GSmooth: Cross-Domain Smoothing Techniques

• GSmooth is a polysemous research term representing diverse smoothing strategies that attenuate high-frequency variations in semantic transformations, optimization landscapes, count data, and Gaussian splatting.
• It encompasses methods like generalized randomized smoothing for robustness, nonlocal gradient estimation for optimization, and Bayesian P-splines for count-data smoothing, each with distinctive mechanisms and performance metrics.
• Applications of GSmooth demonstrate improved certified robustness, efficient gradient descent, and enhanced compression through spectral smoothing, providing actionable insights in multiple research settings.

Searching arXiv for papers using “GSmooth” across its major research senses. GSmooth is a polysemous research term used for several distinct smoothing constructions rather than a single canonical method. In the current literature it denotes, among other things, generalized randomized smoothing for certifying robustness against semantic image transformations, Gaussian-smoothed or anisotropically smoothed gradients for deterministic and stochastic optimization, a count-data smoothing routine built on Bayesian P-splines and Griddy-Gibbs sampling, and, in Gaussian Splatting, procedures for generating smooth 2D GS maps or graph-spectral low-pass representations of GS attributes (Hao et al., 2022, Starnes et al., 2023, Starnes et al., 2023, Starnes et al., 2024, Zhang et al., 2020, Gao et al., 2022, Gressani et al., 2024, Yang et al., 12 Dec 2025, Yang et al., 2024).

1. Scope of the term

The term is used across at least four research settings.

Research setting	Meaning of GSmooth	Representative papers
Certified robustness	Generalized randomized smoothing for semantic transformations	(Hao et al., 2022)
Optimization	Gaussian-smoothed, anisotropic, or directional gradients for GD, SGD, Adam, ES, and random search	(Starnes et al., 2023, Starnes et al., 2023, Starnes et al., 2024, Zhang et al., 2020, Gao et al., 2022)
Count-data smoothing	A smoothing routine for Poisson histograms, Poisson regression curves, and density estimation from counts	(Gressani et al., 2024)
3D Gaussian Splatting	Smooth 2D GS maps or graph-spectral low-pass compression of GS attributes	(Yang et al., 12 Dec 2025, Yang et al., 2024)

Across these usages, GSmooth consistently denotes a mechanism that attenuates high-frequency variation, averages over neighborhoods, or regularizes irregular structure. This suggests a family resemblance centered on nonlocal regularization, even though the mathematical objects being smoothed differ substantially: transformation parameters, optimization landscapes, spline coefficients, or Gaussian Splatting attributes.

2. GSmooth as generalized randomized smoothing for semantic robustness

In the robustness literature, GSmooth is the framework introduced in "GSmooth: Certified Robustness against Semantic Transformations via Generalized Randomized Smoothing" (Hao et al., 2022). The setting replaces additive pixel perturbations by semantic transformations

$$\tau(\theta, x): \mathbb{R}^m \times \mathbb{R}^n \rightarrow \mathbb{R}^n,$$

with a smoothed classifier

$$G(x) = \mathbb{E}_{\theta \sim g(\cdot)}\big[f(\tau(\theta, x))\big].$$

The framework targets both resolvable transformations, such as translation and Gaussian blur, and non-resolvable transformations, such as defocus blur, zoom blur, rotational blur, and pixelate.

A central device is a surrogate image-to-image network of the form

$\tau(\theta, x) \approx H(F_1(\theta) + F_2(x)),$

together with a dimension augmentation strategy. In augmented form, the construction defines

$$\tilde{G}(\tilde{x}) = \mathbb{E}_{\tilde{\theta} \sim \tilde{g}(\cdot)} \big[ \tilde{f}(\tilde{\tau}(\tilde{\theta}, \tilde{x})) \big],$$

where semantic noise is injected into an augmented latent space. For the practical Gaussian case with linear $F_1(\theta)=A_1\theta+b_1$, the certified radius is

$$R = \frac{1}{2 M^*}\big(\Psi(\underline{p_A}) - \Psi(\overline{p_B})\big),$$

with

$$M^{\ast} = \max_{\xi\in P} \sqrt{\frac{1}{\sigma_1^2} + \frac{1}{\sigma_2^2} \left\| \frac{\partial F_2(y_\xi)}{\partial \xi} - A_1 \right\|_2^2 }.$$

