Distortion-Aware Linear Combiners

• Distortion-aware linear combiners are methods that integrate explicit distortion metrics, like MSE and utility loss, to design optimal linear weightings across various systems.
• They employ advanced optimization techniques—including convex programming, heuristics, and neural network-driven methods—to balance signal fidelity with computational and hardware constraints.
• These combiners enhance performance in applications such as beamforming, audio mixing, social choice, and matrix computations, yielding improvements in SNR, throughput, and overall utility.

Distortion-aware linear combiners are linear combining or weighting schemes explicitly designed to optimize signal or decision quality in the presence of various forms of distortion, such as amplifier nonlinearity, quantization error, or constraints on utility aggregation. They arise in domains including wireless communications (beamforming and precoding), audio mixing, social choice, and high-throughput computation, where conventional linear processing can lead to significant suboptimality when neglected distortion effects coherently accumulate or are insufficiently constrained. Central to these methods is the formulation of constraints or objectives that explicitly model distortion, coupled with optimization and algorithmic tools tailored to the underlying signal, channel, or data structure.

1. Mathematical Foundations and Generic Framework

Distortion-aware linear combiners operate by introducing explicit distortion metrics or constraints into the optimization of linear combination weights. Consider the canonical setting:

• Input signals or alternatives $S = [s_1,\dots,s_N]$, or embedded candidates $C = \{c_1, ..., c_m\}$ in a feature space;
• A set of linear combination weights $x$ or $w$ acts on these, forming outputs such as $y = Sx$ or $r = h^T w s$.
• Distortion is represented in various forms: as quantifiable mean-squared error (MSE), a utility loss ratio, physically-induced nonlinearity, or violation of mixture limits.

The distortion-aware combiner design problem is typically cast as either:

• Minimizing an explicit distortion objective (e.g., MSE, worst-case utility gap, in-band nonlinearity) subject to system constraints;
• Satisfying "distortionless" constraints (e.g., maintaining unity gain for the desired signal);
• Optimizing trade-offs between signal fidelity and computational or physical resource utilization.

Optimization takes the form of quadratic programs, convex relaxations, hybrid combinatorial-continuous algorithms, and, in some cases, neural-network-based predictions of combining weights. The precise mathematical description depends on the application domain and the distortion model.

2. Application Domains and System Models

Distortion-aware linear combiners appear in several key contexts, each with distinct system and distortion models:

• Multichannel Audio Mixing: Distortion arises from the reduction of channel gains or non-smooth limiter operation. The goal is to jointly optimize channel gains to minimize perceived loudness reduction or other distortion objectives, under mixture amplitude and box constraints (Luo et al., 9 Jul 2025).
• Wireless Communications (Large-Array Systems, mMIMO): Nonlinear power amplifiers (PAs) introduce third-order or higher-order in-band distortion that can coherently sum at the user in standard precoding such as MRT. Combiners—such as the Z3RO precoders—selectively adjust array weights to null coherent distortion at the receiver, subject to array power or transmit constraints (Rottenberg et al., 2022, Rottenberg et al., 2021, Liu et al., 5 Mar 2025).
• Distributed mMIMO Beamforming: In distributed cell-free massive MIMO, PA distortion creates spatially correlated interference across the network. Distortion-aware beamforming combines local or partially aggregated channel state with models of nonlinearity, optimizing joint or distributed beamforming weights to maximize SINDR (signal-to-interference-noise-and-distortion ratio) (Liu et al., 5 Mar 2025).
• Speech Enhancement: Multiple distortionless beamformers are fused online via neural-predicted convex weights, constrained so their sum is one at each frequency-time point, guaranteeing distortionless speech transmission while optimizing interference suppression (Qian et al., 28 Oct 2025).
• Social Choice and Voting: When candidate utilities are assumed linear in feature or embedding space, distortion-aware combiners correspond to voting rules or lotteries that minimize the worst-case loss of utilitarian welfare, given only ranking (ordinal) reports. The rules are optimized to minimize distortion based on worst-case embedding-aware utility profiles (Ge et al., 22 Oct 2025).
• Computational Linear Algebra and DSP: Matrix multiplication throughput is increased by packing inputs and tolerating controlled distortion (rounding, floating-point errors). The design of the quantization, packing, and unpacking is optimized to maintain error below prescribed distortion bounds for given throughput targets (Anastasia et al., 2011).

3. Representative Designs and Optimization Methods

Several specific distortion-aware combiner designs and solution methodologies are found across these domains:

• Quadratic Programming in Audio Mixing: A convex QP minimizes a quadratic approximation of decibel loss, subject to per-sample and box constraints. Variable reduction via "premixing" and constraint culling by identifying redundant constraints allows efficient real-time solving. Advanced COLA (constant overlap-add) windows are optimized for smooth temporal gain transitions (Luo et al., 9 Jul 2025).
• Distortion-Nulling Precoding in Large-Array Systems: Z3RO precoding imposes a zero third-order distortion constraint $\sum_n h_n w_n |w_n|^2 = 0$ and solves for the weights that maximize array gain. In LOS, closed-form solutions exist by saturating a minority of antennas with negative weighting. In general channels, a heuristic weighting balances array gain and perfect distortion cancellation (Rottenberg et al., 2022, Rottenberg et al., 2021).
• Distributed Optimization in mMIMO: Beamforming weights are optimized in ring or star topologies, utilizing Bussgang decompositions of polynomial nonlinearity. Quadratic-cubic terms are handled via auxiliary variables and penalized MM (majorization-minimization) approaches, with low-dimensional sufficient statistics passed among base stations or CPUs. The solutions iteratively optimize surrogate objectives capturing both signal power and distortion covariance (Liu et al., 5 Mar 2025).
• Neural Network-Driven Fusion in Beamforming: A grouped dual-path RNN architecture takes as input STFT features of beamformer outputs, producing softmax-normalized linear combination weights that are constrained to sum to unity for distortionless fusion. This enables robust adaptation to dynamic environments in real-time (Qian et al., 28 Oct 2025).
• Worst-case Welfare in Social Choice: Linear social choice distortion-aware combiners are constructed by solving LPs or convex programs over polytopes defined by ordinal ranking constraints and candidate embeddings. Instance-optimal deterministic or randomized rules are found by minimizing the maximum welfare distortion on a per-instance basis (Ge et al., 22 Oct 2025).
• Throughput-Distortion Controlled GEMM: Scalar companding and block quantization is employed, followed by floating-point packing/unpacking, with an explicit MSE metric composed of quantization and representation-induced error. Closed-form solutions for optimal quantization factors are derived by balancing these two error sources for a given throughput level (Anastasia et al., 2011).

