Thanks to the discussionAs correctly pointed by Anthony in the comments (Anthony's argument), we already know that this holds for smooth functions $f(z)=\sum a_nz^n$ (equivalentlybelow, it holds for rapidly decaying $a_n$'s)the argument I attempted here originally did not hold water.
So it now suffices to show that an arbitrary $f(e^{ix})=e^{ig(x)}\in H^{1/2}$ can be approximated in $H^{1/2}$ by smooth functions $e^{ih_k(x)}$: we thenThese issues apparently have thatbeen studied quite extensively recently. See for example $\sum n|a_n|^2 = \lim_{k\to\infty} \sum n|a_n^{(k)}|^2$J. Bourgain, One cannot hear the winding number. In particular, and these latter sums are integers.
To doin the introduction to this paper, we recall that $H^{1/2}$ functions (in dimension $1$) can be characterized as those $L^2$ functions $u$Bourgain mentions that also satisfy $$ \int\!\!\!\int dxdy\, \frac{|u(y)-u(x)|^2}{|x-y|^2} < \infty . \quad\quad\quad\quad (1) $$ The $L^2$ condition is automatic here since our functions satisfy $|f|=1$, and (1) for $u=e^{ig}$ implies the same condition foranswer to your question is $g$ after some precautions associated withyes; the non-uniqueness of $g$. More specifically, let's assume that $|u^{-1}(-1)|=0$, and let's then take $|g(x)|<\pi$. Then $|u(y)-u(x)|$result is comparableattributed to $|g(x)-g(y)|$ on every set $A_n=\{ x: |g(x)\mp\pi|\ge 1/n \}$, and these exhaust our domain.
Thus $g\in H^{1/2}$ also, as claimed,Boutet de Monvel and if we now approximate $h_k\to g$ in $H^{1/2}$, then $e^{ih_k}\to e^{ig}$ also, by a similar argument: (1) for $u=e^{ih_k}-e^{ig}$ together with $\|u\|_2$ essentially computes the $H^{1/2}$ norm, but this is controlled by (1) for $u=h_k-g$Gabber.