Bussgang Decomposition: Theory and Applications

• Bussgang decomposition is a probabilistic tool that linearizes memoryless nonlinearities on Gaussian signals, separating the output into a linear term and uncorrelated distortion.
• It underpins efficient designs for quantization in communications, including massive MIMO, channel estimation, and sparse signal recovery.
• Recent extensions adapt the method to non-Gaussian inputs and hybrid estimation strategies, guiding optimal linear estimator performance.

The Bussgang decomposition is a fundamental probabilistic tool in the analysis of nonlinear systems with Gaussian inputs, especially prevalent in the study of quantization effects and hardware impairments in communication, estimation, and signal processing. At its core, the Bussgang theorem provides a rigorously justified decomposition of the output of a memoryless nonlinearity applied to a (joint) Gaussian signal into a linear function of the input plus an uncorrelated distortion term. This linearization principle underpins a wide class of analytic and algorithmic developments in quantized massive MIMO, sparse signal recovery, nonlinear state estimation, and echo cancellation, with recent extensions addressing non-Gaussian regimes and providing optimality conditions for linear estimators.

1. Theoretical Foundation of the Bussgang Decomposition

The classical Bussgang theorem asserts that for any (possibly vector-valued) zero-mean jointly Gaussian input $\mathbf{x}$ and a (possibly nonlinear) memoryless mapping $g(\cdot)$, the output $\mathbf{y}=g(\mathbf{x})$ can be decomposed as

$$\mathbf{y} = \mathbf{A}\,\mathbf{x} + \mathbf{d}, \qquad \mathbb{E}[\mathbf{d}\,\mathbf{x}^H]=\mathbf{0}$$

with the deterministic linear gain

$$\mathbf{A} = \mathbb{E}[\mathbf{y}\,\mathbf{x}^H]\;\mathbb{E}[\mathbf{x}\,\mathbf{x}^H]^{-1}.$$

This decomposition is exact for the first and second-order statistics of any memoryless nonlinearity acting on Gaussian inputs, and $\mathbf{d}$ subsumes all remaining nonlinear distortion, being uncorrelated but generally not independent or Gaussian. This property extends to scalar, complex, and vector settings, including the MIMO case, and is often embedded directly in linear MMSE (LMMSE) estimation architectures for tractable performance analysis (Demir et al., 2020).

For scalar $x \sim \mathcal{N}(0,\sigma_x^2)$ and $y = g(x)$,

$$y = A x + d, \qquad A = \frac{\mathbb{E}[y x]}{\sigma_x^2}, \qquad \mathbb{E}[x d]=0.$$

2. Specialization to Quantization and Hardware Nonlinearities

The Bussgang decomposition is especially crucial for analyzing finite-resolution quantization devices (e.g., low-bit ADCs/DACs, sign quantizers) and other hardware nonlinearities. The decomposition admits efficient closed-form expressions for $A$ in relevant cases:

• One-bit quantization: For $g(\cdot)$0 and $g(\cdot)$1,

$g(\cdot)$2

and the output variance is $g(\cdot)$3, so the distortion variance is $g(\cdot)$4 (Demir et al., 2020).

• L-level uniform quantizer: For a zero-mean Gaussian input $g(\cdot)$5,

$g(\cdot)$6

with output-to-input power ratio and closed-form expressions for the distortion available (Maryopi et al., 2018).

For vector systems with diagonal input covariance, $g(\cdot)$7 becomes a diagonal matrix, enabling per-branch distortion analysis. This is the backbone of linearization in quantized massive MIMO and similar systems (Li et al., 2016, Demir et al., 2020, Atzeni et al., 2021).

3. Application in Channel Estimation, Detection, and Rate Analysis

The Bussgang decomposition fundamentally underpins low-complexity algorithms for channel estimation, detection, and achievable rate analysis in quantized systems:

• Channel Estimation: The linearized model $g(\cdot)$8 enables direct use of LMMSE or BLMMSE estimators. Covariance matrices are constructed from Bussgang parameters, yielding explicit MSE expressions and pilot/detector optimization (Rao et al., 2020, Li et al., 2016, Ding et al., 2024). For one-bit ADCs, the estimator becomes

$g(\cdot)$9

with all moments expressed via the arcsin law.

