⚡ electrical engineering

On two fundamental properties of the zeros of spectrograms of noisy signals

This paper provides formal mathematical arguments using zero intensity and Rouché's theorem to explain how the zeros of noisy spectrograms delineate signal support and form deterministic structures in the presence of interference, thereby validating their utility for signal detection.

Original authors: Arnaud Poinas, Rémi Bardenet

Published 2026-04-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Arnaud Poinas, Rémi Bardenet

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find a specific shape drawn on a foggy window.

Normally, if you look at a window covered in random raindrops (noise), the drops are scattered everywhere. But if someone draws a bright, clear picture (the signal) on the glass, something magical happens: the raindrops seem to avoid the bright parts of the drawing. Instead, they gather around the edges, outlining the shape like a glowing border.

This paper by Arnaud Poinas and Rémi Bardenet is about understanding exactly why this happens and how we can use it to find hidden signals in a world full of noise. They study something called a spectrogram, which is just a fancy map that shows how a sound changes over time and frequency (like a musical score, but for any kind of signal).

Here is the breakdown of their discovery using simple analogies:

1. The "Hole" in the Rain (The Delineation Effect)

Imagine the spectrogram as a field of grass.

The Noise: If there is no signal, just random static, the "zeros" (points where the signal strength is exactly zero) are like dandelions growing randomly all over the field. They are everywhere, evenly spaced.
The Signal: Now, imagine a heavy stone (the signal) is placed on the grass. The grass grows tall and thick under the stone.
The Result: The dandelions (the zeros) cannot grow under the heavy stone because the grass is too thick there. Instead, they crowd around the edge of the stone.

The authors prove mathematically that when you add a signal to noise, the zeros don't just disappear; they get pushed away from the loud parts of the signal and cluster around the boundary of the signal. This creates a perfect outline of the signal, even if the signal itself is buried in noise.

2. The "Valley" Trap (The Trapping Effect)

Sometimes, a signal has a dip or a valley in the middle (like a "hole" in the sound).

The Analogy: Imagine a valley surrounded by high mountains. If you drop a ball (a zero) into that valley, it gets stuck there. Even if you shake the ground (add noise), the ball stays in the valley because the mountains are too high for it to roll out.
The Math: The authors use a famous math rule called Rouché's Theorem to explain this. It basically says: "If the signal is loud enough on the rim of a circle, the number of zeros inside that circle won't change, no matter how much noise you add."
Why it matters: This means that if you see a zero in a specific spot in a noisy recording, you can be very confident that a "hole" or a specific feature of the signal exists there. The zero is "trapped" by the signal.

3. The "Ghost" of Interference (Two Signals Colliding)

The most interesting part of the paper happens when two signals overlap, like two people talking at the same time.

The Scenario: Imagine two parallel lines of sound (chirps) running close to each other.
The Surprise: When these two signals interfere, they create a new, invisible pattern. The zeros don't just sit in the middle; they form a deterministic line (a predictable, straight line) between the two signals.
The "Push": If one signal is louder than the other, the line of zeros gets pushed toward the quieter signal. It's like a tug-of-war where the zeros act as the rope, and the louder signal pulls them to its side.
The Advantage: This is huge for detection. If you look at the "loud spots" (maxima) of the sound, two close signals often merge into one big blob, and you can't tell there are two of them. But the zeros remain distinct and form a clear line, acting like a fingerprint that says, "Hey, there are actually two signals here, not one!"

Why Should We Care?

In the real world, we are constantly bombarded by noise (static, background chatter, interference).

Old Way: We tried to find signals by looking for the loudest parts. This is like trying to find a specific person in a crowd by looking for the tallest head. If two people stand close, you can't tell them apart.
New Way (This Paper): We can now look for the "holes" or the "empty spots" in the noise. Because the zeros naturally outline the signal and get trapped in specific patterns, they are actually better at revealing the true structure of complex, overlapping signals than the loud parts are.

In a nutshell: The authors showed us that the "empty spaces" in a noisy signal aren't random. They are actually a highly organized map that outlines the signal, traps its features, and reveals when multiple signals are fighting for space. By studying these empty spots, we can see through the noise much more clearly.

1. Problem Statement

Time-frequency analysis often relies on the Short-Time Fourier Transform (STFT) and its squared modulus, the spectrogram. While classical methods focus on regions of high energy (local maxima) for signal detection, recent empirical observations suggest that the zeros (roots) of the spectrogram contain significant structural information.

Two specific phenomena have been observed but lacked rigorous theoretical justification:

Delineation: When a deterministic signal is added to white Gaussian noise, the zeros of the noisy spectrogram tend to avoid the signal's support (where the signal energy is high) and instead accumulate around the boundaries of the signal, effectively "delineating" its shape.
Trapping: In the presence of interference (e.g., overlapping chirps), zeros appear to be "trapped" in specific locations (valleys between signal ridges), forming deterministic structures even as noise is added.

Existing literature has treated these as folklore or relied on topological data analysis without providing formal mathematical arguments linking these behaviors to signal parameters like the Signal-to-Noise Ratio (SNR). This paper aims to provide these formal arguments.

