⚡ electrical engineering

A Scaling Law for Bandwidth Under Quantization

This paper derives a scaling law demonstrating that for signals with $1/f^\alpha$ power spectra, each additional ADC bit extends the effective bandwidth by a factor of $2^{2/\alpha}$ , a relationship validated on synthetic and real EEG data with prediction errors under 3% when using theoretical noise floor estimates.

Original authors: Maximilian Kalcher, Tena Dubcek

Published 2026-02-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Maximilian Kalcher, Tena Dubcek

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

The Big Idea: Trading "Precision" for "Reach"

Imagine you are trying to listen to a symphony orchestra playing in a large hall. The music has a special quality: the low notes (like the deep drums) are very loud and clear, but the high notes (like the tiny cymbals) get quieter and quieter as they go up the scale. This is how many real-world signals work, like brainwaves (EEG), ocean waves, or even the hum of a city. In science, we call this a $1/f^\alpha$ signal (a "power-law" signal).

Now, imagine you are recording this music with a digital recorder (an Analog-to-Digital Converter, or ADC). To record sound digitally, the machine has to chop the smooth wave into tiny steps. The number of steps it can take is determined by its bit depth (e.g., 8-bit, 12-bit, 16-bit).

Low bit depth (few steps): The machine is "clumsy." It can't hear the quiet details.
High bit depth (many steps): The machine is "precise." It can hear the faintest whispers.

The Problem: When you use a "clumsy" recorder, the high-pitched sounds get lost in a static hiss (quantization noise). It looks like a low-pass filter: the high frequencies disappear.

The Discovery: This paper found a simple math rule that tells you exactly how much "reach" (bandwidth) you gain for every single step of precision you add.

The Core Analogy: The Floodlight and the Fog

Think of the signal (the music) as a foggy landscape that gets thicker and darker the further you look into the distance (higher frequencies).

The Signal: A thick fog that gets denser as you go further out.
The Noise: A flat, bright floodlight shining from the ground.

When you use a low-bit recorder (a dim floodlight), the light doesn't reach very far before the fog (the signal) becomes darker than the light. The point where the light and the fog meet is your cutoff frequency. Beyond that point, you only see the light (noise), not the fog (signal).

The Magic Rule:
The paper proves that if you make your floodlight twice as bright (which happens when you add just one extra bit of resolution), the distance you can see doesn't just double. It depends on how thick the fog is:

If the fog is "Brownian" (like a random walk, $\alpha=2$ ): Adding one bit of precision doubles your visible range.
If the fog is "Pink" (like brainwaves, $\alpha=1.5$ ): Adding one bit of precision multiplies your visible range by 2.5 times.

In plain English: For brainwave data, upgrading your recorder from 8-bit to 9-bit doesn't just give you a tiny improvement; it suddenly lets you hear frequencies that were previously completely silent, extending your "listening range" by 2.5 times!

The Catch: The "White Noise" Rule

There is a catch. This rule only works if the static hiss from your recorder is "white noise" (random, like TV static).

However, if your recorder is too "clumsy" (too few bits), the static stops being random and starts mimicking the music itself. It becomes "colored noise."

The Threshold: The paper calculates the minimum number of bits needed to ensure the static is truly random.
- For brainwaves ( $\alpha=1.5$ ), you need at least 5 bits.
- For steeper signals ( $\alpha=2.5$ ), you need 10 bits.
Why? If you have too few bits, the high-frequency parts of the signal are so quiet that they get "swallowed" by the quantization steps, making the error look like the signal itself. You need enough bits to ensure the signal is "wiggling" enough to create random noise.

Real-World Example: The Brain (EEG)

The authors tested this on real brainwave data (EEG).

The Situation: Doctors often use 12-bit or 16-bit recorders for brainwaves because they want to be safe. But high-resolution recorders use a lot of battery power.
The Insight: The paper shows that for brainwaves, 6 bits is actually enough to capture all the important medical information (up to the Gamma band, which is crucial for cognition).
The Result: If you are building a low-power wearable device (like a smartwatch for epilepsy monitoring), you don't need a heavy, power-hungry 12-bit chip. You can drop down to 6 bits.
- Benefit: You save massive amounts of battery.
- Cost: You lose nothing in terms of the frequencies doctors care about.

Summary of the "Scaling Law"

Signal Type	What it sounds like	Effect of adding 1 bit
$\alpha = 2$ (Brownian)	Random walk, heavy bass	Doubles your bandwidth (2x).
$\alpha = 1.5$ (Brainwaves)	Pink noise, typical EEG	Multiplies bandwidth by 2.5x.
$\alpha = 1$ (Pink)	Very flat spectrum	Quadruples your bandwidth (4x).

The Takeaway

This paper gives engineers a "cheat code" for designing sensors. Instead of guessing how many bits of precision they need, they can look at the "slope" of their signal's noise.

If you need more bandwidth: Add a bit, and you get a massive jump in performance (especially for brainwaves).
If you want to save power: You can safely lower the bit depth without losing the frequencies you care about, as long as you stay above the "minimum bit" threshold.

It turns a complex engineering problem into a simple multiplication table.

1. Problem Statement

In many practical applications (e.g., electrophysiology/EEG, audio, ocean noise, seismology), signals exhibit power-law power spectral densities (PSD) of the form $S(f) \propto 1/f^\alpha$ . When these signals undergo amplitude quantization (Analog-to-Digital Conversion), practitioners observe a "low-pass filtering" effect where high-frequency content is degraded before low-frequency content.