The framework was proposed because existing certified defenses were insufficient or unable to provably defend against semantic transformations, especially those without closed-form expressions. Its empirical scope includes MNIST, CIFAR-10, and CIFAR-100 with ResNet-110 classifiers and a U-Net-based surrogate. On CIFAR-10, the reported certified accuracies include 67.4% for Gaussian blur at radius $\|\alpha\|_2<4$, 82.2% for translation at radius $\|\alpha\|_2<20$, 65.6% for rotation at $\|\alpha\|_2<10^\circ$, 39.7% for rotational blur at $$G(x) = \mathbb{E}_{\theta \sim g(\cdot)}\big[f(\tau(\theta, x))\big].$$0, 25.0% for defocus blur at $$G(x) = \mathbb{E}_{\theta \sim g(\cdot)}\big[f(\tau(\theta, x))\big].$$1, 44.6% for zoom blur at $$G(x) = \mathbb{E}_{\theta \sim g(\cdot)}\big[f(\tau(\theta, x))\big].$$2, and 45.3% for pixelate at $$G(x) = \mathbb{E}_{\theta \sim g(\cdot)}\big[f(\tau(\theta, x))\big].$$3 (Hao et al., 2022). In this sense, GSmooth is a certification framework for semantic robustness rather than an optimizer.

3. GSmooth in gradient-based optimization

A second major usage refers to Gaussian smoothing of objective functions and the replacement of local gradients by nonlocal smoothed gradients. In "Gaussian smoothing gradient descent for minimizing functions (GSmoothGD)" (Starnes et al., 2023), the smoothed objective is

$$G(x) = \mathbb{E}_{\theta \sim g(\cdot)}\big[f(\tau(\theta, x))\big].$$4

with the GSmoothGD update

$$G(x) = \mathbb{E}_{\theta \sim g(\cdot)}\big[f(\tau(\theta, x))\big].$$5

The corresponding nonlocal gradient admits the function-value formula

$$G(x) = \mathbb{E}_{\theta \sim g(\cdot)}\big[f(\tau(\theta, x))\big].$$6

For convex $$G(x) = \mathbb{E}_{\theta \sim g(\cdot)}\big[f(\tau(\theta, x))\big].$$7-smooth objectives and constant $$G(x) = \mathbb{E}_{\theta \sim g(\cdot)}\big[f(\tau(\theta, x))\big].$$8, the paper gives

$$G(x) = \mathbb{E}_{\theta \sim g(\cdot)}\big[f(\tau(\theta, x))\big].$$9

while for non-convex $\tau(\theta, x) \approx H(F_1(\theta) + F_2(x)),$0-smooth objectives it derives bounds on $\tau(\theta, x) \approx H(F_1(\theta) + F_2(x)),$1 and an $\tau(\theta, x) \approx H(F_1(\theta) + F_2(x)),$2 iteration complexity under square-summable $\tau(\theta, x) \approx H(F_1(\theta) + F_2(x)),$3.

The same program was extended to stochastic optimization and deep learning in "Improved Performance of Stochastic Gradients with Gaussian Smoothing" (Starnes et al., 2023). There the smoothed loss is

$\tau(\theta, x) \approx H(F_1(\theta) + F_2(x)),$4

with $\tau(\theta, x) \approx H(F_1(\theta) + F_2(x)),$5, and the algorithms GSmoothSGD and GSmoothAdam replace $\tau(\theta, x) \approx H(F_1(\theta) + F_2(x)),$6 by $\tau(\theta, x) \approx H(F_1(\theta) + F_2(x)),$7. The paper gives explicit smoothed architectures for feedforward and convolutional networks, including a closed-form smoothed ReLU

$\tau(\theta, x) \approx H(F_1(\theta) + F_2(x)),$8

and shows that the smoothed model can be trained by standard backpropagation rather than zeroth-order Monte Carlo. Under standard assumptions and a decaying smoothing schedule, GSmoothAdam satisfies

$\tau(\theta, x) \approx H(F_1(\theta) + F_2(x)),$9

and also

$$\tilde{G}(\tilde{x}) = \mathbb{E}_{\tilde{\theta} \sim \tilde{g}(\cdot)} \big[ \tilde{f}(\tilde{\tau}(\tilde{\theta}, \tilde{x})) \big],$$0

