On two fundamental properties of the zeros of spectrograms of noisy signals
Arnaud Poinas, Rémi Bardenet · 2025
Read On two fundamental properties of the zeros of spectrograms of noisy signals by Arnaud Poinas, Rémi Bardenet (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-13829.txt.
Introduction
Arnaud Poinas and Rémi Bardenet explain patterns in spectrogram zeros in 2025. A spectrogram maps how a signal energy spreads over time and frequency. Its zeros form patterns that shift when a clean signal hides in white noise. Zeros tend to outline where the signal lives and can trap into shapes under interference.
Detection methods exploit those holes, but formal backing was thin. With toy signals, the authors show two classic math tools explain both effects. Zero density and a winding theorem do the heavy lifting. Signal strength and settings sway the patterns in ways the math predicts.
Catalog record
| CREATOR | Arnaud Poinas, Rémi Bardenet |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 17 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-13829
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
On two fundamental properties of the zeros of spectrograms of noisy signals Arnaud Poinas, Rémi Bardenet Abstract The spatial distribution of the zeros of the spectrogram is significantly altered when a signal is added to white Gaussian noise. The zeros tend to delineate the support of the signal, and deterministic structures form in the presence of interference, as if the zeros were trapped. While sophisticated methods have been proposed to detect signals as holes in the pattern of spectrogram zeros, few formal arguments have been made to support the delineation and trapping effects. Through detailed computations for simple toy signals, we show that two basic mathematical arguments –the intensity of zeros and Rouché’s theorem– allow discussing delineation and trapping, and the influence of parameters like the signal-to-noise ratio. In particular, interfering chirps, even nearly superimposed, yield an easy-to-detect deterministic structure among zeros. I Introduction The short-time Fourier transform (STFT) of a signal, and its squared modulus the spectrogram , are foundational stones of time-frequency analysis ( Coh95 ; Fla98 ; Gro01 ) . Classical approaches to time-frequency detection, denoising, or frequency estimation have relied on identifying regions where the spectrogram is large. Since the seminal letter of Flandrin Fla15 , zeros of the spectrogram have also received a lot of attention. Zeros of the spectrogram of white Gaussian noise spread very uniformly across the time-frequency (TF) plane, and Flandrin Fla15 observed that, in the presence of an additional deterministic signal, the zeros tend to delineate the support of the noiseless signal. Similarly, when one keeps the same realization of white Gaussian noise but increases the signal-to-noise ratio, zeros seem to become trapped in areas surrounded by emerging signal ridges; see Figure 1 . These two foundational claims on delineation and trapping, initially based on empirical observations, have become folklore in subsequent papers, justifying e.g. equating holes in the pattern of zeros to the support of signal components ( BaFlCh18 ) . In this letter, we give simple theoretical arguments to cement these claims, demonstrating delineation and trapping on classical examples of synthetic signals. In particular, we assess the effect of the signal-to-noise ratio and the relative amplitude of signal components. In Section II , we give the intensity of the zeros of the STFT of a noisy signal, and a statement of Rouché’s classical theorem in the context of spectrograms. The intensity illustrates the delineation effect, while Rouché’s theorem is the reason why (i) zeros appear to be trapped in valleys surrounded by signal, and (ii) the position of zeros cristallizes in the presence of interference. In Section III , we apply these two results to a Hermite function and a linear chirp with added noise, while we keep the case of two interfering parallel chirps for Section IV . In particular, we illustrate the delineation and trapping effects, as well as extreme trapping resulting in near-deterministic cristallization of the zeros. The case of parallel chirps of different amplitude also showcases a somewhat counterintuitive effect: informally, zeros that result from the interference between the chirps move to the same side of both chirps. This behavior is in stark contrast with that of the local maxima of the spectrogram, pointing