However, existing literature lacks a predictive formula relating ADC bit depth ( $N$ ) to the effective bandwidth ( $f_c$ ) for such colored signals. Classical Signal-to-Quantization-Noise Ratio (SQNR) formulas (e.g., $6.02N + 1.76$ dB) assume white noise and uniform signal distribution, failing to account for frequency-dependent effects in $1/f^\alpha$ signals.

2. Methodology

The authors derive a theoretical scaling law and validate it through synthetic simulations and real-world data analysis.

Theoretical Model:
- Signal Model: A signal with PSD $S(f) = S_0 f^{-\alpha}$ .
- Noise Model: Quantization error is modeled as additive white noise with a flat PSD floor: $N_q = \Delta^2 / (6f_s)$ , where $\Delta$ is the quantization step size and $f_s$ is the sampling frequency.
- Cutoff Definition: The effective cutoff frequency $f_c$ is defined as the intersection point where the declining signal spectrum equals the flat quantization noise floor ( $S(f_c) = N_q$ ).
- Derivation: By solving for $f_c$ at bit depths $N$ and $N+1$ , the authors derive a ratio that eliminates signal-dependent constants ( $S_0, R, f_s$ ), leaving a relationship dependent solely on the spectral slope $\alpha$ .
Experimental Validation:
- Synthetic Data: Generated $1/f^\alpha$ signals for $\alpha \in \{1.0, 1.5, 2.0, 2.5, 3.0\}$ .
- Noise Characterization: Measured the spectral slope of quantization error to determine the minimum bit depth ( $N_{min}$ ) required for the noise to be considered "white" (a prerequisite for the law).
- Real Data: Applied the analysis to real EEG data from the PhysioNet Motor Movement/Imagery dataset.

3. Key Contributions

A. The Scaling Law

The paper establishes a closed-form scaling law for the effective cutoff frequency:
$\frac{f_c(N+1)}{f_c(N)} = 2^{2/\alpha}$
This implies that each additional bit of resolution extends the usable bandwidth by a factor of $2^{2/\alpha}$ .

For $\alpha = 2$ (Brownian motion): Bandwidth doubles ( $2\times$ ) per bit.
For $\alpha = 1.5$ (Typical EEG): Bandwidth increases by $\approx 2.5\times$ per bit.
For $\alpha = 1.0$ (Pink noise): Bandwidth quadruples ( $4\times$ ) per bit.

B. Validity Threshold ( $N_{min}$ )

The authors identify that the white noise assumption underlying the scaling law is not valid at low bit depths for steep spectra. They empirically determined a minimum bit depth $N_{min}(\alpha)$ required for quantization noise to become approximately white:

$\alpha = 1.0 \rightarrow N_{min} = 4$ bits
$\alpha = 1.5 \rightarrow N_{min} = 5$ bits
$\alpha = 2.0 \rightarrow N_{min} = 7$ bits
$\alpha = 2.5 \rightarrow N_{min} = 10$ bits
For $\alpha \ge 3$ , the noise remains colored even at 12 bits, limiting the law's applicability without dithering.

C. Practical Implications for Low-Power Systems

The law provides a quantitative method to trade ADC resolution for bandwidth. For example, in EEG systems ( $\alpha \approx 1.5$ ), reducing resolution from 8-bit to 6-bit preserves all clinically relevant frequency bands, potentially allowing for significant power savings in wearable devices without sacrificing diagnostic information.

4. Results

Prediction Accuracy:
- Using the theoretical noise floor ( $\Delta^2/6f_s$ ), the scaling law predicts $f_c$ with <3% error across tested $\alpha$ values.
- Using an empirical noise floor (estimated from the quantized signal's spectrum), error increases to ~14% due to spectral leakage from dominant low-frequency components biasing the noise floor estimate.
Robustness:
- The law is robust to estimation errors in $\alpha$ : an error of $\pm 0.1$ in $\alpha$ results in $<5\%$ prediction error.
- The presence of narrowband spectral peaks (e.g., EEG alpha peaks) has minimal impact (<2% error) unless the peak is near the cutoff frequency.
EEG Application:
- Analysis of real EEG data showed that at 4 bits, the Gamma band (>30 Hz) is severely distorted. At 6 bits, all standard bands (Delta to Gamma) are preserved. At 8 bits, the cutoff frequency exceeds the Nyquist limit, rendering the quantization noise negligible for the entire bandwidth.

5. Significance

Bridging Theory and Practice: This work connects classical quantization theory (Bennett's formula) with the specific needs of communities dealing with colored noise (neuroscience, geophysics, audio), which previously relied on high-resolution ADCs to avoid these issues.
Design Optimization: It offers a rigorous mathematical basis for selecting ADC bit depth in power-constrained applications. Engineers can now calculate the exact bit depth required to preserve a specific frequency band for a given signal spectral slope, rather than relying on heuristic over-design.
Dual to Oversampling: The authors note this law is dual to the oversampling theorem; while oversampling trades sampling rate for resolution in white noise, this law describes how resolution translates to bandwidth in colored noise.

In summary, the paper provides a fundamental scaling law that quantifies the trade-off between bit depth and effective bandwidth for $1/f^\alpha$ signals, enabling optimized design of low-power acquisition systems for biological and geophysical signals.