1
$\begingroup$

Let $\theta$ and $f$ be two real analytic non-constant functions defined on $[0,2\pi]$. For simplicity we assume $f$ has just two critical values $m<M$ (in the picture $-1$ and $1$); we index as $\{f_j\}_{j \in\{1,2,...,N\}}$ its invertible branches. Our hypotesis is that $\theta$ and $f$ satisfy $\forall \;a\in(m,M)$ $$ \sum_j \frac{\text{e}^{i \theta(f_j^{-1}(a))}}{|f'(f_j^{-1}(a))|}=0. $$ This is a balancing condition for 2-dimensional vectors represented with complex numbers and involving all points in a given level set (which is finite since $f$ is analytic).


EDIT. As pointed out in the comments it was probably unclear that the hypotesis is about a pair of functions $\theta$ and $f$ that satisfy the condition above. I am not claiming this relation holds for any $\theta$. Such pair of functions can for example be constructed by considering a closed regular curve on $[0,2\pi]$ of the form $$ \gamma(t)=\sum_{j=1}^N a_j \sin(j\cdot t)+i\sum_{j=1}^N b_j \cos(j\cdot t), $$ and observing that for example $\int_0^{2\pi}\gamma(t) \Bigl(\cos((N+1)\cdot t)\Bigl)^n=0$, $\forall n$. Reparametrizing by arc-length $t=v(s)$ and denoting with $\theta(s)$ the turning angle of $\gamma$ we get $$ \int \text{e}^{i\theta(s)} \Bigl(\cos((N+1)\cdot v(s))\Bigl)^n=0. $$ From here one can show that if such property holds for $\cos((N+1)\cdot v(s))$ then it must hold for any composition of this function with $g \in L^1$. Picking $g$ as the characteristic function of $[a,a+\delta]$ and derivating in $\delta$, we get the equality above for $f=\cos((N+1)\cdot v(s))$. You can read more about that in this other question of mine:

Given $\theta$, find $f$ such that $\int_{\mathbb{T}} \text{e}^{i\theta} \cos(h \cdot f) = 0,$ for all $h \in \mathbb{N}$.


function f, only 2 critical values

We observe

$$ 0=\lim_{a \to m^+} \sum_j \frac{\text{e}^{i \theta(f_j^{-1}(a))}}{|f'(f_j^{-1}(a))|} = - 2 \lim_{a \to m^+} \sum_{j \text{ odd }} \frac{\text{e}^{i \theta(f_j^{-1}(a))}}{f'(f_j^{-1}(a))}, $$ where the second equality follows from the analiticity of $f$ at inner points. Similarly, derivating in $a$, one obtains $\forall \; n \in \mathbb{N}$ $$ \lim_{a \to m^+} \frac{d^n}{da^n} \sum_{j \text{ odd }} \frac{\text{e}^{i \theta(f_j^{-1}(a))}}{f'(f_j^{-1}(a))}=0. $$

I would like to prove from these hypoteses that also the sum over odd indices is constantly $0$, that is $\forall \;a\in(m,M)$. $$ \sum_{j \text{ odd }} \frac{\text{e}^{i \theta(f_j^{-1}(a))}}{f'(f_j^{-1}(a))}=0. $$

Note that the fact that all the derivatives are $0$ in the limit is not enough. We could theoretically have something like $\text{e}^{-1/x}$, whose derivatives go to $0$ when $x$ approches $0$, but the function is not constant. Nevertheless, I have experimental evidence that this does not happen in our setting. I tried to exploit the fact that here we are dealing with a single analytic function and that its global behaviour is determined by its Taylor expansion at any point but I could not conclude the implication this question is about.


EDIT 2. Calling our function $F(a):=\sum_{j \text{ odd }} \frac{\text{e}^{i \theta(f_j^{-1}(a))}}{f'(f_j^{-1}(a))}$, proving that it is constantly $0$ is equivalent to prove the existence of $C_\varepsilon>0$ such that $$ \| F^{(n)}(a) \| \leq C_\varepsilon^{n+1} n!, \; \forall n \in \mathbb{N}, \; a \in (m,m+\varepsilon). $$ In this case, the series $0 \equiv \sum F^{(n)}(a) \frac{a^n}{n!}$ would converge uniformly to $F(s)$ in a neighbohood of $0$, providing the desired identity. This implication is explained for example in One-Sided Analyticity Condition Guarantees Analytic Function?.