to a potential advantage of zeros in the identification of interfering sources. Finally, we provide code 1 1 1 https://github.com/APoinas/spectrogram_zeros to reproduce all figures. (a) (b) Figure 1: Spectrogram of a Hermite function with and without noise; zeros are white dots, low values are in yellow, large ones in dark blue. Related work. The study of the zeros of the STFT of white noise was initiated by Fla15 , and their distribution characterized in ( BaFlCh18 ; BaHa19 ) , as well as HaKoRo22 for non-Gaussian windows; see PaBa25 for a tutorial survey. The numerical simulation of zeros is investigated in ( Escudero2021 ) , the distinction between zeros from white noise and zeros from signal interference in ( Meignen ; MACLM24 ) . Detection of signals based on holes can be found in several papers, from the initial Fla15 to more recent topological data analysis-based approaches MTMBG25Sub . We also refer to ( Ghosh ) for a study of level sets of random spectrograms. II Main results Let y:\mathbb{R}\mapsto\mathbb{C} be locally integrable, representing a measured signal. The spectrogram of y with Gaussian window g:t\mapsto 2^{1/4}e^{-\pi t^{2}} is the function \textnormal{Spec}_{g}({y}):z\mapsto\left|\int_{\mathbb{R}}{y}(t){g(t-\tau)}e^{-2i\pi\omega t}\mathrm{d}t\right|^{2} of the complex variable z=\tau+i\omega\in\mathbb{C} . The real and imaginary parts of z are respectively interpreted as time and frequency . The spectrogram can be rewritten as \textnormal{Spec}_{g}(y)(z)=e^{-\pi|z|^{2}}|B(y)(\bar{z})|^{2} , where B(y) is a holomorphic function called the Bargmann transform of y ; see Folland89 ; Gro01 . For the remainder of the paper, we consider y=x+\xi to be the sum of a deterministic signal x and a complex white Gaussian noise \xi ; see BaHa19 for a detailed mathematical definition of \xi . The zeros of \textnormal{Spec}_{g}(x+\xi) are a random configuration of points in \mathbb{C} , i.e. a point process . Our aim is to describe a few properties of that point process that relate to signal detection and reconstruction. II-A Explaining delineation: the intensity of zeros We first wish to explain the general tendency of zeros to avoid locations where the spectrogram of the noiseless signal x is large, while concentrating around that region. For that purpose we compute the intensity function \rho{:\mathbb{C}\rightarrow\mathbb{R}_{+}} of the zeros. Denoting by N(A) the number of zeros of the spectrogram in a bounded Borel set A\subset\mathbb{C} , the intensity function is defined as the density with respect to Lebesgue of the measure \mathbb{E}N(\cdot) , so that \mathbb{E}[N(A)]=\int_{A}\rho(z)\mathrm{d}z . Independently from our work and for a different purpose, an expression of \rho has already been established in ( Escudero2021 ) , as a function of the Bargmann transform of the noiseless signal. We rather use the spectrogram of the signal, which makes \rho easier to interpret. Proposition II.1 . Let z\in\mathbb{C} . Then \rho(z)=\left(1+\textnormal{Spec}_{g}(x)(z)+\frac{\Delta\textnormal{Spec}_{g}(x)(z)}{4\pi}\right)\\ \times\exp\left(-\textnormal{Spec}_{g}(x)(z)\right). (II.1) The proof is in Appendix . The exponential term in ( II.1 ) explains why zeros avoid locations where the spectrogram of the signal is large. Now, when \textnormal{Spec}_{g}(x)(z)\approx 0 , \rho(z) is approximately 1+\Delta\textnormal{Spec}_{g}(x)(z)/4\pi , while \Delta\textnormal{Spec}_{g}(x)(z)\geqslant 0 . Consequently, zeros tends to accumulate around the contours of the signal support, defined as regions where the spectrogram of the noiseless signal x is negligible but \Delta\textnormal{Spec}_{g}(x)(z) is not. Finally, in locations where \textnormal{Spec}_{g}(x)(z) is both negligible and has barely any variation, \rho(z)\approx 1 , which is the intensity of the zeros of the spectrogram of white noise only ( BaFlCh18 ) . II-B Explaining trapping: Rouché’s theorem One specific behavior of zeros of spectrogram of noisy signals that cannot be explained by the intensity function is their tendency to get "trapped". As can be observed in Figures 1 and 4 , when the spectrogram of a signal has a zero surrounded by high values then, when noise is added to the signal, the same zero will still be present at around the same location. This behavior is a direct consequence of Rouché’s théorem, see e.g. Rud87 and Section of the Appendix for a rephrasing with spectrograms. Informally, if \textnormal{Spec}_{g}(x) has large enough values on a closed curve C , then \textnormal{Spec}_{g}(\xi+x) is very likely to have