2. Methodology

The authors employ a probabilistic and complex-analytic approach to model the spectrogram of a signal $y = x + \xi$ , where $x$ is a deterministic signal and $\xi$ is complex white Gaussian noise.

Bargmann Transform: The spectrogram is rewritten using the Bargmann transform $B(y)$ , a holomorphic function. The zeros of the spectrogram correspond to the zeros of $B(x) + B(\xi)$ .
Point Process Theory: The zeros are treated as a random point process. The authors utilize the Kac-Rice formula to compute the intensity function (density) of these zeros.
Complex Analysis: To explain the "trapping" phenomenon, the authors apply Rouché's Theorem, which relates the number of zeros of a sum of functions to the number of zeros of the dominant function within a closed contour.
Specific Signal Models: The theory is applied to three specific cases:
1. Hermite functions (representing localized pulses).
2. Linear chirps (frequency-modulated signals).
3. Superpositions of two parallel linear chirps (interference scenarios).

3. Key Contributions

A. Theoretical Framework for Delineation (Intensity of Zeros)

The paper derives an explicit formula for the intensity $\rho(z)$ of the zeros of the noisy spectrogram:
$\rho(z) = \left( 1 + \text{Spec}_g(x)(z) + \frac{\Delta \text{Spec}_g(x)(z)}{4\pi} \right) \exp(-\text{Spec}_g(x)(z))$

Mechanism: The exponential term $\exp(-\text{Spec}_g(x)(z))$ dictates that zeros are exponentially suppressed in regions where the signal energy is high.
Result: Zeros accumulate where the signal energy is low but the Laplacian (curvature) is non-zero. This mathematically proves that zeros naturally avoid the signal's interior and concentrate on its contours, explaining the delineation effect.

B. Theoretical Framework for Trapping (Rouché's Theorem)

The authors prove that if the noiseless signal's spectrogram dominates the noise spectrogram on a closed curve $C$ , the number of zeros inside $C$ remains invariant.

Mechanism: By applying Rouché's theorem to the holomorphic functions $B(x)$ and $B(\xi)$ , they show that if $\text{Spec}_g(x) > \text{Spec}_g(\xi)$ on a boundary, the zeros of the noisy signal are "trapped" in the same locations as the zeros of the noiseless signal.
Result: This explains why zeros in interference patterns (valleys between ridges) remain stable and deterministic even in the presence of noise, provided the SNR is sufficiently high.

C. Analysis of Interference (Parallel Chirps)

A novel contribution is the analysis of two interfering parallel chirps.

Crystallization: The authors show that zeros form a deterministic line parallel to the chirp axes.
Counter-intuitive Behavior: When the amplitudes of the two chirps differ, the line of zeros shifts toward the weaker signal. This contrasts with local maxima, which tend to merge into a single ridge. This suggests zeros are superior for distinguishing overlapping sources.

4. Key Results

Hermite Functions:
- For a $k$ -th order Hermite function, the average number of zeros inside a specific radius $R = \sqrt{k/\pi}$ is exactly $k$ , independent of the SNR scaling factor $\gamma$ .
- The probability of observing exactly $k$ zeros in this region approaches 1 exponentially fast as the SNR increases.
Single Linear Chirp:
- Zeros are absent along the chirp's axis (where energy is high).
- Two parallel bands of high zero density form on either side of the chirp axis. The distance between these bands increases with the SNR.
Two Interfering Chirps:
- Noiseless Case: Zeros lie on a line defined by $r = \frac{a}{2} - \frac{\log(\gamma_2/\gamma_1)}{4a\pi}$ . If amplitudes are equal, the line is centered; if unequal, it shifts toward the weaker chirp.
- Noisy Case: Under specific conditions on the distance between chirps and the SNR, the probability of finding exactly one zero trapped in each "valley" rectangle approaches 1.
- Detection Advantage: Even when the two chirps are so close that their local maxima merge into a single indistinguishable strip, the zeros remain distinct and separated, offering a robust method for detecting multiple sources.

5. Significance and Implications

Theoretical Validation: The paper moves the study of spectrogram zeros from empirical observation to rigorous mathematical proof, validating the "delineation" and "trapping" hypotheses.
Signal Detection: It suggests that analyzing the holes (regions devoid of zeros) or the trapped zeros in a spectrogram is a more robust method for detecting and separating overlapping signals than analyzing local maxima, especially in low-SNR or high-interference scenarios.
Parameter Estimation: The derived intensity formulas link the spatial distribution of zeros directly to signal parameters (amplitude, frequency slope, separation). This opens the door for new statistical estimators for signal reconstruction based on zero locations.
Limitations: The current framework relies heavily on Gaussian noise properties (Kac-Rice formula, Gaussian tail inequalities). Extending these results to non-Gaussian noise or non-analytic windows remains a challenge for future work.

In summary, the paper establishes that the zeros of a noisy spectrogram are not random artifacts but structured features that encode the geometry and interference patterns of the underlying signal, providing a powerful new tool for time-frequency analysis.