Polynomial and Noise-based Methods

• Polynomial and noise-based methods are frameworks that use polynomial representations and noise modeling to achieve robust estimation, regression, denoising, and variance reduction.
• These techniques optimize algorithmic performance via orthogonal polynomial expansions, minimal-residual approximations, and noise sensitivity bounds in various practical domains.
• Applications include signal processing in clutter, quantum error mitigation, Lattice QCD variance reduction, and high-dimensional trend estimation under heavy-tailed noise.

Polynomial and noise-based methods refer to a diverse group of analytical, algorithmic, and statistical frameworks that exploit polynomial representations and noise modeling in order to perform estimation, regression, denoising, simulation, feature selection, learning, and variance reduction in the presence of stochastic disturbances or measurement errors. These approaches are central in signal processing, computational physics, machine learning, quantum algorithms, numerical analysis, and high-dimensional inference, integrating orthogonal polynomial expansions, polynomial interpolation, minimal-residual polynomials, polynomial phase gates, noise sensitivity bounds, and explicit noise-modeling techniques to optimize algorithmic performance under various noise structures.

1. Orthogonal Polynomial Methods for Signal Processing in Non-Gaussian Noise

The polynomial transformation method enables optimal detection and estimation in situations where classical statistical methods, which assume Gaussian noise, are analytically intractable. Given a discrete-time signal $g[n]$ corrupted by additive noise $w[n]$, the observed data is $x[n] = g[n] + w[n]$. By projecting $x$ onto a low-order orthogonal polynomial basis $\{p_j[n]\}$ via Gram–Schmidt orthogonalization, one constructs a polynomial approximation $y[n]$ that captures "signal-like" structure and leaves a residual or "approximation error" $e[n]$. Under mild conditions—specifically, when $w[n]$ are independent and have finite variance—the central limit theorem implies that $e[n]$ will be approximately Gaussian, even if $w[n]$ is not (Banoth et al., 2014).

Empirical tests such as bicoherence, histogram analysis, and kurtosis confirm that regardless of the initial distribution (Laplacian, uniform, gamma, etc.), the error process after polynomial preprocessing passes strong Gaussianity tests. This "Gaussianization" allows reapplication of classical maximum-likelihood or MMSE estimators.

Main practical aspects:

• The polynomial order $w[n]$0 is selected to minimize empirical error variance on training data.
• Computation involves forming orthogonal polynomials and solving for the coefficients, with complexity $w[n]$1.
• The method is widely applicable, including to radar/sonar processing in clutter, ocean acoustics, and robust RF communications.

2. Noise-based Polynomial Methods in High-dimensional Inference and Learning

Polynomial and noise-based algorithms enable tractable and efficiently learnable solutions to otherwise hard or noisy inverse problems in physical systems and machine learning. In optical physical unclonable function (PUF) modeling, the response to a challenge is linear (or quadratic) in the challenge bits plus subgaussian noise, and it is shown that ordinary least squares recovers the unknown mapping coefficients $w[n]$2 to within any desired accuracy $w[n]$3—in polynomial time and sample complexity—even in the presence of significant noise (Albright et al., 2023). This result contrasts with hardness in Learning-With-Errors (LWE) when modular reduction is present.

Similarly, polynomial threshold function (PTF) analysis leverages tail and anti-concentration inequalities to bound the noise sensitivity and average sensitivity of degree-$w[n]$4 PTFs on both the Boolean hypercube and under Gaussian input. Polynomial regression and $w[n]$5 learning algorithms yield polynomial-time agnostic learning for constant-degree PTFs (0909.5011).

In robust halfspace learning under Tsybakov/adversarial label noise, a reduction from learning to "certifying non-optimality"—combined with polynomial regression—results in the first polynomial-time algorithms under log-concave distributions, improving over prior quasi-polynomial methods (Diakonikolas et al., 2020).

3. Polynomial Noise Subtraction and Variance Reduction

In Lattice QCD, disconnected quark loop evaluations are variance-limited due to the trace estimation of large, sparse matrices. Polynomial noise subtraction replaces matrix inversion with its low-degree polynomial approximant (typically minimal-residual, such as GMRES), subtracts the resulting estimator from the stochastic trace, and adds back an exact, efficiently computed trace correction (Baral et al., 2016). This approach reduces variance, particularly for off-diagonal matrix elements, and can be combined with Hermitian deflation (HFPOLY) for further efficiency.

Recent work extends this to high-degree GMRES polynomials via stable Leja–ordered factorizations, enabling degree-$w[n]$6 polynomials with $w[n]$7 and nearly an order of magnitude variance reduction (Lashomb et al., 2023). A multipolynomial Monte Carlo method is used to evaluate the trace correction, and "double polynomial" (cascade preconditioning) further reduces computational and memory costs.

4. Polynomial Expansion and Denoising Techniques

A recurring theme in denoising and regularization is truncation of polynomial expansions, which suppresses noise-dominated high-frequency components. The polynomial–exponential basis truncation method constructs orthonormal polynomials (e.g., $w[n]$8 in $w[n]$9), expands the noisy function in this basis, and truncates beyond a data-driven cutoff $x[n] = g[n] + w[n]$0 (Nguyen et al., 2023). Differentiation is then performed term-by-term, yielding numerically stable, bias–variance-balanced derivative estimates even under strong noise contamination.

Generalizations to Chebyshev polynomials, especially in the NoisyChebtrunc algorithm, deliver $x[n] = g[n] + w[n]$1 approximate least-squares fits stable at high degree, with degree selection based on Mallows' $x[n] = g[n] + w[n]$2 criterion and high-probability $x[n] = g[n] + w[n]$3 error bounds scaling as $x[n] = g[n] + w[n]$4 under subgaussian or subexponential noise (Matsuda et al., 2024).