4. Distortion Metrics and Performance Guarantees

Distortion-aware combiners are evaluated with detailed metrics reflecting signal integrity, spectral efficiency, or utility loss:

• Signal-to-Noise-and-Distortion Ratio (SINDR) for beamforming/precoding (Liu et al., 5 Mar 2025).
• Array Gain and Distortion-to-Signal Ratio (DSR) in large-array precoding, quantifying the penalty for distortion nulling (Rottenberg et al., 2022).
• Mean-Squared Error (MSE) or SNR degradation in matrix multiplication (Anastasia et al., 2011).
• Worst-Case Utilitarian Distortion in social choice, defined as the maximum ratio between optimum and achieved aggregate utility over all legal utility profiles (Ge et al., 22 Oct 2025).
• Domain-Specific Perceptual Metrics such as f(x), the quadratic-approximated dB loss in audio mixing (Luo et al., 9 Jul 2025).
• Speech Quality/Intelligibility Metrics such as SNR, STOI, SI-SDR, and SIR for speech enhancement tasks (Qian et al., 28 Oct 2025).

Performance guarantees are tightly coupled to system dimension:

• Minimum achievable distortion in linear social choice scales as $O(d^3)$ for deterministic rules and $O(\sqrt{d})$ for randomized rules with embedding access; $O(d)$ for randomized rules without embeddings (Ge et al., 22 Oct 2025).
• In Z3RO precoding, array gain penalty vanishes asymptotically as $N\rightarrow\infty$, maintaining distortion nulling without compromising transmit power (Rottenberg et al., 2022).
• In distributed beamforming, distortion-aware designs restore up to $115\%$ sum-rate gain over naive beamforming in saturation regimes, with significant reductions in computational cost and interconnect demands (Liu et al., 5 Mar 2025).

5. Computational Complexity and Implementation Strategies

Effective real-time deployment of distortion-aware linear combiners necessitates rigorously optimized algorithms:

• Dimensionality reduction via premixing and variable substitution to shrink QP size in audio; constraint culling via geometric occlusion tests (Luo et al., 9 Jul 2025).
• Closed-form solutions and heuristics in Z3RO precoding, yielding $O(N)$ per-antenna computational burden, compatible with large-MIMO arrays (Rottenberg et al., 2022).
• Operator splitting and warm-starting in quadratic programming, accelerating repeated frame-by-frame optimization (Luo et al., 9 Jul 2025, Liu et al., 5 Mar 2025).
• Distributed ring and star topologies for mMIMO, separating local from global computations to optimize backhaul and CPU usage (Liu et al., 5 Mar 2025).
• Neural network architectures that amortize weight prediction costs over fast streaming inference, using small context windows and efficient RNN blocks (Qian et al., 28 Oct 2025).
• Polynomial-time LP- and convex-programming approaches for instance-optimal social choice, leveraging geometric structure in voter and candidate embeddings (Ge et al., 22 Oct 2025).
• Precomputed lookup tables and per-block adaptation for companding and packing parameters in GEMM, allowing blockwise throughput-distortion balancing at negligible overhead (Anastasia et al., 2011).

6. Empirical Results and Comparative Analysis

Extensive experimental evaluations illustrate the practical impact of distortion-aware combiner designs:

• In multichannel audio, coupled mixer–limiter QP designs attain 10–20% lower distortion objectives than decoupled matrix-mixer/limiter approaches, with minimal computational penalty when using occlusion culling and premixed groups (Luo et al., 9 Jul 2025).
• Z3RO precoders, under strong nonlinearity, outperform conventional MRT by several dB in SNDR and allow operation 1–2 dB closer to amplifier saturation for a given link requirement; array gain loss is $<1$ dB for $N\geq32$ (Rottenberg et al., 2022).
• Distributed beamforming in cell-free mMIMO achieves up to $115\%$ sum-rate improvement over distortion-unaware reference at high transmit power, with ring-topology methods cutting GLOPS and backhaul exchange by up to $80\%$ and $60\%$ respectively (Liu et al., 5 Mar 2025).
• In neural beamformer fusion, the BeamFusion method maintains distortionless speech reproduction while significantly exceeding classic adaptive convex combination baselines in interference suppression—e.g., achieving 12 dB SNR improvements at $T_{60}=700$ ms (Qian et al., 28 Oct 2025).
• Linear social choice: LP-based instance-optimal rules consistently outperform classical voting rules in empirical distortion on recommendation and survey datasets, with LSLR closely matching $O(\sqrt{d})$ guarantees and dominating as $d$ grows (Ge et al., 22 Oct 2025).
• Throughput-distortion optimized GEMM consistently enables $130$–$180\%$ of baseline GFLOPS in DSP and neural network tasks with controlled distortion and no accuracy loss, as validated in PCA-based face recognition and neural network training benchmarks (Anastasia et al., 2011).