Although this approach looks reasonable I am still stuck. Does anyone have some hints?


The question is pretty specific but I guess an answer in the positive could connect to properties of analytic functions interesting also from a more general perspective. Thanks.

$\endgroup$
  • 2
    $\begingroup$ Something is crazy as written: if $\theta$ is an arbitrary analytic function, you can just interpolate it to be anything you want at the evaluation points, which will definitely destroy the suggested identity. $\endgroup$ – fedja Jan 27 at 16:22
  • $\begingroup$ @fedja $\theta$ and $f$ are arbitrary as long as they satisfy the condition on the level sets on top, which must hold for all the a. They are a specific pair of functions, which satisfy the given constraint. Is that unclear? $\endgroup$ – Leonardo Jan 27 at 17:01
  • $\begingroup$ Question has been edited. Hope now it makes more sense. Thanks for pointing that out. $\endgroup$ – Leonardo Jan 29 at 15:22
1
+50
$\begingroup$

Something is still strange. Let's take $f=\cos(3x)$. Then the "odd" level set is $t,t+\frac{2\pi}3,t+\frac{4\pi}3$ and the "even" one is the same with $2\pi-$ everywhere for not too large $t$. The derivatives at the pre-images are all the same in absolute value, so we are just asking if the function $g(t)=e^{i\theta(t)}+e^{i\theta(t+\frac{2\pi}3)}+e^{i\theta(t+\frac{2\pi}3)}$ can satisfy $g(t)+g(2\pi-t)=0$ without being identically $0$. Let $\theta(t)=t+\psi(t)$ where $\psi$ is real-analytic, $2\pi$-periodic. We are interested in the Fourier coefficients of $e^{i\theta(t)}$ with indices divisible by $3$. It is enough to obtain some non-trivial sequence satisfying $a_{-k}=-a_k$ for $3\mid k$. This means that we should get some prescribed sequence for the Fourier coefficients of $e^{i\psi(t)}$ with indices $k\equiv -1\mod 3$. The restriction that $\psi$ must be real relates only the pairs of indices $m$ and $-m$, so we have no relation for the indices in our set and the differential (in any reasonable Banach algebra of real-analytic $2\pi$-periodic functions) of the mapping $\psi(\cdot)\mapsto e^{i\psi(\cdot)}$ is just $i$ times the identity, so as a mapping from real functions to the sequences of Fourier coefficients with indices congruent to $-1$ modulo 3 is onto with bounds, which means that we can get any sequence that is sufficiently small, thus refuting your conjecture. Am I missing anything?

$\endgroup$
  • $\begingroup$ Thanks for your answer. I guess you are on point. It is just not completely apparent to me how one handles the condition $\widehat{\psi}(m)=\overline{\widehat{\psi}(-m)}$ when it comes to construct $\text{e}^{i\psi}$ with the desired coefficients using your invertibility result. Your answer in the negative is a bit sad for me ;) since I hoped to use my equality to prove certain periodicity properties of $\theta$ and $f$, in particular that, under the sum condition on level sets, if $\text{e}^{i\theta}$ is $2\pi$-periodic then also $f$ must be. Do you have maybe an intuition about that? $\endgroup$ – Leonardo Feb 3 at 13:05
  • 1
    $\begingroup$ @Leonardo "It is just not completely apparent to me how one handles the condition..." You just force it at every iteration step. If you want to solve $\widehat{e^{i\psi}}(k)=a_k$ for $k\equiv -1\mod 3$ and $\psi_m$ is an approximate solution giving you discrepancies $b_k$, then you just add a real-valued function whose Fourier coefficients with the indices $k$ you are interested in are $-ib_k$ and the coefficients with indices $-k$ are $i\bar b_k$. This function adds just a little to the norm but decreases the discrepancy fixed number of times, so you have a geometric convergence. $\endgroup$ – fedja Feb 3 at 14:21

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.