Fast computation and conditioning are also achieved by expressing Savitzky–Golay–style local polynomial smoothing and differentiation matrices in discrete Chebyshev bases, exploiting their recursive properties and symmetry for $x[n] = g[n] + w[n]$5 algorithms with orders-of-magnitude improvements in numerical accuracy compared to Vandermonde-based methods (Rosso, 8 Apr 2026).

5. Bayesian and Model Selection Approaches for Noisy Polynomial Regression

Model order (degree) selection in polynomial regression under noise is formalized by Bayesian evidence maximization (Mana et al., 2013). Assuming a Gaussian noise model and scale-invariant (Jeffreys) priors, the marginal likelihood can be written explicitly in terms of the sum of squared prediction errors and the total signal energy, naturally penalizing overfitting through an Occam factor in the posterior. This approach extends to bivariate (and multivariate) polynomials and recovers the familiar residual–complexity trade-off in large data limits.

In calibration problems for physical experiments (e.g., 21-cm cosmology), marginalization over polynomial noise parameters with per-calibrator noise variance and analytic integration over coefficients yields robust Bayesian model selection and uncertainty quantification, immune to singularities arising from problematic calibrator geometries (Kirkham et al., 2024).

6. Quantum Simulation, Error Mitigation, and Control via Polynomial Noise Methods

Polynomial-based methods facilitate quantum error mitigation and simulation in noisy intermediate-scale quantum (NISQ) algorithms. The OBPPP (Observable Back-Propagation over Pauli Paths and Polynomials) framework simulates expectation values of noisy quantum circuits by expanding operator evolution along Pauli paths and truncating high-weight contributions, whose amplitudes are exponentially suppressed by single-qubit Pauli noise (Shao et al., 2023). The complexity is polynomial in the number of qubits $x[n] = g[n] + w[n]$6 and circuit depth $x[n] = g[n] + w[n]$7 provided noise rate $x[n] = g[n] + w[n]$8 exceeds $x[n] = g[n] + w[n]$9; otherwise, exponential scaling appears for vanishing noise.

Zero-noise extrapolation (ZNE) employs Richardson extrapolation or polynomial least-squares fits across noise scaling experiments to estimate ideal circuit outputs. Sample-complexity and bias–variance analyses quantify the effect of polynomial order, node structure, and high-frequency amplification of measurement noise, with the least-squares polynomial approach stabilizing against overfitting and delivering superior error-mitigation in practical deployments (Mohammadipour et al., 28 Feb 2025).

Fault-tolerant non-Clifford gates in Gottesman–Kitaev–Preskill (GKP) codes leverage polynomial-phase stabilizer constructions, with on-demand noise-biasing tailored by GKP error correction circuits to ensure the polynomial gate acts with arbitrarily small logical error as squeezing increases (Nguyen et al., 25 Nov 2025). The methodology extends naturally to higher-level polynomial phase gates and multi-mode bosonic encodings.

7. Polynomial Structures in Stochastic Processes and Trend Estimation under Heavy-tailed Noise

In continuous-time inference problems where signals are polynomial functions corrupted by Ornstein–Uhlenbeck noise, explicit least-squares estimators are constructed and shown to be asymptotically efficient, with local asymptotic normality (LAN) proven and explicit CLT/Cramér–Rao bounds attained (Höpfner, 2020). In high-dimensional settings (e.g., stochastic Hodgkin–Huxley models), this yields optimal learning of system and noise parameters from continuous observation.

For nonparametric trend estimation under polynomial-tailed noise, wavelet-based estimators using Haar-like bases and adaptive, scale-dependent thresholding achieve minimax risk rates $x$0 despite weak moment and cumulant conditions on the errors. This approach ensures robust estimation for possibly nonuniformly sampled time series subject to strong, non-Gaussian noise (Neumann et al., 12 Feb 2025).

---

References

• Polynomial transformation for non-Gaussian noise: (Banoth et al., 2014)
• Polynomial-exponential basis for numerical differentiation: (Nguyen et al., 2023)
• Polynomial terrain generation and gradient noise: (Thorimbert et al., 2016)
• Polynomial noise subtraction in Lattice QCD: (Baral et al., 2016, Lashomb et al., 2023)
• System identification via polynomial transformation: (Sircar, 2018)
• NoisyChebtrunc and Chebyshev polynomial regression: (Matsuda et al., 2024)
• Fast orthogonal polynomial denoising algorithms: (Rosso, 8 Apr 2026)
• Quantum circuit simulation via polynomial path integral: (Shao et al., 2023)
• Bayesian polynomial model selection: (Mana et al., 2013)
• Polynomial McKean–Vlasov SDEs: (Cuchiero et al., 26 Feb 2025)
• Polynomial agnostic learning, noise sensitivity bounds: (0909.5011)
• Polynomial-time halfspace learning with Tsybakov noise: (Diakonikolas et al., 2020)
• Generalized polynomial Kalman filtering with arbitrary noise: (Teng et al., 2024)
• Polynomial-tailed noise in nonparametric trend estimation: (Neumann et al., 12 Feb 2025)
• ZNE in quantum error mitigation, polynomial methods: (Mohammadipour et al., 28 Feb 2025)
• Fault-tolerant polynomial-phase GKP gates: (Nguyen et al., 25 Nov 2025)
• Estimation under OU noise, polynomial signals: (Höpfner, 2020)
• Bayesian calibration for 21-cm cosmology with polynomials and noise: (Kirkham et al., 2024)
• Polynomial sample-complexity for learning under noisy optical PUFs: (Albright et al., 2023)