A further generalization appears in "Anisotropic Gaussian Smoothing for Gradient-based Optimization" (Starnes et al., 2024), which defines

$$\tilde{G}(\tilde{x}) = \mathbb{E}_{\tilde{\theta} \sim \tilde{g}(\cdot)} \big[ \tilde{f}(\tilde{\tau}(\tilde{\theta}, \tilde{x})) \big],$$1

for a symmetric invertible matrix $$\tilde{G}(\tilde{x}) = \mathbb{E}_{\tilde{\theta} \sim \tilde{g}(\cdot)} \big[ \tilde{f}(\tilde{\tau}(\tilde{\theta}, \tilde{x})) \big],$$2. The resulting algorithms AGS-GD, AGS-SGD, and AGS-Adam replace isotropic smoothing by directionally adaptive smoothing, often with $$\tilde{G}(\tilde{x}) = \mathbb{E}_{\tilde{\theta} \sim \tilde{g}(\cdot)} \big[ \tilde{f}(\tilde{\tau}(\tilde{\theta}, \tilde{x})) \big],$$3 updated using Covariance Matrix Adaptation. This extension preserves key structural properties: if $$\tilde{G}(\tilde{x}) = \mathbb{E}_{\tilde{\theta} \sim \tilde{g}(\cdot)} \big[ \tilde{f}(\tilde{\tau}(\tilde{\theta}, \tilde{x})) \big],$$4 is $$\tilde{G}(\tilde{x}) = \mathbb{E}_{\tilde{\theta} \sim \tilde{g}(\cdot)} \big[ \tilde{f}(\tilde{\tau}(\tilde{\theta}, \tilde{x})) \big],$$5-smooth, then $$\tilde{G}(\tilde{x}) = \mathbb{E}_{\tilde{\theta} \sim \tilde{g}(\cdot)} \big[ \tilde{f}(\tilde{\tau}(\tilde{\theta}, \tilde{x})) \big],$$6 is $$\tilde{G}(\tilde{x}) = \mathbb{E}_{\tilde{\theta} \sim \tilde{g}(\cdot)} \big[ \tilde{f}(\tilde{\tau}(\tilde{\theta}, \tilde{x})) \big],$$7-smooth; if $$\tilde{G}(\tilde{x}) = \mathbb{E}_{\tilde{\theta} \sim \tilde{g}(\cdot)} \big[ \tilde{f}(\tilde{\tau}(\tilde{\theta}, \tilde{x})) \big],$$8 is convex, then $$\tilde{G}(\tilde{x}) = \mathbb{E}_{\tilde{\theta} \sim \tilde{g}(\cdot)} \big[ \tilde{f}(\tilde{\tau}(\tilde{\theta}, \tilde{x})) \big],$$9 is convex and $F_1(\theta)=A_1\theta+b_1$0; and if $F_1(\theta)=A_1\theta+b_1$1 is $F_1(\theta)=A_1\theta+b_1$2-Lipschitz, then $F_1(\theta)=A_1\theta+b_1$3 is $F_1(\theta)=A_1\theta+b_1$4-Lipschitz and $F_1(\theta)=A_1\theta+b_1$5-smooth (Starnes et al., 2024). In this optimization lineage, GSmooth denotes nonlocal descent on a smoothed landscape, with isotropic and anisotropic variants.

4. GSmooth in derivative-free optimization and evolution strategies

In black-box optimization, GSmooth most naturally refers to Gaussian-smoothing-based gradient estimation. The classical smoothed objective is

$F_1(\theta)=A_1\theta+b_1$6

with Monte Carlo estimator

$F_1(\theta)=A_1\theta+b_1$7

This viewpoint is generalized in "Generalizing Gaussian Smoothing for Random Search" (Gao et al., 2022), which shows that Gaussian perturbations need not be fixed. For the forward-difference estimator, the asymptotic mean-squared error depends on the perturbation variance $F_1(\theta)=A_1\theta+b_1$8 and kurtosis $F_1(\theta)=A_1\theta+b_1$9, and the paper derives three alternative perturbation schemes with provably smaller MSE than Gaussian smoothing: Bernoulli Smoothing (BeS), GS-shrinkage with

$$R = \frac{1}{2 M^*}\big(\Psi(\underline{p_A}) - \Psi(\overline{p_B})\big),$$0

and BeS-shrinkage with

$$R = \frac{1}{2 M^*}\big(\Psi(\underline{p_A}) - \Psi(\overline{p_B})\big),$$1

All three retain the same $$R = \frac{1}{2 M^*}\big(\Psi(\underline{p_A}) - \Psi(\overline{p_B})\big),$$2 direction-sampling complexity as standard Gaussian smoothing.

