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

Chen Tong; Peter Reinhard Hansen

arxiv: 2605.01665 · v2 · pith:D2T4AT7Rnew · submitted 2026-05-03 · 💰 econ.EM · stat.ME

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

Peter Reinhard Hansen , Chen Tong This is my paper

Pith reviewed 2026-05-09 16:59 UTC · model grok-4.3

classification 💰 econ.EM stat.ME

keywords Gauss-Cauchy convolutionVoigt distributionrobust filteringmaximum likelihoodstate-space modelsheavy-tailed noisevolatilityscaled complementary error function

0 comments

The pith

Exact analytical expressions for the Gauss-Cauchy convolution density enable stable maximum likelihood estimation and robust filtering in state-space models.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper derives closed-form expressions for the probability density, score function, Hessian, and conditional moments of the convolution of a Gaussian and a Cauchy distribution. These expressions rely on the scaled complementary error function and permit direct computation of the likelihood without numerical integration or approximations. This structure supports exact inference in models with heavy-tailed measurement errors and leads to a filtering procedure that automatically downweights outliers through a redescending influence function. In an empirical example with realized volatility data, the resulting filter outperforms standard Gaussian and robust alternatives by separating persistent latent states from transient noise.

Core claim

We derive analytical expressions for its density, score, Hessian, and conditional moments using the scaled complementary error function, enabling stable maximum likelihood estimation without numerical convolution, finite-difference derivatives, or pseudo-Voigt approximations. The conditional expectation of the latent Gaussian component is governed by a redescending location score, so extreme observations are automatically discounted rather than propagated. This structure motivates the Gauss-Cauchy Convolution (GCC) filter for state-space models with Gaussian latent dynamics and heavy-tailed measurement errors.

What carries the argument

The scaled complementary error function, which yields exact closed-form formulas for the density, derivatives, and conditional expectations of the Gauss-Cauchy convolution distribution.

If this is right

Maximum likelihood estimation becomes stable and exact without requiring numerical convolution or derivative approximations.
The GCC filter discounts extreme observations automatically rather than propagating them through the state.
In volatility modeling, the filter separates persistent latent variation from transient measurement noise more effectively than Gaussian, Student-t, or Huber alternatives.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Similar analytical derivations could apply to other convolution distributions encountered in signal processing or spectroscopy.
Testing the filter on simulated data with known heavy-tailed noise would confirm its robustness advantages.
The approach may generalize to multivariate settings or non-linear state dynamics if the conditional moments remain tractable.

Load-bearing premise

The observed data are generated exactly from a convolution of Gaussian and Cauchy distributions, and the complementary error function evaluations remain numerically stable for the parameter values arising in estimation.

What would settle it

Numerical verification that the analytical density matches the result of direct numerical convolution to within machine precision for a range of parameter values, or out-of-sample comparison showing superior filtering performance on data simulated from the model.

Figures

Figures reproduced from arXiv: 2605.01665 by Chen Tong, Peter Reinhard Hansen.

**Figure 1.** Figure 1: The left panel shows the Voigt density, V (0, 1, 1), together with the best approximating Student-𝑡 distribution, (𝜎𝑡 , 𝜈𝑡) ≈ (1.45, 1.22), and pseudo-Voigt mixture, (𝜂𝑝, 𝜎𝑝, 𝛾𝑝) ≈ (0.65, 1.62, 1.65). The right panel shows the corresponding Q-Q plots. The approximation parameters are chosen to minimize the Kullback-Leibler discrepancy, but visible discrepancies remain. for N (0, 1), respectively.3 This rep… view at source ↗

**Figure 2.** Figure 2: Conditional density of 𝑍 given 𝑌 = 𝑦, for selected values of 𝑦, where 𝑌 = 𝑍 + 𝑋 with 𝑍 ∼ N (0, 1) and 𝑋 ∼ Cauchy(0, 1) independent. The corresponding conditional expectations, E[𝑍|𝑌 = 𝑦], are indicated with vertical lines. The unconditional density of 𝑍 (standard normal) is represented with the shaded area. strictly positive and smooth density 𝑓𝑌 . Moreover, for all real 𝑦, E[𝑍|𝑌 = 𝑦] = −𝜎 2 𝜕 𝜕𝑦 log 𝑓𝑌 (𝑦… view at source ↗

