Signal Reconstruction from One-Bit Quantized Measurements

368 concepts205 papers1 proposal

In modern signal processing and wireless communications, the hardware components that convert continuous analog signals into digital bits—Analog-to-Digital Converters (ADCs)—present a severe bottleneck. High-precision ADCs consume power that scales exponentially with their bit depth. In applications like massive MIMO or distributed sensor networks, deploying high-resolution ADCs across hundreds of channels is power-prohibitive and expensive.

The extreme solution to this bottleneck is one-bit quantization. By using a simple comparator, the system discards all amplitude information and records only the sign of the signal ($\pm 1$). This reduces hardware complexity to the absolute minimum, but it introduces a glaring mathematical paradox: How can you reconstruct a continuous signal, or infer its underlying parameters, when you have thrown away its scale?

Surprisingly, a rich body of research demonstrates that we can recover high-dimensional signals, estimate channels, and perform precise inference using only one-bit data. The secret lies in leveraging high-dimensional geometry, structural constraints, and statistical variation.

The Geometry of the Sign: Why Direction is Preserved

To understand why one-bit quantization works, it helps to look at the problem geometrically. Imagine a hidden signal vector $x$ in a high-dimensional space. When we take a linear measurement using a sensing vector $a_i$, we compute the dot product $a_i^T x$. A one-bit measurement records only the sign:

$$y_i = \text{sign}(a_i^T x)$$

Geometrically, each sensing vector $a_i$ defines a hyperplane passing through the origin. The one-bit measurement $y_i$ tells us which side of the hyperplane the vector $x$ lies on. It slices the space in half.

If we collect many random measurements, we are effectively slicing the high-dimensional space with dozens of random hyperplanes. The intersection of these half-spaces carves out a narrow, cone-like region. As the number of measurements increases, this cone shrinks around the exact directional vector of $x$.

Because the sign operation is scale-invariant ($\text{sign}(x) = \text{sign}(c \cdot x)$ for any $c > 0$), we completely lose the magnitude (or norm) of the signal, but its direction remains highly preserved. As shown by Lijun Zhang et al., the sample complexity required to recover this directional orientation can theoretically match the efficiency of conventional high-precision compressed sensing. In fact, Martin Slawski et al. demonstrate that if your only goal is to estimate a signal's direction, spending your entire bit budget on one-bit measurements is statistically optimal.

Recovering Scale: The Power of Dithering

If a standard zero-threshold one-bit quantizer completely obliterates amplitude information, how do we recover the absolute scale of a signal when the application demands it? The answer is a technique called dithering—intentionally adding a known noise distribution or a varying threshold to the signal before quantization.

To see why this works, consider a simple scalar example. If a signal has a constant value of $0.1$ or $0.9$, a zero-threshold quantizer outputs $+1$ for both; they are indistinguishable. Now, suppose we add a random dither value uniformly distributed between $-1$ and $+1$ before thresholding.

  • For the $0.1$ signal, the dithered value will cross above zero roughly $55\%$ of the time.
  • For the $0.9$ signal, the dithered value will cross above zero $95\%$ of the time.

By looking at the fraction of $+1$ outputs over multiple samples, the absolute amplitude is mapped directly onto a probability. What was once a deterministic loss of information becomes a statistical estimation problem.

The authors of The Generalized Lasso for Sub-gaussian Observations with Dithered Quantization. proved that combining dithered quantization with a generalized Lasso framework allows for the recovery of both a signal's direction and its precise magnitude, incurring only a minor logarithmic penalty in error compared to infinite-precision data. Similarly, Johannes Maly et al. leverage random dithering alongside Bussgang linearization—a method that decomposes non-linear quantization into a linear gain plus uncorrelated noise—to successfully estimate complex wireless channels.

Exploiting Structure: Sparsity and Low-Rank Constraints

The reason we can recover high-dimensional objects from binary feedback is that real-world signals are rarely arbitrary; they possess underlying structure, such as sparsity or low-rankness. This structure drastically reduces the true degrees of freedom.

In compressed sensing, a sparse signal only has a few non-zero coordinates. In the one-bit domain, this means the hyperplanes do not need to pin down an arbitrary point in space; they only need to locate a point constrained to a few low-dimensional coordinate axes. Phillip North et al. demonstrate that if we possess even partial prior knowledge about which coordinates are non-zero (partial support), the accuracy and efficiency of one-bit signal reconstruction improve dramatically.

This principle extends naturally from vectors to matrices. In matrix completion, we try to fill in a massive table of data from a few observations under the assumption that the underlying matrix is low-rank. When observations are forced into a binary format, Sonia A. Bhaskar proved that enforcing an exact low-rank constraint allows for robust matrix reconstruction, bypassing the need for loose convex relaxations and proving that structural constraints can overcome extreme quantization noise.

Real-World Systems: Spatial Gain and Zero-Crossings

One-bit quantization has found a natural home in massive Multi-User MIMO (MU-MIMO) communication systems and sensor networks. Here, the loss of precision per antenna is compensated for by the massive number of antennas or nodes—a concept known as spatial gain.

In these systems, even though amplitude information is stripped away, the phase and timing relationships of the signals remain perfectly intact within the zero-crossings (the exact moments the signal flips from negative to positive). For spatial tasks like locating an emitter, timing and phase offsets between different antennas are what matter most. Amir Weiss et al. demonstrated that by utilizing the mathematical "arcsine law" to map one-bit correlations back to their unquantized counterparts, direct position determination can achieve near-ideal accuracy at high sample counts, rendering amplitude information largely redundant.

On the downlink side, where a base station transmits signals using one-bit Digital-to-Analog Converters (DACs), the challenge is designing a binary waveform that arrives at the user as a clean, high-order modulated signal (like 16-QAM). Sven Jacobsson et al. showed that non-linear precoding algorithms can pre-distort the transmitted bits such that the constructive interference in the wireless channel naturally reconstructs the required amplitudes at the receiver. To maximize efficiency, Wing-Kin Ma et al. reformulated this precoding task to directly minimize symbol error rates, proving that smart signal processing can completely mitigate the severe distortions of binary hardware.

The Statistical Trade-off

Operating in the one-bit limit changes the nature of inference. As Alon Kipnis notes, compressing measurements down to a single bit inherently introduces a sample size penalty; you need more observations to match the precision of continuous data. However, from a computational and information-theoretic standpoint, this penalty is often surprisingly mild. Dmitriy Kunisky demonstrated that for smooth noise distributions, reducing high-precision data to one-bit values has a negligible impact on the fundamental computational thresholds required for signal detection.

By trading amplitude precision for sample volume, spatial abundance, and structural priors, one-bit systems achieve extreme hardware efficiency without sacrificing the mathematical ability to reconstruct the continuous world.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 214 papers
  2. Version 1 · 2026-10-04 · 147 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_extreme_signal_quantization_e711d68f,
  title = {extreme signal quantization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/extreme-signal-quantization/rdr_e711d68f2d8248929bde535e25f2dc5b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_extreme_signal_quantization_e711d68f_v2,
  title = {extreme signal quantization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/extreme-signal-quantization/rdr_e711d68f2d8248929bde535e25f2dc5b/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake