In digital systems, we constantly translate the continuous real world into discrete bits. Whether we are capturing radio waves in a massive MIMO base station, transmitting model updates in federated learning, or running a massive physics simulation, we must eventually round continuous numbers to a finite precision.
The standard approach is deterministic rounding: mapping a value to its nearest discrete representative. However, deterministic rounding introduces a severe mathematical pathology. If you round $2.3$ to $2$, the error is exactly $-0.3$. If your signal remains near $2.3$, your error remains correlated with the signal. This signal-dependent quantization noise introduces non-linear distortions, systematic biases, and irreducible "error floors"—meaning that even if you take infinite measurements or increase the signal-to-noise ratio (SNR), your accuracy hits a dead end.
To solve this, researchers use a counterintuitive strategy: stochastic dithering. By intentionally injecting controlled random noise into a signal before it is quantized, we can break the correlation between the signal and the rounding error, effectively transforming a harsh, non-linear distortion into benign, independent white noise.
The Core Mechanism: How Noise Restores Linearity
To understand why adding noise improves precision, consider a uniform quantizer with a step size of $\Delta$. Without dithering, the quantization error is a deterministic function of the input. If the input signal changes slowly or stays within a single quantization bin, the error is highly biased and correlated.
Now, imagine adding a random variable $U$, uniformly distributed over $[-\Delta/2, \Delta/2]$, to the signal before rounding. This turns a deterministic rounding operation into a probabilistic one. An input value close to a quantization threshold now has a chance of jumping to the higher level or dropping to the lower level.
As shown by Christos Thrampoulidis et al., this random injection smooths out the discontinuous quantization function. In expectation, the deterministic rounding error is converted into a zero-mean independent random variable. This process, often called quantization noise whitening or dithered noise linearization, fundamentally changes how downstream algorithms perceive the data. Instead of dealing with a complex, non-linear step function, standard optimization frameworks and statistical estimators can treat the quantized observations as simple, noisy linear measurements (Christos Thrampoulidis et al.; Nir Shlezinger et al.).
Nonsubtractive vs. Subtractive Dithering
Dithering strategies generally fall into two categories depending on whether the injected noise can be removed later.
Nonsubtractive Dithering
In nonsubtractive dithering, you add noise before quantization and do nothing else. The noise remains embedded in the final signal, but its presence decorrelates the quantization error from the underlying data.
This is highly effective for preventing performance plateaus in hardware-constrained systems. For example, Sven Jacobsson et al. demonstrate that massive MIMO receivers can bypass complex analog down-conversion hardware by using direct RF-sampling with 1-bit ADCs. By applying nonsubtractive dithering, they eliminate the performance floors that usually plague low-resolution receivers at high SNRs, allowing the system to support high-order modulations like 64-QAM. Similarly, in audio processing, Sneha Das et al. introduce randomization to convert sparse, signal-dependent artifacts into a predictable distribution resembling additive white Gaussian noise, which standard statistical post-filters can easily remove.
Subtractive Dithering
If the encoder and decoder share a source of randomness (such as a synchronized pseudo-random seed), they can perform subtractive dithering. Here, the encoder adds a pseudo-random offset before quantization, and the decoder subtracts the exact same offset after reconstruction (Ran Ben-Basat et al.).
The advantage of this approach is that it completely eliminates rounding bias without permanently degrading the signal with noise. By subtracting the dither signal, the remaining quantization error becomes strictly independent of the input signal (Leila Ben Saad et al.; Nir Shlezinger et al.).
Achieved Consistency: The $O(1/\sqrt{n})$ Guarantee
The most powerful consequence of linearizing quantization error is statistical consistency. In a standard deterministic system with coarse hardware (like 1-bit or few-bit ADCs), taking more samples does not improve accuracy past a certain point because the signal remains trapped in the same rounding bins.
Dithering ensures that the error behaves like zero-mean random noise. Consequently, as you collect more samples ($n$), the errors average out. Sjoerd Dirksen et al. prove that under dithered quantization, the estimation error decreases at a non-asymptotic rate of $O(1/\sqrt{n})$. This means you can achieve arbitrarily high precision using low-resolution hardware simply by increasing the sample count. Wai-Yiu Keung et al. exploit this property to eliminate error floors in Reconfigurable Intelligent Surfaces (RIS), allowing arrays with only 2-bit phase resolution to approach the performance of continuous-phase systems.
Advanced Noise Shaping: Triangular Dithering and Beyond
While uniform dither noise makes the mean of the quantization error zero, it does not completely decouple the variance or higher-order moments of the error from the signal. To achieve higher-order independence, researchers alter the shape of the noise distribution.
Triangular Dithering
Triangular dithering is achieved by adding the sum of two independent uniform random variables before quantization. This triangular noise profile completely cancels out quantization-induced bias in second-order statistics.
Michael Ng et al. leverage this to perform low-rank multivariate regression on quantized covariates, proving that triangular dithering exactly cancels out the bias that typically corrupts the diagonal elements of a covariance matrix. Furthermore, Sjoerd Dirksen et al. demonstrate that adding triangular dither allows subspace and Direction-of-Arrival (DOA) estimation algorithms to accurately recover signal subspaces from coarsely quantized data, even when dealing with highly ill-conditioned signals.
Rejection-Sampled Dithering
In high-dimensional vector quantization, uniform or triangular noise may still leave geometric gaps. To optimize the shape of the error across multiple dimensions, Cheuk Ting Li et al. introduce a rejection-sampling framework for dithered quantization. By generating a sequence of dither signals and selecting the first one that bounds the error within a target geometric set, they shape the high-dimensional quantization error into a perfect sphere, minimizing the maximum possible distortion.
Applications Across Domains
Because dithering transforms a non-linear hardware constraint into a tractable statistical problem, it has found utility across several distinct fields:
- High-Dimensional Sensing: In compressed sensing and corrupted data recovery, dithering allows standard convex optimization tools to maintain their theoretical guarantees. Zhaoqiang Liu et al. use dithered quantization to establish uniform recovery guarantees for corrupted sensing, while Michael Ng et al. show that a robust dithered pipeline preserves minimax optimal error rates when quantizing heavy-tailed datasets containing extreme outliers.
- Privacy-Preserving Distributed Learning: In Federated Learning (FL), distributed nodes must compress their updates to save bandwidth (quantization) and inject noise to protect user privacy (Differential Privacy). Burak Hasircioglu et al. show that subtractive dithering can be tuned so that the resulting quantization error matches the exact noise distribution required for Differential Privacy, providing communication compression and privacy simultaneously. Similarly, Ali Moradi Shahmiri et al. use subtractive dithering to implement the Laplace privacy mechanism without digital rounding bias.
- Physics Simulations: Long-running physical simulations often suffer from artificial energy drift or numerical damping due to deterministic rounding errors. Jiafeng Liu et al. demonstrate that dithering simulation variables before bit-width reduction prevents systematic biases from accumulating, ensuring long-term physical stability and conservation laws remain intact.
Key Foundational Work
Postfiltering with Complex Spectral Correlations for Speech and Audio Coding (2018) — contributed 2 concept elements to this direction.
Massive MU-MIMO-OFDM Uplink with Direct RF-Sampling and 1-Bit ADCs (2019) — contributed 2 concept elements to this direction.
The Generalized Lasso for Sub-Gaussian Measurements With Dithered Quantization (2018) — contributed 1 concept elements to this direction.