A different refinement is "A Novel Evolution Strategy with Directional Gaussian Smoothing for Blackbox Optimization" (Zhang et al., 2020). That work replaces isotropic $$R = \frac{1}{2 M^*}\big(\Psi(\underline{p_A}) - \Psi(\overline{p_B})\big),$$3-dimensional smoothing by one-dimensional directional Gaussian smoothing along $$R = \frac{1}{2 M^*}\big(\Psi(\underline{p_A}) - \Psi(\overline{p_B})\big),$$4 orthogonal directions $$R = \frac{1}{2 M^*}\big(\Psi(\underline{p_A}) - \Psi(\overline{p_B})\big),$$5, defining the DGS gradient operator

$$R = \frac{1}{2 M^*}\big(\Psi(\underline{p_A}) - \Psi(\overline{p_B})\big),$$6

Each directional derivative is estimated by Gauss-Hermite quadrature rather than Monte Carlo,

$$R = \frac{1}{2 M^*}\big(\Psi(\underline{p_A}) - \Psi(\overline{p_B})\big),$$7

which the paper uses to enable large $$R = \frac{1}{2 M^*}\big(\Psi(\underline{p_A}) - \Psi(\overline{p_B})\big),$$8 and low-variance nonlocal exploration in high dimension. The reported experiments show that DGS-ES clearly outperforms competing methods on highly nonconvex Ackley, Rastrigin, and Schaffer benchmarks, while none of the methods, including DGS-ES, solves Schwefel globally (Zhang et al., 2020).

In this derivative-free literature, GSmooth denotes a family of zeroth-order estimators of smoothed gradients. The family includes isotropic Gaussian perturbations, alternative perturbation distributions selected by MSE analysis, and directional constructions designed to retain nonlocal exploration without the variance growth typical of full $$R = \frac{1}{2 M^*}\big(\Psi(\underline{p_A}) - \Psi(\overline{p_B})\big),$$9-dimensional Monte Carlo smoothing.

5. GSmooth as a count-data smoothing routine

In one statistical usage, GSmooth is used generically to mean a smoothing routine for count data, including histograms, Poisson regression curves, and density estimation from counts. The relevant engine is the GGSBPS algorithm introduced in "Griddy-Gibbs sampling for Bayesian P-splines models with Poisson data" (Gressani et al., 2024). The model assumes Poisson counts

$$M^{\ast} = \max_{\xi\in P} \sqrt{\frac{1}{\sigma_1^2} + \frac{1}{\sigma_2^2} \left\| \frac{\partial F_2(y_\xi)}{\partial \xi} - A_1 \right\|_2^2 }.$$0

with log-link

$$M^{\ast} = \max_{\xi\in P} \sqrt{\frac{1}{\sigma_1^2} + \frac{1}{\sigma_2^2} \left\| \frac{\partial F_2(y_\xi)}{\partial \xi} - A_1 \right\|_2^2 }.$$1

where $$M^{\ast} = \max_{\xi\in P} \sqrt{\frac{1}{\sigma_1^2} + \frac{1}{\sigma_2^2} \left\| \frac{\partial F_2(y_\xi)}{\partial \xi} - A_1 \right\|_2^2 }.$$2 is a B-spline basis and $$M^{\ast} = \max_{\xi\in P} \sqrt{\frac{1}{\sigma_1^2} + \frac{1}{\sigma_2^2} \left\| \frac{\partial F_2(y_\xi)}{\partial \xi} - A_1 \right\|_2^2 }.$$3 are spline coefficients. Smoothness is imposed through a difference penalty

$$M^{\ast} = \max_{\xi\in P} \sqrt{\frac{1}{\sigma_1^2} + \frac{1}{\sigma_2^2} \left\| \frac{\partial F_2(y_\xi)}{\partial \xi} - A_1 \right\|_2^2 }.$$4

and Gaussian penalty prior

$$M^{\ast} = \max_{\xi\in P} \sqrt{\frac{1}{\sigma_1^2} + \frac{1}{\sigma_2^2} \left\| \frac{\partial F_2(y_\xi)}{\partial \xi} - A_1 \right\|_2^2 }.$$5

with weak Gamma hyperprior $$M^{\ast} = \max_{\xi\in P} \sqrt{\frac{1}{\sigma_1^2} + \frac{1}{\sigma_2^2} \left\| \frac{\partial F_2(y_\xi)}{\partial \xi} - A_1 \right\|_2^2 }.$$6.