**Figure 3.** Figure 3: The left panel shows the conditional expectation of 𝑍 given 𝑌 = 𝑦, where 𝑌 = 𝑍 + 𝑋 with 𝑍 ∼ N (0, 1) and 𝑋 ∼ Cauchy(0, 1) independent. The right panel shows the corresponding conditional variance. The insets present the same quantities over a wider range of values of 𝑦. The non-monotonicity of the conditional variance is inextricably linked to the redescending nature of the conditional expectation. By diff… view at source ↗

**Figure 4.** Figure 4: The asymptotic standard deviations for 𝜎 and 𝛾 are plotted against the ratio 𝜆 = 𝛾/𝜎 in the left panel for (𝜎, 𝛾) = (1, 𝜆). The right panel shows 𝑅(𝜆), the ratio of the asymptotic standard deviation of 𝜎ˆ to that of 𝛾ˆ. Values above one imply that the Cauchy scale 𝛾 is estimated more precisely than the Gaussian scale 𝜎. for sample sizes ranging from 𝑛 = 100 to 𝑛 = 10, 000. For each replication, we compute … view at source ↗

**Figure 5.** Figure 5: Score functions. The left panel shows the GCC location score, 𝜓(𝑒), for different values of 𝜆 = 𝛾/𝜎. The right panel compares the GCC score with Gaussian, Huber, and Student-𝑡 scores. The GCC score is approximately linear near the origin and redescends in the tails, implying that extreme prediction errors receive little weight in the state update. imposed as a robustification device. Under the Gaussian pre… view at source ↗

**Figure 6.** Figure 6: GCC filtering results for the daily logarithm of realized volatility of the Technology Select Sector SPDR Fund (XLK), December 22, 1998 to November 7, 2025. The left panel shows the observed series 𝑦𝑡 together with the filtered state estimate. The right panel presents the filtered Gaussian and Cauchy measurement-error components by using (6). The same structure also yields a natural robust filtering method… view at source ↗

read the original abstract

The convolution of a Gaussian and a Cauchy distribution, known as the Voigt distribution, is widely used in spectroscopy and provides a natural framework for modeling heavy-tailed measurement noise. We derive analytical expressions for its density, score, Hessian, Fisher information, and conditional moments using the scaled complementary error function, enabling stable maximum likelihood estimation without numerical convolution, finite-difference derivatives, or pseudo-Voigt approximations. The conditional expectation of the latent Gaussian component is governed by a redescending location score, so extreme observations are automatically discounted rather than propagated. This structure leads to the Gauss-Cauchy Convolution (GCC) filter for state-space models with Gaussian latent dynamics and Voigt measurement errors, where the Masreliez Gaussian prediction approximation preserves a Voigt prediction-error density. In an application to log realized volatility for the Technology Select Sector SPDR Fund, the GCC filter separates persistent latent variation from transient measurement noise and attains the highest implemented prediction-error criterion among the Gaussian, Student-$t$, Huber, and related filtering specifications considered.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

They derived closed-form score, Hessian, and conditional moments for the Voigt distribution via the scaled complementary error function, which supports exact MLE and a redescending robust filter.

read the letter

The main takeaway is that this paper supplies usable closed forms for the density, score, Hessian, and conditional moments of the Gauss-Cauchy convolution. The expressions rely on the scaled complementary error function and remove the need for numerical convolution or finite-difference derivatives in estimation and filtering. The conditional expectation of the latent Gaussian component has a redescending score, so large outliers are automatically downweighted rather than passed through. They then use this structure to define the GCC filter for linear-Gaussian state dynamics with heavy-tailed measurement noise. In the realized-volatility application the filter separates persistent variation from transient noise more cleanly than Gaussian, Student-t, or Huber alternatives. The derivations appear to rest on standard properties of the error function rather than on fitted quantities, so there is no obvious circularity. Numerical stability of the erfc evaluations across typical parameter ranges is the main practical question left for referees to check, but the abstract and stress-test note give no sign of instability or derivation gaps. The work is aimed at econometricians and statisticians who already use convolution models or need robust state-space filtering with exact likelihood. It is narrow but technically precise, and the concrete application makes the gain visible. I would send it to peer review; the claims are specific enough that referees can verify them directly from the supplied expressions.

Referee Report

1 major / 3 minor

Summary. The manuscript derives analytical expressions for the density, score, Hessian, and conditional moments of the Gauss-Cauchy convolution (Voigt) distribution using the scaled complementary error function. These closed forms support exact maximum likelihood estimation and motivate the Gauss-Cauchy Convolution (GCC) filter for state-space models with Gaussian latent dynamics and heavy-tailed measurement errors. The conditional expectation exhibits a redescending property that automatically discounts extreme observations. An application to log realized volatility of the Technology Select Sector SPDR Fund compares the GCC filter favorably to Gaussian, Student-t, and Huber alternatives.