• Linear Receivers in Quantized Massive MIMO (BMRC, BZF, BMMSE): The Bussgang decomposition facilitates forming MRC, ZF, and MMSE filters robust to quantization by directly injecting the linearized gain and the quantization-distortion covariance into combiner design (Nguyen et al., 2019, Li et al., 2016).
• Ergodic Rate Analysis: The decomposition provides effective SNR and SINR expressions, leading to explicit achievable rate bounds under quantization (Burr et al., 2018, Atzeni et al., 2021, Maryopi et al., 2018). For example, for user $\mathbf{y}=g(\mathbf{x})$0,

$\mathbf{y}=g(\mathbf{x})$1

where all terms derive from Bussgang parameters.

• Sparse Signal Recovery and State Estimation: Transforming sign-quantized measurements into linear models enables direct use of SBL and Kalman-type filtering under coarse quantization (Zhu et al., 2018, Jung et al., 23 Jul 2025).

4. Extensions: Non-Gaussian Inputs and Generalizations

While the classic Bussgang theorem strictly requires Gaussianity, recent research establishes that the decomposition $\mathbf{y}=g(\mathbf{x})$2 with $\mathbf{y}=g(\mathbf{x})$3 (and $\mathbf{y}=g(\mathbf{x})$4 uncorrelated with $\mathbf{y}=g(\mathbf{x})$5) generalizes to arbitrary zero-mean input distributions, including non-Gaussian sources (Burr et al., 2024). However, second-order performance metrics such as signal-to-distortion-plus-noise ratio (SDNR) can deviate from Gaussian-optimality, and the formulae for $\mathbf{y}=g(\mathbf{x})$6 and for output power $\mathbf{y}=g(\mathbf{x})$7 must reflect the true input distribution:

$\mathbf{y}=g(\mathbf{x})$8

For quantizers and practical signal distributions (e.g., QPSK, 4-PAM), closed-form integrals allow computation of these quantities, illuminating tradeoffs in O-RAN digital fronthaul design (Burr et al., 2024).

5. Optimality of Bussgang-based Linear Estimation

The Bussgang-based linear estimators (e.g., BLMMSE, Bussgang linear estimator) are not universally MMSE-optimal due to the general nonlinearity of the MMSE (conditional mean) estimator in quantized systems. However, there exist important regimes where these linear estimators are provably optimal:

• Univariate Gaussian input, arbitrary SNR: The Bussgang estimator coincides with the conditional mean (Fesl et al., 2022, Ding et al., 2024).
• Multiple pilots, noiseless regime, scalar channel: The Bussgang estimator exactly matches the MMSE estimator when pilots are optimally phased (Fesl et al., 2022).
• MIMO with orthonormal pilots and uncorrelated channels, or matched pilot/covariance structures: The necessary and sufficient condition is that the precision matrix $\mathbf{y}=g(\mathbf{x})$9 (in the orthant-probability expression of the true MMSE estimator) has at most one off-diagonal nonzero entry per row (Ding et al., 2024).

When this optimality does not hold, the error gap emerges, especially for higher dimensions or structured correlations, necessitating nonlinear or hybrid estimation strategies.

6. Algorithmic Embedding and Adaptive Estimation

The decomposition is routinely embedded in adaptive algorithms and learning frameworks:

• Adaptive Bussgang Algorithm in Semi-Blind LMS: Incorporates the Bussgang gain in the update loop for channel tracking under decision-directed settings, yielding robustness to detection errors and faster convergence (Karami et al., 2018).
• Model-based and Data-driven State Estimation: In coarse-quantized or model-mismatched scenarios, the Bussgang gain structure is built into Kalman-type recursions, and in hybrid DNNs (e.g., Bussgang-aided KalmanNet), the gain and distortion corrections are learned directly to mitigate quantization artifacts (Jung et al., 23 Jul 2025).
• Sparse Bayesian Learning from One-bit Data: The Bussgang-like linearization enables classical SBL or EM inference for sparse signal recovery in sign-quantized compressed sensing (Zhu et al., 2018).