The computational issue is that the conditional posterior of $$M^{\ast} = \max_{\xi\in P} \sqrt{\frac{1}{\sigma_1^2} + \frac{1}{\sigma_2^2} \left\| \frac{\partial F_2(y_\xi)}{\partial \xi} - A_1 \right\|_2^2 }.$$7 is non-conjugate under the Poisson likelihood. The paper therefore updates each coefficient through a Griddy-Gibbs step on the univariate conditionals $$M^{\ast} = \max_{\xi\in P} \sqrt{\frac{1}{\sigma_1^2} + \frac{1}{\sigma_2^2} \left\| \frac{\partial F_2(y_\xi)}{\partial \xi} - A_1 \right\|_2^2 }.$$8. Two structural facts are central: the log-conditional $$M^{\ast} = \max_{\xi\in P} \sqrt{\frac{1}{\sigma_1^2} + \frac{1}{\sigma_2^2} \left\| \frac{\partial F_2(y_\xi)}{\partial \xi} - A_1 \right\|_2^2 }.$$9 is strictly concave, so each conditional has a unique mode, and the paper provides a bound locating that mode in a finite interval. GGSBPS then uses a grid-grower to define $\|\alpha\|_2<4$0, evaluates the unnormalized conditional on an equidistant grid, normalizes to a discrete PMF, and samples by inverse CDF. The conditional posterior of $\|\alpha\|_2<4$1 remains conjugate Gamma. The method is described as calibration-free because it avoids Metropolis-Hastings proposal tuning and acceptance-rejection steps (Gressani et al., 2024).

For histogram smoothing, the fitted count curve is

$\|\alpha\|_2<4$2

and the estimated density is

$\|\alpha\|_2<4$3

The paper reports simulation scenarios A–D, reasonable bias and RMSE in density estimation and Poisson regression examples, and real-data illustrations on the Old Faithful geyser and Hidalgo stamp thickness datasets (Gressani et al., 2024). In this usage, GSmooth is not the formal algorithm name; it is a generic label for a count-data smoother whose backend can be GGSBPS.

6. GSmooth in Gaussian Splatting compression

In Gaussian Splatting, GSmooth is used more informally for procedures that make GS attributes smoothly arranged or spectrally smooth so that they compress well under image or video codecs. One line of work is "Lightweight 3D Gaussian Splatting Compression via Video Codec" (Yang et al., 12 Dec 2025). A vanilla 3DGS primitive has position, covariance parameters, opacity, and spherical harmonics coefficients; the details describe vanilla 3DGS as having 59 channels per primitive. Video-based compression converts these primitive attributes into 2D GS maps, and the critical step is generating smooth 2D maps so that neighboring pixels correspond to Gaussians with similar attributes. LGSCV replaces global PLAS sorting by a two-stage Morton scan, applies PCA to the 45 SH AC channels, and refines the layout with block-constrained MiniPLAS. The paper reports over 20% RD gain compared with state-of-the-art methods, 2D map generation time of approximately 1 second, and a 50% reduction in encoding time (Yang et al., 12 Dec 2025). In this context, GSmooth denotes smooth GS maps rather than Gaussian convolution.

A second line is "A Benchmark for Gaussian Splatting Compression and Quality Assessment Study" (Yang et al., 2024), which proposes Graph-based GS Compression (GGSC). GGSC partitions primitives with KDTree, constructs a graph on primitive centers, applies a graph Fourier transform to attribute signals, clips high-frequency coefficients with clipping ratio $\|\alpha\|_2<4$4, and then uses quantization, G-PCC, and adaptive arithmetic coding. The graph is defined by

$\|\alpha\|_2<4$5

and for an attribute signal $\|\alpha\|_2<4$6 the GFT coefficients are

$\|\alpha\|_2<4$7

followed by high-frequency clipping

$\|\alpha\|_2<4$8

This is explicitly a graph-based low-pass smoothing of GS attributes. The same work introduces the GSQA dataset, conducts a subjective experiment with Mean Opinion Scores, and reports that 3SSIM is the best-performing tested objective metric with PLCC $\|\alpha\|_2<4$9, SRCC $\|\alpha\|_2<20$0, and RMSE $\|\alpha\|_2<20$1 (Yang et al., 2024).

Taken together, these Gaussian Splatting papers use GSmooth to denote smooth 2D GS maps, smooth attribute fields, or graph-spectral low-pass representations that favor compression. This suggests a shift in meaning relative to optimization and robustness: the object being smoothed is no longer a loss or a transformation distribution, but the spatial or spectral organization of Gaussian primitives and their attributes.