Significance. If the derivations are correct, the work provides a useful exact-inference tool for Voigt noise models that avoids numerical convolution and finite-difference derivatives. The redescending conditional expectation is a theoretically attractive robustness feature. The empirical illustration in financial volatility data demonstrates potential practical value, though its strength rests on the stability and accuracy of the analytic expressions across relevant parameter regimes.

major comments (1)

[§2] §2 (Gauss-Cauchy Convolution Distribution): The central claim of stable MLE without numerical issues rests on the scaled complementary error function remaining well-behaved. The manuscript should supply explicit numerical checks or bounds demonstrating that the density, score, and Hessian expressions do not suffer from underflow, overflow, or loss of precision for scale and location values encountered in the volatility application (e.g., |z| > 10 in the complex plane).

minor comments (3)

[Abstract] Abstract: The phrase 'pseudo-Voigt approximations' is used without a reference or brief definition; adding one sentence or a citation would improve accessibility for readers outside spectroscopy.
[§4] §4 (GCC Filter): The state-space recursion for the filter is presented clearly, but the transition from the conditional expectation to the filter update equation would benefit from an explicit one-step-ahead prediction formula to aid replication.
[§5] §5 (Empirical Application): Table 1 reports likelihood values and parameter estimates; including standard errors derived from the analytic Hessian (rather than only point estimates) would strengthen the comparison with competing filters.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful review and valuable feedback on our manuscript. We address the major comment below and will incorporate the suggested improvements in the revised version.

read point-by-point responses

Referee: [§2] §2 (Gauss-Cauchy Convolution Distribution): The central claim of stable MLE without numerical issues rests on the scaled complementary error function remaining well-behaved. The manuscript should supply explicit numerical checks or bounds demonstrating that the density, score, and Hessian expressions do not suffer from underflow, overflow, or loss of precision for scale and location values encountered in the volatility application (e.g., |z| > 10 in the complex plane).

Authors: We appreciate the referee's emphasis on verifying the numerical stability of our analytic expressions. The use of the scaled complementary error function is motivated by its known numerical properties that prevent overflow for large complex arguments, as documented in the numerical analysis literature. Nevertheless, to strengthen the manuscript, we will add explicit numerical checks in a new subsection of §2. These will include evaluations of the density, score, and Hessian for parameter values typical of the volatility application, covering |z| > 10 in the complex plane, demonstrating absence of underflow, overflow, or significant loss of precision. We will also provide bounds where possible based on the asymptotic behavior of the erfcx function. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivations are self-contained

full rationale

The paper's central contribution is the derivation of closed-form expressions for the Voigt density, score, Hessian, and conditional moments expressed via the scaled complementary error function (a standard special function equivalent to the real part of the Faddeeva function). These follow directly from the definition of the Gauss-Cauchy convolution and known analytic properties of the erfc function; no step reduces a claimed prediction or uniqueness result to a fitted parameter, self-definition, or load-bearing self-citation. The redescending conditional expectation is a direct consequence of the Cauchy tail and does not rely on data-dependent fitting or prior author results as an unverified axiom. The application to realized volatility is presented as an illustration, not as the source of the analytic forms. The derivation chain is therefore independent of its own outputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on standard analytic properties of the scaled complementary error function and the modeling assumption that measurement errors are exactly Gauss-Cauchy convolutions; no free parameters or new invented entities beyond the proposed GCC filter are indicated in the abstract.

axioms (1)

standard math The convolution integral of Gaussian and Cauchy densities admits an exact closed-form representation via the scaled complementary error function.
Invoked to obtain analytical density, score, and Hessian without numerical integration.

invented entities (1)

GCC filter no independent evidence
purpose: State-space filter that uses the redescending conditional expectation of the latent Gaussian component to discount outliers.
Newly motivated by the conditional-moment property of the Gauss-Cauchy model.

pith-pipeline@v0.9.0 · 5444 in / 1222 out tokens · 29790 ms · 2026-05-09T16:59:11.640253+00:00 · methodology

discussion (0)

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Tweedie's Formula and Score-Driven Updating
econ.EM 2026-05 unverdicted novelty 6.0

Score-driven models equal Bayesian posterior mean updates in conjugate natural exponential families under steady-state precision discounting and inverse-Fisher scaling.