frac{\mathrm{d}}{\mathrm{d}\epsilon}\mathfrak{F}\big[(\mathbf{a}, \mathbf{b})+\epsilon(\boldsymbol{\delta}, \boldsymbol{\gamma})\big]\right|_{\epsilon=0}} \right| \mathrm{d}\epsilon \\ & \le c\left|{\delta \mathfrak{F}}\big[(\mathbf{a}, \mathbf{b})\big](\boldsymbol{\delta}, \boldsymbol{\gamma})\right| \end{split} $$ and therefore the relation $$ \int\limits_0^1\left| \frac{\frac{\mathrm{d}}{\mathrm{d}\epsilon}\mathfrak{F}\big[(\mathbf{a}, \mathbf{b})+\epsilon(\boldsymbol{\delta}, \boldsymbol{\gamma})\big]}{\left.\frac{\mathrm{d}}{\mathrm{d}\epsilon}\mathfrak{F}\big[(\mathbf{a}, \mathbf{b})+\epsilon(\boldsymbol{\delta}, \boldsymbol{\gamma})\big]\right|_{\epsilon=0}} \right| \mathrm{d}\epsilon\le c \label{2}\tag{☆☆} $$ may possibly be a first guess for the sought for constant. Note that the integral on the right side of \eqref{2} can be difficult to estimate: however, despite requiring possibly tedious calculations, you can made explicit its form.

$\endgroup$
2
  • 1
    $\begingroup$ Hallo, thank you for your interesting answer! I'm not very familiar with functional derivatives so I have a question regarding this approach: As I understand it the last inequality holds for every function $\mathfrak{F}$ right? How exactly can one show that the last inequality holds and how can I calculate the constant $c$ dependent on the noise $(\delta, \gamma)$. Is there a general formula for $c$? $\endgroup$
    – Muzi
    Mar 19, 2021 at 21:26
  • $\begingroup$ Hi @Muzi, and thank for your interest in my answer. It is just an intuition and I sketched my approach only very briefly, since I cannot try a full development of the calculations now: however, I will add some more details on the route I'd take to work out an estimate. $\endgroup$ Mar 20, 2021 at 9:09

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.