Reconstructing Signal Scale from Dithered One-Bit Measurements

13 concepts9 papers1 proposal

In signal processing and compressed sensing, hardware constraints often require continuous real-world signals to be heavily quantized before storage or transmission. The most extreme form of this compression is one-bit quantization, where an analog-to-digital converter (ADC) records only the sign of a measurement, reducing a complex floating-point value to a single binary bit ($+1$ or $-1$).

While one-bit quantization drastically reduces memory, power consumption, and bandwidth, it introduces a fundamental mathematical limitation: the complete loss of signal scale. Because $\text{sign}(x) = \text{sign}(c \cdot x)$ for any positive scalar $c$, standard one-bit measurements can only capture the direction of a signal vector, completely erasing its magnitude or norm.

To overcome this limitation, a powerful mechanism has emerged across recent literature: dithered quantization. By intentionally injecting controlled random noise (dither) before the quantization step, the absolute scale of a signal is statistically encoded into the binary output, allowing for full signal and covariance reconstruction.

The Intuition: Dithering as a Shifting Threshold

To understand why dithering works, consider a single scalar measurement $z$ that we want to quantize. In a standard one-bit system, we pass $z$ through a hard threshold at zero:

$$y = \text{sign}(z)$$

If $z = 0.1$, the output is $+1$. If $z = 1000$, the output is also $+1$. The absolute magnitude of $z$ is lost.

Now, imagine we add a random dither variable $\tau$ drawn from a known distribution (such as a Gaussian $\mathcal{N}(0, \sigma^2)$) before quantizing:

$$y = \text{sign}(z + \tau)$$

This is mathematically equivalent to comparing the signal against a random, shifting threshold: $y = +1$ if $\tau > -z$, and $y = -1$ otherwise.

Because $\tau$ is random, the output $y$ becomes a random variable. The probability that $y = +1$ depends directly on the magnitude of $z$:

$$\mathbb{P}(y = +1) = \mathbb{P}(\tau > -z) = \Phi\left(\frac{z}{\sigma}\right)$$

where $\Phi$ is the cumulative distribution function (CDF) of the dither noise.

Standard One-Bit:
z = 0.1    ------------> [ Sign Threshold at 0 ] ------------> +1
z = 1000   ------------> [ Sign Threshold at 0 ] ------------> +1  (Scale lost)

Dithered One-Bit:
z = 0.1    + Noise(τ) -> [ Sign Threshold at 0 ] ------------> Flips between +1 and -1 frequently
z = 1000   + Noise(τ) -> [ Sign Threshold at 0 ] ------------> Almost always +1

If $z$ is very large and positive, even a large negative dither will rarely push $z + \tau$ below zero; the output will almost always be $+1$. If $z$ is close to zero, the dither will frequently flip the sign of the output, resulting in a balanced mix of $+1$ and $-1$.

As noted by Thrampoulidis et al., the signal's norm is effectively encoded in the statistical probability of the dithered measurement crossing the quantization threshold. By observing how often the sign flips across multiple measurements, an algorithm can infer the exact scale of the underlying signal.

Algorithmic Recovery of Direction and Norm

Once the signal's magnitude is embedded into the binary data, the next challenge is designing computationally efficient algorithms to decode it.

Convex Optimization via the Generalized Lasso

A standard tool for recovering structured signals from full-precision data is the Lasso. In The Generalized Lasso for Sub-gaussian Observations with Dithered Quantization., Thrampoulidis et al. demonstrated that the Generalized Lasso (G-Lasso) is remarkably robust when applied directly to dithered binary observations. The random dither serves as a reference level that preserves magnitude information during extreme compression, allowing G-Lasso to recover both direction and scale with only a minor logarithmic performance penalty compared to infinite-precision data.

Expanding on this framework, Stollenwerk et al. proved that a standard least-squares estimator can uniformly recover entire classes of structured signals without needing explicit knowledge of the specific nonlinear quantization function, providing a unified mathematical foundation for nonlinear sensing environments.

One-Shot and Non-Iterative Recovery

While optimization frameworks like G-Lasso are highly accurate, they can be computationally expensive. To address this, Shen et al. introduced a "one-shot" hard thresholding algorithm for 1-bit compressed sensing. By utilizing a single hard-thresholding step on dithered data, the authors achieved near-optimal error rates and superior computational efficiency, bypassing the need for iterative optimization while still providing robust norm estimation.

Integrating Advanced Signal Priors

As compressed sensing has evolved, researchers have shifted from simple sparsity models to more complex structural priors, such as deep generative networks. Dithered quantization integrates seamlessly into these modern architectures.

Qiu et al. analyzed one-bit recovery by leveraging ReLU generative networks as signal priors. They proved that signals can be recovered from limited, dithered single-bit measurements with near-optimal statistical efficiency. Crucially, their global landscape analysis revealed that the underlying non-convex optimization problem is surprisingly well-behaved, completely lacking the spurious local minima that typically plague deep learning tasks.

In a parallel direction, Chen et al. established a unified framework for Generative Compressed Sensing (GCS). The authors demonstrated that adding random noise offsets before nonlinear quantization preserves the signal's norm uniformly across the entire range of a generative model, matching the sample efficiency of non-uniform methods while offering stronger reliability guarantees for real-world applications.

Beyond Vectors: Covariance and Matrix Recovery

The benefits of dithering extend beyond reconstructing individual signal vectors; the mechanism is equally vital for recovering second-order statistical properties, such as covariance matrices.

Without dithering, one-bit quantization only permits the recovery of a signal's correlation matrix (the angles between variables), while the variance information is lost. Dirksen et al. showed that introducing controlled random noise before quantization allows for the mathematical cancellation of quantization distortion. This preserves the full covariance matrix—including variance information—yielding coarse estimators that are nearly as accurate as classical full-precision methods.

Practical Applications and Sample Abundance

This theoretical capacity for covariance recovery has immediate practical applications in sensor array processing. For example, Yeganegi et al. utilized a dither mechanism to perform Direction of Arrival (DoA) estimation using severely quantized sensor measurements. By combining dithering with deep unrolling techniques, the authors successfully reconstructed the signal's covariance matrix, enabling highly accurate spatial estimation from low-resolution hardware.

Furthermore, the high sampling rates natively supported by cheap 1-bit ADCs create a scenario termed "sample abundance." As explored by Eamaz et al., when this abundance is paired with time-varying dither thresholds, the complex task of signal reconstruction can be simplified into a linear feasibility problem. This allows randomized Kaczmarz algorithms to rapidly reconstruct low-rank matrices and sparse signals with significantly lower power consumption than traditional multi-bit architectures.

Key Foundational Work

One-Bit Compressed Sensing via One-Shot Hard Thresholding (2020) — contributed 2 concept elements to this direction.
The Generalized Lasso for Sub-gaussian Observations with Dithered Quantization. (2018) — contributed 2 concept elements to this direction.
The Generalized Lasso for Sub-Gaussian Measurements With Dithered Quantization (2018) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_dithered_quantization_f2b71f02,
  title = {dithered quantization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dithered-quantization/rdr_f2b71f02e7b74973a32265ea83e94772},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_dithered_quantization_f2b71f02_v1,
  title = {dithered quantization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dithered-quantization/rdr_f2b71f02e7b74973a32265ea83e94772/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake