Exact Likelihood Inference and Robust Filtering for Gauss-Cauchy Convolution Models

1. Problem Statement

The paper addresses the challenge of modeling and filtering time-series data where measurement errors exhibit heavy tails (outliers) superimposed on standard Gaussian noise.

• Context: In many fields (e.g., spectroscopy, finance), observations are the sum of a latent signal (often Gaussian) and measurement noise that combines ordinary background noise with occasional extreme shocks.
• The Model: The authors focus on the Gauss-Cauchy Convolution, known as the Voigt distribution ($V$), which is the convolution of a Gaussian $N(\mu, \sigma^2)$ and a Cauchy $C(0, \gamma)$.
• The Bottleneck: Historically, the use of the Voigt distribution in likelihood-based inference has been limited because its density is not an elementary function. Practitioners typically rely on:
  • Numerical convolution (computationally expensive).
  • Pseudo-Voigt approximations (mixture models that lack the additive noise interpretation).
  • Finite-difference derivatives (unstable).
  • Simulation-based methods.
• The Gap: There was a lack of a tractable, exact analytical framework for Maximum Likelihood Estimation (MLE) and robust state-space filtering using the true Voigt density.

2. Methodology

A. Analytical Derivations for the Voigt Distribution

The core methodological breakthrough is the derivation of closed-form expressions for the Voigt density and its derivatives using the scaled complementary error function ($\text{erfcx}$), also known as the Faddeeva function.

• Density Representation: The authors express the Voigt density $f_Y(y)$ using the real part of the scaled complementary error function evaluated on a specific complex line.
• Algebraic Closure: They prove that all finite-order derivatives of the density (score, Hessian, etc.) are algebraic functions of the real and imaginary parts of this special function.
• Implication: This allows for the exact calculation of the likelihood, score vector, and Hessian matrix without numerical integration, quadrature, or approximations.

B. Conditional Moments and Tweedie's Formula

The paper derives the conditional expectation and variance of the latent Gaussian component given an observation.

• Tweedie's Formula: Using Tweedie's formula for Gaussian convolutions, they show that the conditional mean of the latent signal is proportional to the score function of the marginal Voigt density.
• Redescending Property: A critical finding is that the conditional expectation is redescending.
  • For small prediction errors, the relationship is approximately linear (like a Kalman filter).
  • For large prediction errors (extreme outliers), the conditional mean redescends toward zero.
  • Interpretation: The filter automatically identifies extreme observations as likely "noise" (Cauchy component) rather than "signal" (Gaussian component), effectively discounting them rather than propagating them into the state estimate.

C. The Gauss-Cauchy Convolution (GCC) Filter

The authors construct a robust state-space filter for linear systems with Gaussian latent dynamics and Voigt measurement errors.

• Masreliez Approximation: They assume the one-step-ahead prediction density of the latent state is Gaussian.
• Voigt Closure: Because the sum of a Gaussian (state prediction error) and a Voigt (measurement error) is another Voigt distribution, the prediction-error density remains in the Voigt family.
• Update Rule: The state update is driven by the Voigt location score. This replaces the linear update of the Kalman filter with a nonlinear, robust update that naturally handles outliers.
• Smoothing: A Rauch-Tung-Striebel (RTS) smoother is derived based on the filtered moments.

D. Inference and Asymptotics

• MLE: The paper establishes the consistency and asymptotic normality of the MLE for the Voigt parameters, despite the fact that the Voigt distribution itself has no finite moments (due to the Cauchy tail). This is possible because the score and Fisher information are well-defined.
• QMLE for State-Space: For the dynamic model, parameters are estimated via Quasi-Maximum Likelihood (QMLE), treating the Masreliez approximation as the data-generating process for the prediction errors.

3. Key Results

Theoretical and Simulation Results

• Exact Inference: The paper demonstrates that exact MLE is computationally feasible and stable, outperforming pseudo-Voigt approximations which introduce specification errors.
• Finite-Sample Performance: Monte Carlo simulations show that the MLE is unbiased and efficient for moderate sample sizes. The asymptotic standard errors derived from the Fisher information matrix provide accurate approximations.
• Approximation Accuracy: The authors compare the GCC filter against an exact benchmark (numerical density propagation). They find that the Masreliez Gaussian prediction approximation introduces negligible error (Kullback-Leibler divergence $< 10^{-4}$) in the empirically relevant region where the Cauchy component is small relative to the Gaussian component.
• Robustness: The filter successfully separates persistent latent signals from transient heavy-tailed noise, unlike the standard Kalman filter which is sensitive to outliers.

Empirical Application

• Data: Daily log realized volatility for the Technology Select Sector SPDR Fund (XLK) from 1998 to 2025.
• Comparison: The GCC filter is compared against Gaussian (Kalman), Cauchy, Student-$t$, Huber, and Normal-Laplace filters.
• Findings:
  • The GCC filter achieves the highest pseudo-log-likelihood ($-1,306$) compared to the Gaussian benchmark ($-2,465$) and other robust alternatives (Student-$t$: $-1,349$; Huber: $-1,444$).
  • Decomposition: The GCC estimates reveal a dominant Gaussian measurement noise ($\sigma \approx 0.18$) and a small but significant Cauchy component ($\gamma \approx 0.02$).
  • Filtering Behavior: The GCC filter produces a smooth latent volatility series that ignores extreme spikes (e.g., Flash Crash, Dot-com bubble) which are correctly attributed to the Cauchy noise component. In contrast, the Gaussian filter inflates the variance to accommodate these spikes, leading to a noisier state estimate.

4. Key Contributions

1. Tractable Likelihood Theory: The paper resolves the long-standing perception that Voigt-based inference is computationally intractable by providing closed-form analytical expressions for the density, score, and Hessian using special functions.
2. Probabilistic Robust Filtering: Unlike robust filters that impose a redescending score function ad hoc (e.g., Huber loss), the GCC filter derives the redescending property endogenously from the conditional expectation of a latent Gaussian signal observed with Cauchy noise. This provides a rigorous probabilistic micro-foundation for robust filtering.
3. Exact State-Space Filtering: The authors develop a recursive filter that maintains analytical tractability, avoiding the need for numerical integration or particle filters, while nesting the Kalman filter (as $\gamma \to 0$) and the pure Cauchy filter (as $\sigma \to 0$).
4. Empirical Validation: The application to financial volatility demonstrates that the Gauss-Cauchy convolution offers a superior decomposition of persistent dynamics and transient noise compared to standard heavy-tailed alternatives.

5. Significance

• For Econometrics and Finance: The paper provides a powerful tool for filtering financial time series (like volatility) where "black swan" events are common but the underlying dynamics are Gaussian. It offers a principled way to distinguish between structural breaks and measurement outliers.
• For Physics and Spectroscopy: It revitalizes the use of the Voigt profile in statistical modeling by enabling exact likelihood inference, moving beyond the reliance on approximations.
• For Machine Learning: The analytical expressions for the score and Fisher information can be used to refine machine learning estimates of Voigt profiles or to construct uncertainty measures for such models.

In summary, Hansen and Tong demonstrate that the Gauss-Cauchy convolution is not just a heavy-tailed distribution but a computationally tractable framework that unifies exact likelihood inference with robust signal extraction, offering significant improvements over existing methods in both theory and practice.