7. Generalizations beyond Memoryless Nonlinearities and Limitations

The generalised Bussgang decomposition extends the principle to vector, multichannel, or even system-theoretic nonlinearities beyond memoryless mappings, with minimal assumptions (stationarity, invertible covariance) (Roebben et al., 5 Mar 2025). For example, in echo cancellation and enhancement, the echo path can be decomposed into a linearly predictable component and an irreducible uncorrelated residual:

$$\mathbf{y} = \mathbf{A}\,\mathbf{x} + \mathbf{d}, \qquad \mathbb{E}[\mathbf{d}\,\mathbf{x}^H]=\mathbf{0}$$0

Perfect cancellation of the linear part is algorithmically achievable; the residual imposes a fundamental performance limit.

Limitations arise primarily outside the Gaussian regime, for nonlinearities with memory, or when higher-order performance metrics (e.g., BER, information rate) are sensitive to the non-Gaussianity or dependence structure of the residual distortion. In these cases, treating the distortion as independent Gaussian noise can yield only lower bounds (Demir et al., 2020, Burr et al., 2024).

8. Summary Table: Key Bussgang Applications

Application Domain	Mapping/Operation	Decomposition/Estimator Structure
Quantized massive MIMO	$$\mathbf{y} = \mathbf{A}\,\mathbf{x} + \mathbf{d}, \qquad \mathbb{E}[\mathbf{d}\,\mathbf{x}^H]=\mathbf{0}$$1 (ADC/DAC)	$$\mathbf{y} = \mathbf{A}\,\mathbf{x} + \mathbf{d}, \qquad \mathbb{E}[\mathbf{d}\,\mathbf{x}^H]=\mathbf{0}$$2, BLMMSE (Li et al., 2016)
Sparse one-bit CS	$$\mathbf{y} = \mathbf{A}\,\mathbf{x} + \mathbf{d}, \qquad \mathbb{E}[\mathbf{d}\,\mathbf{x}^H]=\mathbf{0}$$3	$$\mathbf{y} = \mathbf{A}\,\mathbf{x} + \mathbf{d}, \qquad \mathbb{E}[\mathbf{d}\,\mathbf{x}^H]=\mathbf{0}$$4, SBL (Zhu et al., 2018)
Kalman state estimation 1-bit	$$\mathbf{y} = \mathbf{A}\,\mathbf{x} + \mathbf{d}, \qquad \mathbb{E}[\mathbf{d}\,\mathbf{x}^H]=\mathbf{0}$$5	Bussgang gain in KF, or learned in DNN (Jung et al., 23 Jul 2025)
Echo cancellation	Nonlinear echo path $$\mathbf{y} = \mathbf{A}\,\mathbf{x} + \mathbf{d}, \qquad \mathbb{E}[\mathbf{d}\,\mathbf{x}^H]=\mathbf{0}$$6	$$\mathbf{y} = \mathbf{A}\,\mathbf{x} + \mathbf{d}, \qquad \mathbb{E}[\mathbf{d}\,\mathbf{x}^H]=\mathbf{0}$$7 (Roebben et al., 5 Mar 2025)

The Bussgang decomposition serves as a universal analytic surrogate for memoryless nonlinearities over Gaussian signals, enabling tractable, robust, and often closed-form solutions for estimation and detection in quantized and nonlinear information processing systems (Demir et al., 2020, Burr et al., 2018, Li et al., 2016, Fesl et al., 2022, Ding et al., 2024, Jung et al., 23 Jul 2025, Burr et al., 2024).